Developing a new methodology for predicting open

stopes’ performance

Thèse

Benoit Mc Fadyen

Doctorat en génie des mines

Philosophiæ doctor (Ph. D.)

Québec, Canada

© Benoit Mc Fadyen, 2024

Developing a new methodology for predicting open stopes’
performance

Thesis

Benoît McFadyen

Under the direction of:

Martin Grenon, supervisor

Yves Potvin, co-supervisor
Résumé
Les chantiers de type chambre ouverte ont été popularisés dans les mines souterraines autour des années 80
dû aux hauts rendements et relativement faibles coûts qu’ils engendrent. Le succès de cette méthode de minage
dépend de la capacité à concevoir une chambre ouverte qui maximise les profits tout en respectant la géométrie
de celle-ci lors du minage. En réalité, des sur-bris (volume de roche minée et non planifiée) et sous-bris (volume
de roche planifié laissée en place) sont générés lors du minage du chantier et peuvent avoir des conséquences
néfastes sur la stabilité du chantier et la rentabilité des opérations. Afin de minimiser ces bris hors-profil et leurs
impacts, des outils empiriques, tels que les abaques de stabilité, ont été développés et améliorés afin de prédire
les volumes de sur-bris par face pour le chantier. Malheureusement, ces outils présentent certaines limitations :
ils intègrent un nombre limité de paramètres (ce qui ne couvre pas toutes les causes possibles des bris hors-
profil pour le site à l’étude), le sous-bris n’est pas considéré et la résolution spatiale des prédictions est limitée
(valeur qualitative ou moyenne par face). Ces limitations empêchent l’intégration de données économiques dans
les prédictions et ainsi l’optimisation globale des chantiers.

L’objectif de cette thèse fut de développer une méthodologie pour les mines en opération qui permet de
comprendre la distribution spatiale et la magnitude des bris hors-profils générés le long de la surface du chantier
et d’utiliser ces connaissances durant la conception du chantier pour prédire la géométrie attendue afin
d’optimiser sa performance. Pour ce faire, une base de données de chantiers fut créée pour trois cas d’étude
(deux en Australie et un au Canada) à une résolution par octree (données géoréférencées mesurées à une
résolution d’environ un mètre cube) qui comprend des paramètres géométriques, géomécaniques, géologiques,
opérationnels et de performance. Des méthodes statistiques univariées, bivariées et multivariées (analyse en
composantes principales et moindres carrés partiels) ont servi à identifier les paramètres critiques. Deux
modèles prédictifs multivariés ont été utilisés par la suite pour prédire la géométrie (moindres carrés partiels et
forêt aléatoire) ainsi qu’un modèle prédictif additionnel (analyse discriminante linéaire) pour prédire la position
des bris hors-profils. Les modèles ont été comparés pour déterminer la meilleure approche. Des données
économiques ont ensuite été intégrées pour caractériser les performances géotechniques (bris hors-profil) et
économiques des chantiers attendus.

L’analyse statistique des paramètres a permis d’identifier les paramètres critiques pour chaque cas d’étude qui
impacte le sur-bris (énergie du tir à l’explosif, surdéveloppement des galeries, position dans la surface et
présence de structures géologiques) et le sous-bris (position dans la surface, distance avec les trous chargés
d’explosifs et l’angle d’interception des trous chargés d’explosifs avec les surfaces). Les paramètres critiques
ont été utilisés pour chaque site afin de construire les modèles prédictifs. Les résultats montrent que la
magnitude des bris hors-profils le long de la surface de conception peut être prédite à l'aide d'un modèle forêt

ii
aléatoire avec moins d’un mètre d’erreur pour 65% de la surface en moyenne. Les modèles de forêt aléatoire
permettent de prédire la géométrie avec succès pour 71 % des faces de chantiers transversaux et 51 % des
faces de chantiers longitudinaux. L’analyse de ces prédictions permet de déterminer les magnitudes maximales
possibles de bris hors-profil et leur position (utilisant l’approche probabiliste) ainsi que d’estimer la performance
économique du chantier.

La résolution des données et l'utilisation de l'analyse multivariée ont permis de prédire la géométrie des
chantiers, allant bien au-delà de la simple prédiction qualitative par face de chantiers fournie par une approche
traditionnelle de l’abaque de stabilité. Ces travaux constituent une étape importante vers un processus de
conception et de planification des chantiers qui minimise les problèmes de stabilité, la dilution et la perte de
minerai à travers le chantier. Étant donné qu’une géométrie de conception maximisant la récupération de minerai
augmente le potentiel de dilution et vice versa, cette approche permet de quantifier le compromis économique
et ainsi maximiser la rentabilité lors de l'exploitation du chantier. Cela marque une contribution importante pour
la compréhension et l’optimisation de la performance des chantiers de type chambre ouverte.

iii
Abstract
Open stopes were popularised in underground mines around the 1980s due to the high yields they generate and
their relatively low cost. The success of this mining method depends on the ability to design a stope that
maximises profits while respecting the geometry during mining. In reality, overbreak (OB, unplanned volume of
rock mined) and underbreak (UB, volume of planned rock left in place) are generated during the mining process
which can have negative consequences on the stability of the stope and the profitability of the operation. To
minimise the OB and UB, and their impact, empirical tools such as the Stability Chart were developed and
improved to predict the volumes of OB per stope faces. Unfortunately, these tools have certain limitations: they
integrate a limited number of variables (which does not cover all the possible causes of OB), UB is not
considered, and the resolution of the predictions is limited (qualitative or average value per face). These
limitations prevent the full integration of economic data into predictions and thus the optimisation of the mines
stopes.

The objective of this thesis is to develop a methodology for operating mines which enables the understanding of
the spatial distribution and magnitude of OB and UB generated along the surface of the stope and to use this
knowledge during stope design to predict the expected geometry to optimise its performance. A stope database
was created for three case studies (two in Australia and one in Canada) at an octree resolution (georeferenced
data measured at an approximately one cubic metre resolution) which includes geometric, geomechanical,
geological, operational and performance variables. Univariate, bivariate and multivariate statistical methods
(principal component analysis and partial least squares) were used to identify the critical variables for each site.
Multivariate and machine learning models were subsequently used to predict the mined geometry (partial least
squares and random forest) and OB and UB location (linear discriminant analysis). The models were compared
to determine the best approach. Economic data was integrated to characterise the geotechnical (OB and UB)
and economic performances of the expected geometry.

The statistical analysis of the variables enables the identification of critical variables for each case study
regarding the OB (observed in areas where there is high blasting energy, where there is a large undercut and
overcut, towards the middle of the faces and in the presence of major geological structure) and UB (observed
near the edges of surfaces, where there is a large distance from the blasting, and where drill holes toe into the
faces). The critical variables were used for each site to build predictive models. The results show that the
magnitude of OB and UB can be predicted using a random forest model with a less than one metre error for
65% of the surface on average. These predictions are used to build the expected geometry of the stope.
Considering the mining context, it is possible to excellently predict the geometry for 71% of the transversal stope
faces and 51% of the longitudinal stope faces. The incorporation of probabilities also enables the estimation of

iv
the maximum magnitudes of OB and UB, their location and the economical estimation of the performance of the
stopes.

The octree resolution of the data and the use of multivariate analysis has enabled the prediction of the stope
geometry, going well beyond the simple qualitative per stope face prediction provided by a traditional Stability
Chart approach. This work is a significant step to advance stope design and the planning process, providing
tools to minimise stability issues, dilution and loss of ore. As designing for higher ore recovery increases the
potential for dilution and vice versa, this approach allows for the value realised from mining a stope to be
maximised by quantifying this inherent economic trade-off. The enhanced resolution of both data and predictions
has allowed for a more comprehensive evaluation of the predicted stope geometry and economic outcomes
during the design phase. This marks a major contribution in the ability to design and plan the optimal stope.

v
Contents
Résumé ............................................................................................................................................................... ii
Abstract............................................................................................................................................................... iv
Contents ............................................................................................................................................................. vi
List of figures ..................................................................................................................................................... xii
List of tables..................................................................................................................................................... xxx
List of abbreviations, symbols and acronyms ............................................................................................... xxxiv
Acknowledgement......................................................................................................................................... xxxvi
Foreword....................................................................................................................................................... xxxvii
Introduction ......................................................................................................................................................... 1
Background .................................................................................................................................................... 1
Problem definition ........................................................................................................................................... 3
Objectives....................................................................................................................................................... 3
Methodology ................................................................................................................................................... 4
Structure of the thesis..................................................................................................................................... 5
Significance and contributions ........................................................................................................................ 7
Chapter 1. Critical review................................................................................................................................ 8
1.1. Introduction........................................................................................................................................ 8
1.2. Quantifying stope OB and UB ........................................................................................................... 8
1.2.1. Per stope resolution .................................................................................................................. 9
1.2.2. Per face resolution .................................................................................................................... 9
1.2.3. Per octree resolution............................................................................................................... 10
1.3. Stope design tools ........................................................................................................................... 12
1.3.1. Numerical modelling ............................................................................................................... 12
1.3.2. Stability Chart ......................................................................................................................... 13
1.3.3. Site-Specific tools ................................................................................................................... 17
1.4. Quantifying critical variables that influence stope OB and UB......................................................... 19
1.4.1. Geometrical variables ............................................................................................................. 19
1.4.2. Geomechanical variables ....................................................................................................... 21
1.4.3. Geological variables ............................................................................................................... 22
1.4.4. Operational variables .............................................................................................................. 23
1.4.5. Seismic variables .................................................................................................................... 26

vi
 1.5. Statistical analysis used for understanding and predicting stope OB and UB ................................. 26
1.5.1. Univariate analysis.................................................................................................................. 27
1.5.2. Bivariate analysis .................................................................................................................... 28
1.5.3. Multivariate analysis ............................................................................................................... 30
1.6. Economic analysis........................................................................................................................... 32
1.7. Conclusion....................................................................................................................................... 35
Chapter 2. Methodology ............................................................................................................................... 37
2.1. Introduction...................................................................................................................................... 37
2.2. Data collection ................................................................................................................................. 37
2.2.1. Data quality verification........................................................................................................... 39
2.2.2. Stope reconciliation ................................................................................................................ 41
2.2.3. Quantifying the variables ........................................................................................................ 42
2.2.4. Data filtering ........................................................................................................................... 48
2.3. Root cause analysis ........................................................................................................................ 48
2.4. Predicting stope geometry ............................................................................................................... 51
2.4.1. Partial least squares model .................................................................................................... 52
2.4.2. Linear discriminant analysis model ......................................................................................... 53
2.4.3. Random forest model ............................................................................................................. 54
2.4.4. Evaluating model performance ............................................................................................... 56
2.5. Assessing predicted stope performance ......................................................................................... 62
2.5.1. Stability assessment ............................................................................................................... 62
2.5.2. Economic assessment ............................................................................................................ 64
2.6. Software .......................................................................................................................................... 69
2.7. Conclusion....................................................................................................................................... 69
Chapter 3. Case studies ............................................................................................................................... 71
3.1. Dugald River.................................................................................................................................... 71
3.1.1. Data overview ......................................................................................................................... 72
3.2. Prominent Hill .................................................................................................................................. 74
3.2.1. Data overview ......................................................................................................................... 75
3.3. Westwood........................................................................................................................................ 79
3.3.1. Data overview ......................................................................................................................... 80
3.4. Conclusion....................................................................................................................................... 83

vii
Chapter 4. Article 1 - Assessing stope performance using georeferenced octrees and multivariate analysis
84
4.1. Résumé ........................................................................................................................................... 84
4.2. Abstract ........................................................................................................................................... 84
4.3. Introduction...................................................................................................................................... 85
4.4. Literature review .............................................................................................................................. 87
4.4.1. Quantifying stope OB and UB................................................................................................. 87
4.4.2. Quantifying critical variables for the analysis .......................................................................... 90
4.4.3. Identification of critical variables through statistical analysis .................................................. 91
4.5. Methodology .................................................................................................................................... 95
4.5.1. Data collection ........................................................................................................................ 95
4.5.2. Variable quantification ............................................................................................................ 97
4.5.3. Root cause analysis.............................................................................................................. 102
4.6. Case study .................................................................................................................................... 103
4.6.1. Data overview ....................................................................................................................... 103
4.7. Results .......................................................................................................................................... 105
4.7.1. PCA analysis ........................................................................................................................ 107
4.7.2. PLS analysis ......................................................................................................................... 113
4.8. Discussion ..................................................................................................................................... 114
4.9. Conclusion..................................................................................................................................... 118
4.10. Chapter conclusion........................................................................................................................ 118
Chapter 5. Replicability of the stope OB and UB analysis with two different case studies ......................... 120
5.1. Prominent Hill ................................................................................................................................ 120
5.1.1. PCA analysis ........................................................................................................................ 121
5.1.2. PLS analysis ......................................................................................................................... 125
5.1.3. Discussion ............................................................................................................................ 128
5.2. Westwood...................................................................................................................................... 132
5.2.1. PCA analysis ........................................................................................................................ 134
5.2.2. PLS analysis ......................................................................................................................... 139
5.2.3. Discussion ............................................................................................................................ 142
5.3. Multivariate analysis summary ...................................................................................................... 148
5.4. Conclusion..................................................................................................................................... 152
Chapter 6. Article 2 – Predicting open stope performance at an octree resolution using multivariate models
154

viii
 6.1. Résumé ......................................................................................................................................... 154
6.2. Abstract ......................................................................................................................................... 155
6.3. Introduction.................................................................................................................................... 155
6.4. Literature review ............................................................................................................................ 157
6.5. Methodology .................................................................................................................................. 161
6.5.1. Octree data ........................................................................................................................... 161
6.5.2. Multivariate and machine learning models............................................................................ 162
6.5.3. Statistical evaluation of model performance ......................................................................... 162
6.6. Case study .................................................................................................................................... 164
6.6.1. Data overview ....................................................................................................................... 165
6.7. Results .......................................................................................................................................... 167
6.7.1. Random forest ...................................................................................................................... 170
6.7.2. Statistical performance overview .......................................................................................... 170
6.7.3. Visual example ..................................................................................................................... 171
6.7.4. Model prediction error ........................................................................................................... 172
6.7.5. Probabilistic approach .......................................................................................................... 173
6.7.6. Limitations ............................................................................................................................ 176
6.8. Discussion ..................................................................................................................................... 177
6.9. Summary and conclusion .............................................................................................................. 179
6.10. Dugald River model performance update ...................................................................................... 180
6.10.1. Random forest model ........................................................................................................... 180
6.11. Chapter conclusion........................................................................................................................ 182
Chapter 7. Replicability of stope geometry predictions with two different case studies (Prominent Hill and
Westwood) 184
7.1. Prominent Hill ................................................................................................................................ 184
7.1.1. Random forest ...................................................................................................................... 188
7.2. Westwood...................................................................................................................................... 197
7.2.1. Random forest ...................................................................................................................... 201
7.3. Applicability of the proposed methodology .................................................................................... 210
7.3.1. Improvement on other prediction methods ........................................................................... 211
7.3.2. Site usage of the prediction method ..................................................................................... 212
7.4. Conclusion..................................................................................................................................... 213
Chapter 8. Article 3 - Optimising stope design through economic and geotechnic assessments of predictions made at
a meter scale resolution using the sites' reconciled data ......................................................................................... 215

ix
 8.1. Résumé ......................................................................................................................................... 215
8.2. Abstract ......................................................................................................................................... 216
8.3. Introduction.................................................................................................................................... 216
8.4. Background on stope optimisation ................................................................................................ 218
8.4.1. Stope reconciliation .............................................................................................................. 218
8.4.2. Predicting stope OB and UB ................................................................................................. 222
8.4.3. Economic analysis ................................................................................................................ 224
8.5. Methodology .................................................................................................................................. 225
8.5.1. Octree data ........................................................................................................................... 225
8.5.2. Predictive model ................................................................................................................... 226
8.5.3. Stope performance assessment ........................................................................................... 229
8.5.4. Root cause assessment ....................................................................................................... 234
8.6. Case study .................................................................................................................................... 234
8.6.1. Data overview ....................................................................................................................... 235
8.6.2. Random forest model overview ............................................................................................ 242
8.7. Results .......................................................................................................................................... 244
8.7.1. Geometry analysis ................................................................................................................ 244
8.7.2. Economic analysis ................................................................................................................ 247
8.7.3. Root cause analysis.............................................................................................................. 252
8.7.4. Exploring other designs ........................................................................................................ 255
8.8. Discussion and summary .............................................................................................................. 260
8.9. Conclusion..................................................................................................................................... 261
8.10. Chapter conclusion........................................................................................................................ 261
Conclusion ...................................................................................................................................................... 262
Summary .................................................................................................................................................... 262
Limitations of the thesis .............................................................................................................................. 264
Recommendations for future work .............................................................................................................. 265
References ..................................................................................................................................................... 267
Appendix A: Dugald River PCA analysis charts .............................................................................................. 274
Appendix B: Prominent Hill PCA analysis charts ............................................................................................ 282
Appendix C: Westwood PCA analysis charts.................................................................................................. 290
Appendix D: Prediction results for the Dugald River PLS and LDA models .................................................... 299
Partial least squares model ........................................................................................................................ 299

x
 Linear discriminant analysis model ............................................................................................................. 305
Appendix E: Prediction results for the Prominent Hill PLS and LDA models .................................................. 310
Partial least squares model ........................................................................................................................ 310
Linear discriminant analysis ....................................................................................................................... 316
Appendix F: Prediction results for the Westwood PLS and LDA models ........................................................ 321
Partial least squares model ........................................................................................................................ 321
Linear discriminant analysis ....................................................................................................................... 327
Appendix G: Model prediction summary ......................................................................................................... 330
Statistical summary ................................................................................................................................ 330
Variables summary ................................................................................................................................ 331
Predicted performance summary ........................................................................................................... 332
Appendix H: Catalogue of predicted geometry that meets the objective set in this thesis .............................. 337

xi
List of figures
Chapter 1
Figure 1. 1: Example of stope overbreak and underbreak (Potvin et al. 2016) ................................................... 8
Figure 1. 2: Example of the equivalent linear overbreak sough (ELOS, Clark 1998) ........................................ 10
Figure 1. 3: Illustration of the subdivision of the 3D space and the faces using octrees (modified from
McFadyen et al. 2021) ...................................................................................................................................... 11
Figure 1. 4: Illustration of stope OB and UB quantified at an octree level by calculating the distance, in a
direction normal to the design surface, between the design surface and the CMS (modified from McFadyen et
al. 2021) ............................................................................................................................................................ 11
Figure 1. 5: Example of how stope OB and UB measured at the octree resolution capture the variation of stope
OB and UB across a surface (McFadyen et al. 2021)....................................................................................... 12
Figure 1. 6: Mathews Stability Chart (taken from Mawdesley et al. 2001) ........................................................ 14
Figure 1. 7: Chart for calculating the A, B and C factors for the stability number (Mathews et al. 1981) .......... 15
Figure 1. 8: Stability Chart generated by: left: Potvin 1988, right: Nickson 1992 .............................................. 16
Figure 1. 9: Example of a relaxation zone created by the drives undercut (Diederichs and Kaiser 1999) ........ 26
Figure 1. 10: Example of a cumulative distribution chart plotting overbreak and underbreak (Potvin et al. 2016)
.......................................................................................................................................................................... 27
Figure 1. 11: Example of a bar chart used to analyse stope OB (Capes 2009) ................................................ 28
Figure 1. 12: Example of the use of a moving average (black line) with a box a whisker plot for analysing the
ERF per octree (Woodward et al. 2019) ........................................................................................................... 29
Figure 1. 13: Example of a box and whisker plot comparing different mining pyramids (McFadyen et al. 2019a)
.......................................................................................................................................................................... 29
Figure 1. 14: Example of a bar chart used for analysing RQD (McFadyen et al. 2019a) .................................. 30
Figure 1. 15: Illustration of the PCA model (Dunn 2023a) ................................................................................ 31
Figure 1. 16: Illustration of the PLS model (Dunn 2023b) ................................................................................. 32

Chapter 2
Figure 2. 1: The four steps of the stope design methodology ........................................................................... 37
Figure 2. 2: The first step of the stope design methodology ............................................................................. 38
Figure 2. 3: Examples of a) wrong surface construction; b) Odd CMS shape caused by a muckpile left in the
stope ................................................................................................................................................................. 39
Figure 2. 4: Example of overlapping drill strings generated for the first blast (in orange) and for the rest of the
stope (in black). ................................................................................................................................................ 40
Figure 2. 5: Example of face delimitation using octrees.................................................................................... 42
Figure 2. 6: The second step of the proposed methodology ............................................................................. 49
Figure 2. 7: Example of a loading chart (a) and score chart (b) that oppose two components ......................... 50
Figure 2. 8: Example of a loading chart using PLS ........................................................................................... 51
Figure 2. 9: The third step of the proposed methodology ................................................................................. 52
Figure 2. 10: Example the classification of two groups using linear discriminant analysis (modified from
Tremblay 2020) ................................................................................................................................................. 53
Figure 2. 11: Random forest diagram for generating a prediction ..................................................................... 54

xii
Figure 2. 12: Example of a prediction interval for a stope design (black) according to the probability (blue =
80% probability the projected distances will be larger; green = median case, 50% probability; red = 80%
probability the projected distance will be smaller) ............................................................................................. 55
Figure 2. 13: Example of the analytical comparison between the predicted face and the actual face. The
cumulative distribution represents the projected distance compiled from smallest (purple colours) to largest
(red colours)...................................................................................................................................................... 60
Figure 2. 14: The fourth step of the proposed methodology ............................................................................. 62
Figure 2. 15: Percentage of octrees above the threshold value for the selected percentile geometry within the
probability interval ............................................................................................................................................. 63
Figure 2. 16: Example of the predicted distance for an octree based on the CDF percentile used .................. 64
Figure 2. 17: Example of the grade block model imported around a stope ....................................................... 65
Figure 2. 18: Example of the calculation of the NSR value per octree (a) and an example of the result per
octree block of the NSR value adjusted with backfill information (b)................................................................. 66
Figure 2. 19: Example of the cumulative distribution of the NSR value for each tonne based on the different
types of volume. Tonnes per volume are indicated as a footnote. The black line represents a fictive cut-off
value ................................................................................................................................................................. 67
Figure 2. 20: 3D view of the NSR value for the octrees located in the OB. Octrees with an NSR value below
the fictive cut-off grade are transparent ............................................................................................................ 68

Chapter 3
Figure 3. 1: The location of Dugald River mine, Queensland, Australia............................................................ 71
Figure 3. 2: 3D Section view of the stope database with the trial area delineated in red .................................. 72
Figure 3. 3: Longitudinal view of the transversal stopes in the database (a) and the stopes forming part of the
third group for the model building process (b) .................................................................................................. 73
Figure 3. 4: Example of stope face delimitation using colour code and orientation .......................................... 73
Figure 3. 5: The location of Prominent Hill mine, South Australia, Australia ..................................................... 75
Figure 3. 6: 3D section view of the stope database for Prominent Hill .............................................................. 76
Figure 3. 7: Stopes at Prominent Hill selected for the analysis ......................................................................... 76
Figure 3. 8: Prominent Hill stopes predicted with the models ........................................................................... 77
Figure 3. 9: Structural model for the studied area (a) and an example of the RQD value for a stope near faults
(b) ..................................................................................................................................................................... 78
Figure 3. 10: The location of Westwood mine, Quebec, Canada ...................................................................... 80
Figure 3. 11: 3D section view of the stope database for Westwood ................................................................. 80
Figure 3. 12: Selected stopes for the Westwood analysis ................................................................................ 81
Figure 3. 13: Westwood stopes predicted with the models ............................................................................... 81

Chapter 4
Figure 4. 1: Stope overbreak and underbreak .................................................................................................. 87
Figure 4. 2: Illustration of the recursive process of octrees as well as the octrees defined along the design
surface (modified from McFadyen et al. 2021).................................................................................................. 88
Figure 4. 3: Illustration of octree stope OB and UB quantified by calculating the distance in a direction normal
to the design surface between the design surface and the CMS for each octree (modified from McFadyen et al.
2021)................................................................................................................................................................. 89

xiii
Figure 4. 4: An example of how stope OB and UB can vary across a stope face (McFadyen et al. 2021) ....... 90
Figure 4. 5: Illustration of the principal component analysis (PCA) structure (Dunn 2023a) ............................. 93
Figure 4. 6: Illustration of the partial least squares (PLS) structure (Dunn 2023b). .......................................... 94
Figure 4. 7: The three steps for understanding stope OB and UB as part of the new stope design methodology.
.......................................................................................................................................................................... 95
Figure 4. 8: Examples of a) Bad mesh; b) Odd CMS shape caused by a muck pile left in the stope. .............. 97
Figure 4. 9: Location of Dugald River mine in Queensland, Australia. ............................................................ 103
Figure 4. 10: 3D section view of the Dugald River stope database with the trial area shown in red. .............. 104
Figure 4. 11: Longitudinal view of the Dugald River transversal stopes in the database. ............................... 104
Figure 4. 12: Example of stope face delimitation using colour code and orientation. ..................................... 105
Figure 4. 13: Normalised cumulative count of octrees projected distance for the Dugald River 49 transversal
stopes. ............................................................................................................................................................ 106
Figure 4. 14: Example of an octree UB quality issue identified for large UB in the corners of the stope. ....... 106
Figure 4. 15: Two stope examples of the Dugald River mine site stope OB and UB at an octree level, with CMS
and designed shape overlaid. View 1 side wall does not show any octrees as they are mined against backfill.
........................................................................................................................................................................ 107
Figure 4. 16: Scree Plot of the variance explained for each component of the Dugald River PCA. ................ 108
Figure 4. 17: Component 2 versus component 3 charts: a) Loading chart; b) Score chart coloured by stope OB
or UB............................................................................................................................................................... 111
Figure 4. 18: Loadings chart for the Dugald River PLS analysis. .................................................................... 114
Figure 4. 19: Critical variable values along Dugald River stope A design shape HW and side wall; (a) ERF;
(b) Blasting energy proxy; (c) Blasthole orientation (◦);( d) Undercut (m); (e) Distance to the closest fault (m);
(f) Dip (◦); (g) Projected distance (m). ............................................................................................................. 115
Figure 4. 20: Critical variable values along Dugald River stope’s FW design shape; (a) Undercut (m); (b)
Blasting energy proxy; (c) Projected distance (m). ......................................................................................... 116

Chapter 5
Figure 5. 1: Prominent Hill transversal stopes used for the analysis .............................................................. 120
Figure 5. 2: Normalised cumulative count of octrees’ projected distances for the 45 Prominent Hill transversal
stopes ............................................................................................................................................................. 121
Figure 5. 3: Scree plot of the variance explained for each Prominent Hill PCA component ........................... 122
Figure 5. 4: Component 1 versus component 2 charts. a: loading chart; b: score chart coloured by stope OB
and UB ............................................................................................................................................................ 124
Figure 5. 5: Loadings chart (W* and C* values) of the first component for the Prominent Hill PLS analysis .. 126
Figure 5. 6: Prominent Hill bivariate charts coloured by point density, with a moving average, of the variables
that do not show a linear trend with the projected distance; a) Dip; b) Octree direction; c) RQD; d) Directional
measure to the drift ......................................................................................................................................... 128
Figure 5. 7: Critical variable values for a Prominent Hill stope design shape HW; (a) Stope design and CMS;
(b) Projected distance (m); (c) Blasting energy proxy; (d) Blasthole orientation (◦); (e) Blasthole standoff (m); (f)
ERF; (g) Undercut (m); (h) Dip (◦). .................................................................................................................. 129
Figure 5. 8: Critical variable values along the Prominent Hill stope’s FW design shape; (a) Stope design and
CMS; (b) Projected distance (m); (c) Blasting energy proxy; (d) Blasthole orientation (◦); (e) Blasthole standoff
(m); (f) ERF; (g) Undercut (m); (h) Dip (◦). ...................................................................................................... 130
Figure 5. 9: Westwood stopes used in the analysis ........................................................................................ 132

xiv
Figure 5. 10: Normalised cumulative count of octree projected distance for the 40 longitudinal stopes ......... 134
Figure 5. 11: Scree plot of the variance explained for each Westwood PCA component ............................... 134
Figure 5. 12: Component 1 versus component 4 charts. a: loading chart; b: score chart coloured by OB-UB; C:
score chart coloured by binned projected distance ......................................................................................... 138
Figure 5. 13: Loadings chart (W* and C* values) of the first component for the Westwood PLS analysis ...... 140
Figure 5. 14: Westwood bivariate charts coloured by point density, with a moving average, of the variables that
do not show a linear trend with the projected distance; a) Dip; b) Directional measure to the drift; c) Directional
measure to the fault; d) Distance to the fault; e) RQD; f) GSI; g) UCS ........................................................... 142
Figure 5. 15: Longitudinal view of the Westwood stope database with octree coloured by a) projected distance;
b) blasting standoff. Stopes with large blasting standoff are boxed in red ...................................................... 143
Figure 5. 16: Westwood stope database with octrees coloured by blasting orientation .................................. 144
Figure 5. 17: Geotechnical variables and projected distance values for the selected stopes that are the cause
of the correlations identified in the PCA analysis ............................................................................................ 145
Figure 5. 18: CMS (orange) and faults (green) for the stoping area where large OB and high geotechnical
values are observed........................................................................................................................................ 145
Figure 5. 19: Critical variable values for a Westwood stope design shape HW; (a) Stope design and CMS; (b)
Blasting energy proxy; (c) Distance to fault (m); (d) GSI; (e) UCS contrast; (g) Projected distance (m); (h)
Undercut (m); (i) ERF; (j) RQD; (k) UCS ......................................................................................................... 147
Figure 5. 20: Critical variable values for a Westwood stope design shape FW; (a) Stope design and CMS; (b)
Blasting energy proxy; (c) Distance to fault [m]; (d) GSI; (e) UCS contrast; (g) Projected distance [m]; (h)
Undercut [m]; (i) ERF; (j) RQD; (k) UCS ......................................................................................................... 147

Chapter 6
Figure 6. 1: Illustration of stope overbreak and underbreak with respect to an open stoping mining method
(McFadyen et al., 2023) .................................................................................................................................. 157
Figure 6. 2: Illustration of equivalent linear overbreak slough (Clark 1998) .................................................... 158
Figure 6. 3: Stability Chart (from Nickson, 1992) ............................................................................................ 159
Figure 6. 4: Illustration of the recursive process of octrees, as well as the octrees defined along the design
surface and the stope OB and UB quantified by calculating the distance in a direction normal to the design
surface between the design surface and the CMS for each octree (modified from McFadyen et al., 2020) ... 160
Figure 6. 5: An example of how stope OB and UB can spatially vary along a stope surface (McFadyen et al.,
2021)............................................................................................................................................................... 161
Figure 6. 6: Location of Dugald River mine, Queensland, Australia ............................................................... 165
Figure 6. 7: 3D section view of the Dugald River transversal stope database (a) and stopes part of group 3
used for predicting and testing the models (b) ................................................................................................ 166
Figure 6. 8: Random forest diagram for generating a prediction ..................................................................... 170
Figure 6. 9: Predictions versus observations for the Dugald River random forest model. The points are
coloured by their confusion matrix classification. The ideal and best linear fit of the data is overlaid. ............ 171
Figure 6. 10: Dugald River prediction example. Stope A observed projected distances of the octrees on the
left, the predicted distances in the middle and the absolute prediction error on the right from the random forest
model. The face with no octree (dark green) is a backfill face. ....................................................................... 172
Figure 6. 11: Prediction absolute error versus the projected distance for the Dugald River random forest model
........................................................................................................................................................................ 173

xv
Figure 6. 12: Dugald River prediction example. Observed projected distances for the octrees, the lower and
upper bound of a 60% prediction interval calculated from the random forest model and the octrees identified in
or out of the prediction interval. The face with no octree is a backfill face. ..................................................... 174
Figure 6. 13: The probability of the predicted projected distances versus the prediction error for the Dugald
River predicted octrees using the random forest model.................................................................................. 175
Figure 6. 14: Example of a cluster of octrees large with PDF values (over 0.4) and large prediction errors
(over 2 m). ...................................................................................................................................................... 175
Figure 6. 15: A Dugald River example of an insufficient prediction with the random forest model. The observed
and predicted projected distances are seen at the top with the absolute prediction error and the lower and
upper bounds of the 60% prediction interval are seen at the bottom with the octrees identified in or out of the
reliability interval. Pale blue octrees on the crest of the HW have a distance of 0 m and were not part of the
predictions. The face with no octree is a backfill face. .................................................................................... 176
Figure 6. 16: 3D view of the HW predictions versus observations for the Dugald River random forest model 181
Figure 6. 17: 3D view of the FW predictions versus observations for the Dugald River random forest model 182

Chapter 7
Figure 7. 1: Prominent Hill stopes in group 3 .................................................................................................. 185
Figure 7. 2: 3D view of the HW predictions versus observations for the Prominent Hill random forest model 189
Figure 7. 3: 3D view of the FW predictions versus observations for the Prominent Hill random forest model 190
Figure 7. 4: Predictions versus observations for the Prominent Hill random forest model. The points are
coloured by their confusion matrix classification. The ideal and best linear fit of the data are overlaid .......... 192
Figure 7. 5: Prediction absolute error versus the projected distance for the Prominent Hill random forest model
........................................................................................................................................................................ 192
Figure 7. 6: Prominent Hill stope random forest model example: observed projected distances of the octrees
on the left, the predicted distances in the middle, and the absolute prediction error on the right.................... 193
Figure 7. 7: Prominent Hill stope example. Observed projected distances for the octrees, the lower and upper
bound of a 60% prediction interval calculated from the random forest model and the octrees identified in or out
of the prediction interval. ................................................................................................................................. 194
Figure 7. 8: Prominent Hill example where the predictions are not adequate with the random forest model. The
observed and predicted projected distances can be seen at the top, while the absolute prediction error and the
lower and upper bounds of the 60% prediction interval can be seen at the bottom with the octrees identified in
or out of the prediction interval. The faces with no octrees are backfill faces ................................................. 195
Figure 7. 9: The cumulative distribution chart of the absolute prediction error, grouped by the rounded
likelihood of observing the predicted distance for the Prominent Hill random forest model ............................ 196
Figure 7. 10: Example of octrees (circled in red) with large prediction errors (over 4 m). a) octrees in UB near a
backfilled face (crown); b) octrees in FW OB .................................................................................................. 197
Figure 7. 11: Westwood stopes in group 3 ..................................................................................................... 198
Figure 7. 12: 3D view of the HW predictions versus observations for the Westwood random forest model ... 201
Figure 7. 13: 3D view of the FW predictions versus observations for the Westwood random forest model.... 202
Figure 7. 14: Predictions versus observations for the Westwood random forest model. The points are coloured
by their confusion matrix classification. The ideal and best linear fit of the data are overlaid ......................... 205
Figure 7. 15: Prediction absolute error versus the projected distance for the Westwood random forest model
........................................................................................................................................................................ 205

xvi
Figure 7. 16: Westwood stope random forest model example: observed projected distances of the octrees on
the left, the predicted distances in the middle, and the absolute prediction error on the right ........................ 206
Figure 7. 17: Observed projected distances for the octrees of the Westwood stope: the lower and upper bound
of a 60% prediction interval calculated from the random forest model and the octrees identified in or out of the
prediction interval ............................................................................................................................................ 207
Figure 7. 18: An example of an insufficient prediction with the Westwood random forest model. The observed
and predicted projected distances can be seen at the top, while the absolute prediction error and the lower and
upper bound of the 60% prediction interval can be seen at the bottom with the octrees identified in or out of the
prediction interval. ........................................................................................................................................... 208
Figure 7. 19: The cumulative distribution chart of the absolute prediction error, grouped by the rounded
likelihood of observing the predicted distance for the Westwood random forest model.................................. 209
Figure 7. 20: 3D view of the Westwood stopes (circled in blue) with large prediction errors (over 4 m) ......... 210
Figure 7. 21: Side view of the Westwood analysed stope with predictions that are inadequate, as well as
adjacent stopes and faults .............................................................................................................................. 210
Figure 7. 22 : Face example of the different predictions using different prediction tools for the HW .............. 212

Chapter 8
Figure 8. 1:Stope overbreak and underbreak with respect to a stoping mining method (McFadyen et al. 2023a)
........................................................................................................................................................................ 219
Figure 8. 2: Illustration of the recursive process of octrees as well as the octrees defined along the design
surface (modified from McFadyen et al. 2021)................................................................................................ 221
Figure 8. 3: Illustration of octree stope OB and UB quantified by calculating the distance in a direction normal
to the design surface between the design surface and the CMS for each octree (modified from McFadyen et al.
2021)............................................................................................................................................................... 222
Figure 8. 4: The proposed four steps for assessing the predicted stope performance as part of the stope
design methodology ........................................................................................................................................ 225
Figure 8. 5: An example of how stope OB and UB can spatially vary along a stope surface (McFadyen et al.
2021)............................................................................................................................................................... 226
Figure 8. 6: Example of a CMS (left) and the predicted mined geometry (right) using a random forest model227
Figure 8. 7: Random forest diagram for generating a prediction ..................................................................... 228
Figure 8. 8: Example of different prediction geometries for a stope design shape (black) according to
probabilities. The blue, green and red surfaces are the 20th, 50th and 80th percentile cases of the prediction
interval, respectively ....................................................................................................................................... 228
Figure 8. 9: Example of the predicted distance for an octree based on the CDF and percentile considered.. 230
Figure 8. 10: Example of the grade block model imported around a stope displaying the NSR value per block
........................................................................................................................................................................ 230
Figure 8. 11: Example of the calculation of the NSR value per octree (left) and an example of the result per
octree block of the NSR value adjusted with backfill information (right).......................................................... 231
Figure 8. 12: Example of the cumulative distribution of the NSR value for each tonne based on the different
categories of volume. Total tonnes per volume are indicated as a footnote and the cut-off value with the black
line .................................................................................................................................................................. 232
Figure 8. 13: 3D view of the NSR value for the OB octrees. Points with an NSR value below the selected cut-
off grade are semitransparent while points above the cut-off grade are opaque ............................................ 233
Figure 8. 14: Location of Prominent Hill mine, South Australia, Australia ....................................................... 235

xvii
Figure 8. 15: 3D section view of the Prominent Hill stope database ............................................................... 236
Figure 8. 16: Prominent Hill first model and prediction database .................................................................... 243
Figure 8. 17: Stopes used for the Prominent Hill model database .................................................................. 243
Figure 8. 18: 3D view of Prominent Hill stope A; a) The design geometry with the drives; b) The octree data
structure coloured per face ............................................................................................................................. 244
Figure 8. 19: The three predicted geometries (20th, 50th and 80th percentiles of the prediction interval) for
Prominent Hill stope A .................................................................................................................................... 245
Figure 8. 20: The three predicted geometries per octree (20th, 50th and 80th percentiles of the prediction
interval) for Prominent Hill stope A ................................................................................................................. 246
Figure 8. 21: 3D view of the probability of exceeding a 0 m or 2 m threshold, and a cumulative distribution
chart of the percentage of octrees above the threshold value for the probability geometry (CDF percentile) for
Prominent Hill stope A .................................................................................................................................... 247
Figure 8. 22: Cumulative distribution of the tonnes in each volume type based on the NSR value for the three
different geometries (20th, 50th and 80th percentiles of the prediction interval) of Prominent Hill stope A. The
mined volume represents the predicted mined ............................................................................................... 251
Figure 8. 23: NSR value per octree and per volume for the three different geometries (20th, 50th and 80th
percentiles of the prediction interval) of Prominent Hill stope A. A cut-off value of $50/tonne is set. Octrees
below the cut-off value are semitransparent ................................................................................................... 252
Figure 8. 24: Variables ordered by variable importance calculated from the Prominent Hill random forest model
........................................................................................................................................................................ 253
Figure 8. 25: Multiple 3D views of the design octrees of Prominent Hill stope A coloured with the different
critical variables .............................................................................................................................................. 255
Figure 8. 26: Original and alternative blast designs proposed for optimisation analysis of Prominent Hill stope A
........................................................................................................................................................................ 256
Figure 8. 27: Original and alternative projected distance and economic predictions for optimisation analysis of
Prominent Hill stope A .................................................................................................................................... 257

Appendix A
Figure A. 1: Component 1 versus component 2 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 274
Figure A. 2: Component 1 versus component 3 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 274
Figure A. 3: Component 1 versus component 4 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 275
Figure A. 4: Component 1 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 275
Figure A. 5: Component 1 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 276
Figure A. 6: Component 2 versus component 3 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 276
Figure A. 7: Component 2 versus component 4 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 277
Figure A. 8: Component 2 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 277

xviii
Figure A. 9: Component 2 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 278
Figure A. 10: Component 3 versus component 4 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 278
Figure A. 11: Component 3 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 279
Figure A. 12: Component 3 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 279
Figure A. 13: Component 4 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 280
Figure A. 14: Component 4 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 280
Figure A. 15: Component 5 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 281

Appendix B
Figure B. 1: Component 1 versus component 2 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 282
Figure B. 2: Component 1 versus component 3 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 282
Figure B. 3: Component 1 versus component 4 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 283
Figure B. 4: Component 1 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 283
Figure B. 5: Component 1 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 284
Figure B. 6: Component 2 versus component 3 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 284
Figure B. 7: Component 2 versus component 4 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 285
Figure B. 8: Component 2 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 285
Figure B. 9: Component 2 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 286
Figure B. 10: Component 3 versus component 4 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 286
Figure B. 11: Component 3 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 287
Figure B. 12: Component 3 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 287
Figure B. 13: Component 4 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 288
Figure B. 14: Component 4 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 288

xix
Figure B. 15: Component 5 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart
coloured by stope OB or UB ........................................................................................................................... 289

Appendix C
Figure C. 1: Component 1 versus component 2 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 290
Figure C. 2: Component 1 versus component 3 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 290
Figure C. 3: Component 1 versus component 4 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 290
Figure C. 4: Component 1 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 291
Figure C. 5: Component 1 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 291
Figure C. 6: Component 1 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 291
Figure C. 7: Component 1 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 292
Figure C. 8: Component 2 versus component 3 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 292
Figure C. 9: Component 2 versus component 4 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 292
Figure C. 10: Component 2 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 293
Figure C. 11: Component 2 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 293
Figure C. 12: Component 2 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 293
Figure C. 13: Component 2 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 294
Figure C. 14: Component 3 versus component 4 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 294
Figure C. 15: Component 3 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 294
Figure C. 16: Component 3 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 295
Figure C. 17: Component 3 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 295
Figure C. 18: Component 3 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 295
Figure C. 19: Component 4 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 296
Figure C. 20: Component 4 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 296

xx
Figure C. 21: Component 4 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 296
Figure C. 22: Component 4 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 296
Figure C. 23: Component 5 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 297
Figure C. 24: Component 5 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 297
Figure C. 25: Component 5 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 297
Figure C. 26: Component 6 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 297
Figure C. 27: Component 6 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 298
Figure C. 28: Component 7 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart
coloured by OB-UB; c: score chart coloured by binned projected distance .................................................... 298

Appendix D
Figure D. 1: Stopes in group 3 used for predicting and testing the Dugald River models ............................... 299
Figure D. 2: Predictions versus observations for the Dugald River PLS model. The points are coloured by their
confusion matrix classification. The ideal and best linear fit of the data are overlaid ...................................... 300
Figure D. 3: Prediction absolute error versus the projected distance for the Dugald River PLS model .......... 301
Figure D. 4: Dugald River stope PLS model example: observed projected distances of the octrees on the left,
the predicted distances in the middle, and the absolute prediction error on the right. The face with no octree is
a backfill face .................................................................................................................................................. 302
Figure D. 5: 3D view of the HW predictions versus observations for the Dugald River PLS model ................ 303
Figure D. 6: 3D view of the FW predictions versus observations for the Dugald River PLS model ................ 304
Figure D. 7: An example stope of an insufficient prediction with the Dugald River PLS model. The observed,
predicted projected distances and the absolute prediction error can be seen. Pale blue octrees on the crest of
the HW have a distance of 0 m and were not part of the predictions. The face with no octrees is a backfill face
........................................................................................................................................................................ 305
Figure D. 8: Stope A observed OB and UB of the octrees on the left, the predicted OB and UB in the middle,
and the prediction error on the right using the LDA model. The face with no octrees is a backfill face .......... 306
Figure D. 9: 3D view of the HW predictions versus observations for the Dugald River LDA model ................ 307
Figure D. 10: 3D view of the FW predictions versus observations for the Dugald River LDA model .............. 308
Figure D. 11: An example of an insufficient prediction with the LDA model. The observed and predicted OB
and UB can be seen as well as the prediction error. Pale blue octrees on the crest of the HW were not part of
the predictions. The face with no octrees is a backfill face ............................................................................. 309

Appendix E
Figure E. 1: Prominent Hill stopes in group 3.................................................................................................. 310
Figure E. 2: Predictions versus observations for the Prominent Hill PLS model. The points are coloured by their
confusion matrix classification. The ideal and best linear fit of the data are overlaid ...................................... 311

xxi
Figure E. 3: Prediction absolute error versus the projected distance for the Prominent Hill PLS model. Outliers
circled in red ................................................................................................................................................... 312
Figure E. 4: Prominent Hill stope PLS model example: observed projected distances of the octrees on the left,
the predicted distances in the middle, and the absolute prediction error. The face with no octrees is a backfill
face ................................................................................................................................................................. 313
Figure E. 5: 3D view of the HW predictions versus observations for the Prominent Hill PLS model............... 314
Figure E. 6: 3D view of the FW predictions versus observations for the Prominent Hill PLS model ............... 315
Figure E. 7: Prominent Hill stope example where the predictions were not adequate with the PLS model. The
observed, predicted projected distances and the absolute prediction error can be seen. The faces with no
octrees are backfill faces ................................................................................................................................ 316
Figure E. 8: Prominent Hill stope LDA model example: observed OB and UB of the octrees on the left, the
predicted OB and UB in the middle and the prediction error on the right. The face with no octrees is a backfill
face ................................................................................................................................................................. 317
Figure E. 9: 3D view of the HW predictions versus observations for the Prominent Hill LDA model .............. 318
Figure E. 10: 3D view of the FW predictions versus observations for the Prominent Hill LDA model ............. 319
Figure E. 11: Prominent Hill example where the predictions were not adequate with the LDA model. The
observed and predicted OB and UB can be seen as well as the prediction error. The face with no octrees is a
backfill face ..................................................................................................................................................... 320

Appendix F
Figure F. 1: Westwood stopes in group 3 ....................................................................................................... 321
Figure F. 2: Predictions versus observations for the Westwood PLS model. The points are coloured by their
confusion matrix classification. The ideal and best linear fit of the data are overlaid ...................................... 322
Figure F. 3: Prediction absolute error versus the projected distance for the Westwood PLS model............... 323
Figure F. 4: Westwood PLS model prediction example: observed projected distances of the octrees on the left,
the predicted distances in the middle, and the absolute prediction error on the right. The face with no octrees is
a backfill face .................................................................................................................................................. 324
Figure F. 5: 3D view of the HW predictions versus observations for the Westwood PLS model .................... 325
Figure F. 6: 3D view of the FW predictions versus observations for the Westwood PLS model ..................... 326
Figure F. 7: Westwood PLS prediction example where the predictions were not adequate. The observed,
predicted projected distances and the absolute prediction error can be seen. The faces with no octrees are
backfill faces ................................................................................................................................................... 327
Figure F. 8: Side view of the Westwood analysed stope with predictions that are inadequate as well as
adjacent stopes and faults .............................................................................................................................. 327
Figure F. 9: 3D view of the LDA prediction error for the predicted Westwood stopes..................................... 328

Appendix G
Figure G. 1 : Cumulative distribution of the 10th to 90 th percentile prediction interval (halved interval distance is
used) for all the predicted octrees for the three mine sites ............................................................................. 335

Appendix H
Figure H. 1: Example no1 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxii
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 337
Figure H. 2: Example no2 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 337
Figure H. 3: Example no3 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 338
Figure H. 4: Example no4 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 338
Figure H. 5: Example no5 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 339
Figure H. 6: Example no6 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 339
Figure H. 7: Example no7 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 340
Figure H. 8: Example no8 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 340
Figure H. 9: Example no9 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 341
Figure H. 10: Example no10 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 341
Figure H. 11: Example no11 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 342
Figure H. 12: Example no12 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxiii
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 342
Figure H. 13: Example no13 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 343
Figure H. 14: Example no14 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 343
Figure H. 15: Example no15 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 344
Figure H. 16: Example no16 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 344
Figure H. 17: Example no17 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 345
Figure H. 18: Example no18 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 345
Figure H. 19: Example no19 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 346
Figure H. 20: Example no20 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 346
Figure H. 21: Example no21 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 347
Figure H. 22: Example no22 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 347
Figure H. 23: Example no23 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxiv
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 348
Figure H. 24: Example no24 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 348
Figure H. 25: Example no25 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 349
Figure H. 26: Example no26 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 349
Figure H. 27: Example no27 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 350
Figure H. 28: Example no28 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 350
Figure H. 29: Example no29 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 351
Figure H. 30: Example no30 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 351
Figure H. 31: Example no31 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 352
Figure H. 32: Example no32 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 352
Figure H. 33: Example no33 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 353
Figure H. 34: Example no34 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxv
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 353
Figure H. 35: Example no35 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 354
Figure H. 36: Example no36 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 354
Figure H. 37: Example no37 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 355
Figure H. 38: Example no38 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 355
Figure H. 39: Example no39 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 356
Figure H. 40: Example no40 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 356
Figure H. 41: Example no41 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 357
Figure H. 42: Example no42 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 357
Figure H. 43: Example no43 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 358
Figure H. 44: Example no44 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 358
Figure H. 45: Example no45 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxvi
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 359
Figure H. 46: Example no46 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 359
Figure H. 47: Example no47 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 360
Figure H. 48: Example no48 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 360
Figure H. 49: Example no49 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 361
Figure H. 50: Example no50 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 361
Figure H. 51: Example no51 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 362
Figure H. 52: Example no52 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 362
Figure H. 53: Example no53 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 363
Figure H. 54: Example no54 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 363
Figure H. 55: Example no55 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 364
Figure H. 56: Example no56 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxvii
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 364
Figure H. 57: Example no57 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 365
Figure H. 58: Example no58 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 365
Figure H. 59: Example no59 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 366
Figure H. 60: Example no60 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 366
Figure H. 61: Example no61 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 367
Figure H. 62: Example no62 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 367
Figure H. 63: Example no63 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 368
Figure H. 64: Example no64 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 368
Figure H. 65: Example no65 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 369
Figure H. 66: Example no66 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 369
Figure H. 67: Example no67 of a predicted geometry that meets the objective set in this thesis. A side-by-side
comparison between the observation and predictions is shown for the projected distance and cumulative

xxviii
distribution of the projected distance. The projected distance for the prediction interval is shown (10th and 90th
percentile) as well as the prediction error ....................................................................................................... 370

xxix
List of tables
Chapter 1
Table 1. 1: Other measures of stope performance ............................................................................................. 9
Table 1. 2: List of variables considered in the different predictive models ........................................................ 18
Table 1. 3: Local characterisation of the stope geometry for per octree analysis (modified from Woodward et al.
2019)................................................................................................................................................................. 20
Table 1. 4: Local characterisation of the geological structures for per octree analysis (modified from Woodward
et al. 2019) ........................................................................................................................................................ 23
Table 1. 5: Local characterisation of the drilling and blasting for per octree analysis (modified from Woodward
et al. 2019) ........................................................................................................................................................ 24
Table 1. 6: Different dilution and loss of ore variables ...................................................................................... 34
Table 1. 7: List of variables influencing dilution compiled by Villaescusa 2014 ................................................ 35

Chapter 2
Table 2. 1: Collected data ................................................................................................................................. 38
Table 2. 2: Local characterisation of stope design variables for per octree analysis (modified from Woodward
et al. 2019) ........................................................................................................................................................ 43
Table 2. 3: Confusion matrix for a threshold of 0 m projected distance. Green represents true predictions and
red represents false predictions (errors) ........................................................................................................... 54
Table 2. 4: Performance metric equations for evaluating the model’s capacity to classify an octree as OB or
UB. Refer to Table 2.3 for the acronyms .......................................................................................................... 56
Table 2. 5: Prediction error brackets for evaluating the model’s statistical performance .................................. 57
Table 2. 6: Different dilution and loss of ore variables used in this thesis ......................................................... 68

Chapter 3
Table 3. 1: Quantified variables for Dugald River mine..................................................................................... 74
Table 3. 2: Quantified variables for Prominent Hill ............................................................................................ 79
Table 3. 3: Quantified variables for Westwood ................................................................................................. 83

Chapter 4
Table 4. 1: Collected data ................................................................................................................................. 96
Table 4. 2: Local characterisation of stope design variables for per octree analysis (modified from Woodward
et al. 2019) ........................................................................................................................................................ 98
Table 4. 3: Variance explained per Dugald River PCA component. ............................................................... 108
Table 4. 4: Contribution, in percentage, of each variable to the variance explained by each component for
Dugald River. The strongest contributing variables for each component are identified in red. The coordinate
sign of each variable for each component are in parentheses. ....................................................................... 109
Table 4. 5: Observations made on the relationships between the variables for Dugald River. ....................... 110
Table 4. 6: Stope OB and UB summary from the Dugald River PCA components. ........................................ 112
Table 4. 7: PLS coefficients of the Dugald River variables for linear regression............................................. 114

xxx
Chapter 5
Table 5. 1: Variables selected for the Prominent Hill mine site root cause analysis ....................................... 121
Table 5. 2: Variance explained per Prominent Hill PCA component ............................................................... 122
Table 5. 3: Contribution, in percentage, of each variable to the variance explained by each Prominent Hill PCA
component. The strongest contributing variables for each component are identified in red (over 10%). The
coordinate sign of each variable for each component are in parentheses ...................................................... 123
Table 5. 4: Observations made on the relationships between the Prominent Hill variables............................ 123
Table 5. 5: Stope OB and UB summary from the Prominent Hill PCA components ....................................... 125
Table 5. 6: PLS coefficients of the Prominent Hill variables for linear regression ........................................... 127
Table 5. 7: Variables selected for the Westwood mine site root cause analysis ............................................. 133
Table 5. 8: Variance explained per Westwood PCA component .................................................................... 135
Table 5. 9: Contribution, in percentage, of each variable to the variance explained by each Westwood PCA
component. The strongest contributing variables for each component are identified in red (over 10%). The
coordinate sign of each variable for each component is in parentheses......................................................... 136
Table 5. 10: Observations made on the relationships between the Westwood variables ............................... 137
Table 5. 11: Stope OB and UB summary from the Westwood PCA components ........................................... 139
Table 5. 12: PLS coefficients of the Westwood variables for linear regression............................................... 141
Table 5. 13: Summary of the PCA .................................................................................................................. 149
Table 5. 14: Summary of the analysis between the variables and the stope OB and UB ............................... 150

Chapter 6
Table 6. 1: Confusion matrix for a threshold of 0 m projected distance. Green represents true predictions and
red represents false predictions (errors). ........................................................................................................ 163
Table 6. 2: Performance metric equations for evaluating the model’s capacity to classify an octree as OB or
UB. Refer to Table 6.1 for the acronyms. ....................................................................................................... 163
Table 6. 3: Prediction error brackets for evaluating the model’s statistical performance ................................ 164
Table 6. 4: Dugald River critical variables as established in McFadyen et al., 2023....................................... 167
Table 6. 5: Summary of the performance metrics for the Dugald River different models ................................ 168
Table 6. 6: Variables ordered by importance in the Dugald River partial least squares and random forest
models ............................................................................................................................................................ 169
Table 6. 7: Dugald River random forest model performance at correctly predicting OB and UB grouped per
projected distance ........................................................................................................................................... 178
Table 6. 8: Summary of the model’s performance at determining the OB and UB location and magnitude
distribution within a face.................................................................................................................................. 180

Chapter 7
Table 7. 1: Critical variables for the Prominent Hill mine site used to build the predictive models .................. 185
Table 7. 2: Summary of the performance metrics for the Prominent Hill models ............................................ 186
Table 7. 3: Summary of the models’ performance at determining the OB and UB location and magnitude
distribution within a face for the Prominent Hill models................................................................................... 187
Table 7. 4: Variables ordered by importance in the Prominent Hill PLS and random forest models ............... 188
Table 7. 5: Prominent Hill random forest model performance at correctly predicting OB and UB grouped per
projected distance ........................................................................................................................................... 191

xxxi
Table 7. 6: Critical variables for the Westwood mine site used to build the predictive models ....................... 198
Table 7. 7: Summary of the performance metrics for the various Westwood models ..................................... 199
Table 7. 8: Summary of the models’ performance at determining the OB and UB locations and magnitude
distribution within a face for the Westwood models ........................................................................................ 200
Table 7. 9: Variables ordered by importance in the PLS and random forest model for Westwood ................. 200
Table 7. 10: Westwood random forest model performance at correctly predicting OB and UB grouped per
projected distance ........................................................................................................................................... 203
Table 7. 11: Westwood random forest model’s performance at correctly predicting over or under 2 m OB
according to the rounded projected distance .................................................................................................. 204

Chapter 8
Table 8. 1: Variables quantified on a per face basis ....................................................................................... 220
Table 8. 2: Different dilution and loss of ore variables .................................................................................... 233
Table 8. 3: Critical variables established through root cause analysis with a description (modified from
Woodward et al. 2019 and McFadyen et al. 2023a) ....................................................................................... 237
Table 8. 4: Prominent Hill predicted per stope summary for the geometries’ 20th, 50th and 80th percentiles of the
prediction interval ............................................................................................................................................ 245
Table 8. 5: Dilution and loss of ore summary for the three different geometries of the prediction interval and the
mined geometry for Prominent Hill stope A..................................................................................................... 248
Table 8. 6: Per stope summary of the NSR values per volume type for the predictions and mined geometry of
Prominent Hill stope A. Profit is calculated according to Equation 8.5 ............................................................ 249
Table 8. 7: Per face summary of the NSR values for the FW and HW OB and UB for the predictions and mined
geometry of Prominent Hill stope A. Profit is calculated according to Equation 8.5 ........................................ 250
Table 8. 8: Comparison of the predicted performance between the original and alternative design for
Prominent Hill stope A .................................................................................................................................... 259

Appendix A
No tables.

Appendix B
No tables.

Appendix C
No tables.

Appendix D
Table D. 1: Results of the performance metrics of the Dugald River PLS model. These metrics do not consider
the octrees with a projected distance larger than -4 m or 6 m ........................................................................ 300
Table D. 2: Results for the performance metrics of the Dugald River LDA model. These metrics do not
consider the octrees with a projected distance larger than -4 m or 6 m .......................................................... 305

xxxii
Appendix E
Table E. 1: Results for the performance metrics of the Prominent Hill PLS model. These metrics do not
consider the octrees with a projected distance larger than -4 m or 6 m .......................................................... 311
Table E. 2: Results for the performance metrics of the Prominent Hill LDA model. These metrics do not
consider the octrees with a projected distance larger than -4 m or 6 m .......................................................... 316

Appendix F
Table F. 1: Results for the performance metrics of the Westwood PLS model. These metrics do not consider
the octrees with a projected distance larger than -4 m or 6 m ........................................................................ 323
Table F. 2: Results for the performance metrics of the Westwood LDA model. These metrics do not consider
the octrees with a projected distance larger than -4 m or 6 m ........................................................................ 328
Table F. 3: Results for the performance metrics of the second Westwood LDA model (over/under 2 m
separation). These metrics do not consider the octrees with a projected distance larger than -4 m or 6 m ... 329

Appendix G
Table G. 1: Summary of the models’ prediction performance for the three mine sites (DR = Dugald River, PH =
Prominent Hill, WW = Westwood). Performance metrics defined below......................................................... 331
Table G. 2: Summary of the model variables’ importance for the three mine sites (DR = Dugald River, PH =
Prominent Hill, WW = Westwood) ................................................................................................................... 332
Table G. 3 : Summary of the models’ performance at determining the OB and UB location and magnitude
distribution within a face for the three mine sites (DR = Dugald River, PH = Prominent Hill, WW = Westwood)
........................................................................................................................................................................ 334

Appendix H
No tables.

xxxiii
List of abbreviations, symbols and acronyms
Abbreviations,
symbols and Definition
acronyms
° Degrees
% Percentage
ACC Accuracy of the Classification
ACG Australian Centre for Geomechanics
Ag Silver
Au Gold
CDF Cumulative Density Function
CMS Cavity Monitoring System
Cu Copper
ELOS Equivalent Linear Overbreak Slough
ELLO Equivalent Linear Lost Ore
ERF Effective Radius Factor
FOB False Overbreak
FUB False Underbreak
FW Footwall
GSI Geological Strength Index
HW Hangingwall
LDA Linear Discriminant Analysis
LiDAR Light Detection and Ranging
LOM Life of Mine
LTP Long-Term Plan
m Metre
mm Millimetre
MCC Matthews Correlation Coefficient
MLR Multiple Linear Regression
MWD Measurement-While-Drilling
NSR Net Smelting Return
OB Overbreak
Pb Lead
PCA Principle Components Analysis
PCR Principal Component Regression
PDF Probability Density Function
PLS Partially Least Squares

xxxiv
 Abbreviations,
symbols and Definition
acronyms
POP Positive Overbreak Predictions
R² Coefficient of Determination
RMR Rock Mass Rating System
RMSE Root Mean Square Error
RQD Rock Quality Designation
STP Short-Term Plan
TOB True Overbreak
TOD True Overbreak Detection
TUB True Underbreak
TUD True Underbreak Detection
UB Underbreak
UCS Uniaxial Compressive Strength
VIP Variable Importance Plot
XRF X-Ray Fluorescence
Zn Zinc

xxxv
Acknowledgement
I would like to thank foremost my supervisor Martin Grenon and co-supervisor Yves Potvin for their confidence
in me as well as their advice and encouragement that made this PhD a pleasant and constructive experience.
I would like to thank the ACG for the technical and financial support for my PhD as well as being gracious hosts
during my stay in Perth. More specifically, I would like to thank Christine Neskudla and Stefania Woodward for
all their help. I would like to thank the mXrap team for their constructive comments and their help with the
software – more precisely Kyle Woodward, Johan Wesseloo, Paul Harris, Matt Heinsen Egan, Liam Niedzielski
and Daniel Cumming-Potvin. A special thank you to Kyle Woodward for his contribution to the project with the
development of the apps. I thank him also, as well as his wife Stefania Woodward, and Daniel Cumming-Potvin
and his wife Jemma for their hospitality, making the stay in Perth comfortable and memorable.

I would like to thank the sponsors. This research would not be possible without the data from the three industry
sponsors and the ongoing collaboration from mine sites personnel. The sponsors were IAMGOLD Corporation,
Westwood Mine; MMG Limited, Dugald River Mine; and BHP, Prominent Hill mine. I gratefully acknowledge both
the corporate and individual support.

I would like to thank my family for their support, advice and the opportunities that they have given me.

I would like finally, and most importantly, to thank my wife, Catherine Kotiuga, and my two dogs, Laika and
Wilson, for all the love and support they have given me during this PhD.

xxxvi
Foreword
The objective of this thesis is to present an innovative and new approach for predicting stope performance,
which goes beyond existing prediction tools. Given the novelty of this work, the research was documented in
three articles to demonstrate different aspects of the proposed methodology and facilitate the sharing of
knowledge acquired during this doctoral project. These articles are the result of a close collaboration between
the following researchers:

Martin Grenon, ing. PhD, Université Laval, Québec, Canada.

Kyle Woodward, PhD, Australian Centre for Geomechanics, Perth, Australia.

Yves Potvin, PhD, Australian Centre for Geomechanics, Perth, Australia.

The three articles are integrated in this thesis in the form of chapters and are as follows:

Chapter 4: “Assessing stope performance using georeferenced octrees and multivariate analysis”
Authors: Benoît McFadyen (first author), Martin Grenon, Kyle Woodward and Yves Potvin
Journal: Mining Technology
[Link]
Article history:
Submitted: June 2022
Accepted: April 2023
Published: May 2023

Chapter 6: “Predicting open stope performance at an octree resolution using multivariate models”
Authors: Benoît McFadyen (first author), Martin Grenon, Kyle Woodward and Yves Potvin
Journal: The Journal of the Southern African Institute of Mining and Metallurgy
[Link]
Article history:
Submitted: May 2023
Accepted: June 2023
Published: June 2023

Chapter 9: “Optimising stope design through economic and geotechnic assessments of predictions made at
a meter scale resolution using the sites' reconciled data”
Authors: Benoît McFadyen (first author), Martin Grenon, Kyle Woodward and Yves Potvin

xxxvii
Journal: The International Journal of Rock Mechanics and Mining Science
[Link]
Article history:
Submitted: October 2023
Accepted: May 2024
Published: June 2024

For all three articles, no modifications other than the format and minor corrections have been made between the
versions included in this thesis and the published versions. As for the work put into these articles, I determined
the scope of each article, compiled the databases, conducted the analyses, produced the results, and redacted
and revised the articles. All three coauthors participated in the revision. Kyle Woodward was involved with the
computer programming of the tools used for the analysis in these articles.

xxxviii
Introduction
Background
Underground mines are designed to extract orebodies lying at depth outside the reach of open pit mines, with
the objectives being to ensure the safety of the employees and maximise profits. To determine how the ore will
be extracted, the orebody geometry, its geological environment and the geomechanical properties of the rock
mass are considered to assess the cost and safety risks of each mining method. Open stoping has become a
popular mining method in hard rock mines due to the high extraction rate and relatively low costs (Potvin and
Hadjigeorgiou 2001), in addition to the safety of the method as a non-entry approach. Mining open stopes
consists of first subdividing the ore lens into vertical stopes which are accessed using drives and crosscuts.
Then, borehole rings are drilled in which explosives are inserted to break the rock and create a muckpile for
extraction at the bottom of the stope. Once the ore extraction has been completed, the open stope is generally
backfilled to facilitate the extraction of adjacent stopes and continue the stope mining sequence. The profitability
of the stopes will depend on the engineer’s capacity to maximise the designed volume while respecting the
designed geometry when executing the mining plan. Any inaccuracy in the execution, uncertainty in the data or
constraints of the rock mass environment can result in overbreak (OB, volume of rock mined outside of what
was designed) or underbreak (UB, planned volume of rock to be mined left behind). OB and UB volumes
generally have negative economic and operational consequences as OB can cause dilution, rehabilitation costs,
loss of equipment or delays (Potvin and Hadjigeorgiou 2001). Dilution (waste rock OB) can impact several steps
of the mining process as it increases mucking time and the efficiency of the processing at the plant. In cases
where excessive OB is causing stability issues, the stope can be prematurely backfilled leaving ore behind. For
UB, ore can be left unbroken and unrecovered, reducing the revenue from the stope. Therefore, to minimise the
OB and UB volumes, variables characterising the geological and mining environment are quantified and used
when designing and mining stopes.

The stope design process consists of four main steps (Potvin et al. 2020) which are completed at different
stages of the mining project. These steps are:

• Life of mine stope planning Feasibility stage

• Stope design
• Operation and execution Operations stage
• Reconciliation

1
The life of mine step is initiated during the feasibility stage of the mine when a generic stope shape and
dimensions need to be set which are later updated periodically. The extraction sequence can be established as
well as the interval between levels. This step is usually accomplished using tools such as the Stability Chart
(Mathews et al. 1981, Potvin 1988, Nickson 1992) where the geomechanical properties of the rock mass and
the geometry of the stope faces are considered to qualitatively estimate the stability of the stope faces (stable,
unstable or caving).

The stope design step occurs closer in time to stope extraction once the mine is in production. At this stage,
additional information from the engineering, geological and planning departments are available and considered
to refine the design. This includes drilling and blasting, ground support, geological structures, mucking, backfill
and adjacent stopes’ OB and UB. The stope faces’ stability can be reassessed using the Stability Chart and new
geotechnical data. From this step, the stope geometry is refined and a final design is approved for mining.

The operation and execution step consists of all the necessary steps for extracting the ore: drilling, blasting
mucking and backfilling. Ideally, during these steps, data is compiled to enable reconciliation.

The final step is reconciliation. This consists of comparing what was mined (obtained using a cavity monitoring
system (CMS) scan of the empty stope [Miller et al. 1992]) with what was planned using the data compiled during
the previous steps. This is usually applied on a per stope and per face basis with recent work providing the
potential to go to an even finer metre-scale resolution, referred to as octree in this thesis (Woodward et al. 2019).
The stope’s performance can be summarised according to each department’s interest (grade, blasting or
stability) and various variables can be quantified to try to understand the root cause of OB and UB. The
information gathered from this step provides feedback into the previous steps for the optimisation of future stope
design.

In reality, based on two workshops conducted in 2019 (one in Canada and one in Australia) the mine sites collect
a large quantity of data related to stope design, extraction and reconciliation, but are conducting limited work in
transforming that data into information (quantifying variables) and even less work in transforming that information
into knowledge (creating design tools) (Potvin et al. 2020). The main reason given by the mine sites is that there
is no method which allows one to consider the large dataset and transform it into predictive tools, making data
manipulation a time-consuming process with no set path to help obtain the required knowledge for stope
optimisation (Potvin et al. 2020). Therefore, the industry still relies on generic empirical charts such as the
Stability Chart during the stope design step, using only a fraction of information collected from their reconciled
data. The Stability Chart provides limited opportunity for stope optimisation as it does not consider UB or any
operational variables that can be modified at the stope design step to minimise OB.

2
The limited impact of the Stability Chart at the stope design step, combined with the limited useable information
generated during the reconciliation step, has highlighted a need in the mining industry for a new stope design
method that can easily extract information from an operational mine site’s data to build knowledge on the root
causes of OB and UB and convert it into a predictive model that is used as feedback to the stope design stage
for predicting and optimising future stopes’ performance.

Problem definition
The approach to designing an open stope in underground mines has seen little change since the introduction of
the Stability Chart by Mathews et al. in 1981. Operational mines rely heavily on the chart, or a version of it, to
predict stope performance during stope design and integrate little data from their reconciliation. The stope design
and reconciliation process shortcomings offer limited insight into the root causes of OB and UB as well as the
ability to develop predictive tools to optimise the design, once reconciliation data becomes available. Current
rock mechanics and mining experiences have identified a wide range of variables that can impact OB and UB
that goes beyond the Stability Chart assessments of stope OB. These variables range from the
three-dimensional geometry of the design and the geomechanical properties of the rock mass, to the major
geological structures (faults, shear zones) and operational variables such as drilling and blasting.

Since the 1990s, great technological advances have been made, allowing mine sites to collect a lot more data.
The increased computational power of computers allows for the handling of large databases with many variables,
enabling the development of new reconciliation tools working at a finer resolution. This allows to characterise
geomechanical, geometrical, geological, operational and performance variables, and to consider their spatial
variation along the design surfaces. These variables can be used to understand OB and UB as well as their
spatial variation. They can also be used to develop new stope design tools for predicting and optimising stope
performance. The increased computational power has also enabled to run multivariate and machine learning
models that are computationally heavy (such as random forest, [Breiman 2001]), offering additional prediction
approaches that were previously not available. These models can consider a wide range of variables. However,
even with these recent breakthroughs, predictive tools do not integrate current mining and statistical knowledge,
as well as the mine site’s own large dataset, which limits the optimisation of stope design and performance.

Objectives
The main objective of this thesis is to develop a methodology for operational mines to understand the spatial
distribution and magnitude of the OB and UB generated across stope surfaces and utilise this knowledge during
the final steps of stope design to predict the expected stope mined geometry to thus optimise stope performance.

More specifically, the following objectives will be met:

3
 1. Use the data collected at the design and reconciliation steps and transform it into pertinent and valuable
information (variables) at an octree resolution, incorporating variables excluded from the Stability Chart
that can significantly influence OB and UB such as drill and blast, major geological structures,
undercutting stope faces etc., and are not a part of the tools currently used.
2. Identify root causes of the stopes’ OB and UB through statistical analysis. The critical variables can
vary from one site to another.
3. Use the critical variables to develop a tool for predicting OB or UB for each octree, giving a detailed
prediction of the location of OB and UB, their magnitude and probability, and transform it into an
expected mine shape (predicted CMS).
4. Characterise the economic value of the predicted OB and UB in terms of net smelting return (NSR).

Methodology
The proposed methodology will permit the conversion of mine site data into information (variables) and
knowledge (predictions) that can be integrated during the final steps of stope design. To develop this
methodology, the research focuses on four key steps that will enable the achievement of the main objective:
1. Data collection and variable quantification.
2. Root cause analysis.
3. Predicting stope geometry.
4. Assessment of the predicted stope performance (OB, UB and economic value).

The data used in this thesis was provided by two Australian and one Canadian mine sites. The multiple mine
sites provide the necessary data to develop the proposed methodology and test its universal applicability. For
each mine site, data related to the stopes (design geometry, CMS, drill rings) as well as the mining and geological
environment (structural and geotechnical model, geometry of the drives intersecting the stopes) are collected.
This data is then used to quantify variables that can influence stope OB and UB. These are operational (drill and
blast, undercut), geometrical (design size and geometry), geological (faults) and geomechanical (RQD) variables
measured at an octree resolution and mapped to the design surface. OB and UB is also quantified through stope
reconciliation.

The variables are then used to understand stope OB and UB through root cause analysis. Univariate, bivariate
and multivariate statistical tools are used to analyse the variance in the data and identify trends between the
variables and OB and UB to determine the critical variables influencing the stopes’ OB and UB. This thesis
focuses on multivariate analysis techniques (principal component analysis [PCA, Pearson 1901] and partial least

4
squares [PLS, Wold 1966]) and how they can increase our understanding of the correlation and independencies
between the variables. The process is repeated for each mine site as the critical variables can vary.

Once the critical variables are identified, a multivariate statistical model is used to predict, at an octree resolution,
the spatial distribution of the OB and UB as well as the magnitude along the design surface. A selection of
different statistical models will be tested and compared with each mine site to identify a statistical model that
integrates various variables and allows the prediction of the geometry, thus meeting the main objective.
Depending on the statistical model, the application of a probabilistic approach will also be explored.

Stope performance is then assessed by looking at the predicted OB and UB and their NSR using a probabilistic
approach. The research for this step of the methodology focuses on building the predicted CMS surface from
the predictions to be able to reconcile the predicted stope as well as quantifying the NSR value in the OB and
UB. This allows to quantify the expected stope performance and work towards the optimal design.

Structure of the thesis

This thesis is divided into eight chapters, with an introduction and a conclusion. The introduction presents the
background, problem, objectives, methodology, and the significance and contribution of this thesis.

Chapter 1 presents a brief critical review on stope performance, from conception to reconciliation. The chapter
covers how stope OB and UB is quantified and the stope design tools that exist, discussing the methods for
predicting OB and UB and their limitations.

Chapter 2 presents a detailed methodology of the proposed stope design approach. First, the data collected is
presented. Then, the process for quantifying the variables from the data, the statistical analysis of these variables
and the predictive model generated using these variables will be presented. Each variable has been selected
and, therefore, will be presented and justified. Univariate, bivariate and multivariate statistical analysis methods
are used in this thesis. Each method will be detailed according to its use and objective. Finally, the economic
analysis and the software used are presented.

Chapter 3 presents the three case studies that will be used throughout this thesis: Dugald River, Prominent Hill
and Westwood mines. It will give an overview of each mines site and present the stope database.

5
Chapter 4 is structured as a scientific article. It presents the assessment of stope OB and UB using octrees
and multivariate analysis (PCA and PLS) for the Dugald River mine. The multivariate statistical approach is
presented as well as the principal results. The impact of each variable on OB or UB is presented, and the critical
variables are identified.

Chapter 5 presents the replicability of the root cause analysis of OB and UB through two additional cases studies
(Prominent Hill and Westwood). For each case study, the principal result from the statistical analysis is presented
(univariate, bivariate, PCA, PLS) and the critical variables are identified. A summary of the impact of each
variable on OB and UB from the different case studies will be presented.

Chapter 6 is structured as a scientific article. It presents the new empirical approach for predicting stope
geometry using Dugald River as the case study. The multivariate and machine learning approaches used, as
well as the different predictive models (PLS, linear discriminant analysis (LDA) and random forest), are
presented. The model’s performance, and the predictive results for each method, are presented and compared.

Chapter 7 presents the replicability of the predictive method developed with Prominent Hill and Westwood. For
each case study, the different predictive models (PLS, LDA and random forest) built for the mine site, as well as
its predictive performance, are presented and compared. The most suitable statistical approach for the data is
determined. A summary of the predictive approach is also presented, discussing the practical interpretation of
the results for site usage, the improvement the proposed methodology offers over existing prediction methods
and how it can be used on site.

Chapter 8 is structured as a scientific article. It presents the interpretation and integration of the predicted stope
performance (OB, UB and economic value) in the stope design step. Multiple scenarios are presented and
discussed. The proposed implementation of the results is detailed. The assessment and use of the predictions
are presented through a case study (Prominent Hill). Only one case study is presented for this step as economic
data was made available by only one of the three mine sites. This case study is used to demonstrate the
approach.

The conclusion gives a detailed summary of the work done in this thesis, the limitations of the work and a list of
recommendations for future research projects.

6
Significance and contributions
The goal of this doctorate thesis is to develop a new methodology for predicting stope geometry that will enable
to go from a qualitative prediction of OB (stable, unstable or caving) or quantitative prediction of ELOS on a per
face basis, to a quantitative prediction of OB and UB at a metre-scale resolution, quantifying the distribution of
the location and magnitude of OB and UB. The proposed methodology will also enable to go from predicting
stope performance using predetermined geomechanical and geometrical variables that influence OB, to a
flexible approach that can integrate any critical variables identified as root causes of OB and UB, covering
geomechanical, geometrical, geological and operational variables. This method can, therefore, be adapted to
each mine site’s context.

The inclusion of operational variables and the use of reconciliation data at an octree resolution for building a
predictive model with probabilistic capabilities makes it a unique method that can be used to optimise stope
design. The resolution of the predictions enables the inclusion of economic data, quantifying the grade value in
the stope OB and UB, as well as the designed and mined geometry, which has been impossible to do at the
design stage in the past. This means the dollar value in reducing the OB and UB is assessed, determining the
profitability of the stope and the financial impact of OB and UB. This represents a major contribution and
advancement in mining.

This new methodology will redefine the existing approach for designing stopes at operational mines which has
seen little change since the 1990s. This new approach to stope design will allow engineers to incorporate in their
stope design process a methodology to predict and understand which variables are critical, and work towards
an optimal design for a stope which maximises the stability and profitability of the mined geometry. This approach
will facilitate evaluation of different scenarios to establish the optimal design. The economic analysis is used to
identify the critical variables to stope profitability and, therefore, improve mine profitability. This doctoral project
will be a significant contribution to the scientific advancement of mining and rock mechanics. It is anticipated that
this thesis will have major impacts on the mining industry as it will help to improve:
• The understanding of stope performance.
• The prediction of the spatial distribution of OB and UB.
• The safety of mine workers, and confidence in the design.
• The profitability of the stopes.
• The planning of the stopes.
To assert and validate the impacts the proposed methodology will have on the mining industry, the whole process
is coded to make the predictions available to the mine site that are providing data for this thesis. This allows
them to test the proposed prediction approach, ensuring the research is not dissociated from the practical world.

7
Chapter 1. Critical review
1.1. Introduction
Analysing, understanding and predicting stope performance for improving stope design have been important
research topics in mining since the late 1970s and early 1980s, when the open stope mining method became
popular. As this thesis builds on the knowledge and tools developed in the last 40 years, it is important to know
what has been done in open stoping and the limitations. This chapter will give a detailed overview of the work
and advancement that has been accomplished relating to stope performance and design. More precisely, this
chapter will provide a critical review of stope design tools, how stope OB and UB are analysed and quantified,
and the status of predictive models and economic analysis.

1.2. Quantifying stope OB and UB

Stope OB and UB are quantified through the reconciliation of the stope design geometry and the mined
geometry. The mined geometry is obtained by scanning the void using a CMS or, as of late, a LiDAR mounted
on a drone (Woolmer et al. 2020). One of the main goals of stope reconciliation is to quantify the stope’s OB and
UB (Figure 1. 1) which are used to understand the variables that control the stope’s OB and UB and to predict
future stope performance. The reconciliation can be done at three different levels of resolution, enabling the
quantification of different variables while being spatially more precise.

Figure 1. 1: Example of stope overbreak and underbreak (Potvin et al. 2016)

8
 1.2.1. Per stope resolution
Common industry practice is to quantify the OB and UB in regard to the whole stope (Potvin et al. 2020). The
common metrics are the volumes of OB and UB or the percentage (divided by the designed volume). At this
resolution, global variables are quantified. Common variables that can be investigated are the mining method
(primary–secondary stoping, for example), the design volume, the mining sequence, the number of production
blasts and the stand-up time (number of days between the final blast and the final CMS). A distinction can also
be made between different materials such as rock and backfill.

Other measures of the stope’s performance not necessarily used for root cause analysis can also be calculated.
These are presented in Table 1. 1. While the recovered volume is a common measure, the others are not
frequently quantified but have been used by Woodward et al. 2019 and McFadyen et al. 2019a (volume
difference only) to characterise the stope performance and identify stopes or sectors that do not perform as
designed.
Table 1. 1: Other measures of stope performance

Measure Explanation Equation

Recovered
Design volume that
volume 𝐷𝑒𝑠𝑖𝑔𝑛 𝑣𝑜𝑙. − 𝑈𝐵 𝑣𝑜𝑙.
was recovered
(m³ or %)
Compliance to The compliance of the 𝐷𝑒𝑠𝑖𝑔𝑛 𝑣𝑜𝑙. − 𝑈𝐵 𝑣𝑜𝑙.
design mined geometry to the ∗ 100
(%) design geometry 𝐷𝑒𝑠𝑖𝑔𝑛 𝑣𝑜𝑙. + 𝑂𝐵 𝑣𝑜𝑙.
The difference in 𝐷𝑒𝑠𝑖𝑔𝑛 𝑣𝑜𝑙. − 𝐶𝑀𝑆 𝑣𝑜𝑙.
Volume difference
volume between design ∗ 100
(%) 𝑑𝑒𝑠𝑖𝑔𝑛 𝑣𝑜𝑙.
and mined
Total OB and UB Total unplanned
𝑂𝐵 𝑣𝑜𝑙. + 𝑈𝐵 𝑣𝑜𝑙.
(m³ or %) volume or percentage

While reconciling at a per stope resolution gives an overview of the mine’s stope OB and UB, it is limited by its
coarse resolution. A limited number of variables can be analysed as mentioned above, therefore limiting the root
cause analysis.

1.2.2. Per face resolution

Also common industry practice, stope OB and UB are quantified at a finer resolution on a per face basis (Potvin
et al. 2020). It is common for only one face to generate most of the OB. Refining the reconciliation on a per face
basis groups the data under similar mining conditions and helps to identify these critical faces and further quantify
variables to identify the root causes. Common metrics for quantifying face OB and UB are the volumes of OB or
UB and the equivalent linear overbreak slough (ELOS, Figure 1. 2) introduced by Clark 1998 and, for the UB,
the equivalent linear lost ore (ELLO). The ELOS and ELLO are obtained by dividing the OB and UB volumes of

9
a face by its area, and represent the average depth of the OB and UB considering the whole surface area. For
faces mined against backfill, rock can be differentiated from the paste or any kind of backfill for the OB (Potvin
et al. 2016). At this resolution, geomechanical, operational and geometrical variables can be quantified for root
cause analysis.

Figure 1. 2: Example of the equivalent linear overbreak sough (ELOS, Clark 1998)

It is also possible to use scale-independent shape measures developed by Cepuritis 2011 to quantify the shape
of the OB; although this is rarely done in practice. These measures describe the circularity and hemisphericity
of the OB per face. The main limitation of the per face resolution is the required averaging of certain variables,
such as the rock quality designation (RQD, Deere et al. 1967), boiling down the information to a single number
for the face, thus limiting the spatial analysis of a face’s performance. The ELOS and ELLO measures cannot
discern where the OB and UB are occurring in the face or its geometry.

1.2.3. Per octree resolution

The final resolution is a recent reconciliation approach developed by Woodward et al. 2019 that looks at creating
point data at a metre-scale resolution by recursively subdividing the three-dimensional space occupied by the
stope using octree blocks until the desired resolution along the design surface (usually <1 m) is obtained (Figure
1. 3). These octrees represent georeferenced points for which variables are quantified to capture the variation
of the stope’s OB and UB and influencing factors along the design surface. In terms of the quantity of information

10
generated, compared to a per stope (one data point) and a per face basis (usually less than 10 data points), a
per octree assessment will generate hundreds to thousands of data points. This exponentially increases the
amount of information that is used which arguably translates into improved understanding of the factors
influencing stope OB and UB.

Figure 1. 3: Illustration of the subdivision of the 3D space and the faces using octrees (modified from McFadyen et al. 2021)

Stope OB and UB at an octree level are quantified by calculating, in the direction normal to the design surface,
the distance to the surface generated by a CMS for each octree (Figure 1. 4). This will be referred to in this
thesis as the projected distance. Compared to the per stope and per face resolution where two separate
variables are calculated for the OB and UB, only one variable is needed for the octrees. Positive values represent
OB, while negative values represent UB.

Figure 1. 4: Illustration of stope OB and UB quantified at an octree level by calculating the distance, in a direction normal to the design
surface, between the design surface and the CMS (modified from McFadyen et al. 2021)

As ELOS or ELLO represent a generalised performance value for a surface, the advantage with the octree
resolution is that it captures how the stope OB and UB actually varies across a surface. Figure 1. 5 provides an
example of how the projected distance measured at each octree captures the variation from OB at the left of the
surface to UB at the right. With the quantification of the factors to be investigated, the octree data enables the

11
consideration of the spatial variation when assessing the impact of the different factors. This type of analysis
was previously impossible on a per stope or per face basis, and it represents a critical technological step towards
improving our understanding and the optimisation of stope performance.

Figure 1. 5: Example of how stope OB and UB measured at the octree resolution capture the variation of stope OB and UB across a
surface (McFadyen et al. 2021)

As octree resolution is a novel approach, limited research has been published (Woodward et al. 2019, McFadyen
et al. 2020) utilising it. Only a selection of geometrical, geological and operational variables have been analysed
so far at this resolution (Woodward et al. 2019). Variables such as undercut, rock mass quality and stress have
not been quantified using octrees, thus limiting the root cause assessment that can be done at an octree
resolution. Furthermore, the exponential number of data points per stope implies a heavier computational
process. The calculation of the projected distance is also more sensitive to the survey qualities as, for example
internal faces from merging solids, can impact the projected distance.

1.3. Stope design tools

To maximise stope stability and profitability, mine site engineers will use stope design tools and acquired
knowledge to help achieve their goals. These theoretical and empirical tools typically include:
• Numerical modelling.
• Stability Chart (Mathews et al. 1981, Potvin 1988, Nickson 1992).
• Site-specific tools.

1.3.1. Numerical modelling

The pre-existing stress state underground gets disrupted once mining occurs. This induced stress can cause
high compressive stress points or relaxation zones around the excavations and will evolve as the mining
progresses. Numerical modelling is a popular mining tool and is used for estimating the stress redistribution

12
around the stope. It is a more sophisticated method that expresses the rock mass behaviour through
mathematical equations. It considers the rock mass failure criterion (Mohr–Coulomb or Hoek–Brown) and its
geological (joints and possibility of faults) and geomechanical properties, as well as stope geometry (complex
or simple) and mining sequence (Potvin 1988).

The model can be used for estimating the principal stress components around the stope, relaxation zone and
the depth of failure. The results can be used for estimating stope stability (Henning and Mitri 2007, Sainsbury et
al. 2015), determine the ideal mining sequence (Castro et al. 2012), as inputs in other design tools (Mathews et
al. 1981, Le Roux 2015) and to assess the impact of stress and design geometry on stope OB (Wang et al.
2007). The variability and uncertainty in the variables can also be considered by using a probabilistic approach
(Idris et al. 2011).

Numerical modelling has limitations, as outlined by Potvin 1988, which are still present today (Shapka-Fels and
Elmo 2022). Assumptions and simplifications are made, thus enabling the modelling of the complex interactions
with the different model inputs. In addition, different choice of models (continuum, discontinuous, three-
dimensional, two-dimensional) and loading conditions (rock, paste) will influence the inputs and results.
Therefore, careful selection and calibration are needed to ensure the model is applicable. Otherwise, it can
provide misleading information. In addition, the models focus on predicting OB and not UB, and operational
variables such as blasting are not included.

1.3.2. Stability Chart

The Stability Chart was first introduced by Mathews in 1981 to help determine the size of a stable stope given
the geomechanical properties of the rock mass (Figure 1. 6). It is an empirical and bivariate tool that is used at
most mine sites (Potvin et al. 2020). It is used for assessing OB by predicting if a stope face will be stable,
unstable or caving. The original data used for this chart comes from stopes mined at a depth less than 1 000 m.
The two variables used in the chart are the stability number calculated from the geomechanical properties of the
rock mass and the hydraulic radius (or shape factor) calculated using the planned geometry.

13
 Figure 1. 6: Mathews Stability Chart (taken from Mawdesley et al. 2001)

The stability number (N’) is obtained using Equation 1.1. It considers the joints, rock quality and stress around
the stope. The Q value (Barton et al. 1974) is modified by replacing the active stress block (JW/SRF) with a value
of 1. The reason being, the stress is measured using factor A which is specifically developed for stopes. Factor
JW which evaluates the effect of water on the stability is not considered as there are no documented cases where
the water pressure was identified as the cause of OB. This is probably due to the fact that following the blasting,
the fractures generated inhibit water from accumulating and developing hydrostatic pressure in the stope walls
(Potvin 2020). Factors A, B and C are calculated using the respective charts (Figure 1. 7).

𝑁 ′ = 𝑄′ ∗ 𝐴 ∗ 𝐵 ∗ 𝐶 [1.1]

Where:
N’= Stability number
Q’= Modified Q value (Barton et al. 1974)
A = Stress factor
B = Joint orientation factor
C = Gravity factor

14
 Figure 1. 7: Chart for calculating the A, B and C factors for the stability number (Mathews et al. 1981)

The hydraulic radius is calculated using Equation 1.2. The hydraulic radius considers the size and shape of the
surface and represents the main factor that can be modified to increase the stability of the stope.
𝑆𝑢𝑟𝑓𝑎𝑐𝑒 𝑎𝑟𝑒𝑎 (𝑚2 )
𝐻𝑦𝑑𝑟𝑎𝑢𝑙𝑖𝑐 𝑟𝑎𝑑𝑖𝑢𝑠 = [1.2]
𝑆𝑢𝑟𝑓𝑎𝑐𝑒 𝑝𝑒𝑟𝑖𝑚𝑒𝑡𝑒𝑟 (𝑚)

The original Stability Chart was built from the data of 26 case histories. Since then, updated Stability Charts
have been proposed, integrating new case studies (Potvin 1988, Stewart and Forysth 1995, Hadjigeorgiou et al.
1995, Mawdesley et al. 2001, Truman et al. 2000) and additional information such as ground support (Nickson
1992, Hutchinson and Diederichs 1996), probabilities of being stable, unstable or caving (Mawdesley et al.
2001), ELOS (Clark and Pakalnis 1997, Capes 2009) and faults (Suorineni et al. 2001a). The most widely used
Stability Charts are those of Potvin 1988 built from 176 cases from 34 mines, and Nickson 1992 which integrates
cable support using 46 cases (Suorineni 2010) (Figure 1. 8). In addition, factors A, B and C have also been
revised, looking to improve the accuracy of the Stability Chart (Diederichs and Kaiser 1999, Bewick and Kaiser
2009, Vallejos et al. 2016). There are uncertainties and variability in the input variables of the chart. A
probabilistic approach can also be used to make risk analysis as demonstrated by (Diederichs and Kaiser 1996)
or to consider the effect of the variability in the geomechanical variables to determine the optimal stope
dimensions (Idris and Nordlund 2019).

15
 Figure 1. 8: Stability Chart generated by: left: Potvin 1988, right: Nickson 1992

The Stability Chart is best used at the feasibility stage, when the stope dimensions need to be determined. It
can also be used during the stope design stage as new geotechnical information is available for a refined
prediction of the stability of the faces or during stope reconciliation by plotting the mined stope faces on the
Stability Chart and assessing the applicability of the method for the mine site. Since the Stability Chart is an
empirical method, its applicability is limited to similar mining conditions in the development database. The use
of the Stability Chart can be inappropriate for severe rockbursting conditions and highly deformable rock mass.
It also presents limitations as it considers only two factors (geomechanical properties and hydraulic radius) that
can affect stope performance while only looking at OB.

Since the introduction of the Stability Chart in the 1980s, research has shown the impact of other variables that
can be critical to stope OB and UB. Research by Clark and Pakalnis 1997 showed that undercut, stress, and
major geological structures such as faults, ground support, drill and blast and stand-up time can affect stope OB
and UB while research by Wang 2004 also highlighted the impact of undercut, drill and blast and stand-up time.
More recently, Potvin et al. 2016, Guido et al. 2017 and McFadyen 2020 showed the impact of operational
variables such as the drilling pattern and standoff for the drilling and blasting (Potvin et al. 2016, McFadyen
2020) as well as the stand-up time and the mining method (Guido et al. 2017, McFadyen et al. 2020). Seismic
variables were also investigated by McFadyen et al. 2019a and showed that the number of events and the
magnitude will impact the stope OB. All the previous research has been done using data at a stope or face
resolution. Analysis conducted by Woodward et al. 2019 at an octree resolution showed the impact of drill and
blast (standoff and orientation), faults, effective radius factor (ERF, Milne 1997) and the geometry of the stope
(dip, convex shape) will impact the stope’s OB and UB. This has led to the development of site-specific tools.

16
 1.3.3. Site-Specific tools
As the stope mining progresses, mine site data becomes available through stope reconciliation and the general
profile of the stope OB and UB can be identified. Engineers will look to understand the root causes of OB and
UB to optimise their design. The collected data, combined with rock mechanics principles, is used to develop an
understanding of the stope OB and UB and construct site-specific prediction tools. The key interest in these tools
is the use of the mine’s own data, as well as the possibility of integrating additional variables and UB to predict
stope OB and UB and determine paths towards an optimal design. These tools can also be more successful
than the Stability Chart. Work done by Guido and Grenon 2018 showed an increase of 20% in the exactness of
the predictions (predicting the OB in the right categories) between a site-specific chart (around 70%) and the
Stability Chart (around 50%). Other work by McFadyen 2020 also shows a high success rate for its respective
tools (58% to 87%).

The site-specific tools often take the form of bivariate charts as most have been developed on a per face basis
and are influenced by the format of the Stability Chart. These tools encompass two or more variables that can
be different to the ones used in the Stability Chart and are developed using statistical methods and models.
Various statistical methods exist and will perform better for a dataset given the form and distribution of the data.
Statistical methods used so far for predicting stope OB are multiple linear regression (MLR) (Wang 2004, Hughes
2011, Guido and Grenon 2018) and principal component regression (PCR) (Guido and Grenon 2018) which
have shown a higher success rate at predicting stope OB than the Stability Chart for the same dataset. In
addition, McFadyen et al. 2020 uses a PLS approach for predicting hangingwall (HW) OB with good success
(58% to 87% faces correctly predicted). Other statistical models have been used for classifying stope OB.
Random forest (Breiman 2001) was used by (Qi et al. 2018) and artificial neural network (McCulloch and Pitts
1943) was used by (Adoko et al. 2022) to classify face OB performance using geometrical and geomechanical
variables. The list of variables used in each model is presented in Table 1. 2.

17
 Table 1. 2: List of variables considered in the different predictive models

Proposed
predictive Considered variables
tools

Average Z coordinate
Guido and Hydraulic Hanging Drilling Fault
stope of the centre Q’
Grenon 2018 radius wall dip method position
width of gravity

Hydraulic Undercut Exposure

Wang 2004 N’
radius factor time

Dilution Dilution
Hughes 2011 Width Length Dip Depth RQD density density
adjacent below
Cumulative Seismic Seismic Adjacent
volume of events events Planned RQD stope Pyramid
Mcfadyen 2020
the mining (local (Richter volume performance ID
pyramid scale) scale)
Dilution
Stope dip, Joint Stope
graph Stress Undercut
Qi et al. 2018 strike and RQD parameters design
factor A, category area
length (Jn, Jr, Ja) method
b and C

Adoko et al. Stability graph variables (hydraulic radius, Q’, N’, A, B,

2022 C)

Statistical methods have also been applied for defining the curves that delimit the boundaries between the
different stability zones. Discriminant methods have been used to minimise the error rate between two classes
of stability given the dataset and variables used. Nickson 1992 employs a discriminant analysis for separating
the stable, unstable and caving zones. A Bayesian likelihood approach is employed by Suorineni et al. 2001b
and Guido and Grenon 2018 that employs a binary logistic regression. Artificial neural network has also been
employed by Clark 1998.

The main limitations with these tools are the quantity of stopes mined needed to identify the trends as well as
the time needed for the reconciliation of the data and the quantification of the variables. While there is no set
number for the minimum number of stopes needed, the more stopes in the database the better as it encapsulates
the general behaviour of stope performance. Work by Guido and Grenon 2018 used 105 stopes while McFadyen
2020 used 166 stopes. Also, workshops held in Australia and Canada with the mining industry in 2019
highlighted the limited processing of the reconciliation data by the mine sites, limiting the analysis that can be
done and the charts that can be created (Potvin et al. 2020). Furthermore, only a single value of OB is predicted

18
per face (none of these methods were applied on UB) as per octree data is only emerging, thus inhibiting the
prediction and interpretation of the spatial distribution of the OB and UB.

1.4. Quantifying critical variables that influence stope OB

and UB
To develop a site-specific approach, or to simply understand stope behaviour, the critical factors that influence
stope OB and UB must be first identified. This is achieved through site knowledge and statistical analysis. The
stability of a stope depends on the geological and mining environment and, therefore, the variables that influence
the OB and UB vary from mine to mine. The factors influencing stope OB and UB are grouped into different
categories:
• Geometrical variables
• Geomechanical variables
• Geological variables
• Operational variables
• Seismic variables
The variables that are quantified and that have been the subject of analysis are based on an understanding of
rock mechanics and mining engineering, as well as the availability of the data and their density.

1.4.1. Geometrical variables

The geometry of a stope impacts its OB and UB as it influences the vulnerability to gravity instability, the
confinement in the rock mass along the faces, and local stress (Hudson and Harrison 1997). The impact of the
shape, size and dip of the design geometry is well documented and integrated into design tools such as the
Stability Chart where the face dip and the hydraulic radius are used to determine OB. On a per face basis, the
hydraulic radius, the dip and the dimensions of a face have been the subject of analyses, either through the
Stability Chart iterations, or through separate analyses (Wang 2004, Guido et al. 2017, Qi et al. 2018, McFadyen
2020).
At a per octree resolution, Woodward et al. 2019 analysed the dip, direction of the design surface, ERF and the
geometry complexity (Table 1. 3) over several mine sites, identifying a correlation between the variables and
stope OB and UB. The impact of these octree variables has, however, not yet been assessed within a predictive
model.

19
 Table 1. 3: Local characterisation of the stope geometry for per octree analysis (modified from Woodward et al. 2019)

Variable Concept Interpretation Example

Dip of the The dip calculated from the Influence of
design normal vector associated with gravity.
surface the octree block (coloured by
degree). Characterise how
OB and UB are
affected by
gravity. Distribution of the dip for each
octree along the design surface
(dots in the image). Warm colours
(90°) mean the design face
attached to the octree is vertical.
Direction of The direction calculated from Influence of
the design the normal vector associated orientation with
surface with the octree block (coloured respect to stress
by degree). and rock mass
properties.

Characterise how Distribution of the direction for each

OB and UB are octree along the design surface
affected by face (dots in the image). The direction
orientation. varies from 0° to 360°.

Effective Calculated by measuring, at Local stress

radius regular radial angles intervals, conditions –
factor the distance from a point bending moment
(ERF) located on the selected of excavated
surface of the stope to the span.
edge of that surface (Milne
1997). A maximum value is Characterise how
obtained towards the centre of OB and UB are
the surface. affected by the
Distribution of the ERF for each
size and geometry
octree along the design surface
of the faces.
(dots in the image). Warm colours
mean the octrees are further away
from the surface edges and have
higher stability concerns.

20
 Distance to The stope design is ‘wrapped’ Local stress
convex hull in a convex hull. conditions.
(stope Distance (coloured by metre)
geometry is found from an octree block Loss of
complexity) to the nearest point on the confinement.
convex hull surface.
Characterise how
OB and UB are Distribution of the distance to the
affected by the convex hull for each octree along
complexity of the the design surface (black lines in
design geometry. the image). Warm colours mean the
design face is further away from the
simplified geometry and can
indicate areas of loss of
confinement due to the design
geometry.

1.4.2. Geomechanical variables

For a given shape and size of a stope face, its stability varies depending on the rock mass properties. As
mentioned by Potvin 1988, “With regards to stability, two of the most important characteristics of the rock mass
are the size and shape of the blocks forming the rock mass matrix.” The geomechanical properties of the rock
mass are obtained through boreholes and mapping of the immediate area around the stope where rock mass
characterisation tools such as RQD, Q-system and RMR (Bieniawski 1973) are most often used to quantify the
quality of the rock mass. Induced stress is most often calculated using numerical modelling and is greatly
influenced by the mining sequence. The induced stress will vary according to factors such as the mining method
(the induced stress is impacted by the geometry of the stope and sequence), the face analysed (the direction of
the face impacts the observed induced stress) and the depth of the stopes (the induced stress varies with the
depth of mining).

The impact of the geomechanical properties of the rock mass and stress on stope OB is well documented on a
per face basis. Mathews et al. 1981 use the stability number, which summarises the information on the
discontinuities and stress, to predict stope OB. The effect of stress was also analysed indirectly by looking at
the sequence, mining method (primary and secondary) and the different faces (Guido et al. 2017, McFadyen et
al. 2020). McFadyen 2020 and Woodward et al. 2019 documented trends between RQD and OB (OB increases
when RQD decreases). The effect of stress has also been analysed indirectly on a per octree basis by looking
at the mining method (primary versus secondary), the faces (HW versus FW versus crown versus side walls)

21
and the depth of the stope (Woodward et al. 2019), but no direct measures of stress or any other geomechanical
variables have been analysed.

Overall, these analyses are limited by the sampling and interpolation, or extrapolation, of the properties of the
rock mass. Limited data will impact the calibration of the numerical model and quality/confidence of the
geotechnical block model. Therefore, the data needs to be used with caution and with a good understanding of
where it comes from. Variables such as the RQD are subjective and the values will fluctuate depending on the
person who calculates it (Sewnun et al. 2019).

1.4.3. Geological variables

The presence of major geological structures, such as faults, can influence a stope’s OB and UB. Their presence
can impact local stress and the quality of the rock mass. Analysis on a per face basis has been documented by
Suorineni 1998, indicating the impact of a daylighting fault versus a parallel fault, the vertical position of intersect
and the distance to the fault. The size of the face will also impact its vulnerability. Analysis on a per octree basis
was performed by Woodward et al. 2019, quantifying the distance and angle to the closest fault (Table 1. 4) and
showed trends with OB for different mine sites. These analyses have certain limitations as they consider the
closest fault. More complex geological environments with multiple faults will not be fully captured by these
variables. The structural model will also be limited by the amount of data used to build it. Certain areas will be
interpreted and, therefore, there is uncertainty regarding the position of the fault.

22
 Table 1. 4: Local characterisation of the geological structures for per octree analysis (modified from Woodward et al. 2019)

Variable Concept Interpretation Example

Distance Distance (coloured by Kinematics – beam
to fault metre) from the octree and wedge
block to nearest formation.
interpreted fault
surface. Characterise how
OB and UB are
affected by the
presence of faults Distribution of the distance to the nearest
near the stope. fault for each octree along the design
surface (dots in the image). Cold colours
mean the octrees are further away from
the fault and less likely to be affected by
the faults.
Angle to The average of the Kinematics – beam
fault inverse distance³ and wedge
weighted angle formation.
(coloured by degree)
between an octree’s Characterise how
design normal vector OB and UB are
and the plane of the affected by the
fault. angle of the fault to Distribution of the angle to the nearest
Convention: 90° = the design surface. fault for each octree along the design
toeing, 0°= parallel. surface (dots in the image). Warm
colours (0°) mean the octrees are parallel
to their fault.

1.4.4. Operational variables

As a stope design is proposed based on its geometry and the rock mass, the mining process chosen will also
influence the stability of the stope. There are different operational steps to mining a stope and, therefore, different
subcategories.

Drill and blast

Rock in open stopes is broken with explosives. The drilling pattern, hole diameter, choice of explosives, delay
and the rock mass will have an impact on the radiation of the energy from the blast and breaking of the rock
(Villaescusa 2014). While variables such as the hole diameter and choice of explosives are decided before the
stope design, the drilling pattern varies from stope to stope. Wang 2004, Potvin et al. 2016, Guido et al. 2017
and McFadyen 2020 documented the impact of the drilling pattern on a per face basis, showing how the standoff

23
and the angle between the stope face and the holes impacts the OB and UB. On a per octree basis, Woodward
et al. 2019 has shown that the standoff, the orientation and the energy (Table 1. 5) influences the stope OB and
UB. The energy distribution is calculated by using the blasthole density as a proxy. Potvin et al. 2016 showed
that the drill and blast is one of the most important aspects influencing UB; however, UB has not been considered
in predictive tools and, therefore, the drill and blast either. When drilling, charging and detonating the explosives,
discrepancies with the design will occur. Blasthole deviation and misfires will happen and are usually not
considered in the variables.

Table 1. 5: Local characterisation of the drilling and blasting for per octree analysis (modified from Woodward et al. 2019)

Variable Concept Interpretation Example

Blasthole Distance (coloured by Blast-induced
standoff metre) from an octree damage.
distance block to the nearest
designed blasthole. Characterise how
OB and UB are
affected by the
standoff distance
Distribution of the distance for each
to the blastholes.
octree along the design surface
(dots in the image) to their nearest
blasthole. Cold colours mean the
octrees are further away from the
blastholes and, generally, UB is
expected.
Blasthole Shortest angle (coloured Blast-induced
orientation by degree) between an damage.
octree’s design normal
vector and the blasthole. Characterise how
Convention: 90° = toeing, OB and UB are
0°= parallel. affected by the
angle of the
blastholes to the
Distribution of the angle for each
design surface.
octree along the design surface
(dots in the image) to their nearest
blasthole. Warm colours (90°) mean
the blastholes are toeing into the
design surface.

24
 Proxy for 𝐵𝑙𝑎𝑠𝑡 𝑒𝑛𝑒𝑟𝑔𝑦 Blast-induced
𝑛
blast L damage.
energy ≈∑
max(0.5, 𝐷)2
0 Characterise how
where: OB and UB are
L = Length of hole affected by the
D = Distance to hole density of
Max. distance of 0.5 m to Distribution of the energy proxy for
blastholes and the each octree along the design
account for stemming near drill pattern.
collars (coloured by energy surface (dots in the image). The
index). energy calculation considers the
density of blastholes around the
octree (inverse distance
relationship) and the length of the
blastholes to determine the spatial
distribution of the blasting energy
according to the blasthole design.
Warm colours mean more energy is
expected in that area due to the drill
design.

Ground support
Analysis of the use of cables for stabilising a stope face by Potvin 1988 and Nickson 1992 showed that the use
of cables in the rock wall can help reduce the OB. Nickson 1992 integrates ground support in his Stability Chart.
This analysis has so far been limited to a per face basis, showing the impact of the cable pattern and density on
OB. The use of cables is also influenced by the rock mass quality and is limited by the access to the desired
area to support.

Time
The effect of time can impact stope OB and UB at two different levels. First is the timing of the stope in the
mining sequence. The stress around the mining area will vary as the mining progresses and, therefore, will affect
the stope OB and UB. The mining sequence has been analysed by Villaescusa 2014, Guido et al. 2017 and
McFadyen 2020. Second, the mucking time will determine how long the stope must stand open. A void
underground will be stable only for a certain period of time before signs of instability start to show and OB issues
arise (Singh and Goel 1999). The effect of stand-up time has been documented by Wang 2004, Guido et al.
2017 and McFadyen 2020. The inclusion of stand-up time in prediction tools, however, remains limited.

Drives
The design of the access to the stope will influence confinement and gravity failures as the drives can overcut
or undercut the stope, leaving hanging rock or thin layers of rock susceptible to instabilities (Diederichs and
Kaiser 1999, Figure 1. 9). The effect of undercutting has been documented by Wang 2004 on a per face basis,

25
quantifying the depth of the undercut by the drives; however, undercut has not yet been quantified on a per
octree basis. Therefore, the local impact of undercut has not yet been analysed.

Figure 1. 9: Example of a relaxation zone created by the drives undercut (Diederichs and Kaiser 1999)

1.4.5. Seismic variables

There is little documentation on the effect of seismicity on stope OB and UB. However, it is reasonable to
assume, based on the impact a seismic event can have on a drive, that the occurrence of seismic events in the
vicinity of an open stope can affect the OB. Analysis by McFadyen et al. 2019a on a per face basis has shown
trends between a HW OB and the magnitude of the seismic event and the quantity of events around the stope.
However, this analysis was conducted for only one mine site and no analysis on a per octree basis has been
documented.

1.5. Statistical analysis used for understanding and

predicting stope OB and UB
As the various levels of resolution of reconciliation will give different information, multiple statistical analyses are
used to understand the stope OB and UB. These statistical analyses can be simple, such as univariate or
bivariate, or more complex, such as multivariate. While statistical analysis technically only identifies trends in
the data, a good understanding of rock mechanics and the mining environment facilitates the selection of
variables being quantified and analysed, which can lead to useful conclusions from these analyses.

26
 1.5.1. Univariate analysis
The first and simplest statistical analysis to be done when assessing stope OB and UB is univariate analysis.
Two main charts are used in root cause analysis. There is the cumulative chart in which a selected variable is
displayed on an XY chart where the X axis represents the selected variable and the Y axis represents the
cumulative distribution of the data for that variable. The data is ordered according to its value for the selected
variable as shown in Figure 1. 10. The chart is used to look at the distribution of different variables and is also
used to compare variables such as OB and UB or the performance of the various faces to identify the leading
performance issues and where they are occurring. Potvin et al. 2016, Guido et al. 2017, McFadyen et al. 2019a
and Woodward et al. 2019 use this type of chart to build the stope OB and UB profile for different mine sites.

Figure 1. 10: Example of a cumulative distribution chart plotting overbreak and underbreak (Potvin et al. 2016)

Bar charts are also used to analyse the distribution of the data for a selected variable and determine the type of
distribution it follows, and to assess the variance in the data (Potvin 1988, Clark and Pakalnis 1997, Wang 2004).
This information is valuable when building a predictive model, as some models perform better depending on the
distribution. It also enables a probabilistic approach as the distribution is considered to generate an interval of
predictions (Diederichs and Kaiser 1996). Bar charts are also used to classify and analyse stope OB and UB
(Capes 2009, Figure 1. 11).

27
 Figure 1. 11: Example of a bar chart used to analyse stope OB (Capes 2009)

The main limitation with univariate analysis is the limited root cause analysis that is achieved when using these
graphs as the variables are not analysed against stope OB and UB. It, however, serves as a first step in
establishing the distribution in the data and the broad trends in OB and UB.

1.5.2. Bivariate analysis

Correlation or trend between two selected variables or with stope OB and UB is identified using a bivariate
analysis chart. The data is plotted on an XY chart with the selected variables plotted on the X axis and Y axis.
Statistical methods are overlaid on the chart to interpret the data. The method chosen will depend on if the data
is quantitative or qualitative. The main statistical methods used in root cause analysis are regressions, moving
averages, and box plots or bar charts.

Regression is used to quantify the linear correlation between the two variables by plotting the linear trend on the
chart and calculating the coefficient of determination (R²) and the root mean square error (RMSE) to assess the
quality of the linear trend. It can also be validated using a T-test (Student 1908) by determining if the slope of
the linear trend is statistically different to 0. This type of chart has been used by Clark 1998, Wang 2004, Milne
et al. 2004, Capes 2009 and McFadyen et al. 2019a for identifying correlations in stope OB and UB.

A moving average is used to identify the general trend in data that is non-linear. To do so, the selected variable
is ordered from the smallest to largest value, and the stope performance metric average is calculated for each
point by considering the previous observations up to a certain given number. This method has only recently seen
use in stope OB and UB analysis (Woodward et al. 2019), as analysing octree data involves looking at thousands
of points in the chart overlaying each other. This method gives a smoother line for large data sets, making it

28
easier to identify the trend. Before the use of this method became commonplace, data points were simply
connected with lines, as only a few points were displayed in the chart (Milne 1997, Clark 1998). This tool is used
to complement the other statistical methods that are mentioned in this section (Figure 1. 12).

Figure 1. 12: Example of the use of a moving average (black line) with a box a whisker plot for analysing the ERF per octree
(Woodward et al. 2019)

If the trend in the data is non-linear or qualitative variables are analysed, a whisker and box plot is used. The
data is binned, and the maximum and minimum using the whiskers as well as the first and third quartile using
the box, is plotted for each bin. The median is also shown, and notches can be plotted around it to display the
95% confidence range in which it is situated (Chambers et al. 1983). The different box and whiskers are then
compared to assess how the stope OB and UB fluctuates and what the distribution is for each category (Figure
1. 13). The notches are used to determine if two bins have statistically different medians. This type of chart has
been used by Guido et al. 2017, McFadyen et al. 2019a and Woodward et al. 2019 to assess stope OB and UB.

Figure 1. 13: Example of a box and whisker plot comparing different mining pyramids (McFadyen et al. 2019a)

The bar chart is used to bin the data and plot the frequency of observations in a bin exceeding a certain
performance threshold. The information is then used to assess the fluctuation of the stope OB and UB; more
specifically, cases with a value over a given threshold. These types of charts have been used by Guido et al.
2017 and McFadyen et al. 2019a (Figure 1. 14) to identify the frequency of OB that exceeds a critical threshold
given the selected variable.

29
 Figure 1. 14: Example of a bar chart used for analysing RQD (McFadyen et al. 2019a)

A third dimension can also be added using marker styles for these charts. A good example where marker styles
are used is the Stability Chart. The hydraulic radius and the stability number are plotted against each other with
the data points marked according to the OB class (Mathews et al. 1981) to determine the separation between
the different categories. The markers are used to show groups of data in the chart, enabling a separate analysis
for each group to be performed, and the different trends to be identified.

The main limitation with bivariate analysis is the capability of analysing only two or three variables at a time. The
reality is that multiple factors will have an impact on OB or UB and the correlations might differ if all the influential
factors are considered in the same analysis.

1.5.3. Multivariate analysis

Underground mines are complex environments, and the outcome of a stope can be controlled by a multitude of
factors. Being able to analyse all the variables together can prove beneficial to better understanding of stope
OB and UB instead of looking at each variable separately using bivariate analysis. For analysing stope OB and
UB, PCA and PLS has been used by McFadyen 2020 to determine the critical factors. Multivariate predictive
models presented in Section 1.3.3 also offer insight into the critical factors through the weights (MLR, PCR and
PLS models) or importance (random forest models) of the variables in the model.

Principal component analysis

PCA was first introduced by Pearson 1901 with the goal of analysing the variance for a selection of observation
and variables. More precisely, the goal was to calculate a set of parameters called components (T) that are the
result of the linear combination of the original variables which maximises the variance explained in the data.
PCA is used to analyse the correlation and independence of the variance of each variable. With the linear
combination, a weight (P) is given to each variable and for each component (Equation 1.3). Figure 1. 15 gives
an illustration of the different variables of PCA. The matrix X represents the input data with K variables and N
observations, matrix P represents the weights for each k variable and a component, and matrix T represents the
principal component matrix with A component and N observations.

30
 𝑇𝑎 = ∑𝑘=𝐾
𝑘=1 𝑃𝑎,𝑘 ∗ 𝑋𝑘 [1.3]
Where:
Ta = Score column for component a in matrix T
Pa, k = Weight for variable K and component A
Xk = Observation values (centered and scaled) for variable K

Figure 1. 15: Illustration of the PCA model (Dunn 2023a)

Each component is independent of one another and will explain part of the variance in the data, with the first
component explaining the most and so forth. The sum of the variance explained by each component will equal
the total variance in the data. However, the benefit of PCA is the ability to reduce the number of variables, or
components, to analyse based on the variance explained by each component. To be able to use the results,
data is first centred and scaled by taking out the mean value and dividing by the standard error for the given
variable. This is done because not all variables have the same units, and this transformation will give the same
weight to each variable when analysing the variance.

McFadyen 2020 used the PCA method to analyse and identify the correlations and independencies between the
variables on a per face basis, which revealed valuable information. Stope OB and UB was included in the
analysis as a marker style when analysing the charts which allows the identification of distribution patterns. PCA
is also used as a quality control step by identifying observations with abnormally high or low values for variables
that contribute to the component being analysed. PCA has, however, not been used at an octree resolution and,
therefore, the spatial distribution of OB and UB has not yet been assessed using PCA.

Partial least squares

PLS was first introduced by Wold 1966 and is similar to PCA with the difference being that the goal is to analyse
the covariance between the input and output variables. Weights (W’ and C’) and components (T and U) are

31
calculated for the input (X) and output (Y) variables (Figure 1. 16). Additionally, for PLS, the output variable, or
in this case the stope OB and UB, is included in the input component’s calculation.

Figure 1. 16: Illustration of the PLS model (Dunn 2023b)

For PLS, the linear combination of the variables maximises the covariance explained between the input and
output variables. The covariance is a product of the correlation between the input and output variables and the
variance of each one (Equation 1.4).
1
𝐶𝑜𝑣 (𝑡𝑎 , 𝑢𝑎 ) = 𝐶𝑜𝑟𝑟(𝑡𝑎 , 𝑢𝑎 ) ∗ √𝑡 ′ 𝑎 𝑡𝑎 ∗ √𝑢′ 𝑎 𝑢𝑎 ∗ 𝑁 [1.4]
Where:
a = Chosen component
Corr = Correlation between T and U
𝑡 ′ 𝑎 𝑡𝑎 = Variance of T
𝑢′ 𝑎 𝑢𝑎 = Variance of U
N = Number of observations

McFadyen 2020 used PLS to indicate the importance of each variable with regard to stope OB and UB and how
they relate to OB and UB on a per face basis, revealing valuable information. PLS was also used as a predictive
model (McFadyen 2020). PLS has, however, not been used at an octree resolution and, therefore, the correlation
of the spatial distribution of OB and UB with the different variables has not yet been analysed through PLS.

1.6. Economic analysis

The economic performance of a stope depends on the mine’s careful planning process and the mining of the
stope. The planning process of a mine orbits around the economic value that will be extracted and is divided
into three main steps (Steffen 1997):

32
 • Life of mine plan (LOM).
• Long-term plan (LTP).
• Short-term plan (STP).
At the LOM step, the inventory and spatial location of the ore reserve are defined, as well as the infrastructure
requirement and capital cost. This step allows one to determine the cut-off cost and limits of the mine’s orebody.
The next step is the LTP. In this step, the operation and mining strategy for the next few years is defined. The
stope design is defined (quantifying the metal in the stopes) as well as the mining sequence and rate of mining.
The projected sale price and cost of production are considered during this step. The goal is to maximise the
LOM and value for shareholders. The final step is the STP (one-year period). The goal is to apply grade and
cost control and manage productivity, handle day-to-day scheduling and to ensure long-term goals are met.

During these mine planning steps, variables such as the sale price and production cost will be refined as they
fluctuate in time and space, impacting the profitability of the stopes. Given the multiple variables, mines will boil
down most of the economic information to two numbers: the cut-off grade and the NSR. The cut-off grade
represents the minimum grade required for the ore to generate positive value, separating ore and waste. It is
calculated as a metal content, percentage or grams per tonne, depending on the metal, or also as dollars per
tonne if multiple metals are mined. The cut-off grade can be calculated in different ways and can consider the
following information (Lane 1988):

• Recovery.

• Concentrator capacity.

• .Refining capacity.

• Fixed costs.

• Sale price.

• Concentrator costs.

• Refining costs.

• Present value.

• Discount rate.

33
The NSR represents the estimate value of the ore, calculated in $/tonne, and is based on the sale price of the
metals minus the transport and treatment costs at the mill. The NSR considers all the metals being mined and
sold, and is calculated on a point basis along with the grades of all the metals in that location. This information
is presented as a block model.

The cut-off grade, in combination with a grade block model, is used at the LTP stage to propose a design
geometry by establishing where the boundary between ore and waste is (Wang and Webber 2012). Since there
are uncertainties in the grade block model, a probabilistic approach can be used to optimise the position of the
stope geometry and maximise profitability (Grieco and Dimitrakopoulos 2007; Tholana and Musingwini 2022).

Unfortunately, the mined geometry rarely matches the design geometry. Factors such as recovery and
fragmentation can fluctuate from stope to stope. The generated OB and UB can result in unplanned dilution or
loss of ore, which is determined using the cut-off grade. The loss of ore occurs when there is ore left behind in
the UB that is unmined. Dilution occurs when waste rock is mixed in with the ore. The dilution is either planned
(due to the geometry of the ore lens and access) or unplanned (from OB) and will incur direct (blasting, mucking,
crushing etc.) and indirect (metal recovery) costs (Villaescusa 2014). Based on literature and from what is seen
in mine site practices, there are several ways of calculating the economic impact of OB and UB as listed in Table
1. 6.
Table 1. 6: Different dilution and loss of ore variables

Variables Formula
Dilution (Pakalnis 1986) 𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑 𝑊𝑎𝑠𝑡𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑
or
𝑂𝑟𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑 𝑂𝑟𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑+𝑊𝑎𝑠𝑡𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑

Planned dilution (Scoble and Moss 1994) 𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
Unplanned dilution (Scoble and Moss 1994) 𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑂𝐵
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
Loss of ore (Villaescusa 2014) 𝑂𝑟𝑒 𝑖𝑛 𝑈𝐵
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
Recovered ore 𝑂𝑟𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

The effects of dilution and loss of ore on stope profitability will vary given the volume and grade value of the
stope, and the OB and UB. It will also depend on the dilution material. For example, dilution from backfill can
affect metal recovery from the milling process. Dilution is caused by a variety of factors that occur at different
stages of the mining process (Table 1. 7).

34
 Table 1. 7: List of variables influencing dilution compiled by Villaescusa 2014

Orebody delineation Drilling and blasting

Undersampling of orebody boundaries Poor initial markup of holes
Errors in decisions regarding cut-off
Setup, collaring and deviation of blastholes
grades
Incorrect choice of blasting patterns, sequences and explosive
Downhole survey errors
types
Lack of geotechnical characterisation
Production stages
Design and sequencing Mucking of backfill floors
Poorly designed infrastructure Mucking of falloffs and stope wall failures
Poor stope design (dimensions) Contamination of broken ore by backfill
Lack of proper stope sequencing Leaving broken ore inside the stopes
Lack of economic assessment Poor management of waste rock (tipped into the ore stream)

Stope development Mine management

Non-alignment of sill horizons Lack of supervision and communication
Poor geological control during mining Excessive turnover of personnel
Mining not following geological
Limited time for planning
markups
Inappropriate reinforcement schemes Lack of stope performance reviews
No documentation and proper training
Performance indicators based on quantity (focus on tonnes as
opposed to metal content)
Lack of leadership and vision

Different methods are used to limit the impact of dilution, such as strategic control (mining method chosen, stope
design) and tactical control (cable bolt, blasting, backfill, draw control) (Potvin and Wesseloo 2022). Determining
the root causes of OB and UB, and predicting the dilution or loss of ore can also be used to determine the tactical
control needed. In the current state of practice, predicting the dilution and loss of ore is limited as current
predictive models do not associate a grade value to the OB and do not predict UB. This is because per face
predictions cannot depict the OB geometry. Although a volume of unplanned dilution could be predicted if the
ELOS is predicted and the design surface lies on the cut-off limit. Therefore, there is currently no method to fully
assess the economic impact of predicted OB or UB in terms of dilution, loss of ore and recovery, thus limiting
opportunities for complete stope optimisation.

1.7. Conclusion
A literature review on stope design and performance was presented. This chapter covered the quantification of
stope OB and UB, stope design tools, stope OB and UB analysis, predictive models and economic analysis.
Current methods for understanding or predicting stope OB and UB lack one or more of the three key elements
for the stope optimisation methodology proposed in this thesis. These are the fine resolution (octree data

35
structure), the multivariate analysis/model and the understanding/prediction of OB and UB. With the exception
of Woodward et al. 2019, the analysis and predictive tools use per face data, limiting the characterisation of the
spatial distribution of stope OB and UB. Multivariate analysis and models enable better understanding of stope
OB and UB and incorporate a variety of variables. However, these statistical tools have only been used on per
face data, limiting the variables that can be used. Finally, the predictive models focus mainly on OB and not UB,
limiting the insight and possible optimisation.

This thesis builds on the current knowledge to advance the field of open stope optimisation by developing a new
predictive approach for stope design and optimisation. It incorporates the three key elements (octrees,
multivariate analysis, and OB and UB data) to improve our understanding and the prediction of stope OB and
UB. The methodology is discussed in the next chapter where the detailed step-by-step approach is presented.

36
Chapter 2. Methodology
2.1. Introduction
The proposed stope prediction methodology is an empirical approach that uses recent developments in stope
reconciliation (octree resolution) and advanced statistical methods (multivariate analysis) to understand and
predict stope geometry, thus filling the gaps highlighted in the literature review and augmenting mines’ abilities
to predict and optimise their stope design. There are four main steps to the method (Figure 2. 1). This chapter
explains each step that allows the extraction of valuable information from the reconciled data and transform it
into knowledge that can be used as feedback to optimise stope performance at the short-term design stage.

Data collection Root cause Predicting stope Assessing

•Data quality analysis of stope geometry predicted stope
verification OB and UB •Build a multivariate or performance
•Stope reconciliation machine learning
•Univariate, bivariate •Reconcile predicted
•Variable quantification and multivariate (PCA model to predict the geometry
•Data filtering and PLS) statistical geometry (PLS and
•Quantify grade value
analysis of OB and UB random forest) and OB
and UB location (LDA) •Assess probabilistic
•Critical variable stope geometry
identification •Evaluate model
outcomes
performance
•Identify critical
•Quantify standard
location of OB and UB
error

Figure 2. 1: The four steps of the stope design methodology

2.2. Data collection

The first step is to gather all the necessary data and quantify variables for root cause analysis (Figure 2. 2). This
step is time consuming as preparing the data took up approximately 60% of the time dedicated to this thesis to
complete. This is about the same amount of time as other types of data analysis projects (Pyle 1999). The data
is obtained from several mining departments and is grouped into four main categories, detailed in Table 2. 1.
This data allows to characterise the geometry of the stope, the geomechanical properties of the rock mass, the
major geological structures and the operational aspects of the stope. The selection is based on our
understanding of rock mechanics and mining, and the possible root causes of OB and UB as well as the data
that was readily available at mine sites.

37
Data collection Root cause Predicting stope Assessing
•Data quality analysis of stope geometry predicted stope
verification OB and UB •Build multivariate or performance
•Stope reconciliation machine learning
•Univariate, bivariate •Reconcile predicted
•Variable quantification and multivariate (PCA model to predict the geometry
•Data filtering and PLS) statistical geometry (PLS and
•Quantify grade value
analysis of OB and UB random forest) and OB
and UB location (LDA) •Assess probabilistic
•Critical variable stope geometry
identification •Evaluate model
outcomes
performance
•Identify critical
•Quantify standard
location of OB and UB
error

Figure 2. 2: The first step of the stope design methodology

Table 2. 1: Collected data

Categories Description Sought after data

• Cavity monitoring system scan
Georeferenced 3D shapes (Miller et al. 1992)
Geometrical • Design/planned geometry
of the mining geometries.
• Drift surveys

Georeferenced rock mass • Geotechnical block model

Geomechanical properties surrounding the • Borehole rock quality
stope. designation data
Georeferenced structures, • Structural model
Geological such as faults and • Lithological contacts
lithological contacts.
• Drill and blast data
Planning and mining o Blast dates
properties of the stopes, o Drill rings
Operational such as drill and blast, o Charging
ground support and mining o Delays
sequence. • Ground support strings

38
 2.2.1. Data quality verification
Once the data is obtained, it needs to be validated to ensure a high quality of analyses and exactness of models.
The possible quality issues will vary for each category and type of data. For the geometrical data, the quality of
CMS surveys is the major concern. The volumes and projected distances computed during stope reconciliation
are affected by errors in the mesh, such as internal faces, non-manifold faces or a wrong surface construction
when built from the point cloud (Figure 2. 3a). These are identified through visual analysis or by using existing
mining software with built-in survey quality tools (for example Deswik [Deswik 2016] and mXrap [Harris and
Wesseloo 2015]). Bad meshing can occur when shapes are merged (sometimes several scans are conducted
at different stages of mining), or when the scans are of poor quality. The CMS mesh is generally validated on
site and surveying artefacts in the CMS repaired, but based on feedback from different mine sites, this is not
always done. Another relatively common issue is shapes that are the result of muckpiles still present in the stope
during the final CMS scan (Figure 2. 3b). The CMS will not be representative of the completely empty stope, and
this information should be treated with caution. This can be validated with mine site personnel.

a) b)
Figure 2. 3: Examples of a) wrong surface construction; b) Odd CMS shape caused by a muckpile left in the stope

Geomechanical data comes from either block models (built from boreholes and underground mapping) or directly
from boreholes. Information on how the block model was built needs to be gathered, as well as the source of
the data and the density. This will allow the determination of the resolution of the model and how it can be
integrated with the octree points. Areas of low confidence can then be identified.

For geological data, similar consideration is made for the structural model to evaluate the confidence in the
position and the presence of structures. The structural data represents a special case as it is not necessarily
present for each stope. There is not necessarily a fault in the vicinity of a stope and, therefore, octrees will have
absent data for these structural variables. Also, not all structures cause stability problems for the stopes.

39
Regarding stope OB and UB, structures are treated as a source of OB potentially affecting stope performance.
Only the faults that influence the stope OB are of interest. Mine site knowledge is required to select influential
structures.

For operational data, the main quality issue noticed during the data processing was overlapping drill string
designs (Figure 2. 4). This occurs when there is more than one firing for the stope and a separate drill string
design is generated for each firing. The final drill string design that covers the whole stope will mask the previous
drill strings. The additional data will affect the result of some of the drilling variables and, therefore, need to be
identified and filtered out.

Figure 2. 4: Example of overlapping drill strings generated for the first blast (in orange) and for the rest of the stope (in black).

Data measurement accuracy

In the dataset, the quantified values that are used for each variable can differ from their true values depending
on the type of data (drill and blast data, geomechanical and geological data, stope OB and UB measurements).
These quantification errors are associated either with the tools used or the human interpretation of the data and
will impact the prediction errors, so it is important to know what the possible quantification errors in this dataset
are.

For drilling and blasting variables, the drill strings design is used as, compared to the actual drill strings, it is
easily accessible at any mine site. In reality, deviation occurs when drilling and, therefore, the real location of
the drill strings will vary compared to the design. Usually, the stope is designed to keep the maximum deviation
down to 2° (Villaescusa 2014), but it will vary based on the length drilled and equipment used. Accessing this
information is not always possible as the drill holes are often not surveyed and, if they are, mines do not
necessarily generate a new drill design. The impact of the deviation is, therefore, not measured in this thesis
and will be accepted as a source of uncertainty.

40
For the geomechanical variables, the geotechnical block models and structural models are built using data
gathered from borehole logging and underground mapping. The geotechnical (RQD, RMR etc.) and geological
(structures) properties are calculated and positioned where the data was measured. The block model is
populated by creating the grid point and calculating the geotechnical properties to each point location. These
calculations can be deterministic, such as using the inverse distance weighting or the nearest neighbour. They
can also be calculated through geostatistics such as kriging (Krige 1951 and Matheron 1963). Kriging offers an
advantage for directional variables as it can consider the variation of the data according to different directions.
The confidence in the block model values will vary according to the density of data, nearest neighbour and the
direction of the data. The level of confidence will vary according to each region of the mine, as data collection
varies spatially. Some of the information that is gathered also varies based on the person doing the logging, as
there is subjectivity in calculating the rock mass classifications (for example RQD, Sewnun et al. 2019). The
quality of the block model is considered by attributing different levels of confidence based on the input data.
Boreholes can also deviate from the design, but methods such as directional drilling exist, that use
measurement-while-drilling (MWD) systems to keep track of the borehole locations.

Finally, for the measurement of OB and UB, the CMS scan will impact the calculated projected distance and
volume of OB and UB. The precision for a CMS scan is ±20–30 mm according to various providers. The precision
of the projected distance is, therefore, to the decimetre. For a 10×20×30 m void, the CMS precision can have a
±34 m³ impact on the stope volume, which represents half a percent. The quality of the scan and the meshing
will impact the depiction of the void by the CMS. Blind spots caused by the position of the CMS and rocks
obstructing the path of the scan will lead to missing information or interpretation of the actual surface. This will
lead to uncertainty in the measurement of the projected distance and volume.

Overall, these inaccuracies impact the quantified variables, the analysis and predictive models that will be built.
They will be considered when analysing and predicting the data.

2.2.2. Stope reconciliation

Once all the necessary data has been gathered and validated, the stope database is built by reconciliating the
design and mined geometry for the three different levels of resolution (per stope, per face and per octree). During
this process, the octree data structure is constructed, enabling the quantification of the different octree variables.
The octrees that intersect the design surface are then used to identify and delineate the different stope faces
(Figure 2. 5). Finally, OB and UB are quantified for each level of resolution using the different metrics presented
in Section 1.2. All the information is then exported into a common database.

41
 Figure 2. 5: Example of face delimitation using octrees

2.2.3. Quantifying the variables

Once the reconciliation is complete, variables are quantified and associated to each face or octree depending
on the resolution of the data. The variables used in this thesis are selected based on general knowledge,
understanding and principles of rock mechanics and stope stability discussed in Section 1.4.1. At an octree
resolution, the selected variables measure the effect of blasting, structures, geotechnical variables and stope
geometry on the stope OB and UB. A total of 10 variables were quantified and analysed by Woodward et al.
2019, and linked to stope OB and UB using bivariate analysis. Additional variables are considered in this thesis
(undercut, directional measure to structures, geotechnical variable and dissimilarity). Table 2. 2 provides the
concept, interpretation, and an example of the variables which will be part of the multivariate analysis. A stope
is used for each example, displaying the octree variables according to the legend attached, and the variables
are described to help characterise and understand them. These variables require the CMS, stope design
geometry, drift surveys, drill rings, and geotechnical and structural model to be available for each stope.

42
 Table 2. 2: Local characterisation of stope design variables for per octree analysis (modified from Woodward et al. 2019)

Variable Concept Interpretation Example

Blasthole Distance (coloured by Blast-induced
standoff metre) from an octree damage.
distance block to the nearest
designed blasthole. Characterise how OB
and UB are affected
by the standoff
distance to the Distribution of the distance for each
blastholes. octree along the design surface (dots
in the image) to their nearest
blasthole. Cold colours mean the
octrees are further away from the
blastholes and, generally, UB is
expected.
Blasthole Shortest angle Blast-induced
orientation (coloured by degree) damage.
between an octree’s
design normal vector Characterise how OB
and the blasthole. and UB is affected by
Convention: 90° = the angle of the
toeing, 0° = parallel. blastholes to the
design surface.
Distribution of the angle for each
octree along the design surface (dots
in the image) to their nearest
blasthole. Warm colours (90°) mean
the blastholes are toeing into the
design surface.
Proxy for 𝐵𝑙𝑎𝑠𝑡 𝑒𝑛𝑒𝑟𝑔𝑦 Blast-induced
𝑛
blast energy L damage.
≈∑
max(0.5, 𝐷)2
0 Characterise how OB
where: and UB are affected
L = Length of hole by the density of
D = Distance to hole blastholes and the
Max. distance of Distribution of the energy proxy for
drill pattern.
0.5 m to account for each octree along the design surface
stemming near collars (dots in the image). The energy
(coloured by energy calculation considers the density of
index). blastholes around the octree (inverse
distance relationship) and the length
of the blastholes to determine the

43
 spatial distribution of the blasting
energy according to the blasthole
design. Warm colours mean more
energy is expected in that area due to
the drill design.
Dip of the The dip calculated Influence of gravity.
design from the normal
surface vector associated Characterise how OB
with the octree block and UB is affected by
(coloured by degree). gravity.

Distribution of the dip for each octree

along the design surface (dots in the
image). Warm colours (90°) mean the
design face attached to the octree is
vertical.
Direction of The direction Influence of
the design calculated from the orientation with
surface normal vector respect to stress and
associated with the rock mass properties.
octree block
(coloured by degree). Characterise how OB
and UB are affected Distribution of the direction for each
by face orientation. octree along the design surface (dots
in the image). The direction varies
from 0° to 360°.
Effective Calculated by Local stability
radius factor measuring, at regular conditions –
(ERF) radial angle intervals, deformation of
the distance from a excavated span.
point located on the
selected surface of Characterise how OB
the stope to the edge and UB are affected
of that surface (Milne by the size and
1997). A maximum geometry of the
value is obtained faces. Distribution of the ERF for each
towards the centre of octree along the design surface (dots
the surface. in the image). Warm colours mean
the octrees are further away from the
surface edges and have higher
stability concerns.

44
Distance to Only the stope design Local stress
convex hull is ‘wrapped’ in a conditions.
(stope convex hull.
geometry Distance (coloured by Loss of confinement.
complexity) metre) is found from
an octree block to the Characterise how OB
nearest point on the and UB are affected
convex hull surface. by the complexity of
Distribution of the distance to the
the design geometry.
convex hull for each octree along the
design surface (black lines). Warm
colours mean the design face is
further away from the simplified
geometry and can indicate areas of
loss of confinement due to the design
geometry.
Distance to Distance (coloured by Kinematics – beam
fault metre) from the and wedge formation.
octree block to
nearest interpreted Characterise how OB
fault surface. and UB are affected
by the presence of
faults near the stope.
Distribution of the distance to the
nearest fault for each octree along the
design surface (dots in the image).
Cold colours mean the octrees are
further away from the fault and less
likely to be affected by the faults.
Angle to fault The average of the Kinematics – beam
inverse distance³ and wedge formation.
weighted angle
(coloured by degree) Characterise how OB
between an octree’s and UB are affected
design normal vector by the angle of the
and the plane of the fault to the design
fault. surface. Distribution of the angle to the
Convention: 90° = nearest fault for each octree along the
toeing, 0°= parallel. design surface (dots in the image).
Warm colours (0°) mean the octrees
are parallel to their fault.

45
Undercut The stope design and Local stress
the drifts are conditions.
‘wrapped’ in a convex
hull. Loss of confinement.
Distance (coloured by
metre) is found from Characterise how OB
an octree block to the and UB are affected
nearest point on the by the complexity of
convex hull surface. the geometry and the
cutting of the stope
by the drives.

In comparison with Distribution of the undercut for each

the stope geometry octree along the design surface (dots
complexity variable, in the image). Warm colours indicate
the undercut that, due to the presence of the drives
considers the full and/or the shape of the stope, the
geometry of the void octrees are further away from the
and the impact the simplified shape that wraps the stope
drives have. and drives, and highlight areas where
there can be a loss of confinement
due to how the drives cut the stope.
Directional Directional measure Deconfinement and
measures to the drift or faults in gravity effect on
to structures regard to its position design surface areas
and distance. near the drifts.
Convention: 180° =
octrees situated over Characterise how OB
their nearest drift and and UB are affected
close to the drift, by gravity caused by
0° = octrees situated the presence of
under their nearest structures such as
structure and close to drives and faults. Distribution of the directional measure
the drift. 90° = to the drives for each octree along the
octrees situated away design surface (dots in the image).
from the structures Cold colours (0°) mean the octrees
(over 2 m). are under and close to the nearest
drive and warm colours (180°)
indicate the octrees are over and
close to the nearest drive. 90°
indicates the octrees are further away
from the drives.

46
 Geotechnical Value from the Characterisation of
variables nearest block of the how OB and UB are
(RQD, GSI, geotechnical block affected by the rock
RMR etc) model mapped to the mass quality.
octree.

Distribution of the RQD for each

octree along the design surface (dots
in the image). Warm colours highlight
areas where there is poor rock mass
quality.
Dissimilarity The average Influence of the
difference between spatial fluctuation of
the adjacent octrees the critical variables
and the selected on OB and UB.
octree using
normalised values of Characterise how OB
the critical variables. and UB are affected
by the spatial
fluctuation, over a
short distance, of all
the variables used in
the root cause
analysis.
Distribution of the dissimilarity for
each octree with its adjacent octrees
along the design surface (dots in the
image). Cold colours mean the
octrees have similar values to the
adjacent octree for the selected
variables, and warm colours indicate
that one or more variables for one or
more of the adjacent octrees differ
from the selected octree.

While ground support data is collected, no variables characterising the cable bolts are used in this thesis. Given
the novelty of the octree data, no ground support variable at an octree resolution existed yet and work done on
creating one during this thesis has not yielded a suitable variable that can be used for analysis. Therefore,
ground support is excluded. Other root cause factors, such as stress, are excluded for the same reasons.

47
 2.2.4. Data filtering
Once the database has been built and all the variables are calculated, a final filtering of the octrees is done to
ensure the quality of the following analysis and models. This means filtering any stopes or portion of stopes with
data concern highlighted section 2.2.1. It also means filtering out octrees that represents drives or backfill. This
ensures that the data analysed has the same correlation structure and are representative of the stope
performance we are trying to understand and predict. This process is done one stope at a time, filtering out the
identified octrees.

2.3. Root cause analysis

Once the stope database is built, the variables are analysed to identify the root causes of OB and UB, and to
develop a good understanding of the mine site’s stope OB and UB (Figure 2. 6). Statistical analysis is used to
identify the general trends and relationships between the variables. The focus of this thesis is to analyse stope
OB and UB at an octree resolution; therefore, analysis on a per stope and per face basis will not be explored in
detail. However, these resolutions provide valuable ‘first pass’ information for directing the focus of the analysis
by identifying the critical faces, sectors or type of stope at the mine (longitudinal versus transversal). As such, a
general analysis on a per stope and per face basis will be done, working towards the finer resolution (per octree)
where the root cause analysis is performed. Per stope and per face analysis is done using univariate statistics,
such as cumulative chart. The distribution of the stope OB and UB is analysed and grouped, for example, per
face, to compare two different groups on the same chart.

48
 Data collection Root cause Predicting stope Assessing
•Data quality analysis of stope geometry predicted stope
verification OB and UB •Build multivariate or performance
•Stope reconciliation machine learning
•Univariate, bivariate •Reconcile predicted
•Variable quantification and multivariate (PCA model to predict the geometry
•Data filtering and PLS) statistical geometry (PLS and
•Quantify grade value
analysis of OB and UB random forest) and OB
and UB location (LDA) •Assess probabilistic
•Critical variable stope geometry
identification •Evaluate model
outcomes
performance
•Identify critical
•Quantify standard
location of OB and UB
error

Figure 2. 6: The second step of the proposed methodology

Multivariate analysis is used per octree (observation) for the root cause analysis. This approach allows the
identification of the complex relationships between 12 quantified causative variables and the stope OB and UB
variables in order to understand the correlation structure. The two multivariate methods used are PCA and PLS,
presented in Section 1.5.3. For each method, the variables are normalised and scaled to a mean of 0 and a
standard deviation of 1. This is due to the variables having different units. Normalising the variables gives them
the same weight in the analysis and enables the comparison of the impact of the variance of each variable on
stope OB and UB.

The PCA model is used to understand the interconnection and independence between the root cause variables.
The number of components kept will depend on the data and the variance of each component. Different rules
exist for eliminating components. In this case, the Kaiser’s rule is used (Kaiser 1960), where only the components
with a variance higher or equal to one are kept. Since the variables are standardised to a variance of 1,
components with a variance higher than or equal to 1 indicate a larger variance than a single variable, indicating
the components characterise a correlation between variables. The results are analysed by looking at the quality
and contribution tables or the loading and score charts. The quality table describes the proportion of the
variable’s variance described by a given component, while the contribution table describes each variable’s
contribution to a component’s variance. These tables are used to establish which components the variables are
associated to and interpret the correlation and independence. This information is also analysed using graphs

49
(Figure 2. 7). The loading chart (Figure 2. 7a) provides similar information to the tables, plotting only two
components at a time. The closer the variables are to the outer circle, the more they contribute to the component,
and the closer they are to the centre, the less they contribute to these components. Two variables pointing in
opposite directions means there is a negative correlation. The score chart (Figure 2. 7b) plots the component
value for each observation and is coloured using a marker style (OB and UB classes, for example). These charts
are used to assess the distribution of the OB and UB according to the two components being plotted and
determine if there is a trend in the OB and UB data that can be associated to a component. Therefore, the
variance of the variables shown by these components indicates whether they have an impact on OB and UB.

Figure 2. 7: Example of a loading chart (a) and score chart (b) that oppose two components

The PLS model is used to identify linear relationships between the root cause variables and stope performance.
The number of components is selected based on the number of performance variables, which, in this case, is
only one (only the projected distance is used for the performance variables). A loading chart is used to interpret
the linear relationships (Figure 2. 8). The weight of the variables is interpreted as the correlation with the stope
OB and UB and, therefore, the larger the weight, the stronger the correlation is to the projected distance given
the combination of variables.

50
 Figure 2. 8: Example of a loading chart using PLS

Since PCA and PLS are linear methods, non-linear relationships could be missed. Bivariate charts are used to
complete the root cause analysis by looking at variables that have shown no correlations in the PLS analysis.
This multivariate approach enables the identification of the critical variables, providing a better understanding of
stope OB and UB as well as identifying where OB and UB will occur in a face.

2.4. Predicting stope geometry

The variables identified as the root causes of OB and UB are used to predict the stope geometry (Figure 2. 9).
The use of the octree data structure means a projected distance, either OB or UB, is predicted for each octree
along the design surface, enabling the prediction of the stope geometry. The predicted stope geometry,
therefore, is represented by thousands of predicted points that are spatially georeferenced. Multivariate and
machine learning models are used to generate these predictions. These models enable an empirical approach,
using the previously mined stope data, as well as using multiple input variables, to predict future stopes’
geometry. Various statistical models based on different mathematical concepts exist. Depending on the type of
data used, for a given dataset, some models perform better than others at predicting geometry. For this
doctorate, two models are tested and presented in this thesis: PLS and random forest. An additional model is
tested as well; it allows the prediction of the location of OB and UB, but not the magnitude: LDA. PLS and LDA
are selected given their previous use and ability to predict stope OB (McFadyen 2020 and Nickson 1992).
Random forest is selected as it is a non-linear tree-based approach and has fewer constraints regarding the

51
data requirements in terms of variance and structure. All three models are relatively simple to build and learn,
facilitating the implementation of the methodology on site.

Data collection Root cause Predicting stope Assessing

•Data quality analysis of stope geometry predicted stope
verification OB and UB •Build multivariate or performance
•Stope reconciliation machine learning
•Univariate, bivariate •Reconcile predicted
•Variable quantification and multivariate (PCA model to predict the geometry
•Data filtering and PLS) statistical geometry (PLS and
•Quantify grade value
analysis of OB and UB random forest) and OB
and UB location (LDA) •Assess probabilistic
•Critical variable stope geometry
identification •Evaluate model
outcomes
performance
•Identify critical
•Quantify standard
location of OB and UB
error

Figure 2. 9: The third step of the proposed methodology

2.4.1. Partial least squares model

The first method is PLS, introduced in Section 1.5.3 which is used to build and test a linear predictive approach.
A weight is calculated for each variable which enables the construction of a linear predictive model, as given in
Equation 2.1. The importance of each variable in the model is also calculated and used to understand the
influence of each variable on stope OB and UB. The importance of the variable is, however, model dependent
and, therefore, if the model is inaccurate, limited information can be derived from it.

52
 𝑌 = 𝑇 ∗ 𝐶′ =
(𝑋 ∗ 𝑊 ′ ) ∗ 𝐶′ [2.1]
Where:
W’ = Weight of the input variable
X = Octree’s variable value
T = X component
C’ = Weight of the performance variable
𝑌 = Predicted octree projected distance

2.4.2. Linear discriminant analysis model

The second method is LDA (Fisher 1936). The goal of this discriminant approach is to predict, or classify, the
octrees as OB or UB. The magnitude of OB or UB is not considered. This model makes certain hypothesis about
the distribution of the data (follows a normal distribution) and finds the axis (p) in the dataset which best
distinguishes the two populations of OB and UB (Figure 2. 10).

Figure 2. 10: Example the classification of two groups using linear discriminant analysis (modified from Tremblay 2020)

The results are then classified using a confusion matrix to evaluate the quality of the classification (Table 2. 3).
If the octrees are well classified, the mean value of each population for each variable is analysed to understand
when OB and UB occur. The model is used to classify the performance of an octree which can then be used for
predicting the magnitude of OB and UB. If the octrees are not well classified, the causes can vary from the
selection of variables to the fact that the model is not well suited for this dataset.

53
 Table 2. 3: Confusion matrix for a threshold of 0 m projected distance. Green represents true predictions and red represents false
predictions (errors)

PREDICTED UB PREDICTED OB
ACTUAL UB True UB (TUB) False OB (FOB)
ACTUAL OB False UB (FUB) True OB (TOB)

2.4.3. Random forest model

The final method is the random forest model (Breiman 2001) which is used to build and test a tree-based
approach. This method generates multiple trees to obtain a predictive value (Figure 2. 11). Compared to a single
tree, this approach reduces the model’s variance, thus making the results and analysis more stable. This is due
to the number of trees and how each tree and node is processed. A number of observations are selected at
random for each tree and a specified number of input variables are selected at random for each node, thus
reducing the correlation between each tree. For each selection of variables, based on a specified criterion, the
variable and its value that best distinguishes the projected distance is selected for the node. This variable and
given value are used to separate the group of observations into two sub-groups, creating two new nodes. This
process is repeated until a certain end criterion is met (minimum number of observations in a terminal node or
depth of the tree). The prediction is obtained by calculating the average projected distance for the group of
observations in the final node for which the predicted octree falls into. Then, the average value for all the trees
is calculated.

Figure 2. 11: Random forest diagram for generating a prediction

54
Given the structure of the model, this method could perform better at predicting than the two other methods
presented for this dataset, as it does not depend on linear relationships to generate predictions. Its non-linear
approach makes it well suited for data that can present complex structures. For each model, the variable’s
importance and the standard error for each prediction is outputted. The variable’s importance is related to the
decision-making process of the model and is used to understand stope OB and UB if the model predictions are
deemed good. The standard error is calculated using the infinitesimal Jackknife approach (Wager et al. 2014)
and is then used to develop a probabilistic approach (the term probabilistic is not used here according to its
statistical meaning, but rather as a methodology to take into account the existing variation in the projected
distance obtained in the stopes). The standard error is used to calculate the probability density function (PDF),
which is used to determine the likelihood of observing a given projected distance and the cumulative density
function (CDF, used for determining the probabilities of exceeding a given distance). A prediction interval is
outputted using the different percentiles (Figure 2. 12). Different levels of confidence can be used. For example,
for a 60% probability interval, there is a 60% likelihood that the actual projected distance of an octree will be
within its given interval.

Figure 2. 12: Example of a prediction interval for a stope design (black) according to the probability (blue = 80% probability the
projected distances will be larger; green = median case, 50% probability; red = 80% probability the projected distance will be smaller)

Of the three methods presented, LDA is the only one that cannot be used to predict the stope geometry. It is a
classification method used to predict OB and UB location, excluding the magnitude. While it cannot be used
directly to meet the main objective, it can still provide valuable information, given its performance, on the stope
OB and UB location and can be used as a variable in the other models.

55
 2.4.4. Evaluating model performance
Once the predictions are generated, the model’s performance is quantified to determine the quality of the
predictions. The goal of quantifying the model performance is to provide metrics and characterise the predictions
to determine if the predicted geometry is realistic and can be used to optimise the design. To do so, the data is
separated chronologically into three groups: the first group represents the first 50% of the data, the second group
represents the following 25% and the final group represent the final 25% of the data. The model is trained with
the first group of data. The second group is used to test the predictive performance and optimise the model’s
parameters. The third group is used to evaluate the final model. A similar process could also be done with only
two groups if the amount of data is limited.

The model’s predictive statistical performance is assessed by determining the prediction error (difference
between the predicted distance and the actual projected distance) and whether an octree is correctly predicted
as OB or UB. The model’s capacity to classify the octrees as OB or UB is measured by looking at the confusion
matrix presented in Table 2. 3 in Section 2.4.2. Using this matrix, different performance metrics are calculated,
such as the classification accuracy (ACC, percentage of octrees correctly predicted), the positive OB predictions
(POP, percentage of OB predictions that are correct), the true OB detection (TOD, percentage of overbroken
octrees predicted as OB) and the true UB detection (TUD, percentage of underbroken octrees predicted as UB).
These metrics enable the evaluation of a specific quality of the model. To evaluate the global quality of the
classifications, the Matthews correlation coefficient (MCC, Matthews 1975) is used, with 1 being perfect
predictions, 0 being random and -1 being all incorrect predictions. Table 2. 4 gives the equation for these metrics.
Table 2. 4: Performance metric equations for evaluating the model’s capacity to classify an octree as OB or UB. Refer to Table 2.3 for
the acronyms

Performance metric Equation

Classification accuracy 𝑇𝑈𝐵 + 𝑇𝑂𝐵
(ACC) 𝑇𝑈𝐵 + 𝑇𝑂𝐵 + 𝐹𝑈𝐵 + 𝐹𝑂𝐵
Positive OB predictions 𝑇𝑂𝐵
(POP) 𝑇𝑂𝐵 + 𝐹𝑂𝐵
𝑇𝑂𝐵
True OB detection (TOD) 𝑇𝑂𝐵+ 𝐹𝑈𝐵
𝑇𝑈𝐵
True UB detection (TUD)
𝑇𝑈𝐵 + 𝐹𝑂𝐵
Matthews correlation 𝑇𝑈𝐵 ∗ 𝑇𝑂𝐵 − 𝐹𝑈𝐵 ∗ 𝐹𝑂𝐵
coefficient (MCC) √(𝑇𝑂𝐵 + 𝐹𝑂𝐵) ∗ (𝑇𝑂𝐵 + 𝐹𝑈𝐵) ∗ (𝑇𝑈𝐵 + 𝐹𝑂𝐵) ∗ (𝑇𝑈𝐵 + 𝐹𝑈𝐵)

The prediction error is measured by calculating the absolute difference between the predicted value and the
actual value of the projected distance for the octree. These errors are placed into arbitrarily determined brackets
to characterise the spread in the prediction error (Table 2. 5). In addition, the RMSE (equation 2.2) is calculated
from the predicted and observed distances.

56
 (𝑦̂𝑖 −𝑦𝑖 )2
𝑅𝑀𝑆𝐸 = √∑𝑛𝑖=1 𝑛
[2.2]

Where:
n = number of observation
i = Selected observation
𝑦i = Observed value for i
𝑦̂i = Predicted value for i

Table 2. 5: Prediction error brackets for evaluating the model’s statistical performance

Absolute error brackets

% <0.5 m
% <1 m
% <2 m
% >2 m
% >3 m
% >4 m

The prediction error is also evaluated analytically through a bivariate chart of the predictions versus the observed
values for each octree. For a perfect model, the predicted CMS would match exactly the actual CMS. The data
would follow a linear trend with all the points falling on a 45° line. This, however, never occurs due to the
complexity of the underground geological and mining environment where the variation in the geomechanical
properties of the rock mass, as well as deviation from the planned data during mining, cannot be fully captured,
thus causing inherent uncertainty. The drill and blast methods are also not an exact method for breaking the
rock with metre precision, causing additional uncertainties. The general trend of the predictions is captured using
a linear regression fit and compared with the linear trend of a perfect model.

Even if the predictions do not follow the linear fit of a perfect model, the predictions still provide valuable
information about the location of OB and UB, and the expected magnitude, since this innovative method predicts
the detailed geometry of the expected void. Given the fact that the location is not considered in the model, the
capacity of the model to predict the geometry correctly cannot be measured using only the prediction error or
the confusion matrix. The predicted geometry needs to be verified locally as well to see if it matches the mined
geometry, allowing one to verify if the predictions can be used by the engineers for decision-making, based on
information going beyond a simple ELOS prediction. Therefore, the capacity of the model to predict the geometry
correctly is evaluated in two steps.

The first step is to quantify the three-performance metrics used for quantifying the performance of the predictions.
Each metric characterises either the magnitude of the projected distance, the shape of the face or the OB and

57
UB locations and allows to set minimum criteria that should be met for determining if a prediction of the geometry
is acceptable. The criteria are arbitrarily selected, based on the understanding of mine planning and the impact
OB and UB will have, and the tolerance to these impacts at the different mining operations. An example of criteria
that could be adopted is as follows:
• Prediction error criterion:
o Characterises the capacity of the model to correctly predict the magnitude of OB and UB by
quantifying the prediction error for each octree.
o To meet the criterion, the prediction error needs to be within X m for at least 50% of the octrees
and within Y m for at least 80% of the octrees. X and Y vary according to the mine site.
• Shape criterion:
o Characterises the capacity of the model to correctly predict the shape of the mined stope by
converting the projected distances of each octree in a face to a cumulative count (percentage
value) from the smallest projected distance (0%) to the largest (100%).
o The cumulative distribution of the projected distance in the face is a scale-independent shape
measure that allows to discard the magnitude of the projected distance or the OB-UB concept
and focus on the shape of the face.
o To meet the criterion, at least 70% of the octrees in the surface need to have the same
cumulative position (±10%) in the predicted and mined geometry.
• OB-UB location criterion:
o Characterises the capacity of the model to correctly predict where OB and UB will occur by
quantifying the percentage of octrees within the face that are correctly predicted as OB or UB.
o To meet the criterion, at least 70% of the octrees in the surface need to be correctly identified
as OB or UB.

The thresholds used for assessing the prediction error may vary between sites due to the impact OB and UB
have on the economic performance of different stopes. For example, 1 m OB for a longitudinal stope that is 2 m
to 5 m wide with clear waste-ore boundaries will see dilution increase by 20% to 50% for every metre of OB.
Therefore, the tolerated error for the predicted geometry should be small (under a metre) to allow the mine site
personnel to get an exact estimate of their stope performance. On the other hand, for low disseminated grade
using transversal stopes that are 20 m wide, dilution would be limited to only 5% or less for every metre of OB.
Therefore, the tolerated error for the predicted geometry can be larger to allow the mine site personnel to get an
adequate estimate of stope performance. The thresholds for the prediction error will be defined for each site
during the analysis. A threshold of 70% is used for the OB and UB location and the distribution of the projected
distance. This number was selected arbitrarily. However, this means that the majority of the surface is correctly

58
predicted, leaving only 30% or less up for further interpretation. Furthermore, deduction can be made of the
remaining 30% as the majority of the predictions can be trusted. These incorrect octrees are not necessarily
adjacent to each other and can be spread out across the face, thus reducing the impact they have on the overall
stope performance assessment.

Given that the geometry of the stope is predicted using octrees, different categories of prediction performance
are set depending on the capacity of the predictions to capture the geometry, magnitude of the projected distance
or the location of OB and UB. While the goal of this thesis is to predict the geometry of the stope, the predictions
can still provide information regarding the general shape of the stope or the OB and UB location. Therefore, the
geometry assessment is arbitrarily classified into four categories, allowing different prediction performance of
the geometry:
• Excellent prediction:
o The prediction error criterion and shape criterion are met.
o The predictions identify the magnitude and distribution of the projected distance across the
face within a set threshold, allowing the prediction of the mined geometry.
o The prediction error threshold depends on the stope being analysed.
o These predictions can be used by the engineers to assess the expected stope geometry.
• Good prediction:
o The shape criterion is met, but not the prediction error criterion.
o The predictions identify the distribution of the projected distance across the face, allowing
engineers to determine the general shape of the mined stope. However, the magnitude of the
projected distance cannot be identified across the face within a set threshold.
o The predictions allow engineers to determine where the larger OB and UB will occur, but not
its magnitude.
• Limited prediction:
o The OB-UB location criterion is met, but not the prediction error criterion and shape criterion.
o The predictions allow engineers to identify where OB and UB will occur, but do not identify the
magnitude or distribution of the projected distance for the majority of the surface.
• Insufficient prediction:
o None of the criteria are met.
o The predictions do not allow the determination of the distribution of the projected distance,
and where OB and UB will occur, and its magnitude for the majority of the surface.

59
After the predictions have been categorised, the second step is to validate if the criteria placed the face in the
proper category as these criteria are not perfect and fringe cases between two categories could change. This is
done analytically in the 3D space on the design surface by comparing side-by-side the predictions with the
observed projected distance for each face. The visual assessment is made by comparing side-by-side the
predictions with the observed projected distance for each face using three different marker styles: the projected
distance, the absolute prediction error and the cumulative distribution of the projected distance (Figure 2. 13).
The cumulative distribution represents the projected distance compiled from smallest (purple colours) to largest
(red colours). The projected distance and prediction error are used to assess the predicted magnitudes of OB
and UB and how they compare with the actual OB and UB. The cumulative distribution of the projected distance
is used to assess how the predicted geometry or shape matches the geometry/shape of the actual mined stope.
It is possible for the predictions to have enabled the depiction of the geometry, but for the prediction error to
exceed a tolerated threshold, meaning the magnitudes do not match. The combination of these three markers
enables the assessment of the performance of the predictions.
Absolute prediction error Cumulative distribution
Projected distance (m)
(m) of the projected distance

Observations

Predictions

Figure 2. 13: Example of the analytical comparison between the predicted face and the actual face. The cumulative distribution
represents the projected distance compiled from smallest (purple colours) to largest (red colours)

There is one final factor to consider; the probabilistic approach that is enabled by the random forest model. A
prediction interval is given with each prediction which allows to assess different possibilities for the mined
geometry. While the predicted face may not be termed excellent with the median case (prediction outputted by
the model), the prediction interval might still allow to adequately capture the geometry. Therefore, an example
of criterion for the random forest model that integrates the prediction interval is as follows:

60
 • Excellent prediction:
o The prediction error criterion and shape criterion are met using the median case.
Or
o The shape criterion is met, but not the prediction error criterion using the median case.
However, the actual projected distance is within the 10th and 90th percentiles of the prediction
interval for 70% or more of the octrees.
• Good prediction:
o The shape criterion is met, but not the prediction error criterion using the median case.
Or
o The OB-UB location criterion is met, but not the shape criterion using the median case.
However, the actual projected distance is within the 10th and 90th percentiles of the prediction
interval for 70% or more of the octrees.
o The probability interval can still be used for assessing the possibility of large OB or UB.
• Limited prediction:
o The OB-UB location criterion is met using the median case.
Or
o None of the criteria are met using the median case. However, the actual projected distance is
within the 10th and 90th percentile of the prediction interval for 70% or more of the octrees.
o The probability interval can still be used for assessing the possibility of large OB or UB.
• Insufficient prediction
o None of the criteria are met.
o The actual projected distance is within the 10th and 90th percentiles of the prediction interval
for less than 70% of the octrees.
o There is the potential to still use the probability interval for assessing the possibility of large
OB or UB.

The use of the probability facilitates the interpretation of the results by quantifying the variability in the results,
offering a prediction interval that captures the actual geometry, in parts or completely. The probabilistic criterion
is determined arbitrarily. The 70% threshold is used for the same reasons discussed earlier in this section. The
10th and 90th percentiles are used as they give a prediction interval that should cover 80% of the octrees. The
probabilistic criterion changes how the different categories are met. For example, if the distribution is depicted
correctly for at least 70% of the surface using the median case, but the prediction error is over X m for 50% of
the octrees and/or over Y m for 80% of the octrees, then the face would be categorised as ‘good prediction’.
However, if the actual projected distance is within the 10th and 90th percentiles of the prediction interval for 70%

61
or more of the octrees, then it is assumed the actual geometry can be captured within the prediction interval and,
therefore, the face is categorised as ‘excellent prediction’. The inclusion of the probabilistic criterion allows to
bump a predicted face up one category if it meets the probabilistic criterion.

Following the model performance assessment, the capabilities of the proposed methodology to predict the mined
geometry and meet the objectives are determined, considering the prediction results and site-specific factors.

2.5. Assessing predicted stope performance

Once the predictions are obtained, the predicted geometry is analysed to understand where the OB and UB will
occur in the stope, their magnitude, the economic impact it will have and the possible scenarios (Figure 2. 14).
The goal is to determine the expected OB and UB, the grade value and the probabilities for OB or UB to exceed
an acceptable threshold. This information will then be used by engineers for planning and possibly modify certain
steps of the mining process to minimise the OB or UB volumes or their impact.

Data collection Root cause Predicting stope Assessing

•Data quality analysis of stope geometry predicted stope
verification OB and UB •Build multivariate or performance
•Stope reconciliation machine learning
•Univariate, bivariate •Reconcile predicted
•Variable quantification and multivariate (PCA model to predict the geometry
•Data filtering and PLS) statistical geometry (PLS and
•Quantify grade value
analysis of OB and UB random forest) and OB
and UB location (LDA) •Assess probabilistic
•Critical variable stope geometry
identification •Evaluate model
outcomes
performance
•Identify critical
•Quantify standard
location of OB and UB
error

Figure 2. 14: The fourth step of the proposed methodology

2.5.1. Stability assessment

The predicted stope geometry is reconciled with the design, allowing to quantify the predicted OB and UB. Given
the octree resolution, the areas where OB and UB are expected are determined, as well as the probability for
the OB or UB to exceed a given threshold at each octree location. The tolerance for OB or UB is dictated by the
geological and mining environment and, therefore, their impact on the profitability and stability of the stope will
vary from one mine site to another. To assess the OB and UB, the predicted stope and the octrees are plotted

62
in a 3D view, colouring the octrees using markers to show the predicted distance or the probability of exceeding
a given threshold of OB or UB. This allows to determine the areas of concern and the probability of experiencing
critical OB or UB. Using each octree’s CDF, the prediction interval for each octree is calculated to give a
prediction interval for the expected geometry. This allows to determine the proportion of the design surface
(considering all the octrees of the stope) that exceeds a set projected distance for each percentile of the
prediction interval and is analysed by plotting the information in a chart (Figure 2. 15). In this example, just over
30% of the octrees exceed the threshold for the median scenario (50th percentile).

Figure 2. 15: Percentage of octrees above the threshold value for the selected percentile geometry within the probability interval

For the prediction interval, three different geometries (for example the 25th, 50th and 75th percentile of the CDF)
are generated based on the projected distance of each octree for the three different selected percentiles of the
CDF, thus giving the prediction interval (Figure 2. 16). For the example, in Figure 2. 16, there is a 50% probability
(prediction interval between 25% and 75%) that the projected distance for the octree will be between 1.3 m and
2.7 m. The 50th percentile represents the geometry generated by the model (median case) and is used to
determine where the OB and UB will most likely occur. Percentiles over 50 are used to assess worst-case
scenarios for OB (projected distance increases with the percentile) and percentiles under 50 are used to assess
worst-case scenarios for UB (projected distance decreases when the percentile decreases). The different
geometries generated using the probabilities are then reconciled with the stope design to calculate the predicted
OB and UB volumes and perform an economic assessment.

63
 Figure 2. 16: Example of the predicted distance for an octree based on the CDF percentile used

2.5.2. Economic assessment

An economic assessment is made using data from a grade block model where the grades for each economic
metal and the NSR are given per block. The NSR designates the value of the ore in $ per tonne after the mill
and is often used as the economic variable for designing the stope as it encompasses all the metals. The NSR
value is used for the analysis in this thesis. Figure 2. 17 shows an example block model with the NSR given per
block. The goal is to characterise the economic value of the predicted geometry, enabling engineers to use
economic data for the design and optimisation process.

64
 Figure 2. 17: Example of the grade block model imported around a stope

The grade value for the OB, UB or other volumes is obtained by mapping the block model data to the octree
data structure. The weighted average and the total metal value in each octree are calculated from the grade
blocks that intersect the octree block (Figure 2. 18a). Backfill information is also considered, adjusting the grade
value in consequence. The areas with low or high grade are identified by plotting the data in a 3D view for the
different percentiles (Figure 2. 18b). The data is separated into the different volume types (OB, UB design, mined
and recovered), calculating for each volume the average and total NSR per face and per stope. This information
is calculated for the different probabilities to assess and compare the economic impact of the OB and UB
magnitude on the profitability of the stope.

65
 a b
Figure 2. 18: Example of the calculation of the NSR value per octree (a) and an example of the result per octree block of the NSR value
adjusted with backfill information (b)

Since the economic data is calculated and mapped to the octree data structure, instead of only averaging the
value of each block within each octree, the blocks are also subdivided according to the octree structure with
information such as the volume type (design, mined, OB and UB) mapped to each sub-block. This allows to
calculate the distribution of the tonnes in each volume based on their economic value and be plotted on a
cumulative distribution chart of the grade value for each tonne in the stope or specific volume (Figure 2. 19).
With this chart, the distribution of the grade, or in this case the NSR, in each volume (design, mined, OB and
UB) is compared and used to determine if the predicted geometry follows the design NSR tonne distribution. It
is also used to calculate the number of tonnes under the cut-off grade for each volume type which is overlaid in
this chart. The economic value of the OB and UB is assessed, evaluating the proportion of waste and ore to
determine how the OB and UB impact the stope dilution and loss of ore. The tonnes for each volume is indicated
as a footnote to show the relative importance of the cumulative volume.

66
 Figure 2. 19: Example of the cumulative distribution of the NSR value for each tonne based on the different types of volume. Tonnes
per volume are indicated as a footnote. The black line represents a fictive cut-off value

At this point in the economic assessment, the overall distribution and economic value of the stope has been
quantified, informing if the predicted stope performance is susceptible to dilution or loss of ore that is critical to
the mine site. To optimise the stope, the critical areas susceptible to the dilution or loss of ore are identified. This
is done analytically by visualising the data in the 3D view to identify their spatial location. Figure 2. 20 is an
example of a selection of OB octrees plotted in 3D and coloured using their NSR value. Octrees with an NSR
value bellow cut-off are semitransparent to highlight the areas where dilution is expected to occur.

67
Figure 2. 20: 3D view of the NSR value for the octrees located in the OB. Octrees with an NSR value below the fictive cut-off grade are
transparent

The dilution and loss of ore are calculated using the selected cut-off grade and are used for assessing the
profitability of the stope. Table 2. 6 lists the different economic metrics that will be used in this thesis to
characterise dilution and loss of ore. The selection is based on literature (Section 1.6) and from what is seen on
mine sites. These values are compared to the different predicted geometries to assess the impact of OB and
UB on dilution.
Table 2. 6: Different dilution and loss of ore variables used in this thesis

Variables Formulae
𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑
Predicted dilution
𝑂𝑟𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑
𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
Planned dilution
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑂𝐵
Unplanned dilution
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
𝑂𝑟𝑒 𝑖𝑛 𝑈𝐵
Loss of ore
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛
𝑂𝑟𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑
Recovered ore
𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

Following the assessment of the stope performance, the chances for problematic OB and UB are determined. If
the predicted stope performance is satisfactory (economically and geometrically) and the probabilities for critical

68
OB or UB are within the mine site’s risk tolerance, engineers can proceed with the final planning of the stope. If
concerns about the OB or UB have been flagged during the stope performance assessment, mitigations can be
applied to optimise the stope performance.

2.6. Software
The main software used to develop this method is mXrap (Harris and Wesseloo 2015). mXrap is a geotechnical
software originally developed for seismicity in underground mines. It is developed by the Australian Centre for
Geomechanics (ACG), and is used by over 70 mine sites, as well as by universities and consulting firms across
the world. The main advantage of mXrap is the capability it provides for developing tools and apps for the
proposed design methodology for mine sites to test. mXrap also currently has the only tools available for
performing reconciliation on an octree basis. The stope reconciliation and analysis app (Woodward and Harris
2019) is used to build the database for each mine site.

Multivariate analysis is done using the R software (R Core Team 2021). This free platform is developed for
statistical computing and graphics, and has powerful plugin scripts to run various statistical analyses. Different
scripts are employed based on the statistical method used. For PCA, the package FactoMineR (Le et al. 2008)
is used, and for PLS, the package PLS is used (Mevik and Wehrens 2007). The statistical models are computed
in R as well. PLS uses the PLS package, LDA uses the MASS package (Venables and Ripley 2002), and random
forest uses the Ranger package (Wright and Ziegler 2017).

The predicted CMS surface is built using CloudCompare (Girardeau-Montaut 2022). This free software is
developed for the manipulation and analysis of point clouds and triangular mesh and has the necessary tools to
build a surface from a point cloud.

2.7. Conclusion
A detailed step-by-step description of the methodology was presented. This chapter covered the data collection,
data analysis, model building, predicted stope performance assessment, and the software used. The
methodology enables one to build a good understanding of stope OB and UB through data at an octree resolution
and multivariate statistics. This information is then used to predict the projected distance, OB and UB at the
octree resolution, capturing the spatial variation of the stope OB and UB, thus allowing to determine the expected
mined geometry at the stope design stage. The use of random forest enables a probabilistic approach, predicting
multiple scenarios as well as including a variety of variables, while the use of octrees enables to make an
economic assessment of the predicted stope, thus providing additional knowledge for the design and planning
of the stope.

69
This methodology will be applied on three different sites. The following chapter presents the case studies used
for testing the proposed methodology.

70
Chapter 3. Case studies
The proposed methodology is used with data from three different mine sites to demonstrate the flexibility and
capabilities of this site-specific approach. The three case studies presented in this chapter are located in
Australia and Canada, and were selected based on the varying geological and mining conditions as well as the
different mining methods used (transversal, longitudinal, sub-level), the quality of the data and the number of
stopes in the database. The impact the number of stopes has on the analyses or predictions is not assessed in
this thesis. However, since there is no set number of stopes required to run the prediction models, an arbitrary
minimum number of 40 stopes is set for this thesis. This ensures that at least 40 stopes are used for the analysis
and 30 stopes are used for building the models. Contrary to a per face resolution, this ensures that at least 100
000 octrees are considered for each mine site, which is deemed a reasonable amount of data. The raw data for
these mine sites (design, CMS, drill strings, drifts, and fault surveys) have been compiled and stored on an
external hard drive given the size and format, and will be stored on a repository at Université Laval.

3.1. Dugald River

Dugald River mine is located in Queensland, Australia (Figure 3. 1) where an underground Zn-Pb-Ag deposit is
mined. They mine longitudinal and transversal stopes. Stope dimensions will vary depending on the area of the
mine, but they are typically around 25 m high between levels and 15 m to 30 m long (Hassell et al. 2015). Stope
widths will depend on the thickness of the orebody and vary between 2 m and 35 m. The orebody reaches a
depth of approximately 1 km.

Figure 3. 1: The location of Dugald River mine, Queensland, Australia

71
 3.1.1. Data overview
The mine database comprises 173 stopes mined between 2014 and 2021 (Figure 3. 2). Of these 173 stopes,
41 are a part of the north area of the mine and 132 are a part of the south area of the mine. Nineteen stopes are
also part of the 2014 trial stoping stage that was conducted to test the validity of the geotechnical and mining
variables (Hassell et al. 2015). For the analysis, the 2014 trial stoping, as well as stopes with data quality
concerns (bad mesh, missing data or root cause of OB not characterised by the variables) have been removed,
resulting in a total of 130 stopes. Stopes from the 2014 sector have previously been part of a study by McFadyen
et al. 2019b which conducted root cause analysis. Univariate and bivariate analysis was done at a per stope,
per face and per octree level, revealing that stopes generate larger OB than UB with no specific face associated
to the OB. The octree analysis identified the blasting energy proxy, blasting standoff, ERF, geometry complexity
and distance to fault as variables that showed trends towards OB or UB for each or specific faces.

Figure 3. 2: 3D Section view of the stope database with the trial area delineated in red

From the 130 stopes, 49 are transversal stopes and 81 are longitudinal stopes. Since longitudinal and transversal
stopes can behave differently, they will be treated as two different datasets. For this thesis, only the transversal
stopes will be discussed as they are grouped in one area of the mine and have been mined more recently. These
stopes are shown in Figure 3. 3a. When building the models, the data is divided into three groups, with the third
group representing the stope predicted. These stopes are shown in Figure 3. 3b. After filtering out drives, backfill
and muckpiles, close to 396 000 octrees make up the dataset.

72
Figure 3. 3: Longitudinal view of the transversal stopes in the database (a) and the stopes forming part of the third group for the model
building process (b)

Reconciliation for each stope is performed on a per stope, per face and per octree basis. Depending on the
geometry, stopes are typically separated into six faces: the hangingwall (HW), the footwall (FW), the floor, the
crown and the two stope ends (Figure 3. 4). During the survey building process, drifts are cut out of the design
and CMS. For this reason, stope faces that represent drift surfaces (floors and crowns mostly) are not part of
the analysis. Additionally, faces mined against backfill are not part of the analysis.

Figure 3. 4: Example of stope face delimitation using colour code and orientation

The variables included in the analysis are presented in Table 3. 1. For the structural variables, there are absent
data since there is not always a fault in front of the design surfaces. When the distance to fault and angle to fault
are being calculated, a maximum search distance of 30 m from the octrees is used. This distance was set based
on the work by Woodward et al. 2019. For octrees with absent fault data, a unique value needs to be given to
include these octrees in the model (the software package used does not allow for missing data). This unique
value needs to be outside of the range of values in the dataset for these variables. Arbitrary choices of a distance

73
of 40 m and an angle of 100° are set to separate them from the octrees with fault data and keep them in the
model.

Table 3. 1: Quantified variables for Dugald River mine

Categories Variables Unit Description

ERF – 2D measure of the hydraulic radius
Dip of the design face at the octree
Dip °
location
Direction of the design face at the octree
Geometrical Direction °
location
Stope geometry complexity m Measure of convex shapes in the design
Measure of stope undercut by drifts using a
Undercut m
convex hull that includes the drifts
Geomechanical None – –
Distance between the octree and the
Distance to fault m
nearest fault
Angle between the octree and the nearest
Geological Angle to fault °
fault
Metric describing the position and distance
Fault directional measure °
to the nearest fault
Blasting energy proxy – Energy estimated from the drill design
Angle between the octree and the nearest
Blasting orientation °
blasthole
Operational Distance between the octree and the
Blasting standoff m
nearest blasthole
Metric describing the position and distance
Drift directional measure °
to the nearest drift
Characterisation of the similarity between
Spatial Dissimilarity –
the octree variables
Distance between the octree and the CMS
Performance Projected distance m
in a normal direction to the design

3.2. Prominent Hill

Prominent Hill is a mine located in South Australia (Figure 3. 5) where an underground copper deposit with gold
and silver is mined. They mine sub-level transversal stopes to extract the ore using a primary–secondary
sequence. Stopes dimensions vary depending on sectors, but are typically 20 m wide by 20 m long or 30 m by
30 m and up to 90 m high. The orebody reaches depth of approximately 1 km.

74
 Figure 3. 5: The location of Prominent Hill mine, South Australia, Australia

3.2.1. Data overview

The stope database is comprised of 73 stopes mined between 2018 and 2022, inclusive (Figure 3. 6).
Reconciliation on a per stope, per face and per octree basis was conducted for each stope. To ensure the quality
of the analysis, any octrees that coincide spatially to drives are excluded as they represent existing voids, as
well as backfill faces due to their different strength and stress conditions.

75
 Figure 3. 6: 3D section view of the stope database for Prominent Hill

Of the 73 stopes, 18 of them diverge from the typical design geometry as they are smaller or thinner than design
and are cut longitudinally. Since the controlling factors will most likely vary for these stopes, they are excluded
from the following analysis. Also, due to data quality issues (missing data), four stopes are excluded from the
analysis. Therefore, the analysis will focus on the 51 transversal stopes (Figure 3. 7). After filtering out drives,
backfill and muckpiles, close to 253 000 octrees make up the dataset used.

Figure 3. 7: Stopes at Prominent Hill selected for the analysis

As mentioned before, the model building process is done by separating the data into three groups following the
sequence. For this mine site, the first group comprises 26 stopes, and the last two comprise 12 and 13 stopes

76
(separated following the mining sequence). The third group is used for evaluating the quality of the predictions.
These stopes are presented in Figure 3. 8.

Figure 3. 8: Prominent Hill stopes predicted with the models

The variables quantified for the Prominent Hill mine site are presented in Table 3. 2. A structural model is
available for the quantification of the fault variables for this mine site. However, limited faults are present in the
mining area with only 10% of the octrees having fault data (four stopes, Figure 3. 9a). Discussions with site
personnel have identified these faults as non-critical to stope OB and UB and are, therefore, excluded from the
analysis. However, RQD data is available through a geotechnical block model and correlations can be made
between the presence of the faults and the RQD value (Figure 3. 9b). The block model is built using borehole
data. The nearest RQD value is mapped to the block with the nearest distance varying between 0 m and 30 m
for the whole block model. 20% of the blocks are within 5 m, 60% within 10 m and 90% within 15 m. The RQD
is directly logged from the core. Blasting data consists of the design for blastholes which are to be fired for each
stope.

77
Figure 3. 9: Structural model for the studied area (a) and an example of the RQD value for a stope near faults (b)

78
 Table 3. 2: Quantified variables for Prominent Hill

Categories Variables Unit Description

ERF – 2D measure of the hydraulic radius
Dip of the design face at the octree
Dip °
location
Direction of the design face at the octree
Geometrical Direction °
location
Stope geometry complexity m Measure of convex shapes in the design
Measure of stope undercut by drifts using a
Undercut m
convex hull that includes the drifts
Geomechanical RQD % Measure of the rock mass quality
Distance between the octree and the
Distance to fault m
nearest fault
Geological
Metric describing the position and distance
Fault directional measure °
to the nearest fault
Blasting energy proxy – Energy estimated from the drill design
Angle between the octree and the nearest
Blasting orientation °
blasthole
Operational Distance between the octree and the
Blasting standoff m
nearest blasthole
Metric describing the position and distance
Drive directional measure °
to the nearest drive
Distance between the octree and the CMS
Performance Projected distance m
in a normal direction to the design

The dissimilarity variable is not used for this mine site. It is an experimental variable, and a decision was made
to include it in the first article in Chapter 4. The objective of using this variable was to integrate the spatial aspect
in the analysis, but showed limited results. Therefore, it is only used with the Dugald River mine site data.

3.3. Westwood
Westwood mine is located in Quebec, Canada (Figure 3. 10) where an underground Au-Cu-Zn-Ag deposit is
mined. The deposit is comprised of multiple lenses divided into the east and west area. They use longitudinal
stopes to extract the ore. Stope dimensions vary depending on the area of the mine, but they are around 30 m
high, between levels, and 10 m to 20 m long. Stope widths vary between 2 m and 4 m, depending on the orebody
thickness. The orebody reaches a depth of more than 1 km.

79
 Figure 3. 10: The location of Westwood mine, Quebec, Canada

3.3.1. Data overview

The stope database is comprised of 219 stopes mined between 2014 and 2023, inclusive (Figure 3. 11).
Reconciliation on a per stope, per face and per octree basis was conducted for each stope. To ensure the quality
of the analysis, any octrees that coincide spatially to drives are excluded as they represent existing voids, as
well as backfill faces due to their different strength and stress conditions.

Figure 3. 11: 3D section view of the stope database for Westwood

80
Given the different mining areas, the analysis will focus on one area of the mine; the WW28 area. Of the 219
stopes, 42 are in the WW28 area, which is the most for any area. It is also the only area with a complete dataset.
Due to data quality issues (missing drill rings and incomplete CMS), two stopes are excluded from the analysis.
Therefore, the analysis will focus on the 40 longitudinal stopes (Figure 3. 12). After filtering out drives, backfill
and muckpiles, close to 203 400 octrees make up the dataset used.

Figure 3. 12: Selected stopes for the Westwood analysis

As mentioned before, the model building process is done by separating the data into three groups following the
sequence. For this mine site, the first group comprises 20 stopes and the last two groups comprise 10 stopes
each. The third group is used for evaluating the quality of the predictions. The stopes are presented in Figure 3.
13.

Figure 3. 13: Westwood stopes predicted with the models

81
The variables quantified for the Westwood mine site are presented in Table 3. 3. Blasting data consists of the
design for blastholes which are to be fired for each stope. A structural model is available for the quantification of
the fault variables for this mine site. For octrees with absent fault data, an arbitrary choice of a distance of 40 m
is set and an angle of 100° to separate them from the octrees with fault data and keep them in the model (same
values used for Dugald River). This is due to the software package not allowing to use observations with missing
data. Two geotechnical block models are available for this mine site. Site engineers built the 4 m by 4 m block
models using a kriging method from geotechnical logging of boreholes and underground mapping data. These
block models were validated by the site engineers and are used as is for the purpose of this thesis. Further
validation is not a part of the objectives set. The first block model includes RQD and GSI (which consider the
alteration, schistosity and structures). The data doesn’t distinguish the different directions of the boreholes, and
part of the data is extrapolated from qualitative information (older boreholes where quantitative measures were
not taken). The second block model includes the uniaxial compressive strength (UCS) and the contrast UCS.
The UCS values are estimated from geochemistry analysis of the rocks based on a general trend found between
UCS values and X-ray fluorescence (XRF) measures for around 100 samples (Martel and Tremblay 2023).
These UCS values represent an estimate and are used by the mine site to get a general idea of the fluctuation
of the rock strength for seismicity risk assessment purposes. The contrast UCS represents the difference
between the selected block’s UCS value and the UCS value of the adjacent block to the north. The north–south
direction is used as it is the direction perpendicular to the lithology and foliation, which are the main controls of
the UCS spatial variation. The block model was populated using the kriging method. For the WW28 area, the
data sampling per block varies from two samples of a single borehole to 20 samples of 12 boreholes. The
average distances of the sampling per block vary also, from 2 m to 200 m (96 m at the 50% mark). The two
block models were merged, creating new blocks that contain information for the four variables.

82
 Table 3. 3: Quantified variables for Westwood

Categories Variables Unit Description

ERF – 2D measure of the hydraulic radius
Dip of the design face at the octree
Dip °
location
Direction of the design face at the octree
Geometrical Direction °
location
Stope geometry complexity m Measure of convex shapes in the design
Measure of stope undercut by drifts using a
Undercut m
convex hull that includes the drifts
RQD % Measure of the rock mass quality
GSI Measure of the rock mass quality
Geomechanical UCS MPa Estimate of the rock strength
Estimate of rock strength contrast with the
UCS contrast MPa
surrounding rock
Distance between the octree and the
Distance to fault m
nearest fault
Geological
Metric describing the position and distance
Fault directional measure °
to the nearest fault
Blasting energy proxy – Energy estimated from the drill design
Angle between the octree and the nearest
Blasting orientation °
blasthole
Operational Distance between the octree and the
Blasting standoff m
nearest blasthole
Metric describing the position and distance
Drive directional measure °
to the nearest drive
Distance between the octree and the CMS
Performance Projected distance m
in a normal direction to the design

The dissimilarity variable is not used for this mine site. It is an experimental variable, and a decision was made
to include it in the first article in Chapter 4. The objective of using this variable was to integrate the spatial aspect
in the analysis, but this showed limited results. Therefore, it is only used with the Dugald River mine site data.

3.4. Conclusion
The following case studies were presented in this chapter: Dugald River mine, Prominient Hill mine and
Westwood mine. Two mines are located in Australia and one in Canada, and all three present a wide range of
mining conditions and stope geometry. The data from the three mine sites are used in the next chapters to apply
the proposed methodology. The results for the root cause analysis for Dugald River is presented in the next
chapter in the form of an article, while Prominent Hill and Westwood are presented in Chapter 5.

83
Chapter 4. Article 1 - Assessing stope
performance using georeferenced octrees and
multivariate analysis
Benoît McFadyena, Martin Grenona, Kyle Woodwardb and Yves Potvinb

aDépartement de génie des mines, de la métallurgie et des matériaux, Faculté des sciences et de génie,
Pavillon Adrien-Pouliot, Université Laval, Québec, Canada;
bAustralian Centre for Geomechanics, The University of Western Australia, Perth, Australia

Keywords: stope design, stope reconciliation, overbreak, underbreak, multivariate, octree.

In order for mines to reach a better understanding of the root causes of their stopes OB and UB, this thesis
proposes a multivariate analysis approach that utilises the octree database compiled for each mine site. The
objective is to demonstrate, through this approach, that the relationship between variables and their impact on
the magnitude and spatial distribution of OB and UB can be identified, enabling sites to reach a better
understanding of what is happening when they mine their stopes. In the same time, the selection of variables
will be validated. This chapter presents the first scientific article that is integrated in this thesis. The article
discusses the root cause analysis for the Dugald River mine site.

4.1. Résumé
Cet article présente les premières étapes vers une nouvelle approche pour la conception de chantier de type
chambre ouverte, où le sur-bris (volume de roche non planifié qui est miné) et le sous-bris (volume de roche
planifié laissé en place) sont évalués à l’aide d’analyses statistiques multivariées (analyse en composantes
principales et moindres carrés partiels) et des données géoréférencées mesurées à une résolution d’environ un
mètre cube (octrees). Les résultats ont montré que le sur-bris est observé dans les zones du chantier où il y a
une grande énergie de dynamitage, une grande distance de surdéveloppement des galeries et vers le milieu
des surfaces. Le sous-bris a tendance à être observé près de la bordure des surfaces, là où il y a une grande
distance avec le dynamitage, et lorsque les trous de forage interceptent les surfaces. La résolution des données
et l’utilisation de l’analyse multivariée ont permis d’interpréter avec une grande précision la variation de la
performance du chantier le long de sa surface, allant au-delà de la simple analyse par face.

4.2. Abstract
This paper presents the first steps towards a new stope design approach, where stope overbreak (OB; rock
mined outside of the design) and underbreak (UB; planned rock left behind) is assessed using statistical

84
multivariate analysis (PCA and PLS) and georeferenced data measured at an approximately cubic metre
resolution (octrees). Results showed that OB is observed in areas of the stope where there is a large blasting
energy proxy, large undercut and towards the middle of the face. UB tends to be observed near the border of
the faces and where there is large blasting standoff and toeing of the rings. The resolution of the data (per octree)
and the use of multivariate analysis have enabled the depiction of the variation of stope performance along the
design surface, going beyond the simple per face analysis.

4.3. Introduction
Designing and mining open stopes are complex processes in underground mines. They necessitate input from
the engineering, geological and planning departments to account for all the critical information used in designing,
executing and achieving a stable open stope that is as close to the desired planned geometry as possible. Mining
which deviates from the planned geometry is categorised as overbreak (OB; rock mined outside of the design)
and underbreak (UB; planned rock left behind). Both can have major operational and economic impacts on the
mine.

The stope design process can be divided into four main steps, with the information gathered from the final step
providing feedback for the previous steps along with future stope design (Potvin et al. 2020). These steps are:

1. Life of Mine stope planning.

2. Stope design.
3. Operation and execution.
4. Reconciliation.

Mine sites will look to maximise stope performance by using tools which typically include rock mass classification
charts, numerical modelling and the Stability Chart (Mathews et al. 1981). These tools are used during the first
two steps of the design process to facilitate decision-making. Rock mass classification charts developed by
Barton et al. 1974 (Q-system) and Bieniawski 1973 (RMR system) were among the first tools developed for
quantifying the properties of the rock mass. Despite these tools originally being developed for civil tunnelling
applications, since their inception, they have been widely used in mining and, more specifically, stope design.
Numerical modelling is used for estimating the stress redistribution around the stope. It considers geological and
geomechanical properties, as well as the stope geometry and mining sequence (Potvin 1988).

The Stability Chart introduced by Mathews et al. 1981 is an empirical and bivariate tool that is used on most
mine sites to predict stope stability on a per face basis (Potvin et al. 2020). It considers a stability number which
is calculated from the geomechanical properties of the rock mass and the hydraulic radius which is derived from
the planned geometry. Decisions can be made regarding stope dimensions or the need for ground support based
on the results. Since its creation, many authors have revisited the Stability Chart to improve its predictive

85
capabilities by adding more stopes to the database or revisiting the stability number. Widely used versions of
the Stability Chart are Potvin 1988, Nickson 1992 and Clark and Pakalnis 1997. Site-specific versions are also
possible by using their own reconciliation data. Despite the evolution of this method, the use of the Stability Chart
is best suited for use at the feasibility stage (or the Life of Mine stope planning step) when the dimensions of the
stopes need to be estimated. The use of the Stability Chart as an optimisation or predictive tool once mining has
started remains limited by its fundamental approach as further research has shown the impact of other variables
that influence stope stability and can provide the largest opportunity to optimise mining performance. These
variables include faults, undercut, stand-up time and blasting (Clark and Pakalnis 1997; Wang 2004; Potvin et
al. 2016; Guido et al. 2017; McFadyen 2020). Perhaps more importantly, it does not consider UB which,
depending on the mine, can have a greater economic impact than OB.

Workshops conducted in Australia and Canada in 2019 showed that mine sites still use the Stability Chart as
the main design tool, even 30 years after its creation (Potvin et al. 2020). The advancement of technology has
increased the amount of data being collected at mine sites as well as the computational potential of computers.
This technological progression has increased our analytical capabilities to understand the root cause of stope
OB and UB. However, the analysis undertaken on site hasn’t matched this progression (Potvin et al. 2020).
Evidently, there is a need to develop an approach to stope design which is fundamentally different and enables
incorporation of the additional variables that are critical to the stope OB and UB. This method should capture
the influence of the variables which underpin the prediction and optimisation of stope performance (OB and UB)
at the stope design and operation and execution stages. Increased computational capabilities have enabled us
to run more powerful statistical analyses, such as multivariate analysis (Guido and Grenon 2018; McFadyen
2020), and create reconciliation tools at a significantly finer resolution (per octree) that allow us to extract spatially
correlated information that is insightful to stope performance (Woodward et al. 2019).

This paper will present the first steps towards a new stope design methodology that combines the recent
technological advancement in stope reconciliation (data measured at an approximately cubic metre resolution,
referred to as octrees in this paper) with statistical multivariate analysis to obtain an in-depth understanding of
any mine site stope OB and UB that could not be reached before due mainly to limitations in computational
power. This knowledge can then be used at the design stage for the optimisation of stope performance and the
planning of mining. The novelty of this approach lies in the wide range of variables that can be considered which
are known to impact stope OB and UB, as well as the resolution of the analysis (per octree). The use of an
octree data structure allows georeferenced points along the design surface to be efficiently obtained and
provides much better representation of the spatial variation of the influencing factors and stope OB and UB
compared to a per face analysis. This allows for a significant leap in our understanding of what degree different
factors control stope behaviour. The reconciliation and statistical analysis approach are detailed in the following
sections and the results of a case study are presented and discussed.

86
 4.4. Literature review
To understand the root causes of stope OB and UB at a mine site, an empirical approach is usually taken. The
performance of previously mined stopes is analysed and compared with selected variables to get a better
understanding of how OB and UB are being generated. To do so, a database needs to be built for each mine
site using data available from previously mined stopes. This consists of quantifying stope OB and UB which
requires the reconciliation of a designed stope shape and the survey of a mined void. Additionally, various
geological and mining variables are collected, quantified and referenced with respect to the octree. This provides
a data structure which facilitates statistical assessment of the stope OB and UB. In summary, this approach
consists of three essential steps:

1. Quantifying stope OB and UB.

2. Quantifying various variables for analysis.
3. Identifying the critical variables that influence stope OB and UB.

The following sub-sections will detail previous work, knowledge and approaches for these three essential steps.

4.4.1. Quantifying stope OB and UB

Stope OB and UB data are obtained through the reconciliation of the stope’s planned geometry and the mined
stope which is captured through a cavity monitoring system (CMS) scan (Miller et al. 1992). Current industry
practice quantifies stope OB and UB by looking at volumes of OB and UB (Potvin et al. 2020) (Figure 4. 1).

Figure 4. 1: Stope overbreak and underbreak

87
The common levels of resolution for stope reconciliation are on a per stope or per face basis (Potvin et al. 2020).
At these resolutions, OB and UB are quantified as volumes and percentages. For faces only, an equivalent linear
overbreak slough (ELOS) or equivalent linear lost ore (ELLO) is assessed (Clark 1998).

A new advancement in stope reconciliation assesses OB and UB at a finer level of resolution by quantifying
stope OB and UB on a per octree basis (Woodward et al. 2019). An octree is obtained by recursively subdividing
the three-dimensional space into blocks (octrees) until the desired size of the octrees along the design surface
(usually <1 m) is obtained (Figure 4. 2). Variables are then quantified for each georeferenced octree along the
design surface which captures the variation of OB and UB and influencing factors.

Figure 4. 2: Illustration of the recursive process of octrees as well as the octrees defined along the design surface (modified from
McFadyen et al. 2021)

At an octree resolution, the OB and UB is quantified by calculating the distance in the direction normal to the
design surface and a surface generated by a CMS on a point basis (Figure 4. 3). This measure of OB and UB
is referred to as the projected distance. Positive values represent OB while negative values represent UB.

88
 Figure 4. 3: Illustration of octree stope OB and UB quantified by calculating the distance in a direction normal to the design surface
between the design surface and the CMS for each octree (modified from McFadyen et al. 2021)

Figure 4. 4 provides an example of how stope OB and UB can vary across a stope face and how the octree
resolution captures it. In comparison, a per face analysis would have a given single OB and UB measurement
for the whole face with average values for the variables to analyse and would not have considered the location
of the OB and UB across the face. The implementation of an octree data structure provides hundreds to
thousands of data points, addressing the loss of information caused by averaging values over a stope face and
enables the spatial variation of values to be captured along the design surface. For example, this enables the
spatial variation of blasting energy to be quantified along the design surface and to assess if the OB (left of the
face) and UB (right of the face) are influenced by this factor. This type of analysis was not possible before on a
per stope or per face basis and it represents a critical technological step towards improving our understanding
and the optimisation of stope performance. Reconciliation can thus be done at different levels of resolution (per
stope, per face and per octree). With the octree resolution, different variables can be quantified, increasing the
level of analysis that can be done and the knowledge that can be obtained.

89
 Figure 4. 4: An example of how stope OB and UB can vary across a stope face (McFadyen et al. 2021)

4.4.2. Quantifying critical variables for the analysis

The variables quantified for the analysis are selected based on general knowledge, understanding and principles
of rock mechanics and stope stability. The goal of open stoping is to remove a volume of rock which is
economically viable to mine, creating an excavation where the walls are generally unsupported rock mass or
backfill. This ultimately leads to ground failure over time, where the time frame of stability is dependent on the
complex geological and mining environments. Therefore, there are four main categories of variables to consider.
Firstly, there are geometrical variables. These include the shape, size and dip of the designed geometry which
influence the local stress and confinement in the rock mass along the faces, as well as the vulnerability to gravity
instability (Hudson and Harrison 1997). Stope geometry is a critical variable in the Stability Chart. This method
considers a stope face’s hydraulic radius which is a dimensionless variable, and the face dip which establishes
the face stability (Mathews et al. 1981).

The stability of a specific stope geometry varies depending on the rock mass quality, which leads to the second
category: the geomechanical variables. In the Stability Chart, the second critical variable is summarised by the
stability number, which is defined using the geomechanical properties of the rock mass (discontinuities and
stress) with respect to stope geometry orientation. As mentioned by Potvin 1988, “With regards to stability, two
of the most important characteristics of the rock mass are the size and shape of the blocks forming the rock
mass matrix.” Rock mass characterisation tools such as the rock quality designation (RQD) (Deere and Deere
1988) and the Q-system (Barton et al. 1974) characterise the discontinuities and can spatially be associated
with georeferenced octree data for investigation.

90
The third category is the geological variables. In addition to small-scale discontinuities in the rock mass, major
geological structures such as faults also influence the stope stability. The interaction between the fault and the
stope face will determine how the stope stability will be affected. For example, stability will be different if the fault
is daylighting into the stope or if it is parallel to the stope (Suorineni 1998). Suorineni et al. 2001a created a
version of the Stability Chart that incorporates a fault variable. Correlations between faults and stope stability
have also been made by Woodward et al. 2019.

The last category is the operational variables. The mining process chosen influences the stability of the stope.
This can be separated into different subcategories:

• The mining sequence chosen, meaning how the induced stress will be managed through stope
sequencing (Villaescusa 2014; Guido et al. 2017; McFadyen 2020).
• The drifts undercutting the walls have an impact on wall confinement and therefore on stope stability
(Wang 2004).
• The ground support used for stabilising faces (Nickson 1992).
• The mucking time influence how long the stope is standing unsupported (Wang 2004; Guido et al.
2017; McFadyen 2020).
• The design of blasting pattern and sequence. Research by Wang 2004, Potvin et al. 2016, Guido et
al. 2017, McFadyen 2020 and Woodward et al. 2019 has shown that the drilling pattern, the angle at
which the holes hit the face and the energy influences stope OB and UB. It is also one of the most
important aspects influencing UB.
These various variables can be quantified at one or more resolutions (per stope, per face and/or per octree) and
compared with stope OB and UB. On a per stope and per face basis, broad variables or variables that are
averaged per face can be analysed. The novelty with doing analysis on a per octree basis is the stope OB and
UB measured as a projected distance can be compared to a local characterisation of stope design factors. This
analysis is only limited by the amount and the quality of data collected on site, meaning the wide variety of
variables that are critical to stope OB and UB can be considered. The flexibility of this approach allows for the
characterisation of data from various sources that are available and relevant to a specific site. The various octree
variables are assigned to an octree block on the stope design surface by spatially referencing the data which is
queried for the desired variables.

4.4.3. Identification of critical variables through statistical analysis

Once stope reconciliation and variable quantification are complete, this information can be used to develop
knowledge on the critical factors that influence stope OB and UB. To do so, statistical analysis can be used to
characterise the relationship between variables. These analyses should be done sequentially for the three
different levels of resolution (per stope, then per face and finally per octree) to extract the most information.

91
Analysis can be simple, such as univariate and bivariate analysis. Univariate analysis can be used to describe
the distribution of OB and UB through cumulative charts. This analysis is done on a per stope and per face basis
(Potvin et al. 2016; McFadyen et al. 2019a; Woodward et al. 2019). Additionally, broader variables such as
comparing longitudinal and transversal stopes can be compared to stope OB and UB using cumulative charts.
Bivariate analysis can be used to evaluate the impact of input variables on OB and UB using an X–Y chart to
plot the two variables being analysed. Statistical methods such as regressions, whiskers and box plot, or bar
charts are then overlaid to statistically interpret the information. Bivariate analysis can be done on a per stope,
per face and per octree basis (Woodward et al. 2019; McFadyen et al. 2020).

While univariate and bivariate analyses provide a great deal of information on stope OB and UB, interpretation
is inherently limited as only one or two variables are analysed at a time. Stope performance is the result of a
complex interaction of different variables including geometry, geomechanics, operation and geology. Multivariate
analysis offers a number of methodologies which can provide further insight into the relationships the various
variables have and the impact that they have, or do not have, on stope OB and UB. Two different methods of
analysis are presented in the following sub-sections and are used in this paper.

Principal component analysis

The first method is the principal component analysis (PCA), first introduced by Pearson, 1901. The goal of PCA
is to calculate principal components, or a new set of latent parameters that are the result of a linear combination
of variables which maximises the variance explained in the data. Integrating PCA in the root cause analysis
means it is possible to quantify and analyse how the variance of each variable is correlated. Figure 4. 5 illustrates
the variables of PCA, where the matrix X represents the input data with k variables, or k octree input variables,
and N observations, or in this case N octrees. Matrix P represents the weights for each k variable and a
component. Matrix T represents the principal component matrix with N observation and A components. Each
component explains part of the variance in the data, with the first component explaining the most variance and
so forth. The explained variance of each component sums to the total variance in the data, so if all the
components are used in the analysis, then A equals K.

92
 Figure 4. 5: Illustration of the principal component analysis (PCA) structure (Dunn 2023a)

Equation 4.1 shows the linear combination of variables used to calculate a component.

𝑇𝑎 = ∑𝑘=𝐾
𝑘=1 𝑃𝑎,𝑘 ∗ 𝑋𝑘 [4.1]
Where:

Ta = Score column for component a in matrix T

Pa, k = Weight for variable k and component a
Xk = Octree observation values for variable k

Each variable contributes to the variance being explained by the component to a certain degree. Work by
McFadyen 2020 has shown that the PCA components can be analysed to identify correlations and
independencies between the calculated stope face variables. PCA analysis used for root cause analysis
revealed valuable information. Output variables, which in this case is the stope OB and UB, can be included in
the analysis as marker styles when assessing results. This allows the identification of distribution patterns and
provides additional information, but it does not contribute to the calculation of the components.

Further analysis is done by using PCA results, for example, as a quality control step to identify observations with
abnormally high or low values for variables that contribute to the component being analysed. However, this will
not be discussed in this paper. Instead, it will extend previous work by applying the PCA on a per octree basis,
allowing further understanding of the root causes of stope OB and UB.

Partial least squares method

The second method is the partial least squares (PLS), first introduced by Wold in 1966. PLS is similar to PCA
with the difference being that the weights (W’ and C’) and components (T and U) are calculated for the input (X)
and output (Y) variables (Figure 4. 6). Additionally, for PLS, the output variable, or in this case the projected
distance, is included in the input component’s calculation.

93
 Figure 4. 6: Illustration of the partial least squares (PLS) structure (Dunn 2023b).

For PLS, the goal is to find a linear combination that maximises the covariance explained between the input
variables and the output variables. The covariance is a product of the correlation between the input and output
variables and the variance of each one (Equation 4.2).
1
𝐶𝑜𝑣 (𝑡𝑎 , 𝑢𝑎 ) = 𝐶𝑜𝑟𝑟(𝑡𝑎 , 𝑢𝑎 ) ∗ √𝑡 ′ 𝑎 𝑡𝑎 ∗ √𝑢′ 𝑎 𝑢𝑎 ∗ [4.2]
𝑁

Where:

a = Chosen component

Corr = Correlation between T and U

𝑡 ′ 𝑎 𝑡𝑎 = Variance of T

𝑢′ 𝑎 𝑢𝑎 = Variance of U

N = Number of observations

Work by McFadyen (2020) has shown that analysing the PLS components can be indicative of the importance
of each variable with regards to stope performance and how they relate to OB and UB on a per face basis. PLS
can also be used as a predictive model (McFadyen 2020). As such, the coefficients can be calculated for the
linear regression used for the predictive model (Equation 4.3) and can generate further insight into the impact of
each variable on stope OB and UB.

𝑌 = ∑ 𝑋𝑖 ∗ 𝐵𝑖 [4.3]

Where:

Y = Predicted value

Xi = Value of the input parameter

Bi = Parameter coefficient

94
This type of analysis was shown to be valuable for providing quantitative information for stope performance
assessment. This paper will extend previous work by applying the PLS on a per octree basis, allowing further
understanding of the root causes of stope OB and UB.

4.5. Methodology
This new stope design approach is an empirical method that uses multivariate statistical analysis and data at an
octree resolution to understand the mine site specific stope OB and UB. This methodology enables the
identification of critical variables and uses this information to understand the stope OB and UB which is critical
for the optimisation of stope design and the operational planning of future mining. The approach can be
described in three steps (Figure 4. 7).

Step 1: Data collection Step 2: Variable quantification Step 3: Root cause analysis

• Assembling previously mined • Stope reconciliation and • Use of univariate, bivariate

stope data octree generation and multivariate statistical
• Data quality verification • Input variables quantification analysis
for the different level of • Identification of critical
resolution variables on stope OB and UB

Figure 4. 7: The three steps for understanding stope OB and UB as part of the new stope design methodology.

4.5.1. Data collection

The first step in this method is to collect the necessary data to be able to quantify the variables for octree
analysis. Data must be obtained and validated to ensure good quality data will be processed at the next step.
Four main categories of data are collected with each data, detailed in Table 4. 1.

95
 Table 4. 1: Collected data

Categories Description Sought after data

• Cavity monitoring system scan
Georeferenced 3D shapes (CMS, Miller et al. 1992)
Geometrical
of the mining geometries • Design/planned shape
• Drift surveys
Georeferenced rock mass • Geotechnical model
Geomechanical properties surrounding the • Borehole rock quality
stope designation (RQD) data

Georeferenced structures • Structural model

Geological such as faults, lithological
contacts and other • Lithological contacts

• Drill and blast data

Planning and mining o Blast dates
properties of the stopes, o Drill rings
Operational such as drill and blast,
o Charging
ground support, sequence
and other o Delays
• Ground support strings
One of the most critical steps is the collection of data and subsequent validation. Preparing the data for analysis
can be time consuming in a project and can take up to 60% of the time (Pyle 1999); however, it is a necessary
step to ensure the quality of the analysis and the reliability of the results. This step is accomplished using a
combination of software tools and mine site knowledge.

Each category of data has its own quality issues. For the geometrical data, the quality of the mesh needs to be
verified as it affects the volumes computed in stope reconciliation. Bad meshing can occur when shapes are
merged, or in the case of the CMS, when the scans are of poor quality. Figure 4. 8(a) shows an example of bad
meshes. Some mine sites do validate the CMS mesh on site and repair surveying artefacts in the CMS but
based on feedback from different mine sites, this is not done everywhere. For the CMS, odd shapes should also
be identified. For example, some stopes can be closed before mucking is finished due to excessive OB. In these
cases, the CMS will not be representative of the completely empty stope and information from this stope should
be treated with caution. Figure 4. 8(b) shows an example of a CMS where the stope still has a muck pile.

96
 a) b)
Figure 4. 8: Examples of a) Bad mesh; b) Odd CMS shape caused by a muck pile left in the stope.

When considering the geomechanical data incorporated in the mine block model, knowing how it was built and
with what data helps to determine at what resolution the model can be integrated to ensure octree points have
representative values. The source of the data and its density correspond to the reliability of the block model and
should be considered to determine areas of low confidence. Similar considerations should be made for the
structural model.

The structural model represents a special case of data as it is not necessarily present for each stope. There is
not always a fault in the vicinity of the stope and results in absent data for some or all octrees. In addition, some
of these faults are not a cause for concern for the stope OB. Since the presence of geological structures is
treated in these analyses as a source of instability affecting stope OB, only the faults that influence the stope
OB should be used. The selection of influential structures can generally be done using mine site knowledge.

An issue concerning operational data is overlapping drill strings designs. This can often occur when there is
more than one firing for the stope, however, final designs containing the complete drill strings have been
generated for each firing. This will affect the results of some of the variables and should therefore be identified
and filtered out. Once the data have been validated, stope reconciliation can be done to obtain the OB and UB
data and quantify the various variables.

4.5.2. Variable quantification

For the second step in the method, the variables that will be quantified for the analysis are selected based on
general knowledge, understanding and principles of rock mechanics and stope stability which were discussed
in Section 4.4.2. The variables are quantified through the mXrap software (Harris and Wesseloo 2015) using the
stope reconciliation and analysis apps (Woodward and Harris 2019).

97
Geometrical, structural, and drilling and blasting variables are captured for each mine site stope. Woodward et
al. 2019 previously quantified a total of 10 variables on a per octree basis. These variables have been tested
and validated through bivariate analysis as variables that can be linked to stope OB and/or UB depending on
the mine site. In addition, three structural variables have been added as well as a spatial variable. Table 4. 2
provides the concept, interpretation and an example of the variables that will be a part of the multivariate
analysis. A stope is used for each example, displaying the octrees variables value according to the legend
attached. Additional text is added to help characterise and understand the variables. These variables require
the CMS, stope design geometry, drift surveys, drill rings and structural model to be available for each stope.

Table 4. 2: Local characterisation of stope design variables for per octree analysis (modified from Woodward et al. 2019)

Variable Concept Interpretation Example

Blasthole Distance (coloured by Blast-induced
standoff metre) from an octree damage.
distance block to the nearest
designed blasthole. Characterise how
OB and UB is
affected by the
standoff distance
to the blastholes.
Distribution of the distance for each
octree along the design surface (dots in
the image) to their nearest blasthole.
Cold colours mean the octrees are
further away from the blastholes and
generally, UB is expected.
Blasthole Shortest angle Blast-induced
orientation (coloured by degree) damage.
between an octree’s
design normal vector Characterise how
and the blasthole. OB and UB is
Convention: 90⁰ = affected by the
toeing, 0⁰ = parallel angle of the
blastholes to the
design surface.
Distribution of the angle for each octree
along the design surface (dots in the
image) to their nearest blasthole. Warm
colours (90°) mean the blastholes are
toeing into the design surface.
Proxy for 𝐵𝑙𝑎𝑠𝑡 𝑒𝑛𝑒𝑟𝑔𝑦 Blast-induced
𝑛
blast energy L damage
≈∑
max(0.5, 𝐷)2
0 Characterise how
where: OB and UB is
L = Length of hole affected by the
D = Distance to hole density of

98
 Max. distance of 0.5 m blastholes and the Distribution of the energy proxy for
to account for drill pattern. each octree along the design surface
stemming near collars (dots in the image). The energy
(coloured by energy calculation considers the density of
index). blastholes around the octree (inverse
distance relationship) and the length of
the blastholes to determine the spatial
distribution of the blasting energy
according to the blasthole design.
Warm colours mean more energy is
expected in that area due to the drill
design.
Dip of the The dip calculated from Influence of
design the normal vector gravity.
surface associated with the
octree block (coloured Characterise how
by degree). OB and UB is
affected by the
effect of gravity.
Distribution of the dip for each octree
along the design surface (dots in the
image). Warm colours (90°) mean the
design face attached to the octree is
vertical.
Direction of The direction calculated Influence of
the design from the normal vector orientation with
surface associated with the respect to stress
octree block (coloured and rock mass
by degree). properties.

Characterise how
OB and UB is Distribution of the direction for each
affected by face octree along the design surface (dots in
orientation. the image). The direction varies from 0°
to 360°.
Effective Calculated by Local stability
radius factor measuring, at regular conditions—
(ERF) radial angles intervals, deformation of
the distance from a excavated span.
point located on the
selected surface of the Characterise how
stope to the edge of OB and UB is
that surface (Milne, affected by the
1997). A maximum size and geometry
value is obtained of the faces.
Distribution of the ERF for each octree
towards the centre of
along the design surface (dots in the
image). Warm colours mean the
octrees are further away from the
surface edges and have higher stability
concerns.

99
 the surface.

Distance to Only the stope design Local stress

convex hull is ‘wrapped’ in a conditions.
(stope convex hull.
geometry Distance (coloured by Loss of
complexity) metre) is found from an confinement.
octree block to the
nearest point on the Characterise how
convex hull surface. OB and UB is
affected by the
complexity of the Distribution of the distance to the
design geometry. convex hull for each octree along the
design surface (black lines). Warm
colours mean the design face is further
away from the simplified geometry and
can indicate areas of loss of
confinement due to the design
geometry.
Distance to Distance (coloured by Kinematics—
fault metre) from the octree beam and wedge
block to nearest formation.
interpreted fault
surface. Characterise how
OB and UB is
affected by the
presence of faults
near the stope. Distribution of the distance to the
nearest fault for each octree along the
design surface (dots in the image).
Cold colours mean the octrees are
further away from the fault and less
likely to be affected by the faults.
Angle to fault The average of the Kinematics—
inverse distance³ beam and wedge
weighted angle formation.
(coloured by degree)
between an octree’s Characterise how
design normal vector OB and UB is
and the plane of the affected by the
fault. angle of the fault
Convention: 90⁰ = to the design Distribution of the angle to the nearest
toeing, 0⁰= parallel. surface. fault for each octree along the design
surface (dots in the image). Warm
colours (0°) mean the octrees are
parallel to their fault.

100
Undercut The stope design as Local stress
well as the drifts is conditions.
‘wrapped’ in a convex
hull. Loss of
Distance (coloured by confinement.
metre) is found from an
octree block to the Characterise how
nearest point on the OB and UB is
convex hull surface. affected by the
complexity of the
geometry and the
cutting of the
stope by the
drives.

In comparison Distribution of the undercut for each

with the stope octree along the design surface (dots in
geometry the image). Warm colours indicate that
complexity due to the presence of the drives
variable, the and/or the shape of the stope, the
undercut octrees are further away from the
considers the full simplified shape that wraps the stope
geometry of the and drives, and highlights areas where
void and the there can be a loss of confinement due
impact the drives to how the drives cut the stope.
have.
Directional Directional measure to Deconfinement
measures the drift or faults in and gravity effect
to structures regards to its position on design surface
and distance. areas near the
Convention: 180° = drifts.
octrees situated over
their nearest drift and Characterise how
close to the drift, OB and UB is
0° = octrees situated affected by the
under their nearest effect of gravity
structure and close to caused by the
the drift. presence of
structures such as Distribution of the directional measure
drives and faults. to the drives for each octree along the
design surface (dots in the image).
Cold colours (0°) mean the octrees are
under and close to the nearest drive
and warm colours (180°) indicate the
octrees are over and close to the
nearest drive. 90° indicates the octrees
are further away from the drives.

101
 Dissimilarity The average difference Influence of the
between the adjacent spatial fluctuation
octrees and the of the critical
selected octree using variables
normalised values of On OB and UB.
the critical variables.
Characterise how
OB and UB is
affected by the
spatial fluctuation,
over a short
distance, of all the
variables
(selected from the
Distribution of the dissimilarity for each
ones presented in
octree with its adjacent octrees along
this table) used in
the design surface (dots in the image).
the root cause
Cold colours mean the octrees have
analysis.
similar values to the adjacent octree for
the selected variables and warm
colours indicate that one or more
variables for one or more of the
adjacent octrees differ from the
selected octree.

4.5.3. Root cause analysis

The third step in the method uses statistical analysis to identify the critical variables influencing stope OB and
UB. The objective of the statistical analysis is to identify the general trends and relationships between the
variables for a selected group of stopes to identify variables critical to stope OB and UB. Since this method looks
at developing knowledge using the octrees and the variables quantified at this resolution, analysis conducted at
a per stope and per face resolution will not be presented. Despite this, valuable information can be extracted
from these analyses and should not be excluded when conducting root cause analysis. In general, root cause
analysis should start at the larger resolution (per stope) and work towards a finer resolution (per octree), with
each assessment providing valuable information that focuses on the most critical causative factors.

Multivariate statistical analysis considers the 13 variables previously presented and allows for the identification
of complex relations between the variables. The two multivariant methods used are PCA and PLS presented in
Section 4.4.3. Since the variables used have different units, they first need to be normalised to have the same
weight in the analysis. In this case, the distribution of a variable is transformed to have a mean of 0 and a
standard deviation of 1. For the PCA analysis, the variance explained by each component diminishes with each
component, containing less and less information pertinent for the global understanding of the variable’s
relationship. Therefore, components are excluded from the analysis. For the PLS analysis, the covariance

102
between the variables and projected distance is analysed and because there is only one output variable, there
is therefore only one component to analyse.

Bivariate charts can be used to highlight the relationship with the stope OB and UB for critical variables identified
from multivariate methods. These critical variables provide a better understanding of stope OB and UB, and can
be used, stope by stope, to identify where OB and UB will occur in the faces. All the statistical analyses are
made using the software R (R Core Team 2021). For PCA, the package FactoMineR (Le et al. 2008) is used
and for PLS, the package PLS is used (Liland et al. 2020).

4.6. Case study

The data used in this paper are from the underground Dugald River mine in Queensland, Australia (Figure 4. 9).
Longitudinal and transversal stopes are used to extract the Zn-Pb Ag ore. Stope dimensions can vary depending
on the area of the mine but are approximately 25 m high and 15–30 m long (Hassell et al. 2015). Stope width
varies between 2 m and 35 m depending on the thickness of the orebody. The orebody reaches depth of
approximately 1 km.

Figure 4. 9: Location of Dugald River mine in Queensland, Australia.

4.6.1. Data overview

The mine database comprises 173 stopes mined between 2014 and 2021 (Figure 4. 10). Of these 173 stopes,
41 are a part of the north area of the mine and 132 are a part of the south part of the mine. Nineteen stopes
were also part of the 2014 trial stoping stage that was conducted to test the validity of the geotechnical and
mining variables (Hassell et al. 2015). For the analysis, the 2014 trial stoping, as well as stopes with data quality

103
concerns, has been removed resulting in a total of 130 stopes. Stopes from the 2014 sector have previously
been part of a study by McFadyen et al. 2019b which conducted root cause analysis using only univariate and
bivariate analysis at a per stope, per face and per octree level. These are longitudinal stopes. The analysis
revealed that stopes generate larger OB than UB with no specific face associated to the OB. The octree analysis
identified the blasting energy proxy, blasting standoff, effective radius factor (ERF), distance to convex hull and
distance to fault as variables that showed trends towards OB or UB for each or specific faces. The analysis,
however, didn’t consider the directional measure to structures, undercut and the dissimilarity, and doesn’t
consider multivariate analysis.

Figure 4. 10: 3D section view of the Dugald River stope database with the trial area shown in red.

From the 130 stopes, 49 are transversal stopes and 81 longitudinal stopes. Since longitudinal and transversal
stopes can behave differently, they will be treated as two different datasets. Only the transversal stopes will be
discussed in this paper. These stopes are shown in Figure 4. 11.

Figure 4. 11: Longitudinal view of the Dugald River transversal stopes in the database.

Reconciliation for each stope was performed on a per stope, per face and per octree basis. Depending on the
geometry, stopes are typically separated into six faces: the HW, the FW, the floor, the crown, and the two sides

104
(Figure 4. 12). During the survey building process, drifts are cut out of the design and CMS. For this reason,
stope faces that represent drift surfaces (floors and crowns mostly) are not part of the analysis. Additionally,
faces mined against backfill are not part of the analysis.

Figure 4. 12: Example of stope face delimitation using colour code and orientation.

The variables used for analysis were presented in Section 4.5.2, Table 4. 2. A structural model is available for
the quantification of the fault variables for this mine site. Mine site knowledge was used to process the structural
model and filter out the faults that do not cause stability issues in the stopes. When distance to fault and angle
to fault are being calculated, a maximum search distance of 30 m from the octree is used. This distance was set
based on the work by Woodward et al. 2019. For octrees with absent fault data, a distance of 40 m is set and
an angle of 100° to separate them from the octrees with fault data. This is necessary as the computer approach
used for the statistical analysis does not process observations with missing values. Borehole RQD data are
available, but density does not allow for it to be used at an octree resolution. Blasting data consists of the design
for blasthole which are to be fired for each stope.

4.7. Results
The 49 transversal stopes were reconciled at an octree resolution varying between 0.4 m and 0.8 m, with most
stopes at a 0.5 m resolution. This translated into almost 400 000 octree points along these stope design surfaces
that were analysed. The projected distances for these octrees varied between -9 m and 15 m (Figure 4. 13) with
60% of octrees having a projected distance between -1 m and 1 m and 55% of octrees being UB and 45% OB.
This means that for around 60% of the design surfaces being analysed, the final surface captured with the CMS
was within a 1 m distance of the planned surface.

105
 Figure 4. 13: Normalised cumulative count of octrees projected distance for the Dugald River 49 transversal stopes.

The large UB distances (-4 m and under) are often situated in the corners of the stope. Due to the proximity of
two faces with perpendicular directions, large UB projections intersect the CMS further into the opposite face
(Figure 4. 14). These extreme projections aren’t characteristic of the factors which are in the locality of the octree
and are removed. The removal of these points doesn’t significantly influence the degree to which the stope UB
is captured as this UB is already being characterised by octree points on the perpendicular face.

Figure 4. 14: Example of an octree UB quality issue identified for large UB in the corners of the stope.

The extreme OB projections (over 6 m) are typically the result of an unravelling of the rock mass as discussed
with site personnel, which is not captured by the current selection of variables. These occurrences are
uncommon in this database and as analysis is focused on investigating typical stope OB and UB, these octrees
are also excluded from the analysis. Therefore, to ensure the representability of the analysis, it was deemed
adequate that all projected distances smaller than -4 m and larger than 6 m are filtered out. These extreme
values represent a small percentage of the dataset (less than 5%).

106
Figure 4. 15 provides an example of stope OB and UB at an octree resolution for two stopes. In the following
sub-sections, a root cause analysis of these projected distances of OB and UB will be presented, providing the
knowledge to understand why these magnitudes of OB and UB were observed in these specific locations.

Figure 4. 15: Two stope examples of the Dugald River mine site stope OB and UB at an octree level, with CMS and designed shape
overlaid. View 1 side wall does not show any octrees as they are mined against backfill.

4.7.1. PCA analysis

The PCA analysis enables to understand the relationship of the variance between the input variables. To
determine to components used in the analysis, a decision is made using a Scree Plot (Figure 4. 16). From this
chart, the first six components will be used in the analysis. This arbitrary decision is based on the fact that the
variance of the variables is 1 since they have been normalised. The first six components have a higher variance
than 1 meaning they explain a larger variance than the original variables. These six components explain a total
of 73% of the variance in the data (Table 4. 3). In the case that all the variables would vary independently, the
percentage of variance for each component would be of 8%. Since none of these components predominantly
explains the variance in the dataset, the variance of the variables must be, in the most part, independent.

107
 Figure 4. 16: Scree Plot of the variance explained for each component of the Dugald River PCA.

Table 4. 3: Variance explained per Dugald River PCA component.

Components 1 2 3 4 5 6 7 8 9 10 11 12 13
Variance 2.1 2.0 1.8 1.5 1.1 1.0 0.9 0.7 0.6 0.5 0.5 0.4 0.0
Percentage of
16 15 14 11 9 8 7 6 5 4 4 3 0
variance
Cumulative
16 31 45 56 65 73 79 85 89 93 97 100 100
percentage

The contribution of each variable to each component is presented in Table 4. 4 with the sum of each column
being 100%. This table can be used to identify variables that can be independent or correlated by looking at how
much they contribute to each component. Variables identified for the same component can be interpreted as
variables that vary at the same time. The correlation between the variables can be positive or negative. This is
determined using the coordinate sign in parentheses. If the signs are the same, the correlation is positive (both
variables increase or decrease) but if not, it is negative (one variable increases while the other decreases).
Independent variables will contribute to different variables, meaning their variance is not correlated. Analysis of
the variance in the data allows a number of observations to be made. These observations are presented in Table
4. 5.

108
Table 4. 4: Contribution, in percentage, of each variable to the variance explained by each component for Dugald River. The strongest
contributing variables for each component are identified in red. The coordinate sign of each variable for each component are in
parentheses.

Variables Comp. 1 Comp. 2 Comp. 3 Comp. 4 Comp. 5 Comp. 6

Blasting energy
11.84 (-) 14.52 (+) 4.52 (+) 4.98 (+) 0.02 (-) 4.05 (-)
proxy
Blasting
8.33 (+) 2.07 (-) 5.45 (+) 16.99 (+) 2.00 (+) 2.94 (-)
orientation
Blasting standoff 6.56 (+) 8.85 (-) 1.37 (-) 0.02 (-) 16.69 (+) 0.11 (+)
Dip 13.57 (-) 1.91 (+) 8.58 (-) 0.44 (+) 6.49 (-) 6.39 (+)
Direction 2.59 (+) 3.18 (-) 0.24 (-) 23.58 (+) 14.01 (+) 0.66 (-)
ERF 9.29 (-) 5.85 (+) 0.05 (+) 0.03 (-) 35.26 (+) 0.12 (+)
Stope geometry
0.00 (-) 0.00 (+) 35.54 (+) 1.71 (+) 0.12 (+) 4.28 (+)
complexity
Undercut 0.11 (-) 0.24 (+) 35.01 (+) 3.89 (+) 0.21 (+) 1.83 (+)
Angle to fault 20.62 (+) 27.20 (+) 0.34 (-) 0.11 (+) 0.00 (+) 1.45 (+)
Distance to fault 0.03 (-) 0.84 (+) 5.14 (+) 37.45 (-) 0.66 (-) 2.34 (+)
Fault directional
20.49 (+) 26.51 (+) 0.87 (-) 0.84 (+) 0.00 (+) 1.05 (+)
measure
Drift directional
0.36 (+) 4.49 (-) 0.12 (-) 1.13 (+) 0.10 (-) 70.07 (+)
measure
Dissimilarity 6.19 (+) 4.33 (-) 2.76 (+) 8.83 (+) 24.43 (-) 4.71 (-)

109
 Table 4. 5: Observations made on the relationships between the variables for Dugald River.

Components Observations
• Mainly describe the variance of the fault’s angle to octree as well as the directional
measure for the fault.
• These two variables are highly positively correlated, meaning when the fault is
perpendicular to the face, it tends to daylight in the face implying a variance in the
directional measure.
• The variance in the dip and the blasting energy proxy is also correlated to the faults.
• The energy proxy for a portion of octrees is positively correlated to the fault
variables (component 2) while another portion is negatively correlated
1 and 2
(component 1). This is an influence of the drilling pattern on the energy proxy in the
faces as well as the position of the intersect of the fault in the stope, meaning the
two components are describing different faces or different faults.
• For component 1, the faults tend to toe into faces with a smaller dip, like a HW and
FW. To a smaller degree, the energy proxy tends to be higher as the ERF gets
larger, meaning there is usually more energy in the middle of the face. The energy
proxy is also negatively correlated to the blasting orientation and standoff, meaning
there is more energy when the rings are close to the face and parallel.
• Describes the stope geometry complexity and undercut.
• These two variables are highly positively correlated as they are measured in a
3 similar way with the only difference being that the undercut uses the stope design
as well as the drifts shape for building the convex hull.
• These variables vary independently of the others.
• Describes how faults tend to be further to the design face when there is toeing of
the drill holes (positively correlated).
4 • There is also a negative correlation between the octree direction and the distance
to fault. This is caused by the fact that the faults tend to be in front of the HW of the
stope, therefore the distance generally increases when looking at the other faces.
• Described by the blasting standoff, the octree direction, ERF and the octree
dissimilarity.
• The first three are positively correlated, meaning the standoff increases towards
5
the middle for some of the stopes faces. This is caused by the drill design.
• The dissimilarity number decreases meaning these octrees tend to be more similar
(small fluctuation in the drill design).
• Describes the directional measure for the drift, with its variance being independent
6
of the other variables.

To summarise, the PCA analysis has helped to identify that stope geometry and undercut are independent of
the blasting and faults. Correlations have been shown between the blasting and the faults, but they do not
necessarily represent dependencies, meaning the drill rings do not necessarily change according to the
presence faults.

110
While the projected distance wasn’t used in PCA analysis, it was used as a marker in the score charts to look at
how the stope OB and UB varies across the component’s axis. This is usually analysed two components at a
time. Since there are six components, there are 15 charts to look at. It was observed that components 1 through
4 show some separation between OB octrees and UB octrees, for example Figure 4. 17 shows two charts
comparing component 2 to component 3. Figure 4. 17(a) shows the loading chart and Figure 4. 17(b) shows the
score chart. These two charts combined show that OB is generally observed when components 2 and 3 increase,
meaning the larger the blasting energy proxy is, the larger the undercut is. When standoff increases, UB is
generally observed.

Figure 4. 17: Component 2 versus component 3 charts: a) Loading chart; b) Score chart coloured by stope OB or UB.

111
This observational analysis can be conducted for each component and is summarised in Table 4. 6. The charts
analysed are located in Appendix A.
Table 4. 6: Stope OB and UB summary from the Dugald River PCA components.

Component Stope OB and UB analysis Variables analysis

combination

1–2 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy and ERF increases, as
well as the fault angle and the directional
measure to fault.

1–3 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy and the undercut
increases.

1–4 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy increases and more UB
than OB when the distance to fault and the
blasting orientation increases.

1–5 Gradual separation of OB and UB along More OB than UB is observed when the
the component 1 axis. blasting energy proxy and ERF increases.

1–6 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy, the ERF and the
directional measure to drift increases.

2–3 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy and the undercut
increases.

2–4 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy, the fault angle and the
directional measure to fault increase and UB
when the distance to fault and the blasting
orientation increase.

2–5 Gradual separation of OB and UB along More OB than UB is observed when the
the component 1 axis. blasting energy proxy, the fault angle and the
directional measure to fault increase.

2–6 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy the fault angle, the
directional measure to fault and the directional
measure to drift increase.

3–4 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. blasting energy proxy, the fault angle, and the
directional measure to fault increase and the

112
 distance to fault and the blasting orientation
decrease.

3–5 Gradual separation of OB and UB along More OB than UB is observed when the
the component 1 axis. undercut increases.

3–6 Gradual separation of OB and UB in a More OB than UB is observed when the

diagonal way. undercut and the directional measure to drift
increase.

4–5 Gradual separation of OB and UB in a Separation of OB and UB function of the design

diagonal way. face ID.

4–6 Small distinction of OB and UB in a More OB than UB is observed UB when the

diagonal way. fault distance increases.

5–6 No distinction between OB and UB.

4.7.2. PLS analysis

Now that we understand the relationships, we can use a PLS analysis to quantify the covariance between the
variables and the stope OB and UB measure (projected distance) and determine the importance each variable
have on the generated OB and UB. The analysis is concentrated on the loadings bar chart (Figure 4. 18) and
coefficient table (Table 4. 7). Each variable has a weight, or loading, in the calculation of the projected distance
and is displayed as a bar chart where the longer the bar, the stronger the correlation between the variance of
the variable and the projected distance. In Figure 4. 18, the projected distance is the last variable. This figure
shows that the blasting energy proxy, the ERF, the stope geometry complexity (convex hull) and undercut are
strongly positively correlated to the octree’s projected distance while the blasting standoff is strongly negatively
correlated. The blasting orientation, the angle to fault and the fault directional measure show a moderate
correlation. The rest of the variable variance shows little correlation with the projected distance. In Table 7, the
same information from Figure 18 can be assessed by comparing the numbers. In addition, the strongly correlated
variables have higher coefficients which vary between 0.15 and 0.26. A coefficient of 1 would indicate that a
change of 1 unit for a variable would change the projected distance by 1 m (1 for 1 impact). The strongly
correlated variables rather have a smaller incremental impact on the fluctuation of stope OB and UB (4-10 units
for a 1 m variation for the projected distance), but the combining effect can imply a metre scale fluctuation of
stope OB and UB across a stope surface. PLS remains, however, a linear approach. The relationship between
the variables and stope OB and UB can be non-linear and therefore the coefficients need to be analysed with
care.

113
 Figure 4. 18: Loadings chart for the Dugald River PLS analysis.

Table 4. 7: PLS coefficients of the Dugald River variables for linear regression.

Variable Coefficient
Blasting energy proxy 0.26
Blasting orientation -0.08
Blasting standoff -0.17
Dip -0.04
Direction 0.08
ERF 0.21
Convex hull 0.15
Undercut 0.20
Angle to fault 0.09
Distance to fault -0.06
Fault directional measure 0.09
Drift directional measure 0.05
Dissimilarity -0.06

4.8. Discussion
A multivariate analysis approach at an octree resolution has enabled a thorough root cause analysis of the
transversal stope’s OB and UB. This approach to statistical analysis builds an understanding of the global critical
factors influencing the transversal stopes OB and UB.

The following discussion focuses the results for an individual stope to demonstrate the practical outcomes of the
multivariate analysis. We refer to stope A presented in Figure 4. 15. Figure 4. 19 shows the octree points

114
coloured by the critical variables for this stope’s HW and side wall. The side wall sustained small OB and UB
(Figure 4. 19g), most likely due to the fact that the face is vertical (Figure 4. 19f) and the drill rings are parallel
(Figure 4. 19c). The area over the drift generated larger OB (Figure 4. 19[g]) due to the undercutting of the HW
by the drift (Figure 4. 19d). The top of the HW generated large OB (Figure 4. 19g). This area has a shallow
dipping angle around 50° (Figure 4. 19f) with a high blasting energy proxy around the drift (Figure 4. 19b). In
addition, the top drift in the stope caused overcut of this area (Figure 4. 19d) and there is a fault 5–6 m in front
of that area of the stope according to the structural model (Figure 4. 19e). These combining effects explain the
large OB that is observed. The bottom part of the HW under the crest shows UB (Figure 4. 19g), most likely due
to a low blasting energy proxy (Figure 4. 19b), coupled with the effect of the edge of the crest on the face stability
(low ERF, Figure 4. 19a).

Figure 4. 19: Critical variable values along Dugald River stope A design shape HW and side wall; (a) ERF; (b) Blasting energy proxy;
(c) Blasthole orientation (◦);( d) Undercut (m); (e) Distance to the closest fault (m); (f) Dip (◦); (g) Projected distance (m).

Figure 4. 20 shows the octree points coloured by the critical variables for this stope’s FW. The FW shows UB in
the top part (Figure 4. 20c), which can be associated to a low blasting energy proxy (Figure 4. 20b). The drift
undercutting the bottom of the FW (Figure 4. 20a), combined with a higher energy proxy (Figure 4. 20b), is most
likely the cause of the OB.

115
 Figure 4. 20: Critical variable values along Dugald River stope’s FW design shape; (a) Undercut (m); (b) Blasting energy proxy; (c)
Projected distance (m).

The multivariate analysis was conducted using a selection of geometrical, structural and operational variables
to understand their relationship as well as identifying the controlling factors. The PCA analysis highlighted the
different relationships. It was observed that for the most part, the variables seem to vary in an independent
fashion, and while some correlations were identified, they did not necessarily represent dependencies between
them.

Dependencies that could be observed are the drill and blast as well as the stope geometry being adapted to the
geological structures position. The consistent fanning pattern used for drilling in the HW and FW results in toeing
of these holes into these faces. Typically, blastholes are parallel to the side walls. While the designed drill rings
were used in this analysis, charging data were made available for some of these stopes. It was observed that
on occurrence, the charging near the HWs with faults at close proximity was modified, meaning blasting is
sometimes adjusted to the presence of faults. However, it is not captured with the current variables. The
structures tend to be closer to the HW and discussion with site personnel confirmed that on occurrence, the
design of the HW was modified to include a critical fault.

The blasting energy proxy is inversely correlated to the blasting standoff, which is due to these variables being
calculated using similar data. The blast energy proxy shows a stronger correlation with stope OB and UB.
Therefore, the blasting standoff could be ignored when extending these results to predictive modelling. PCA
analysis showed that blast energy seemed to be higher towards the middle of the faces as this variable was
correlated to ERF.

116
The stope geometry complexity and undercut are strongly correlated due to how they are calculated, and since
the undercut has a stronger correlation to stope OB and UB, the stope geometry complexity (distance to convex
hull) could be excluded from further analysis or predictive modelling.

The octree data were plotted using a projected distance marker style for a PCA components chart and this
enabled us to identify trends by examining the separation of OB and UB. These findings were supported with
the PLS analysis.

In summary, multivariate analysis of the transversal stope OB and UB provided the following key insights:

o OB tends to be observed when the following variables increase:

o The blasting energy proxy.
o ERF (towards the middle of the face).
o Undercut.
o The angle to the fault (toeing).
o The directional measure to the fault or drift (octree situated over the structure).
o OB tends also to be observed when the following variables decrease:
o Blasting orientation (parallel rings).
o Standoff.
o UB tends to be observed near:
o Borders of the faces (small ERF).
o Octrees with small blasting energy proxy and undercut.
o Octrees situated under structures.
o Octrees with large blasting standoff.
o When there is toeing of the rings.

The coefficient calculated from the PLS indicates that a fluctuation of the variables between 4 and 10 units can
result in an additional metre of OB or UB, with the combining effect adding up to possibly 4 m to 5 m.

To obtain the best assessment of the stope OB and UB, it can be necessary to separate mining sectors, or in
this case, transversal and longitudinal stopes, as the impact of each variable can vary with factors such as the
geological setting or the mining approach. To use these results for optimising the stope OB and UB of future
mining, analyses should be considering similar stopes and conditions. This can easily be done with the built
database.

Variables that are usually considered critical, such as the rock mass quality, were not considered in this analysis
due to the density of geomechanical data for this mine site not being sufficient to be modelled at the octree

117
resolution. However, the inclusion of a variable such as the RQD could be beneficial in the root cause analysis
and future steps.

This novel approach has shown that it can be an efficient and complete approach to extract the root causes of
stope OB and UB and increase our understanding of the conditions which cause OB and UB. The use of
multivariate analysis enables us to understand the interaction between the multiple critical variables and stope
OB and UB. The use of georeferenced data at a metre scale resolution permits the quantification and
understanding of the spatial variation of stope OB and UB along the design surface, opening the possibility for
optimisation.

4.9. Conclusion
To conclude, a root cause analysis of stope OB and UB, taking benefit from the rapid developments in computer
hardware, was conducted using a fine resolution octree data structure and the application of multivariate and
bivariate statistics. Geometrical, geological, operational and spatial variables were quantified and analysed using
PCA and PLS. The relationship between the individual input variables and the stope OB and UB was quantified
and identified. The blasting energy proxy, orientation and standoff, as well as the shape of the design, the
undercutting and the presence of faults influence the transversal stopes’ OB and UB magnitude and location on
the stope design surface. This multivariate approach enables us to quantify the complex relationship between
the variables and analyse how it affects OB and UB.

This empirical approach is easily applied at any mine site and helps to build a complete and good quality
database of their stope, allowing to extract valuable information and knowledge that can be used for
understanding and optimising stope design. The knowledge acquired from the multivariate root-cause analysis
can form the basis for developing an efficient predictive approach at a fine resolution.

4.10. Chapter conclusion

This chapter presented the root cause analysis for the Dugald River mine site. The goal of the root cause
assessment is to identify the critical variables that control the OB and UB and to help predict the stope geometry
with minimal error. The proposed methodology has enabled this. Following this analysis, the critical variables
that will be used for predicting the stope geometry are:

118
 • Undercut.
• ERF.
• Blasting energy proxy.
• Dip of the face.
• Blasting orientation.
• RQD.
• Direction of the face.
• Drift directional measure.
• Fault directional measure.
• Distance to fault.
• Angle to fault.
• Dissimilarity.

The results provide a first step for operational mines to understand how their stopes behave and work towards
an optimal design. Since the controlling variables will vary from site to site given the mining and geological
conditions, the root cause analysis step is repeated for the Prominent Hill and Westwood case studies. This is
presented in the next chapter.

119
Chapter 5. Replicability of the stope OB and UB
analysis with two different case studies
One of the objectives of the proposed methodology for predicting stope geometry is to develop a root cause
assessment through statistical analysis that can be applied at any mine site for identifying the critical variables
needed in the predictive model. To demonstrate the replicability and versatility of the root cause analysis
approach, two additional case studies are presented in this chapter (Prominent Hill and Westwood). In addition,
a summary of the root cause variables of OB and UB is presented and discussed.

5.1. Prominent Hill

The following analysis focuses on the 45 transversal stopes (Figure 5. 1) presented initially in Section 3.2 and
uses the quantified variables also presented in Section 3.2 with the exclusion of the fault variables (Table 5. 1).

Figure 5. 1: Prominent Hill transversal stopes used for the analysis

120
 Table 5. 1: Variables selected for the Prominent Hill mine site root cause analysis

Variables selected for analysis

Blasting standoff
Blasting orientation
RQD
Drift directional measure
Direction of the face
Dip of the face
Blasting energy proxy
ERF
Undercut

The projected distances for these octrees from 45 transversal stopes varied between -15 m and 13 m (Figure 5.
2) with 60% of octrees having a projected distance between -1 m and 1 m and 51% of octrees being UB and
49% OB. This means, that for 60% of the design surfaces being analysed, the final surface captured with the
CMS is within a 1 m distance of the planned surface.

Figure 5. 2: Normalised cumulative count of octrees’ projected distances for the 45 Prominent Hill transversal stopes

5.1.1. PCA analysis

A scree plot (Figure 5. 3) is used to determine the components in the analysis. From this chart, the first five
components are used in the analysis based on Kaiser’s rule presented in Section 2.3 (variance higher than 1).

121
These five components explain a total of 73% of the variance in the data (Table 5. 2). In the case that all the
variables would vary independently, the percentage of variance for each component would be 10%. With the
exception of the first component, none of these components predominantly explain the variance in the dataset;
therefore, the variance of the variables must be, in the most part, independent.

Figure 5. 3: Scree plot of the variance explained for each Prominent Hill PCA component

Table 5. 2: Variance explained per Prominent Hill PCA component

Components 1 2 3 4 5 6 7 8 9 10
Variance 2.6 1.5 1.2 1.0 1.0 0.9 0.7 0.5 0.4 0
Percentage of variance 26 15 12 10 10 10 7 5 4 0
Cumulative percentage 26 41 53 63 73 83 90 96 100 100

The contribution of each variable to each component is presented in Table 5. 3, with the sum of each column
being 100%. This table is used to identify variables that are independent or correlated by looking at how much
they contribute to each component. Variables identified for the same component are interpreted as variables
that vary at the same time. The correlation between the variables is either positive or negative. This is determined
using the coordinate sign in parentheses. If the signs are the same, the correlation is positive (both variables
increase or decrease) but if not, it is negative (one variable increases while the other decreases). Independent
variables will contribute to different components, meaning their variance is not correlated.

122
 Table 5. 3: Contribution, in percentage, of each variable to the variance explained by each Prominent Hill PCA component. The
strongest contributing variables for each component are identified in red (over 10%). The coordinate sign of each variable for each
component are in parentheses

Variables Comp. 1 Comp. 2 Comp. 3 Comp. 4 Comp. 5

Blasting energy proxy 12 (+) 16 (-) 6 (-) 4 (+) 0 (+)
Blasting orientation 8 (-) 9 (+) 8 (+) 22 (+) 0 (+)
Blasting standoff 5 (-) 21 (+) 12 (+) 12 (-) 0 (+)
Geometry complexity 32 (+) 6 (+) 3 (+) 0 (-) 0 (+)
Drift directional measure 0 (+) 0 (-) 0 (-) 42 (+) 1 (+)
Undercut 32 (+) 6 (+) 3 (+) 0 (-) 0 (+)
ERF 8 (+) 1 (-) 32 (+) 1 (+) 0 (+)
Direction 4 (+) 11 (+) 25 (-) 4 (-) 0 (+)
Dip 0 (+) 29 (+) 11 (-) 14 (+) 1 (-)
RQD 0 (-) 0 (+) 0 (-) 0 (-) 97 (+)

Analysis of the variance in the data allows a number of observations to be made (Table 5. 4).

Table 5. 4: Observations made on the relationships between the Prominent Hill variables

Components Observations
• Describes the stope geometry, with its variance correlating with the blasting energy
1 proxy.
• The energy proxy is positively correlated to the undercut and geometry complexity.

• Describes the different faces and how it correlates to blasting.

2 • The energy proxy is inversely correlated to the standoff, the direction and the dip.

• Describe the variance related to face size.

3 • ERF is positively correlated with the standoff and inversely correlated with the
direction and dip.

• Describe the variance related to the blasting around the drives.

4 • The blasting orientation and the drift directional measures will vary according to
the dip.

• Describes the rock mass.

5 • RQD is independent from the other variables.

To summarise, the PCA analysis identified correlations between the drilling pattern, the stope geometry and how
the drives intersect the stopes. In addition, the blasting energy tends to increase where there are larger undercut

123
and dipping faces. The RQD, on the other hand, didn’t show any correlations with the other variables implying
that neither the drill rings or the stope geometry account for the RQD.

While the projected distance wasn’t used in the PCA analysis, it was used as a marker in the score charts to
look at how the stope OB and UB varies across the component’s axis. This is analysed two components at a
time. Since there are five components, there are 10 charts to look at which are presented in Appendix B. It is
observed that components 1 through 4 show some separation between OB octrees and UB octrees. For
example, Figure 5. 4 shows two charts comparing component 1 to component 2. Figure 5. 4a shows the loading
chart and Figure 5. 4b shows the score chart. These two charts combined show that OB is generally observed
when component 1 increases and 2 decreases, meaning the larger the blasting energy proxy is, the larger the
undercut is. When standoff increases, UB is generally observed.

a b
Figure 5. 4: Component 1 versus component 2 charts. a: loading chart; b: score chart coloured by stope OB and UB

This observational analysis is conducted for each component. The combinations with the best separation are
summarised in Table 5. 5.

124
 Table 5. 5: Stope OB and UB summary from the Prominent Hill PCA components

Component Stope OB and UB analysis Variable analysis

combination
1–2 Gradual separation of OB and More OB when the blasting energy and undercut
UB in a diagonal direction. increases. More UB when the standoff increases and
the blastholes are toeing.
1–3 Gradual separation of OB and More OB when the blasting energy, undercut and ERF
UB along the component 1 axis. increases and for certain faces. More UB when the
standoff increases and the blastholes are toeing.
1–4 Gradual separation of OB and More OB when the blasting energy and undercut
UB along the component 1 axis. increases. More UB when the standoff increases and
the blastholes are toeing.
1–5 Gradual separation of OB and More OB when the blasting energy and undercut
UB along the component 1 axis. increases. More UB when the standoff increases and
the blastholes are toeing. No distinctions with the RQD
2–3 Gradual separation of OB and More OB when the blasting energy increases. More
UB in a diagonal direction. UB when standoff increases and for vertical faces.
2–4 Gradual separation of OB and More OB when the blasting energy increases and
UB in a diagonal direction. depending on the position of the drive. More UB when
standoff increases and the dip of the face increases.
2-5 Gradual separation of OB and More OB when the blasting energy increases. More
UB along the component 2 axis. UB when standoff increases and for vertical faces.
Some separation in regard to RQD.
3-4, 3-5, 4-5 No distinct separation of OB and
UB.

5.1.2. PLS analysis

Now that we understand the relationships, a PLS model is used to identify variables that, linearly, will vary with
the stope performance measure (projected distance). A PLS model is built with one component. The analysis is
concentrated on the loadings bar chart (Figure 5. 5). Each variable has a weight, or loading, in the calculation of
the stope OB and UB and is displayed as a bar chart where the longer the bar, the stronger the correlation
between the variance of the variable and the stope OB and UB. In this figure, the stope OB and UB is the last
variable identified as projected distance.

125
 Figure 5. 5: Loadings chart (W* and C* values) of the first component for the Prominent Hill PLS analysis

Figure 5. 5 shows that the blasting energy proxy, the ERF, the stope geometry complexity (convex hull) and
undercut are strongly positively correlated to the octree’s projected distance, while the blasting standoff and
orientation are strongly negatively correlated. In Table 5. 6, the same information from Figure 5. 5 is assessed
by comparing the numbers. In addition, the strongly correlated variables have higher coefficients, which vary
between 0.13 and 0.28. A coefficient of 1 would indicate that a change of 1 unit for the combination of variables
would change the projected distance by 1 m (1 for 1 impact). The strongly correlated variables, rather, have a
smaller incremental impact on the fluctuation of stope OB and UB (4 to 8 units for a 1 m variation for the projected
distance), but the combining effect can imply a metre-scale fluctuation of stope OB and UB across a stope
surface. PLS remains, however, a linear approach. The relationship between the variables and stope OB and
UB can be non-linear and, therefore, the coefficients need to be analysed with care.

126
 Table 5. 6: PLS coefficients of the Prominent Hill variables for linear regression

Variable Coefficient
Blasting standoff -0.15
Blasting orientation -0.13
RQD -0.06
Dip 0.01
Drift directional
measure 0.05
Direction 0.06
Blasting energy proxy 0.23
ERF 0.26
Geometry complexity 0.28
Undercut 0.28

The rest of the variable variance shows little correlation with the projected distance. Since PLS is a linear
approach, a bivariate analysis is done before eliminating these variables as the relationship might not be linear.
Each variable is plotted versus the projected distance with a moving average (Figure 5. 6). Figure 5. 6a plots
the dip against the projected distance and shows no linear relationship between the two. It is noticed that the
larger OB tends to happen around a dip of 10–30°, although there is a large spread in the data. Figure 5. 6b
plots the octree direction against projected distance. There is a large spread of the projected distance, but there
are two directions where there is larger OB (around 100° and 270°) which correspond to the FW and HW.
Figure 5. 6c plots the RQD against projected distance. There is a large spread of OB and UB and no clear
tendencies when looking at the moving average, except for RQD values over 90 where the average projected
distance decreases. Figure 5. 6d plots the directional measure to the drift against projected distance. There is
no linear correlation and there is a large spread of the data. The larger OB and UB distances don’t usually occur
near the drifts. What is noticed, however, is the tendency to OB around the drives (values near 0 or 180).

127
 a) b)

c) d)
Figure 5. 6: Prominent Hill bivariate charts coloured by point density, with a moving average, of the variables that do not show a linear
trend with the projected distance; a) Dip; b) Octree direction; c) RQD; d) Directional measure to the drift

5.1.3. Discussion
A multivariate analysis approach at an octree resolution has enabled a thorough root cause analysis of the
transversal stopes OB and UB. This approach to statistical analysis builds an understanding of the global critical
factors influencing the transversal stopes OB and UB.

The following discussion focuses on the results of an individual stope to demonstrate the practical outcomes of
the multivariate analysis. Figure 5. 7 shows the octree points coloured by the critical variables for this stope’s
HW. The sidewalls are mined against backfill and are excluded from the analysis. The majority of the HW
generated large OB (Figure 5. 7b). This is a large face (Figure 5. 7f) which has a shallow dipping angle of
approximately 50° (Figure 5. 7 h) with a 7 m undercut due to the geometry (Figure 5. 7g). These combining
effects explain the large OB that is observed. The bottom part of the HW near the floor sees OB that is most
likely in part due to the OB above and due to the blasting pattern around the drive (higher energy proxy (Figure 5.
7c), low standoff (Figure 5. 7e and blastholes parallel to the wall (Figure 5. 7d)).

128
Figure 5. 7: Critical variable values for a Prominent Hill stope design shape HW; (a) Stope design and CMS; (b) Projected distance (m);
(c) Blasting energy proxy; (d) Blasthole orientation (◦); (e) Blasthole standoff (m); (f) ERF; (g) Undercut (m); (h) Dip (◦).

Figure 5. 8 shows the octree points coloured by the critical variables for this stope’s FW. The FW shows UB
towards the edges of the face (Figure 5. 8b), which is associated to a low blasting energy proxy (Figure 5. 8c)
and toeing of the rings in the middle of the face (Figure 5. 8d). The drift undercutting the bottom of the FW
(Figure 5. 8g), combined with a higher energy proxy (Figure 5. 8c) and the small standoff (Figure 5. 8e), is most
likely the cause of the OB.

129
 Figure 5. 8: Critical variable values along the Prominent Hill stope’s FW design shape; (a) Stope design and CMS; (b) Projected
distance (m); (c) Blasting energy proxy; (d) Blasthole orientation (◦); (e) Blasthole standoff (m); (f) ERF; (g) Undercut (m); (h) Dip (◦).

The multivariate analysis was conducted using a selection of geometrical, geomechanical and operational
variables to understand their relationship and identify the controlling factors. The PCA analysis highlighted the
different relationships. It was observed that, for the most part, the variables seem to vary in an independent
fashion, and while some correlations were identified, they did not necessarily represent dependencies between
them.

The blasting energy proxy is inversely correlated to the blasting standoff, which is due to these variables being
calculated using the same data. The blasting energy proxy shows a stronger correlation with stope OB and UB.
Therefore, the blasting standoff could be ignored when extending these results to predictive modelling. PCA
analysis showed that blasting energy seemed to be higher towards the middle of the faces as this variable was
correlated to ERF.

The stope geometry complexity and undercut are strongly correlated due to how they are calculated, and since
the undercut has a stronger correlation to stope OB and UB, the stope geometry complexity (distance to convex
hull) could be excluded from further analysis or predictive modelling.

The octree data were plotted using a projected distance marker style for a PCA components chart (Appendix B)
that enabled us to identify trends by examining the separation of OB and UB. These findings were supported by
the PLS analysis.

130
In summary, multivariate analysis of the transversal stope OB and UB provided the following key insights:

o OB tends to be observed when the following variables increase:

o The blasting energy proxy.
o ERF (towards the middle of the face).
o Undercut.
o OB tends also to be observed when the following variables decrease:
o Blasting orientation (parallel rings).
o Standoff.
o UB tends to be observed near:
o Borders of the faces (small ERF).
o Octrees with small blasting energy proxy and undercut.
o Octrees with large blasting standoff.
o When there is toeing of the rings.
The coefficient calculated from the PLS indicates that a fluctuation of the variables between 4 and 8 units can
result in an additional metre of OB or UB, with the combining effect adding up to possibly 4 m to 5 m.

Variables that are usually considered critical, such as the rock mass quality, were considered through the RQD.
No clear trend was observed or correlated to the RQD. In general, the rock mass is of good quality (varying
between 70 and 100) with lower RQD values correlating to the presence of faults.

The goal of the root cause assessment is to identify the critical variables that control the OB and UB. It will also
help predict the stope geometry with minimal error. Following this analysis, the critical variables that will be used
for predicting the stope geometry are:

• Undercut.
• ERF.
• Blasting energy proxy.
• Blasting standoff.
• Dip of the face.
• Blasting orientation.
• RQD.
• Direction of the face.
• Drift directional measure.

131
 5.2. Westwood
The root cause analysis for the Westwood mine site focuses on the 40 longitudinal stopes mined in the WW28
area (Figure 5. 9) presented in Section 3.3 and use the quantified variables also presented in Section 3.3 (Table
5. 7). This section presents and discusses the results of the statistical analysis for Westwood.

Figure 5. 9: Westwood stopes used in the analysis

132
 Table 5. 7: Variables selected for the Westwood mine site root cause analysis

Variables selected for analysis

Blasting standoff
Blasting orientation
Blasting energy proxy
Drive directional measure
Direction of the face
Dip of the face
ERF
Undercut
Stope geometry complexity
RQD
GSI
UCS
UCS contrast
Distance to fault
Fault directional measure

The projected distances for these octrees varied between -2.8 m and 10 m (Figure 5. 10) with 66% of octrees
having a projected distance between 0 m and 2 m, and 14% of octrees being UB and 86% OB. This means that,
for 66% of the design surfaces being analysed, the final surface captured with the CMS was within 0 m and 2 m
of the planned surface. Given the small thickness of the stopes and the high grade of ore, the mine’s goal is to
leave no ore behind, explaining the little UB that is observed. 21% of the octrees have a projected distance of
over 2 m, which is two to three times more octrees than the other two sites (less than 10%).

133
 Figure 5. 10: Normalised cumulative count of octree projected distance for the 40 longitudinal stopes

5.2.1. PCA analysis

To determine to components used in the analysis, a decision is made using a scree plot (Figure 5. 11). From
this chart, the first eight components will be used in the analysis based on the rule of Kaiser presented in Section
2.3 (variance higher than 1). These eight components explain a total of 74% of the variance in the data (Table
5. 8). In a case where all the variables would vary independently, the percentage of variance for each component
would be 6%. With the exception of the first three components, none of these components predominantly explain
the variance in the dataset; therefore, the variance of the variables must be, in the most part, independent.

Figure 5. 11: Scree plot of the variance explained for each Westwood PCA component

134
 Table 5. 8: Variance explained per Westwood PCA component

Components 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Variance 2.5 1.9 1.5 1.4 1.3 1.1 1.1 1.0 0.9 0.8 0.8 0.6 0.4 0.3 0.2 0.2
Percentage
15 12 10 9 8 7 7 6 5 5 5 4 3 2 1 1
of variance
Cumulative
15 27 37 46 54 61 68 74 79 84 89 93 96 98 99 100
percentage

The contribution of each variable to each component is presented in Table 5. 9, with the sum of each column
being 100%. Each variable was described in Section 3.3. This table is used to identify variables that are
independent or correlated by looking at how much they contribute to each component. Variables identified for
the same component are interpreted as variables that vary at the same time. The correlation between the
variables is either positive or negative. This is determined using the coordinate sign in parentheses. If the signs
are the same, the correlation is positive (both variables increase or decrease), but if not, it is negative (one
variable increases while the other decreases). Independent variables will contribute to different variables,
meaning their variance is not correlated.

135
Table 5. 9: Contribution, in percentage, of each variable to the variance explained by each Westwood PCA component. The strongest
contributing variables for each component are identified in red (over 10%). The coordinate sign of each variable for each component is
in parentheses

Variables Comp. 1 Comp. 2 Comp. 3 Comp. 4 Comp. 5 Comp. 6 Comp. 7 Comp. 8

Blasting
15.4 (+) 6.7 (-) 8.5 (+) 4.7 (+) 1.4 (-) 2.2 (+) 1.1 (+) 0.0 (+)
energy proxy
Blasting
2.0 (-) 23.8 (+) 11.1 (+) 10.1 (+) 3.2 (-) 0.4 (+) 1.8 (+) 0.0 (+)
orientation
Blasting
0.5 (-) 12.9 (+) 6.1 (-) 14.8 (-) 1.7 (+) 7.0 (+) 11.4 (-) 7.3 (+)
standoff
Geometry
28.9 (+) 2.1 (+) 0.2 (-) 0.0 (-) 0.1 (+) 4.3 (-) 0.5 (-) 0.4 (-)
complexity
Drift directional
0.0 (+) 0.1 (-) 0.0 (+) 3.4 (-) 2.5 (+) 12.2 (-) 20.9 (+) 29.1 (+)
measure
Drift distance to 0.2 (+) 0.1 (-) 1.2 (+) 7.3 (-) 0.1 (-) 29.7 (+) 30.2 (+) 0.4 (+)
Undercut 26.9 (+) 1.0 (+) 1.9 (+) 0.6 (-) 1.9 (+) 3.6 (-) 2.3 (-) 1.4 (-)
ERF 16.3 (+) 0.5 (-) 2.7 (+) 2.6 (-) 2.3 (+) 15.6 (+) 2.2 (-) 5.8 (+)
Fault directional
1.2 (+) 0.8 (+) 19.2 (-) 19.7 (+) 0.1 (+) 4.2 (+) 0.2 (-) 5.0 (+)
measure
Fault distance to 3.7 (+) 5.8 (+) 20.1 (-) 6.4 (+) 0.2 (-) 0.0 (-) 2.5 (+) 0.3 (+)
Direction 0.4 (-) 2.6 (-) 1.9 (+) 3.9 (+) 3.0 (-) 0.0 (+) 10.2 (-) 43.7 (+)
Dip 0.0 (+) 33 (+) 15.0 (+) 1.6 (+) 0.3 (-) 0.0 (+) 0.2 (+) 0.7 (+)
RQD 0.7 (-) 0.9 (+) 3.4 (+) 0.5 (-) 26.3 (+) 10.8 (-) 0.0 (-) 2.3 (+)
GSI 2.3 (-) 0.6 (+) 1.3 (+) 0.0 (+) 20.1 (+) 8.4 (+) 5.7 (-) 3.5 (-)
UCS 1.1 (+) 8.0 (+) 3.2 (-) 21.3 (-) 15.5 (-) 1.0 (-) 0.1 (+) 0.0 (-)
UCS contrast 0.4 (+) 1.1 (+) 4.1 (-) 3.0 (+) 21.3 (+) 0.5 (+) 10.7 (+) 0.2 (-)

136
Analysis of the variance in the data allows a number of observations to be made (Table 5. 10).

Table 5. 10: Observations made on the relationships between the Westwood variables

Components Observations
• Describes the stope geometry, with its variance correlating with the blasting
energy proxy.
1 • The energy proxy is positively correlated to the ERF, undercut and geometry
complexity, meaning more energy towards the middle of the HW and FW.

• Describes the different faces and how they correlate to the blasting pattern.
• The dip is positively correlated with the orientation and standoff of the
2 blastholes, meaning there is more standoff in the vertical faces (side walls)
and more toeing scenarios in the HW and FW.

• Describes the position of the faults to the stopes.

3 • Faults tend to be near HW or FW.

• Describes the rock strength.

4 • The UCS is positively correlated to the blasting standoff and negatively
correlated to the blasting orientation and the fault directional measure.

• Describes the rock mass and rock properties.

5 • The RQD, GSI and UCS contrast are positively correlated, while the UCS is
negatively correlated.

• Describes the drives.

• Positive correlation between the ERF and the distance to the drives since
6 the drives are located in the floor and crown only. Negative correlation
between the distance to the drive and the RQD means higher RQD near the
drives, in some cases.

• Describes the drives.

7 • The distance to drive is positively correlated with the UCS contrast and
negatively correlated with blasting standoff and direction.

• Describes the direction of the faces.

8 • Some correlation with the directional measure to the drives.

To summarise, the PCA has helped identify correlations between the blasting pattern and the stope geometry.
This is the result of the narrow stopes and parallel drilling. The blasting energy tends to increase where there
are larger undercut and dipping faces. The RQD and GSI, on the other hand, didn’t show any correlations with
the other variables, implying that neither the drill rings or stope geometry consider the rock mass properties.

While the projected distance is not used in PCA analysis, it is used as a marker in the score charts to look at
how the stope OB and UB vary across the component’s axis. This is analysed two components at a time. Since

137
there are eight components, there are 28 charts to look at (charts presented in Appendix C). It is observed that
components 1 through 6 show some separation between different OB distances, and between OB octrees and
UB octrees. For example, Figure 5. 12 shows three charts comparing component 1 to component 4. Figure 5.
12a shows the loading chart and Figure 5. 12b shows the score chart with an OB-UB marker. Given the small
quantity of UB data as desired by the mine site, little information is taken from Figure 5. 12b. Instead, Figure 5.
12c is used, showing the score chart with a binned projected distance marker. This chart allows to separate
larger and smaller OB and analyse the trends in the OB. These three charts combined show that OB is generally
larger when component 1 increases and 4 decreases, meaning the larger the UCS, the ERF and the undercut
are.

a b c

Figure 5. 12: Component 1 versus component 4 charts. a: loading chart; b: score chart coloured by OB-UB; C: score chart coloured by
binned projected distance

This observational analysis is conducted for each component. The combinations with the best separation are
summarised in Table 5. 11. These results will be discussed later in this section.

138
 Table 5. 11: Stope OB and UB summary from the Westwood PCA components

Component Stope OB and UB analysis Variable analysis

combination
1–2 Mostly OB for Larger Mostly OB is observed when blastholes
component 2 values. are toeing in shallow dipping faces.
1–3 Some separation vertically for More OB is observed when higher
larger component 1 values. blasting energy, ERF, undercut.
1–4 Some separation diagonally. More OB is observed when higher ERF
and undercut and toeing. Larger OB is
observed for stope with higher UCS
values in the face.
1–5 Some separation diagonally. More OB when UCS contrast and ERF
increases.
2–3, 2–4 Some separation diagonally. Mostly OB when the dip increases.
2–5 Some separation vertically. More OB observed where the RQD, GSI
and UCS contrast is higher.
3–4 Some separation diagonally. More OB when closer to faults.
4–5 Some separation diagonally. Larger OB is observed where there is
higher RQD and UCS.
4–6 Some separation diagonally. Larger OB observed further from the
drives (middle of the stope).
5–6 Some separation diagonally. Larger OB observed in middle of stope
with high GSI and UCS contrast.
1–6, 1–7, 1–8, 2–6, 2– No distinct separation of OB and
7, 2–8, 3–5, 3–6, 3–7, UB.
3–8, 4–7, 4–8, 5–7, 5–
8, 6–7, 6–8, 7–8

5.2.2. PLS analysis

Now that we understand the relationships, a PLS model is used to identify variables that linearly vary with the
stope OB and UB measure (projected distance). A PLS model is built with one component. The analysis is
concentrated on the loadings bar chart (Figure 5. 13). Each variable has a weight, or loading, in the calculation
of the stope OB and UB and is displayed as a bar chart where the longer the bar, the stronger the correlation
between the variance of the variable and the stope OB and UB. In this figure, the stope OB and UB is the last
variable identified as projected distance.

139
 Figure 5. 13: Loadings chart (W* and C* values) of the first component for the Westwood PLS analysis

Figure 5. 13 shows that the blasting standoff, the ERF, the stope geometry complexity (convex hull) and undercut
are strongly positively correlated to the octree’s projected distance and the blasting energy proxy. The distance
to drive and the UCS contrast are also positively correlated, while the blasting orientation and the direction are
inversely correlated. In Table 5. 12, the same information from Figure 5. 13 is assessed by comparing the
numbers. In addition, the strongly correlated variables have higher coefficients, which vary between 0.1 and
0.34. A coefficient of 1 would indicate that a change of 1 unit for a variable would change the projected distance
by 1 m (1 for 1 impact). The strongly correlated variables, rather, have a smaller incremental impact on the
fluctuation of stope OB and UB (3 to 10 units for a 1 m variation for the projected distance), but the combining
effect can imply a metre-scale fluctuation of stope OB and UB across a stope’s surface. PLS remains, however,
a linear approach. The relationship between the variables and stope OB and UB can be non-linear and,
therefore, the coefficients need to be analysed with care.

140
 Table 5. 12: PLS coefficients of the Westwood variables for linear regression

Variable Coefficient
Direction -0.09
Fault distance to -0.08
Blasting orientation -0.07
Fault directional measure -0.06
UCS -0.03
RQD -0.02
Blasting energy proxy -0.01
Drift directional measure 0.01
Dip 0.05
GSI 0.07
Drift distance to 0.09
Geometry complexity 0.1
UCS contrast 0.12
Undercut 0.2
Blasting standoff 0.21
ERF 0.34

The rest of the variable variance shows little correlation with the projected distance. Since PLS is a linear
approach, a bivariate analysis is done before eliminating these variables as the relationship might not be linear.
Each variable is plotted versus the projected distance with a moving average (Figure 5. 14). Figure 5. 14a plots
the dip against projected distance and shows no linear relationship between the two. It is noticed that the larger
OB tends to happen around a dip of 10° to 20°, although there is a large spread in the data. Figure 5. 14b plots
the directional measure to the drift against projected distance. There is no linear correlation and there is a large
spread of the data. The larger OB and UB projected distances don’t usually occur near the drifts but rather at
the midpoint of the stope height (values of around 90°). Figure 5. 14c plots the directional measure to the fault
against projected distance. There is no linear correlation, and there is a large spread of the data. Figure 5. 14d
plots the distance to the fault against projected distance. There is no linear correlation. It is noticed that the larger
OB tends to happen at a distance of between 0 m and 5 m, although there is a large spread in the data. Figure 5.
14e plots the RQD against projected distance. There is a large spread of OB and UB and no clear tendencies
when looking at the moving average, except for RQD values over 80 where the average projected distance
decreases. Figure 5. 14f plots the GSI against projected distance. There is a large spread of OB and UB, and
no clear tendencies are noted when looking at the moving average. Figure 5. 14g plots the UCS against

141
projected distance. There is a large spread of OB and UB with an increase of around 125 MPa and over 175
MPa.

a) b)

c) d)

e) f)

g)
Figure 5. 14: Westwood bivariate charts coloured by point density, with a moving average, of the variables that do not show a linear
trend with the projected distance; a) Dip; b) Directional measure to the drift; c) Directional measure to the fault; d) Distance to the fault;
e) RQD; f) GSI; g) UCS

5.2.3. Discussion
A multivariate analysis approach at an octree resolution has enabled a thorough root cause analysis of the
longitudinal stope’s OB and UB in the WW28 area. The multivariate analysis was conducted using a selection
of geometrical, geomechanical and operational variables to understand their relationship and identify the
controlling factors. The PCA analysis highlighted the different relationships. It was observed that, for the most
part, the variables seem to vary in an independent fashion, and while some correlations were identified, they did
not necessarily represent dependencies between them.

142
The blasting energy proxy is inversely correlated to the blasting standoff, because these variables are calculated
using the same data. The blasting standoff showed a strong positive correlation with the stope OB and UB,
which differs from the other sites and is unusual given the general knowledge that OB tends to decrease further
away from blastholes. Figure 5. 15 shows the stope database with the octrees coloured by projected distance
and blasting standoff. The stopes with large standoff and OB are boxed in red. These stopes are wider than the
others, extending further into the HW, and no additional blastholes were added in the HW, explaining the larger
standoff. In these cases, the standoff wouldn’t be attributed to the large OB, but rather other factors, such as
undercut and faults. Therefore, this correlation in the data should be ignored as it reflects design adjustments
made by the mine sites. The blasting standoff could be ignored when extending these results to predictive
modelling. PCA analysis showed that blasting energy seemed to be higher towards the middle of the faces as
this variable was positively correlated to ERF.

Figure 5. 15: Longitudinal view of the Westwood stope database with octree coloured by a) projected distance; b) blasting standoff.
Stopes with large blasting standoff are boxed in red

The PCA analysis showed correlations between the blasthole orientation and the dip. Parallel drilling is used
mainly for these narrow stopes; therefore, the HW and FW are parallel (orientation close to 0°) to the blastholes
(Figure 5. 16) and blastholes are toeing in the crowns and floors. Given that the crowns and floors are excluded
from the analysis since they represent drifts, the octrees in the database with orientations over 40° are on the
edges of the faces. The inverse correlation observed with the PLS analysis between the projected distance and
the blasthole orientation is most likely a reflection of the position of the octrees where UB tends to occur on the
edges of the faces.

143
 Figure 5. 16: Westwood stope database with octrees coloured by blasting orientation

The stope geometry complexity and undercut are strongly correlated due to how they are calculated, and since
the undercut has a stronger correlation to stope OB and UB, the stope geometry complexity (distance to convex
hull) could be excluded from further analysis or predictive modelling.

Rock mass quality and rock strength were considered through the RQD, GSI, UCS and UCS contrast. Trends
varied in function of the stopes considered, but trends were observed with each variable. These trends, however,
are in contrast to the literature assessment presented in Section 1.4.2 as the analysis indicates larger OB is
observed when there are good ground conditions (high RQD, GSI or UCS) instead of having OB associated with
poor ground conditions. Figure 5. 17 shows the geotechnical variables and the projected distance for the stopes
where the RQD, GSI and UCS values are high and there is large OB. In reality, the OB for these HWs is most
likely controlled by the structures (Figure 5. 18). The mined HW geometry of these stopes stops on the faults;
therefore, the correlation observed between the geotechnical variables and the OB in the PCA analysis is
discarded as a root cause of OB.

144
 Figure 5. 17: Geotechnical variables and projected distance values for the selected stopes that are the cause of the correlations
identified in the PCA analysis

Figure 5. 18: CMS (orange) and faults (green) for the stoping area where large OB and high geotechnical values are observed

The octree data were plotted using a projected distance marker style for a PCA components chart. This enabled
the identification of trends by examining the separation of OB and UB. These findings were supported with the
PLS analysis. In summary, multivariate analysis of the longitudinal stope OB and UB provided the following key
insights:

145
 o OB tends to be observed when the following variables increase:
o The blasting energy proxy.
o ERF (towards the middle of the face).
o Undercut.
o Distance to the drifts (towards the middle of the face).
o UCS contrast.
o OB tends also to be observed when the following variables decrease:
o Distance to fault.
o UB tends to be observed near:
o Borders of the faces (small ERF).
o Octrees with small blasting energy proxy and undercut.
The coefficient calculated from the PLS indicates that a fluctuation of the variables between 3 and 10 units will
result in an additional metre of OB or UB, with the combining effect adding up to even more.

To visualise these statements, the result for an individual stope is used to demonstrate the practical outcomes
of the multivariate analysis. Figure 5. 19 shows the octree points coloured by the critical variables for this stope’s
HW. The majority of the HW generated OB that is larger towards the top of the face (Figure 5. 19a and g). The
HW has a high energy proxy towards the middle of the face (Figure 5. 19b). The HW is close to a fault (Figure 5.
19c) and is being overcut by the top drift (Figure 5. 19h) which most likely explains the larger OB at the top of
the HW. The rock mass overall is of fair to good quality (Figure 5. 19d and j), but there is a variation between
the top of the face and the rest of the face which can be observed for the rock strength as well (Figure 5. 19e
and k). The combination of these variables negatively impacts the OB. The UB in the bottom right corner can be
associated to the blasting and good rock mass quality.

146
Figure 5. 19: Critical variable values for a Westwood stope design shape HW; (a) Stope design and CMS; (b) Blasting energy proxy; (c)
Distance to fault (m); (d) GSI; (e) UCS contrast; (g) Projected distance (m); (h) Undercut (m); (i) ERF; (j) RQD; (k) UCS

Figure 5. 20 shows the octree points coloured by the critical variables for this stope’s FW. The FW shows OB
for the majority of the FW with larger OB near the top drift which can be associated to the overcut (Figure 5.
20h). Again, there is a high energy proxy towards the middle of the face (Figure 5. 20b). The FW is close to a
fault (Figure 5. 20c) with the OB distance matching the distance to the fault, indicating it most likely broke to a
fault. The rock mass quality is poorer for the FW than the HW (Figure 5. 20d and j).

Figure 5. 20: Critical variable values for a Westwood stope design shape FW; (a) Stope design and CMS; (b) Blasting energy proxy; (c)
Distance to fault [m]; (d) GSI; (e) UCS contrast; (g) Projected distance [m]; (h) Undercut [m]; (i) ERF; (j) RQD; (k) UCS

The Westwood mine site is a geologically complex environment that is seismically active. Variables such as
stress have not been analysed or discussed directly for this mine site as no data was made available, but it
would be interesting to look at as it can provide further insight on the OB generated.

147
The goal of the root cause assessment is to identify the critical variables that control the OB and UB and will
help predict the stope geometry with minimal error. Following this analysis, the critical variables that will be used
for predicting the stope geometry are:

• Undercut.
• ERF.
• Blasting energy proxy.
• Dip of the face.
• Blasting orientation.
• RQD.
• UCS contrast.
• Direction of the face.
• Distance to fault.

5.3. Multivariate analysis summary

Root cause analysis using statistics was conducted with data from three different mine sites. Transversal and
longitudinal stopes were analysed. The objective was to identify the root causes of OB and UB for each mine
site using the proposed approach (through univariate, bivariate, and more importantly, multivariate statistics)
and to demonstrate the flexibility of the approach to each site. These critical variables will be used to build the
predictive models and predict the mined geometry. This was achieved for each site as the key factors were
identified. The results allowed multiple observations regarding the root causes of OB and UB to be made. The
methods used are summarised in this section.

Four main categories of variables were analysed:

• Geometrical.
• Geomechanical.
• Geological.
• Operational.
The PCA enabled the identification of the correlations between the variables to be made (excluding the
projected distance). The summary is presented in Table 5. 13.

148
 Table 5. 13: Summary of the PCA

Variables that show correlation Interpretation

ERF, undercut and blasting energy Generally, the energy increases towards the middle of the faces
proxy. which correlates with concave faces where there is undercut.

Direction, dip and the blasting Since the blasting pattern usually doesn’t change, depending on
pattern. the direction, the holes will be parallel or toeing into the face when
a fan drill pattern is used.

Rock mass quality and faults. The rock mass quality will vary around and near certain faults
(generally the quality decreases).

Direction, dip and faults. Some sites will observe the faults in the wall of specific faces.

Blasting standoff and the blasting The two variables are inversely correlated. This is a result of how
energy proxy. they are calculated as the distance to the hole is considered for
both variables.

Geometry complexity and undercut. The two variables are positively correlated. This is a result of how
they are calculated as it is the same approach with the difference
being the inclusion or exclusion of the drifts.

Drift and blasting pattern. Since the drilling is made from the drifts, correlation can be made
between blasting standoff, the energy and the distance to the
drifts.

A summary of the correlation between the variables and stope OB and UB based on the PLS analysis is
presented in Table 5. 14. The variables are ordered according to the linear correlation range obtained with the
models. The main focus here is the interpretation of these factors that were made, which also considers the
bivariate analysis and stope examples.

149
 Table 5. 14: Summary of the analysis between the variables and the stope OB and UB

PLS coefficient
Trend with stope OB and
Variables range (absolute Interpretation
UB
values)
Stronger linear correlation.
Projected distance (PD) OB is generally observed towards
ERF 0.21–0.34
increases when increased. the middle of the face and UB
towards the edge.
Stronger linear correlation. Chance
PD increases when of
Undercut 0.2–0.28
increased. OB increases when undercut
increases.
Stronger linear correlation.
Geometry PD increases when
0.1–0.28 Chance of OB increases when the
complexity increased.
geometry complexity increases.
Stronger linear correlation.
Blasting PD increases when OB is generally observed where
0.02–0.26
energy proxy increased. there is a higher energy and UB
when there is a lower energy.
Presence of a linear correlation.
Blasting PD decreases when
0.07–0.17 UB is generally observed when the
standoff increased.
standoff increases.
Presence of a linear correlation.
Blasting PD decreases when OB is generally observed when the
0.07–0.13
orientation increased. holes are parallel to the wall and UB
when toeing.
Will depend on the Weak linear correlation.
Direction of
0.06–0.09 direction of the critical Larger OB (over 2 m) is observed in
the face
faces. the HW and FW.

150
 Weak linear correlation.
Distance to PD decreases when OB is observed when there is the
0.06–0.08
the fault increased. presence of faults near the face
(within 5 m).
Fault Weak linear correlation.
PD increases when
directional 0.06–0.09 OB is observed when a fault is near
increased.
measure and below the octree.
Weak linear correlation.
PD decreases when
RQD 0.02–0.12 OB is observed when RQD
increased.
decreases.
Weak linear correlation.
PD decreases when
GSI 0.07–0.09 OB is observed when GSI decreases
increased.
(only analysed for one mine site).
Weak linear correlation.
PD increases when Potential OB increases when the
UCS contrast 0.01–0.12
increased. UCS contrast increases
(only analysed for one mine site).
Weak linear correlation.
PD increases when
UCS 0.03–0.07 No clear trends
increased.
(only analysed for one mine site).
Drift No linear correlation.
PD increases when
directional 0.01–0.05 OB is generally observed near and
increased.
measure over the drift.
No linear correlation.
PD decreases when
Dip of the face 0.01–0.05 Higher chance for larger OB (over 2
increased.
m) when dip between 0° and 50°).

While the OB and UB is a result of the combination of these variables, some variables are more critical than
others and determine if OB or UB is observed in a specific location in the stope. These factors are:
• Blasting:
o The pattern use determines the distribution of the blasting energy and is one of the main
factors for determining UB. A higher energy proxy implies OB, but a low energy implies UB.
This was seen for the large transversal stopes where a fan drill pattern is used, creating areas

151
 with low energy and UB. Also, stopes in the database with blasting issues, such as misfires,
resulted in UB, as discussed with site personnel.
• Stress:
o The induced stress from mining impacts the presence and magnitude of OB and was analysed
indirectly using the direction, undercut and ERF. Generally, the HW faces observed larger OB
than the other faces. Sites such as Westwood where longitudinal stopes are mined are
strongly affected by the undercut from the drifts. The stress gets redistributed in the face and
leads to OB. The rock mass quality and the rock strength also impact the stress or the effect
of the stress redistribution on the stope OB and UB. Westwood has a complex geological
setting where stress and faults are the main issues affecting stope OB and UB. Unfortunately,
no direct measure of stress was available to further analyse and support these statements.
• Faults:
o The presence of faults will impact the stress redistribution and the quality of the rock mass.
Given that faults are not always present, its presence implies potential OB. It was observed
that, for some stopes, the rock broke and stopped on a fault, controlling the magnitude of OB.

Following the analysis of the stope OB and UB, two statements regarding the methods used are made. First,
the multivariate statistical methods used (PCA and PLS) resulted in a good understanding of the correlations
between the variables and how stope OB and UB are impacted when these variables vary. However, the trends
identified do not automatically imply they are the causative factors of the OB or UB. A good understanding of
rock mechanics and mining is needed for interpreting the results.

Second, to obtain the best assessment of the stope OB and UB, it can be necessary to separate mining sectors,
or in this case, transversal and longitudinal stopes, as the impact of each variable will vary with factors such as
the geological setting or the mining approach. To use these results for optimising the stope OB and UB of future
mining, analyses should consider similar stopes and conditions. This is easily done with the database that was
built.

5.4. Conclusion
This novel approach was tested at two mine sites in this chapter and has shown that it is an efficient and complete
approach to extract the root causes of stope OB and UB and increase our understanding of the conditions which
cause OB and UB. The use of multivariate analysis permits to understand the interaction between the multiple
critical variables and stope OB and UB. The use of georeferenced data at a metre-scale resolution enables the
quantification and understanding of the spatial variation of stope OB and UB along the design surface, opening
the possibility for optimisation. This chapter demonstrated the root cause approach proposed in the

152
methodology, meeting the set objectives of identifying the critical variables, and allowing the selection of
variables for the next step, which is to build the predictive model. The next chapter presents, in the form of an
article, predictive models built for the Dugald River mine site using the identified critical variables.

153
Chapter 6. Article 2 – Predicting open stope
performance at an octree resolution using
multivariate models
Benoît McFadyena, Martin Grenona, Kyle Woodwardb and Yves Potvinb

aDépartement de génie des mines, de la métallurgie et des matériaux, Faculté des sciences et de génie,
Pavillon Adrien-Pouliot, Université Laval, Québec, Canada;
bAustralian Centre for Geomechanics, The University of Western Australia, Perth, Australia

Keywords: stope design, stope reconciliation, overbreak, underbreak, multivariate, prediction, random forest

To help mines work towards an optimal stope design, this thesis proposes a methodology for predicting the
expected mined geometry that utilises the octree database compiled for each mine site and a
multivariate/machine learning approach. The objective of this chapter is to apply the proposed methodology for
predicting the expected mined geometry on a case study and analyse the prediction performance. The
predictions will enable operational sites to assess the expected stope geometry during the final stage of the
stope design step. This chapter presents the second scientific article that is integrated in this thesis. The article
discusses the predictive approach developed and the prediction performance of the three models (PLS, LDA
and random forest) for the Dugald River mine site.

6.1. Résumé
Cet article présente une étape vers une nouvelle approche de conception de chantiers dans laquelle le sur-bris
et le sous-bris des chantiers sont mesurés et géoréférencés à l'aide d'octrees à une résolution d'environ un
mètre cube et prédits à l'aide de modèles statistiques multivariés (moindres carrés partiels, analyse
discriminante linéaire et forêt aléatoire). Les résultats montrent que l'emplacement du sur-bris et du sous-bris le
long de la surface de conception ainsi que leur ampleur sont prédits avec une bonne précision à l'aide d'un
modèle forêt aléatoire (65% avec moins d’un mètre d’erreur). Ces prédictions sont utilisées pour construire la
géométrie attendue du chantier. La résolution des données et l'utilisation de l'analyse multivariée ont permis de
prédire la variation des performances des chantiers le long de la surface de conception, allant bien au-delà de
la simple prédiction qualitative par face de chantier fournie par une approche traditionnelle de l’abaque de
stabilité.

154
 6.2. Abstract
Open stoping has become a popular mining method in hard rock mines, not only due to the safety of the method
as a non-entry approach, but also because of the high extraction rate and low costs. At mine sites, stope
performance is evaluated by calculating stope overbreak using the Stability Chart. However, limitations of the
Stability Chart regarding the resolution of the predictions, non-consideration of factors such as the influence of
blasting, and the exclusion of underbreak have led to non-optimal designs. The capabilities of today’s computers
have increased the amount of data being collected and the power of models being built. This article presents a
step towards a new stope design approach where stope overbreak and underbreak are measured and
georeferenced using octrees at an approximately cubic metre resolution and predicted using multivariate and
machine learning models (partial least squares, linear discriminant analysis, and random forest). Results show
that overbreak and underbreak location along the design surface and their magnitude are predicted with minimal
error using a random forest model (65% with less than 1 m error). These predictions are used to build the
expected geometry of the open stope. The octree resolution of the data and the use of multivariate analysis has
enabled the prediction of variation in stope OB and UB along the design surface, going well beyond the simple
qualitative per stope face prediction provided by a traditional Stability Chart approach.

6.3. Introduction

When designing and mining an open stope, the goal is to create a final void geometry which is stable and closely
matches the designed geometry. However, what commonly occurs when extracting open stopes is the
unintentional mining of volumes of overbreak (OB; rock or backfill material mined outside of the design volume)
and underbreak (UB; rock in the design volume left behind). This can have a strong operational and economic
impact on the mine as it can cause stability issues which extend to the adjacent mining excavations, dilution of
ore with waste material, the loss of ore reserves that are no longer recoverable, or problems at the mill which is
optimised for expected grades and rock chemistry. Therefore, identifying the root cause factors of stope OB and
UB enables the mine’s engineers to understand stope OB and UB. This knowledge is then used when designing
open stopes to minimise stope OB and UB and to maximise the value realised from mining.

The Stability Chart is a popular predictive tool used in current stoping practices (Mathews et al., 1981; Potvin,
1988; Nickson, 1992). This chart is an empirical and bivariate tool that qualitatively predicts the stability of a
stope face and assesses a stability number, which is calculated from the geomechanical properties of the rock
mass and the hydraulic radius, which is derived from the planned geometry. Based on the results, a decision is
made regarding stope dimensions or the need for ground support. Once mining has started, the use of the
Stability Chart as an optimisation tool is limited as it does not quantify operational variables that can be modified

155
for optimising stope OB and UB, e.g. blasting variables. The Stability Chart is not capable of assessing UB,
which is largely controlled by operational variables not considered by the method. Therefore, the use of the
Stability Chart is best suited for use at the feasibility stage (or the life-of-mine stope planning step) when the
approximate dimensions of the stopes and sequence need to be decided.

Site-specific tools can also be developed using their own reconciliation data. However, workshops conducted in
Australia and Canada in 2019 highlighted that root cause analysis is done for OB at all participating mine sites,
but only some mines do it for UB. Furthermore, they still use the Stability Chart developed over 40 years ago to
assess stope stability and do not use any tools to predict UB (Potvin et al., 2020). Important variables that have
a major impact on the stope OB and UB, such as faults, undercut, stand-up time, and blasting, are not considered
in the standard Stability Chart (Clark and Pakalnis, 1997; Wang, 2004; Potvin et al., 2016; Guido, Grenon, and
Germain, 2017; McFadyen, 2020). Despite the evolution of technology increasing the amount and type of data
collected at mine sites and the computational methods for quantifying these variables, these variables are not
currently used for prediction of stope OB and UB.

The common mining practices described above highlight the need for developing a new stope design approach
that incorporates the different variables critical to stope OB and UB in order for mines to understand, predict,
and optimise stope OB and UB at the design stage. A novel approach to predict and optimise the outcome of a
planned stope can be developed, thanks to a combination of increased computational capabilities, powerful
multivariate statistical analysis (Guido and Grenon, 2018; McFadyen, 2020), and reconciliation tools at a
significantly finer resolution (Woodward et al., 2019).

Building towards a new stope design methodology, this paper will show how multivariate statistics and machine
learning, as well as data measured at an approximately cubic metre resolution (referred to here as octrees) is
used to predict stope OB and UB, giving information and knowledge at the design stage for the optimisation of
stope geometry and mine planning. This approach presents many novelties due to the resolution of the analysis
(per octree) and the wide range of variables considered which are known to impact stope OB and UB.
Georeferenced points along the design surface are efficiently obtained with the octree data structure, providing
much better predictions and spatial representation of stope OB and UB compared to a per face prediction. This
research represents a significant leap for the capability of mine sites to optimise stope geometry at the design
stage. The following work details the statistical analysis approach and presents the results of a case study.

156
 6.4. Literature review
The stope design process can be divided into four main steps (Potvin et al., 2020):
• Life-of-mine stope planning
• Stope design
• Operation and execution
• Reconciliation

Predicting stope OB and UB is an empirical process that utilises information gathered from the final reconciliation
step to develop a tool that can be integrated into the first two steps of the design process. Stope reconciliation
compares the design and mined geometry to quantify the OB and UB (Figure 6. 1).

Figure 6. 1: Illustration of stope overbreak and underbreak with respect to an open stoping mining method (McFadyen et al., 2023)

This step is typically done at two levels of resolution, on a per stope and per face basis. The total volume and
percentage are quantified and, at a per face resolution, the linear equivalent overbreak slough first introduced
by Clark, 1998 (ELOS; Figure 6. 2) and the linear equivalent loss ore (ELLO) are also quantified.

157
 Figure 6. 2: Illustration of equivalent linear overbreak slough (Clark 1998)

The mined surface is generated by a CMS (Miller, Potvin, and Jacob, 1992). A CMS uses a laser rangefinder
mounted on a head that tilts and rotates 360° to survey a cavity from one of the entry points. The scan generates
a point cloud which can then be meshed for analysis. The level of details captured by the scan will depend on
the density of points set by the surveyors. Multiple CMS scans can be done at different times to follow the
evolution of the void. However, the CMS used for the reconciliation is the final scan done once mucking is
finished. The precision is typically 2 cm, but the scan can be affected by external factors such as fog, dust,
ground support mesh, or irregular surfaces. These will cause shadowed areas for part of the stope, preventing
adequate interpretation of the actual geometry. It is possible to get around these limitations by doing multiple
scans from different entry points.

The predictive tools can be based on the mine’s own data as mining progresses or on a stope database built
from other mines that employ similar methods and have analogous conditions. The latter approach is comparable
to the Stability Chart (Figure 6. 3), which is the most widely used predictive tool at the life-of-mine stope planning
step. The Stability Chart qualitatively estimates OB on a per face basis. This tool is well adapted for this step as
no reconciliation data is available from the mine and only geomechanical and geometrical data are quantified.
The Stability Chart is also commonly used at the stope design stage, although it does not offer possibilities for
optimisation.

158
 Figure 6. 3: Stability Chart (from Nickson, 1992)

As stope extraction progresses, data from the reconciliation of the mined stopes becomes increasingly available
and can be used to develop a site-specific predictive tool. These tools use a statistical approach for a given
resolution of reconciled data. Existing predictive tools generate predictions on a per face basis, as this is the
finest resolution being processed for most mines (Potvin et al., 2020). Mines generally predict an OB volume as
the ELOS using variables averaged over the whole face. They also tend to focus on specific faces such as the
HW (Potvin et al., 2020). UB tends not to be considered during these assessments. A new and finer detailed
reconciliation process where data is reconciled on a per octree basis allows for a more refined and complete
prediction (Woodward et al., 2019).

Octrees represent georeferenced blocks which cover the design and mined three-dimensional (3D) space. A
finer resolution of blocks is obtained along the design and mined surfaces by recursively subdividing the 3D
space until the desired size or resolution of the octrees (usually ≤1 m) is obtained (Figure 6. 4). The stope OB
and UB is quantified for each octree block on the design surface by calculating a projected distance, which is
defined as the distance in the direction normal to the design between the design and mined surfaces. Positive
values represent OB while negative values represent UB.

159
Figure 6. 4: Illustration of the recursive process of octrees, as well as the octrees defined along the design surface and the stope OB
and UB quantified by calculating the distance in a direction normal to the design surface between the design surface and the CMS for
each octree (modified from McFadyen et al., 2020)

Different statistical methods exist for building a predictive model. The ability of a model to generate excellent
predictions will depend on the type of data and its distribution, along with the relationship between the variables.
The statistical model can be obtained by using a supervised method, which means, in this specific case, the
stope OB and UB data is used for building the model. The method used to build the model can be linear,
discriminant, tree-based, additive, or a neural network (Hastie et al. 2017). The model can also be obtained
using an unsupervised method, meaning, in this specific case, the stope OB and UB is not considered when
building the model but can be used later to analyse the applicability of the model to distinguish stope OB and
UB. These methods are usually based on cluster analysis or principal component analysis (Hastie et al., 2017).

Current statistical approaches for predicting stope OB and UB vary from a bivariate to a multivariate decision-
making process. In all cases, a supervised method is used to calculate the predictions, but some do utilise
concepts of unsupervised methods to create the charts used for predictions. The Stability Chart is a bivariate
approach, where the decision is made from two variables (stability number and hydraulic radius). The different
zones (stable, unstable, caving) can be separated using a discriminant method (Nickson, 1992). Mines can also
create their own bivariate chart if the critical variables are known and quantified. However, these tend to be
limited in terms of predictive capabilities since there is most likely more than one critical variable affecting stope
OB and UB. A multivariate approach is therefore the most suitable method for generating excellent predictions.
MLR (Wang, 2004; Hughes, 2011; Guido and Grenon, 2018) and PCR (Guido and Grenon, 2018) were shown
to improve the OB predictions exactness over the Stability Chart on a per face basis. McFadyen (2020) has
shown that a PLS model can be used for predicting OB in the HW (58% to 87% success rate). Random forest
(Breiman, 2001) was used by Qi et al. (2018) and an artificial neural network (McCulloch and Pitts, 1943) was
used by Adoko et al. (2022) to classify a face OB performance using geometrical and geomechanical variables.

160
 6.5. Methodology
This article proposes a new empirical stope design approach to help engineers with optimising stope
performance. This approach uses data at an octree resolution to understand (through multivariate statistical
analysis) and predict (through a multivariate or machine learning model) stope geometry at the mine site. The
methodology for understanding and identifying the critical variables using multivariate analysis is detailed in
McFadyen et al. 2023a. The critical variables are used for predicting stope OB and UB on a point basis, giving
detailed and spatial information of the OB and UB. Furthermore, this allows for the generation of a predicted
shape of the final void (referred to as 'predicted CMS' in this article). The acquired knowledge and predicted
CMS can then be used during stope design for optimisation and planning. This article details the prediction
process from building to validating the models.

6.5.1. Octree data

The new stope design approach is based on data being quantified and reconciled at an octree resolution. Using
this resolution means we are quantifying and predicting the spatial variation of the stope OB and UB along the
designed stope shape. Figure 6. 5 gives an example of the spatial variation of stope OB and UB captured using
octrees. For this example, predicting on a per face basis would give the thickness of OB and UB (ELOS and
ELLO), but would not allow identification of where the OB and UB would occur in the face, or the magnitude of
OB or UB in specific locations. Using octrees, the stope OB and UB is measured at thousands of georeferenced
points, spatially capturing where OB and UB occur, as well as the spatial distribution of the variables being
analysed. We can therefore predict the magnitude of OB or UB in specific locations across the designed stope
surface.

Figure 6. 5: An example of how stope OB and UB can spatially vary along a stope surface (McFadyen et al., 2021)

Another advantage of using data quantified at an octree resolution is that geometrical, geological,
geomechanical, and operational variables can be quantified, offering a wider range of variables that can be

161
considered than on a per face basis. It also allows for optimisation as operational variables such as blasting can
be integrated into the model.

6.5.2. Multivariate and machine learning models

Multivariate and machine learning models are used to generate the predictions. These models use multiple input
variables and previously mined stope data to predict the OB and UB of the future stopes. While many different
models exist, three models were tested and are discussed in this paper: PLS (Wold, 1966), LDA (Fisher, 1936),
and random forest (Breiman, 2001). PLS is a linear approach that quantifies the covariance between the input
variables and stope OB and UB to make predictions. LDA is a discriminant approach that aims to correctly
classify the octrees as OB or UB (for this paper) by finding the axis that best separates the two. Random forest
is a tree-based method where multiple trees are generated to obtain a predictive value.

These models were selected due to previous use and performance on a per face basis (PLS: McFadyen, 2020,
and LDA: Nickson, 1992) as well as the random forest model’s ability to capture a complex data structure. The
three models were developed through the free software environment R (R Core Team, 2021). PLS used the
PLS package (Liland et al., 2020), LDA used the MASS package (Venables and Ripley, 2002), and random
forest used the Ranger package (Wright and Ziegler, 2017).

6.5.3. Statistical evaluation of model performance

The data is separated chronologically into three groups to test the quality of the models during their construction.
The first group represents the first 50% of the data, the second represents the following 25%, and the final group
represents the final 25% of the data. This approach enables the user to train a model with the first group of data
and then test the predictive performance on the second group, optimise the model variables and finally predict
the third group of data. This allows verification of the quality of the predictions and determines which model is
best suited for the dataset, all while minimising overfitting by the model.

When evaluating the model’s predictive statistical performance, there are two aspects to consider. First, is
whether an octree is correctly predicted as OB or UB, and secondly, is the predicted distance close to the actual
projected distance. Based on the model’s performance, the variables can be analysed to understand their
relationship with OB and UB.

For the first case, the capacity of the model to correctly predict if OB or UB is generated at the octree’s location
is measured by using the confusion matrix presented in Table 6. 1. This table separates the data into four
categories, allowing the user to evaluate the error rate for OB and UB.

162
 Table 6. 1: Confusion matrix for a threshold of 0 m projected distance. Green represents true predictions and red represents false
predictions (errors).

PREDICTED UNDERBREAK PREDICTED OVERBREAK

(UB) (OB)

ACTUAL UB True UB (TUB) False OB (FOB)

ACTUAL OB False UB (FUB) True OB (TOB)

Different performance metrics will be calculated, such as the ACC (percentage of octrees correctly predicted),
the POP (percentage of overbroken octrees predicted as OB), the TOD (percentage of OB predictions that are
correct), and the TUD (percentage of UB predictions that are correct). Each of these metrics evaluates a specific
quality of the model. There is also the MCC (Matthews, 1975), which is a more complex metric but gives a global
measure of the quality of the classifications, with 1 being perfect predictions, 0 being random, and -1 being all
wrong predictions. Table 6. 2 gives the equations for these metrics.
Table 6. 2: Performance metric equations for evaluating the model’s capacity to classify an octree as OB or UB. Refer to Table 6.1 for
the acronyms.

Performance metric Equation

Classification accuracy 𝑇𝑈𝐵 + 𝑇𝑂𝐵

(ACC) 𝑇𝑈𝐵 + 𝑇𝑂𝐵 + 𝐹𝑈𝐵 + 𝐹𝑂𝐵

Positive OB predictions 𝑇𝑂𝐵

(POP) 𝑇𝑂𝐵 + 𝐹𝑂𝐵

𝑇𝑂𝐵
True OB detection (TOD)
𝑇𝑂𝐵 + 𝐹𝑈𝐵

𝑇𝑈𝐵
True UB detection (TUD)
𝑇𝑈𝐵 + 𝐹𝑂𝐵

Matthews correlation 𝑇𝑈𝐵 ∗ 𝑇𝑂𝐵 − 𝐹𝑈𝐵 ∗ 𝐹𝑂𝐵

coefficient (MCC) √(𝑇𝑂𝐵 + 𝐹𝑂𝐵) ∗ (𝑇𝑂𝐵 + 𝐹𝑈𝐵) ∗ (𝑇𝑈𝐵 + 𝐹𝑂𝐵) ∗ (𝑇𝑈𝐵 + 𝐹𝑈𝐵)

TUB = true underbreak, TOB = true overbreak, FUB = false underbreak, FOB = false overbreak

For the second case, the error of the prediction is measured by calculating the difference between the predicted
value and the actual value of the projected distance for the octree. The overall RMSE will be calculated from the
predicted and observed distances as well as the distribution of the absolute value of the errors. This allows
calculation of the percentage of octrees for which the predictions are within a certain distance. In the case of
stope performance, six brackets have been established, as shown in Table 6. 3. These brackets have been
determined using engineering judgment to characterise the spread in the prediction error.

163
 Table 6. 3: Prediction error brackets for evaluating the model’s statistical performance

Error brackets

% <0.5 m

% <1 m

% <2 m

% >2 m

% >3 m

% >4 m

It is also possible to evaluate the prediction error by plotting the predictions versus the actual values for each
octree. A perfect model would have the predictions follow a linear trend with all the points falling on a 45° line,
meaning the predicted CMS would look exactly like the actual CMS. This rarely occurs due to the inherent
uncertainty of measuring causative factors arising from the complexity of the underground geological and mining
environment where the variation in the geomechanical properties of the rock mass, as well as the variation
between the mined and planned data, cannot be fully captured. Furthermore, blasting is not an exact cutting tool
compared to a tunnel-boring machine, implying some uncertainty around the position of the final surface. A linear
regression fit will be passed through the data to visualise the general trend of the predictions. The coefficient of
determination (R²) will also be calculated to analyse the statistical quality of the regression. While the slope for
the best linear fit in the data may not match the slope of the linear fit for a perfect model, it does not mean the
predictions cannot be used at the design stage. The predictions can still indicate where OB and UB will likely
occur and the expected magnitude. The reduced statistical significance arises from the model prediction error.
An analytical analysis should be done in the end by comparing the predictions with the actual stope OB and UB
to evaluate the performance of the model as the statistical evaluation does not describe spatial performance.

6.6. Case study

The Dugald River Pb-Ag mine is located in Queensland, Australia (Figure 6. 6). Longitudinal and transversal
stopes are used to extract the ore. Stope dimensions will vary depending on the area of the mine, but stopes
are around 25 m high, between levels, and 15 m to 30 m long (Hassell et al., 2015). Stope widths vary between
2 m and 35 m depending on the orebody thickness. The orebody reaches a depth of approximately 1 km.

164
 Figure 6. 6: Location of Dugald River mine, Queensland, Australia

6.6.1. Data overview

This article focuses on 49 transversal stopes mined between 2017 and 2020 (Figure 6. 7a). Each stope is
reconciled on a per stope, per face, and per octree basis. Drifts are cut out of the design and CMS geometry by
site personnel during the building and processing of the geometries. For this reason, stope faces identified as
drift surfaces (floors and crown mostly) are excluded from the analysis. Additionally, stope faces that are mined
against backfill are excluded from analysis due to their vastly different strength and stress conditions. The 49
transversal stopes were separated into three groups chronologically. The first group contained 24 stopes, the
second 13 stopes, and the last 12 stopes. The models were built using data from the first group, which consisted
of 200 000 points, representing half the database. Since the final testing of the model was done with the final
group, the results for group 3 will be presented (Figure 6. 7b).

165
 Figure 6. 7: 3D section view of the Dugald River transversal stope database (a) and stopes part of group 3 used for predicting and
testing the models (b)

The variables included in the analysis are presented in Table 6. 4. These variables correspond to the critical
variables identified in McFadyen et al., 2023 while assessing stope OB and UB. Absent data occurs for the
structural variables since not all stopes have a fault in their vicinity. A maximum of 30 m between the fault and
octrees is used to compute the distance and angle to fault. This distance was set based on the work by
Woodward et al., (2019). For octrees with absent fault data a value must be given to include these octrees in
the model since the model does not tolerate absent data. An arbitrary choice of a distance of 40 m is set and an
angle of 100° to separate them from the octrees with fault data and keep them in the model.

166
 Table 6. 4: Dugald River critical variables as established in McFadyen et al., 2023

Categories Variables Unit Description

Equivalent radius
– Two-dimensional measure of the hydraulic radius
factor

Dip ° Dip of the design face at the octree location

Geometrical
Direction ° Direction of the design face at the octree location

Measure of stope undercut by drifts using a convex

Undercut M
hull that includes the drifts

Distance to fault M Distance between the octree and the nearest fault

Angle to fault ° Angle between the octree and the nearest fault
Geological
Fault directional Angle function of the position and distance to the
°
measure nearest fault

Blasting energy proxy – Energy estimated from the drill design

Blasting orientation ° Angle between the octree and the nearest drill hole
Operational
Drift directional Angle function of the position and distance to the
°
measure nearest drift

Characterisation of the similarity between the octree’s

Spatial Dissimilarity –
variables for the adjacent octrees

Distance between the octree and the CMS in a normal

Performance Projected distance m
direction to the design

6.7. Results
The overall statistical performance of the three predictive models is presented in Table 6. 5. Comparing the
models, the random forest model performs better than the LDA and PLS based on the capacity to correctly
predict octrees as OB and UB (67%) and the prediction error. It has a higher MCC (0.3) and a higher portion of
octrees with a prediction error smaller than 0.5 m (39%) and 1 m (65%). Based on these statistical metrics, the
random forest model performs well at predicting where OB and UB will occur as well as its magnitude, and would
bring beneficial insight to the engineers.

167
 Table 6. 5: Summary of the performance metrics for the Dugald River different models

Performance metrics Random forest LDA PLS

Classification Accuracy (ACC) 67% 66% 65%

Positive OB predictions (POP) 63% 53% 56%

True OB detection (TOD) 57% 57% 50%

True UB detection (TUD) 70% 74% 70%

Matthews correlation coefficient (MCC)
0.3 0.27 0.25

% <0.5 m 39% – 34%

% <1 m 65% – 60%

% <2 m 88% – 86%

% >2 m 12% – 14%

% >3 m 3% – 3%

% >4 m 0.75% – 0.4%

RMSE 1.26 – 1.31

After evaluating the statistical performance of the model, the predictions and the observations were analytically
compared in 3D space on the design surfaces. This innovative method predicts the detailed geometry of the
expected void and allows for comparison of the spatial variation of the predictions to the actual CMS geometry.
This is important as it will enable local verification of whether the predicted OB or UB matches the stope OB and
UB. It also allows verification if the predictions would help the engineers in their decision-making since the
prediction location is not considered in the statistical verification of the model’s performance. Overall, with the
PLS model, around 50% of the faces predicted would lead the engineer to the right interpretation of the location
of OB and UB in the face based on the analytical comparison of the predictions with the observed projected
distance (location and magnitude of OB and UB). Good and bad prediction examples will be provided later in
this paper. Similar to the PLS method, around 50% of the faces predicted with the LDA model would lead the
engineer to the right interpretation of the location of OB and UB in the face. This assessment is based on a side-
by-side comparison of the predictions with the observed projected distance for each face. These results indicate
little benefit of the PLS and LDA models in stope design as the results are as likely to be wrong as right. For the
random forest, around 75% of the faces predicted would lead the engineer to the right interpretation of the
location of OB and UB in the face based on the analytical comparison. The PLS and random forest models take

168
two different approaches to predicting stope geometry. While PLS is a linear approach, random forest is a non-
linear approach and is better adapted for the complex interaction between blasting, the rock mass properties,
the stope geometry, and stope performance. This is seen by the variables’ importance in the models (Table 6.
6), which change according to the predictive model. The PLS model assigns more importance to the ERF
variable and this is seen in the predictions as the projected distance increases towards the middle of the faces.
For the random forest model, the stope geometry and orientation are important variables for making a prediction.
Since the model generates useful predictions, the variables’ importance is used to interpret, to a certain degree,
which variables seem to influence the stope OB and UB more. In this case, the blasting energy proxy, the stope
geometry and orientation, and how the drifts cut the stope faces are probably the most important variables for
determining the stope OB and UB, followed by the presence of faults.

Table 6. 6: Variables ordered by importance in the Dugald River partial least squares and random forest models

The success rate of the random forest model indicates that stope OB and UB are predicted at an octree
resolution using a more complex statistical model with good confidence. Given the superiority of the random
forest model at predicting stope geometry, the following section will further explore the results of the model by
giving visual examples and discussing the exactness, probabilities, and limitations of the approach. Results from
the PLS and LDA models are presented in appendix D.

169
 6.7.1. Random forest

Predictions from the random forest model represent the average prediction of all trees in the model (Figure 6.
8). The model performance is optimised for a given dataset by varying the number of trees generated, the
number of variables randomly picked at each node (from which, based on a specified criterion, the variable that
maximises the separation of the projected distance is then selected for that node), and the weight of octrees
with a large predicted distance (over 2 m OB or UB) when randomly picking the octrees used for each tree due
to the lower frequency of these distances. Preliminary testing of these three variables was done using the second
group of stopes to minimise the prediction error as much as possible with this method. Based on the preliminary
testing, the model is built using 500 trees, randomly picking two variables at each node and assigning a 3:1
weight ratio to the larger projected distances to ensure a balanced representation of the different projected
distances in the model. This model was used for the prediction of the stopes in group 3 and to evaluate its
performance.

Figure 6. 8: Random forest diagram for generating a prediction

6.7.2. Statistical performance overview

The predictions generated for group 3 are plotted against the observations in Figure 6. 9. The points have been
coloured according to their confusion matrix classification. Linear regressions have been overlaid, with the blue
line representing the ideal linear fit for predictions and the red line representing the best linear fit that is obtained
from the predictions. The linear regressions help show that the data does have a linear trend, but the slope is
smaller than what is desired.

170
 Figure 6. 9: Predictions versus observations for the Dugald River random forest model. The points are coloured by their confusion
matrix classification. The ideal and best linear fit of the data is overlaid.

The statistical metrics presented in Table 6. 5 indicates the octrees were correctly predicted as OB or UB 67%
of the time. This means 67% of the design surface would be correctly depicted as OB or UB. The MCC value is
0.3, meaning the classification is not perfect (MCC = 1), but is not random either (MCC = 0), indicating the
selected variables are critical for determining where OB and UB will occur, and the model allows the user to
make the distinction between the two. The model performs better at predicting if an octree is underbroken (70%)
compared to overbroken (57%). The RMSE is 1.26, meaning the average error is around 1.26 m. The distribution
of the prediction error indicates that just under 40% of the octrees are predicted within 0.5 m of the actual
projected distance, 65% under 1 m, and 88% under 2 m. Given that the median is under 1 m error, the higher
average error is due to localised areas of the stope where OB or UB is underpredicted. Also, given the prediction
error distribution, when looking at the predicted stope surface it can be assumed that most of the surface is
within 1 m of its actual position and the majority within 2 m. Knowing that the projected distance has a range of
10 m (projected distance between –4 m and 6 m), the model allows the expected value range to be decreased
to 1 to 2 m much of the time and up to 4 m for the majority of instances. The practical outcomes of these projected
distances for a 25×20×20 m stope with a 0.5 m error across the whole stope would represent 1 250 m³ error on
the final volume (or 12.5%), while a 2 m error would represent 5000 m³ error on the final volume (or 50%). This
volume is, however, separated between OB and UB. Overall, the random forest model with the selected critical
variables at an octree resolution has enabled us to determine where OB and UB will occur for the majority of the
surface, as well as the projected distance within 1 m in many cases, and 2 m for the majority. These are
considered significant contributions to prediction of stope geometry, as the following visual example and
probabilities will show.

6.7.3. Visual example

To better visualise the exactness of the predictions, a stope example (stope A) is presented for which the
predictions match the statistical results presented in Table 6. 5 and discussed in the previous section. Figure 6.
10 presents two views of stope A, showing the actual values, the predictions, and the prediction error side-by-

171
side. The visual example shows that the prediction error of the model allows visual depiction of where OB and
UB will occur and in the most part, the magnitude. The following conclusions are drawn from this figure.

• The model predicts that the bottom of the HW will be underbroken and the top part overbroken. The
sidewall will be overbroken for the most part. The bottom of the FW will be overbroken and the top part
underbroken. UB will be observed in most of the corners.
• When compared with the actual CMS, we observe a similar trend overall, predicting OB where there
will be OB and UB where there will be UB.
• The main difference between the predictions and the actual values is in the prediction of large
magnitudes of OB and UB (over 2 m or under 1 m). As we can see from the prediction error (areas with
errors over 2 m), the predictions underpredict the OB or UB.
• These discrepancies occur, for the most part, towards the side of the stope that is against backfill, which
is not considered in the model. The probabilities discussed in the next sections will provide further
information.

Figure 6. 10: Dugald River prediction example. Stope A observed projected distances of the octrees on the left, the predicted distances
in the middle and the absolute prediction error on the right from the random forest model. The face with no octree (dark green) is a
backfill face.

6.7.4. Model prediction error

The overall prediction error of the model indicated in Table 6. 5 and the visual analysis indicate reliable results
can be obtained for predicting stope geometry through a random forest model using the critical variables
identified at the mine site. Further analysis indicates that the prediction error from the model will vary depending

172
on the observed projected distance. Figure 6. 11 shows that the majority of octrees with a projected distance
between -2 m and 2 m have smaller errors than larger distances. This is due in part to the fact that there are
fewer octrees with a projected distance over 2 m and under -2 m. Since the predictions from random forest are
the result of averaging, they are less sensitive to extreme values and can underpredict the larger projected
distances. However, for each prediction, the standard error is calculated as the standard deviation of the
predictions based on the prediction of each tree. A probabilistic approach is therefore used to assess larger OB
and UB.

Figure 6. 11: Prediction absolute error versus the projected distance for the Dugald River random forest model

6.7.5. Probabilistic approach

From the standard error and predicted value, the PDF is obtained as well as the probability density curve. A
normal distribution is assumed in this case. Using the PDF, we calculate the probability of observing a given
value or the probability that an octree’s projected distance will be larger or smaller than a specified distance. We
also calculate a prediction interval within which the projected distance of the predicted octree is most likely to be
for a given level of confidence. This means that for a level of 95% (calculated using the standard score), there
is a 95% probability that the actual projected distance will be within the interval based on the model and selected
observations.

Figure 6. 12 shows the observed projected distance and the 60% prediction interval. The upper bound
represents a probability that 80% of the projected distance will be smaller than this value and the lower bound
represents a probability that 80% of the projected distance will be larger. These limits highlight the possibilities
of large OB and UB. The projected distance of 63% of the octrees falls within the maximum and minimum
projected distance values of the 60% prediction interval. Large OB can be expected in the HW and UB in the

173
bottom of the HW and top of the FW. Octrees that fall within the interval are also highlighted. In addition to the
observations made in Figure 6. 12 the probabilistic approach allows the user to determine that the HW will likely
see OB over 2 m, the top of the FW, UB over 2 m and the top of the bottom drive in the sidewall will see OB
over 1.5 m. Since the location is not considered in the model, adjacent octrees can show a different prediction
interval. In such cases, engineering judgment should be used for interpreting these predicted distances.

Figure 6. 12: Dugald River prediction example. Observed projected distances for the octrees, the lower and upper bound of a 60%
prediction interval calculated from the random forest model and the octrees identified in or out of the prediction interval. The face with
no octree is a backfill face.

Given the standard error of the predicted octree, the probability of observing the predicted value will vary (the
probability decreases as the standard error increases) and reflects the level of confidence we can attribute to
the prediction. Figure 6. 13 shows the binned cumulative probability of the predicted projected distances versus
the prediction error. The prediction error increases when the probability of observing the predicted distance
decreases, indicating the model’s confidence in the prediction (higher probability), reflects the exactness that
can be expected based on the probability of the prediction. Small probabilities (larger standard error) for octrees
are observed when some of the trees in the model will predict a wide range of projected distance. This would
mean large and small OB or UB has been observed previously for octrees with similar parametric values, and
therefore the probability of large OB or UB is higher for these octrees and hence the standard error is larger.
The probability groups merge into one line for prediction errors over 4 m. These errors are associated with
octrees with large OB and UB distances and can be due in part to the model averaging; and the fact that the
selected variables do not fully capture the cause for these large OB and UB distances and that some critical
variables do not match the reality or are missing.

174
 Figure 6. 13: The probability of the predicted projected distances versus the prediction error for the Dugald River predicted octrees
using the random forest model.

There are, however, some octrees for which the model would be misleading as the probability of the prediction
is high (over 0.4, indicating a small standard error, usually under 2 m), but for which the prediction error is large
(error over 2 m). These octrees were plotted in a 3D view to identify the causes for the model’s inaccuracy. Two
observations were made. First, some of these octrees are randomly distributed and isolated, meaning the
surrounding octrees were well predicted. These octrees would have limited effects on the visual interpretation
of the predictions. Second, some of these octrees are grouped in clusters, meaning an area of a stope is not
well predicted. From a visual analysis, these clusters are located in parts of the stopes that behaved abnormally.
In the Figure 6. 14 example, the large UB seems to be a muckpile left in the stope and therefore reflects the
quality of the data and not the model.

Figure 6. 14: Example of a cluster of octrees large with PDF values (over 0.4) and large prediction errors (over 2 m).

175
 6.7.6. Limitations

The quality of the predictions can vary from stope to stope and is attributed to the data quality, such as fault
position being wrong or the OB and UB being controlled by different variables than those that the model
considers as critical. Sometimes an individual wall is not well predicted but the rest of the stope is. Figure 6. 15
gives an example of inaccurate predictions where OB is expected for an UB region and vice versa. The
probabilities also failed to highlight the possibility of large OB in the FW. From discussions with mine personnel,
OB was expected due to the presence of a nearby fault in the HW. Poor ground conditions were expected, which
would have caused the rock to break towards the fault. The different outcome indicates information about the
fault and its influence was inaccurate in this instance, which can occur because part of the structural model is
inferred.

Figure 6. 15: A Dugald River example of an insufficient prediction with the random forest model. The observed and predicted projected
distances are seen at the top with the absolute prediction error and the lower and upper bounds of the 60% prediction interval are seen
at the bottom with the octrees identified in or out of the reliability interval. Pale blue octrees on the crest of the HW have a distance of
0 m and were not part of the predictions. The face with no octree is a backfill face.

176
 6.8. Discussion
The new proposed design approach enables the prediction of stope OB and UB at an octree level. This means
the location and magnitude of OB and UB are predicted, allowing the user to visualise the expected CMS in 3D
and calculate volumes of OB and UB from it. This is a step change from current practice, which can only predict
performance at individual stope surface scale without identifying the location of OB, while UB is ignored. Three
separate statistical models were tested for predicting the projected distance of OB and UB. When choosing the
best model, three criteria need to be considered: the classification of the octrees as OB or UB, the prediction
error (how close is the prediction to the actual measure) and how the predictions compare to the actual CMS in
a 3D view. It is important to validate the last criterion if the model is used in the stope design process to provide
trustworthy information for engineers to make decisions.

The ACC of each model is similar. The random forest model performs better than the other two models when
looking at the statistical metrics (Table 6.5). For the three cases, the model performed better at classifying UB
than OB. Similar statistical trends were observed for the three models, but the random forest model performed
better statistically and visually. The PLS model would lead the engineer to the right assessment 50% of the time
compared to 75% of the time for the random forest when analysing stope face geometry, which indicates a high
success rate for the random forest model. The model would bring valuable insight to the engineers during the
design, unlike anything before.

Table 6. 7 presents the random forest model’s performance at correctly classifying OB and UB according to the
rounded projected distance. This allows the comparison of the model’s OB and UB classification performance
according to the projected distance that was measured with the post-mining CMS. The distance was rounded to
the lowest unit, creating 10 classes. The model is less performant in classifying octrees with small OB (0 m to 1
m) as OB (61%) but performs very well for classifying octrees with a larger projected distance of OB or UB (over
80%).

177
 Table 6. 7: Dugald River random forest model performance at correctly predicting OB and UB grouped per projected distance

This indicates that the locations where large OB and UB occur have certain characteristics quantified by the
variables that indicate that OB and UB will occur in these locations and will reliably identify these octrees as OB
or UB. Therefore, given the selection of variables used, critical portions of the stope are efficiently classified as
OB or UB. These locations are key for optimising the design. These numbers also reflect the quality variables
selected.

In addition, probabilities are generated for the predictions with random forest which enables the user to develop
a probabilistic approach to stope design. This means different geometry of the mined shape can be built. The
engineer can build a mean and worst-case geometry for assessing the expected stope geometry. Therefore, the
random forest model is the recommended statistical model for generating predictions based on the models
presented in this paper.

Testing of the random forest model showed that the predictions are quickly generated (less than five minutes).
The process can be interactive, where the engineer generates a prediction and, based on the results, modifies
certain variables to optimise stope OB and UB and obtain a new set of predictions. This can be done for one
stope or multiple stopes at once if a certain area or period of mining needs to be analysed. The predictive model
is flexible and can be generated for different sectors, types of stope, and be updated over time. Stopes can be

178
added to the model or removed if they are outdated, making sure the stopes in the model are relevant to the
stopes being predicted.

This new stope design approach started with collecting mine site data and quantifying variables at an octree
resolution. Through multivariate analysis, an understanding of the stope OB and UB was established and the
critical variables were identified (McFadyen et al., 2023). Stope geometry was predicted at an octree’s resolution,
as demonstrated in this paper, using a machine learning model. The good quality results of this first attempt at
a complete prediction of the stope geometry using a random forest predictive model show the power and
possibilities of using machine learning to predict stope OB and UB and spatially visualise what the CMS shape
could look like. The standard error is used to create probability shells. As further work is done to refine and
define input variables at an octree resolution, the model can be refined, improving the predictions of the stope
geometry.

The predictions could be implemented at the two main stages in the stope design process. There is the long-
term design where the design geometry is established and the short-term design, closer to the mining date,
where the operational variables such as the drill rings are established. The selected variables for the model
would vary.

6.9. Summary and conclusion

Multivariate and machine learning models were built using variables quantified at an octree resolution to predict
the stope geometry; more precisely, predicting the location and magnitude of overbreak (OB) and underbreak
(UB) at each point along the design surface. This novel approach significantly increases the capability of mine
sites to optimise stope performance at the design stage due to the resolution of the analysis (per octree) and
the wide range of variables considered which are known to impact stope OB and UB. Of the three tested
statistical models (PLS, LDA and random forest) random forest is suitable for predicting stope OB and UB at an
octree resolution, correctly classifying octrees as OB or UB 67% of the time with the predictions being 65% of
the time within 1 m of the actual surface in that location. From the 3D view analysis, 75% of the stope faces
(HW, FW, sidewalls and, crown) that were predicted would have led an engineer to a correct interpretation of
the face OB and UB in terms of location and magnitude. These predicted results can be included in the stope
design step for optimising stope geometry.

This methodology results in a quality site-specific predictive model which is available at the final steps of the
stope design stage. The multivariate and machine learning approach enables the user to consider multiple
critical variables and their complex relationship to understand the effect on OB and UB locally, and to predict the

179
stope geometry (CMS shape). This empirical approach is easily applied at any mine site, allowing a complete
and good quality database of the sites’ stopes to be built at the same time, making it quick and easy to predict
stope geometry by simply importing a design geometry or drill ring design and running the model.

6.10. Dugald River model performance update

The Dugald River mine site was the first case study to be analysed. Two articles based on this case study were
published, presenting the analysis (Chapter 4) and prediction process (Chapter 6). However, after the redaction
of these articles, modifications and improvements were made to the way the model performance is characterised
and quantified. The way the model performance is determined in the article (Chapter 6) is by conducting an
analytical side-by-side comparison in 3D view of the projected distances for the actual and predicted faces.
Revision of the performance metric detailed in Section 2.4.4 resulted in a revision of the Dugald River model
performance, as follows (Table 6. 8). Results from the PLS and LDA models are presented in detail in Appendix
D.
Table 6. 8: Summary of the model’s performance at determining the OB and UB location and magnitude distribution within a face

3D prediction
PLS RF LDA
representation
Excellent 59 76 -
Good 14 12 -
Limited 4 2 59
Insufficient 22 10 41

6.10.1. Random forest model

The predictions generated for the stopes in group 3 are plotted against the observations in 3D view (Figure 6.
16 and Figure 6. 17). As can be seen, although there are variations between observed and predicted CMS, the
general trends of OB and UB are captured and the predicted geometry is realistic. Overall, with the random
forest model, through the per face analytical assessment (presented in Section 2.4.4) it was determined that,
looking at the geometry of the 48 faces, 76% are excellently predicted, 12% are well predicted (good prediction),
2% show limited results (limited prediction) and 10% are insufficient.

180
Figure 6. 16: 3D view of the HW predictions versus observations for the Dugald River random forest model

181
 Figure 6. 17: 3D view of the FW predictions versus observations for the Dugald River random forest model

The statistical criterion used for determining if a geometry is predicted excellently is 50% of the octrees with a
prediction error within 1 m and 80% within 2 m and/or 70% of the octrees have their projected distance fall within
the prediction interval (10th to 90th percentile). These values were selected arbitrarily based on the size of the
stope and the impact 1 m OB or UB would have on the economic performance. The statistical criterion for the
OB and UB location (limited prediction) is 70% of the octree correctly predicted as OB or UB. Overall, the random
forest model allowed the generation of excellent predictions of the stope geometry with a high success rate
(76%) as well as generating minimal cases that predict insufficiently the stope geometry (10%).

6.11. Chapter conclusion

This chapter presented three different approaches for building a multivariate and machine learning model and
compared them to assess their capabilities of predicting OB and UB, as well as the expected mined geometry
(PLS and random forest only). Random forest was identified as the better method for predicting the stope
geometry. The proposed methodology enabled the construction of a machine learning model with minimal error
(65% of octrees within a metre) that can be used in operational mines during the final stages of the stope design
to predict the mined geometry (76% excellently). Since the controlling variables vary from site to site given the

182
mining and geological conditions, the predictive approach needed to be flexible. The predictive approach is
repeated and tested for the other two mine sites to evaluate the flexibility of the approach and the model’s
performance in different mining conditions. Once the methodology has been applied on all three sites, it will be
determined if the methodology can meet the objectives set in this thesis. This is presented in the next chapter.

183
Chapter 7. Replicability of stope geometry
predictions with two different case studies
(Prominent Hill and Westwood)
The objective of this thesis is to develop a methodology for operational mines to understand the spatial
distribution and magnitude of the OB and UB generated across stope surfaces and utilise this knowledge during
the final steps of stope design to predict stope mined geometry to optimise stope performance. The methodology
proposed utilises an octree database compiled for each mine site and a multivariate/machine learning approach
for predicting the mined geometry. The approach has the benefit of being flexible, offering the possibility to
incorporate the critical variables (geometrical, geological, geomechanical or operational) dictating the OB and
UB. The objective of this chapter is to demonstrate, through two additional case studies (Prominent Hill and
Westwood), that the prediction methodology enables operational sites to predict the expected stope geometry
during the final stage of the stope design step (when the design geometry is refined and the blasting pattern is
set) and work towards an optimal stope design. To do so, this chapter replicates the analyses presented in
Chapter 6, presenting for the two case studies the prediction results for the different predictive models (LDA,
PLS and random forest). Based on the model’s performance (quantified using set site-specific criteria presented
in Section 2.4.4), the best model for each site is identified from which the ability of the proposed methodology to
predict the mined geometry is determined.

7.1. Prominent Hill

A predictive model is built and tested for the 51 transversal stopes presented in Section 3.2.1. For this case
study, the variable inputs for the statistical models are the same as for the previous case study presented in
Chapter 6. Therefore, the first two groups are used for building the model (38 stopes) and the remaining 13
stopes for testing the predictions (group 3, Figure 7. 1). The stope selection follows the extraction sequence in
a chronological order. Following the root cause analysis presented in Section 5.1, the critical variables were
identified and are presented in Table 7. 1.

184
 Figure 7. 1: Prominent Hill stopes in group 3

Table 7. 1: Critical variables for the Prominent Hill mine site used to build the predictive models

Variables selected for prediction

Undercut
ERF
Blasting energy proxy
Blasting standoff
Dip of the face
Blasting orientation
RQD
Direction of the face
Drift directional measure

Multivariate and machine learning models were tested for predicting the projected distance of OB and UB (PLS
and random forest) and an additional model for predicting the OB and UB locations (LDA). When choosing the
best model for this mine site, three factors were considered: the classification of the octrees as OB or UB, the
prediction error (how close the prediction is to the actual measure) and the capability of predicting the stope
geometries by comparing the predictions to the actual CMS in 3D view on a face per face basis. The last factor
is important for validating if the model can be used in the stope design process to provide information for
engineers to make sound decisions. The overall performance of the three predictive models is presented in
Table 7. 2 and Table 7. 3. Comparing the models, the random forest model and PLS model showed similar
performance when classifying OB-UB (Table 7. 2), with the PLS performing slightly better at classifying the OB
and UB (2% difference in some metrics). The LDA model performs better at detecting OB than UB, while the

185
other two models perform better at detecting UB than OB. Since the LDA model does not perform better overall
than the other two models, the model is discarded as it does not predict a projected distance, but only if OB or
UB is observed at the octree’s location. When looking at the prediction error (Table 7. 2), the random forest
model and the PLS model perform the same. 57% to 59% of octrees are predicted within a metre of the actual
value and 33% within half a metre. Based on the analytical visual assessment and using the minimum criteria
set (see Section 2.4.4), 67% are excellently predicted for the random forest model compared to 54% for the
PLS, 15% compared to 23% are well predicted (good prediction), 6% compared to 8% have only the OB and
UB location predicted adequately (limited prediction) and 13% compared to 15% are insufficient. The statistical
criteria used for categorising the predictions was presented in Section 2.4.4.

Table 7. 2: Summary of the performance metrics for the Prominent Hill models

Performance metrics PLS Random forest LDA

Classification accuracy (ACC) 67% 66% 69%

Positive OB predictions (POP) 81% 79% 66%

True OB detection (TOD) 63% 62% 74%

True UB detection (TUD) 73% 71% 64%

Matthews correlation coefficient (MCC)
0.39 0.32 0.38

% <0.5 m 33% 33% –

% <1 m 59% 57% –

% <2 m 83% 83% –

% >2 m 17% 17% –

% >3 m 6% 6% –

% >4 m 1% 1% –

RMSE 1.45 1.47 –

186
 Table 7. 3: Summary of the models’ performance at determining the OB and UB location and magnitude distribution within a face for
the Prominent Hill models

3D prediction categories PLS RF LDA

Excellent 54 67 –
Good 23 15 –
Limited 8 6 63
Insufficient 15 13 38

The difference in performance between the PLS and random forest models lies in part in the probabilities that
allow to determine the likelihood of larger OB or UB to occur, thus increasing the capabilities of the model to
capture the geometry. Trends are also observed for the insufficient predicted faces. For the random forest model,
they are scattered between the different face types (HW, FW and side walls), while for the PLS model, the
insufficient predicted faces are all FW. The PLS and random forest models are two different approaches to
predicting stope geometry. While PLS is a linear approach, random forest is a non-linear approach and seems
better suited for this dataset. The variables’ impact on the model’s prediction varies between the two models
(Table 7. 4). The PLS model gives more importance to undercutting stope surfaces, ERF and blasting variables,
and it is seen in the predictions as the projected distance increases towards the middle and around drives. For
the random forest model, undercutting and dip/direction are important variables for making a prediction, followed
by blasting. The dip and direction are non-linear variables and are used to distinguish the different faces. Since
the model generates excellent predictions, the variables’ importance is used to interpret, to a certain degree,
which variables seem to influence more the stope OB and UB. In this case, blasting energy proxy and orientation,
stope geometry (face distinction), and how the drifts cut the stope faces are the most important variables for
determining the stope OB and UB.

187
 Table 7. 4: Variables ordered by importance in the Prominent Hill PLS and random forest models

Given the higher percentage of predicted faces that are in the excellent category for the random forest model,
the following section will further explore the results of the model, giving visual examples and discussing the
predictions, probabilities and limitations of the approach. The results of the PLS and LDA models are presented
in Appendix E.

7.1.1. Random forest

The goal of the random forest model is to see if a tree-based model should be used for predicting stope geometry
at an octree level. The model used 500 trees, picks two variables at random for each node and triples the weight
for octrees with a projected distance over 2 m and under -2 m to balance out the larger and smaller distances
when picking them at random for each tree. These parameters were decided after preliminary testing of the
model, generating the best predictions. The weighting thresholds were determined from the distribution of the
projected distance presented in Section 5.1. The predictions generated for the stopes in group 3 are plotted in
3D view (Figure 7. 2 and Figure 7. 3) showing a HW and FW view. As can be seen, although there are variations
between observed and predicted CMS, the general trends of OB and UB are captured and the predicted
geometry are realistic.

188
Figure 7. 2: 3D view of the HW predictions versus observations for the Prominent Hill random forest model

189
 Figure 7. 3: 3D view of the FW predictions versus observations for the Prominent Hill random forest model

The statistical metrics were presented in Table 7. 2. The octrees are correctly predicted as OB or UB 66% of the
time. This means 66% of the design surface is correctly depicted as OB or UB. The MCC value is 0.32, meaning
the classification is not perfect (MCC = 1), but not random either (MCC = 0). The model performs better at
predicting if an octree is overbroken (79%) compared to underbroken (71%). The capacity of the model at
classifying OB and UB will vary according to the projected distance. Table 7. 5 presents the model’s performance
at correctly classifying OB and UB according to the rounded projected distance for the random forest model.
This allows a comparison of the model’s OB and UB classification performance according to the projected
distance that was measured with the real CMS. The distance was rounded to the lowest unit, creating nine
classes. It is more difficult for the model to classify UB than OB and classifying octrees with small UB (-1 m to 0
m) as UB (40%), but it performs very well when classifying octrees with a larger projected distance of OB or UB
(over 80%). These results indicate that the locations where large OB and UB occur, and OB in general, can be
identified as OB and UB locations the majority of the time (over 80%) using the selected variables. This indicates
that the selected variables are representative of the root causes dictating where the larger OB and UB will occur.
In this case, the undercut and blasting are dictated by the different face direction and geometry. Therefore, given
the selection of variables used, critical portions of the stope can be efficiently classified as OB or UB. These

190
locations are key for optimising the design. In addition, the reduced ACC in locations of small UB most likely
means there is no clear distinction between the properties of small UB and OB as their characteristics might be
similar. Further analysis would be needed to validate this. The misclassification of octrees with a projected
distance between -1 m and 1 m does not imply a large prediction error. The predictions can still provide key
information on the geometry of the mined stope.

Table 7. 5: Prominent Hill random forest model performance at correctly predicting OB and UB grouped per projected distance

The predictions generated for group 3 are plotted against the observations in Figure 7. 4. The points have been
coloured according to their confusion matrix classification. Linear regressions are overlaid, with the blue line
representing the ideal linear fit for predictions, and the red line representing the best linear fit that is obtained
from the predictions. The linear regressions help show that the data does have a linear trend, but the slope is
smaller than what is expected. The RMSE is 1.47, meaning the average error is around 1.47 m. In terms of
prediction error, just under 33% of the octrees are predicted within half a metre of the actual projected distance,
57% under 1 m and 83% under 2 m. This means that, when looking at the predicted stope’s surface, it is
assumed that at least half of the surface is within a metre of its actual position and the majority within 2 m.
Knowing that the projected distance has a range of 10 m (-4 m to 6 m), the model allows us to narrow down the

191
expected value range to 1 m to 2 m most of the time, up to 4 m for the majority. Given that the median is under
a 1 m error, the higher average error is due to localised areas of the stope where OB and UB are underpredicted.

Figure 7. 4: Predictions versus observations for the Prominent Hill random forest model. The points are coloured by their confusion
matrix classification. The ideal and best linear fit of the data are overlaid

The prediction error varies depending on the observed projected distance. From Figure 7. 5, octrees with a
projected distance between -2.3 m and 2.5 m (distances between the dotted lines) tend to have smaller
prediction errors than larger OB and UB distances. This is due in part to the fact that there are less octrees with
a projected distance over 2.5 m and under -2.3 m. Since the predictions from the random forest model are the
result of averaging, they are less sensitive to extreme values and will underpredict the larger projected distances.
However, the standard error is used to identify the probability for areas in the stope where large OB or UB could
occur.

Figure 7. 5: Prediction absolute error versus the projected distance for the Prominent Hill random forest model

192
To better visualise the statistical information, a stope example is presented for which the predictions match the
statistical results presented in Table 7. 2. Figure 7. 6 presents two views of the stope, showing the actual value,
the predictions, and the prediction error side-by-side. The following conclusions are drawn from this figure:
• The HW and FW are excellently predicted.
• The model predicts overall OB for the stope, with larger OB over the drives.
• When compared with the actual CMS, a similar trend is observed overall with the difference being in
the magnitude of OB and UB.
• Looking at the prediction error, the predictions tend to underpredict the OB.

Figure 7. 6: Prominent Hill stope random forest model example: observed projected distances of the octrees on the left, the predicted
distances in the middle, and the absolute prediction error on the right.

From the standard error and predicted value, the prediction interval is calculated, within which the projected
distance is most likely. Figure 7. 7 shows the observed projected distance and the 60% prediction interval
(arbitrarily selected). The upper bound represents a probability that 80% of the projected distance will be smaller
than this value and the lower bound represents a probability that 80% of the projected distance will be larger
than this value. These limits highlight the possibility of large OB and UB. Large OB is expected in the top of the
HW and over the bottom drives. Octrees that fall within the interval are also highlighted.

193
 Figure 7. 7: Prominent Hill stope example. Observed projected distances for the octrees, the lower and upper bound of a 60%
prediction interval calculated from the random forest model and the octrees identified in or out of the prediction interval.

The quality of the predictions varies from stope to stope and even from face to face. A stope is shown in Figure 7.
8 as an example of inaccurate predictions where OB is expected for an UB region of the FW. The probabilities
identify the large OB in the HW, but do not capture the complete magnitude of the OB. From discussions with
mine site personnel, inclined FW do not generally get OB; however, analysing the stopes in the model, multiple
stopes have OB in their FW, which explains the predictions.

194
Figure 7. 8: Prominent Hill example where the predictions are not adequate with the random forest model. The observed and predicted
projected distances can be seen at the top, while the absolute prediction error and the lower and upper bounds of the 60% prediction
interval can be seen at the bottom with the octrees identified in or out of the prediction interval. The faces with no octrees are backfill
faces

For this prediction set, nine stopes out of the 13 have 60% or more of their octrees’ projected distance fall within
the 60% probability interval (the worst stope has 43% of its octrees fall within the 60% interval) and three out of
the 13 stopes have 95% or more octrees fall within the 95% probability interval (the worst stope has 82% of its
octrees fall within the 95% interval).

From the standard error and predicted value of an octree, the PDF is calculated. This step enables the
determination of the likelihood of observing a projected distance inside the prediction interval for a given octree.
The likelihood fluctuates given the selected percentile of the prediction interval. The maximum likelihood for an
octree is obtained with the initial predicted distance (median case) and reflects the confidence that can be
attributed to observing the predicted distance for a selected octree. This should be reflected in the prediction

195
error, where octrees with a larger likelihood value for their predicted distance have a smaller prediction error.
Therefore, it is used to assess the quality of the model. To analyse the link between the likelihood and prediction
error, Figure 7. 9 shows, using all the octrees, the cumulative prediction error grouped by their binned likelihood
values (values rounded to the first decimal). These groups are arbitrarily determined, but are used to distinguish
the spread in the likelihood values. The prediction error decreases when the likelihood value increases,
indicating that the model’s confidence in the prediction reflects the expected margins of error. Small likelihood
values (under 0.3) for octrees often occur when some of the trees predict large OB or UB (over 2 m). This would
mean large OB or UB has been observed previously for octrees with similar parametric values and, therefore,
the probability of large OB or UB is higher for these octrees. The likelihood groups merge into one line for
prediction errors over 4 m. These errors are obtained for large OB and UB distances and can be attributed to
selected variables not fully capturing the cause of these large OB and UB distances, and because some critical
variables are missing.

Figure 7. 9: The cumulative distribution chart of the absolute prediction error, grouped by the rounded likelihood of observing the
predicted distance for the Prominent Hill random forest model

If the octrees with errors over 4 m are selected and plotted in the 3D view, three observations are made. First,
some of these octrees are randomly distributed and isolated, meaning the surrounding octrees have a smaller
prediction error. These octrees would have limited effect on the visual interpretation of the predictions. Second
are the octrees located near the edge of the faces that are adjacent to backfilled faces and/or large UB. In the
Figure 7. 10a example, the large UB is against the crown that is backfilled and can be due to uphole blasting
issues, reflecting the quality of the data and not the model. However, the impact of the sequence and backfill
would need to be further investigated and accounted for. The third observation is the cluster of octrees located
in the FW of certain stopes with large OB (Figure 7. 10b). These octrees underpredict the OB, meaning possibly
the controlling factor is not captured by the model in this instance.

196
Figure 7. 10: Example of octrees (circled in red) with large prediction errors (over 4 m). a) octrees in UB near a backfilled face (crown);
b) octrees in FW OB

Overall, the random forest model allowed the generation of excellent predictions of the stope geometry with a
high success rate (67%), and generated minimal cases that insufficiently predict the stope geometry (13%).
Therefore, the random forest model is the recommended statistical model for generating predictions based on
the models presented for this mine site.

7.2. Westwood
A predictive model was built and tested for the 40 longitudinal stopes of the WW28 area presented in Section
3.3.1. For this case study, the variable inputs for the statistical models are the same as for the previous case
studies. Therefore, the first two groups are used for building the model (30 stopes) and the remaining 10 stopes
for testing the predictions (group 3, Figure 7. 11). The selection follows a chronological order. Following the root
cause analysis presented in Section 5.2, the critical variables were identified and are presented in Table 7. 6.

197
 Figure 7. 11: Westwood stopes in group 3

Table 7. 6: Critical variables for the Westwood mine site used to build the predictive models

Variables selected for prediction

Undercut
ERF
Blasting energy proxy
Blasting orientation
Dip of the face
Direction of the face
Distance to fault
RQD
UCS contrast

A multivariate and machine learning model were tested for predicting the projected distance of OB and UB (PLS
and random forest) as well as an additional model for predicting the OB and UB location (LDA). When choosing
the best model for this mine site, three factors were considered: the classification of the octrees as OB or UB,
the prediction error (how close the prediction is to the actual measure) and the capability of predicting the stope
geometries by comparing the predictions to the actual CMS in a 3D view on a face per face basis. The last factor
is important for validating if the model can be used in the stope design process to provide information for
engineers to make sound decisions. The overall performance of the three predictive models is presented in

198
Table 7. 7 and Table 7. 8. Comparing the models, the random forest model and PLS model showed similar
performance when classifying OB (Table 7. 7). The random forest model performs better at classifying UB than
the other two models. Since the LDA model does not perform better than the other two models, the model is
discarded as it does not predict a projected distance, but only if OB or UB will be observed at the octree’s
location. When looking at the prediction error (Table 7. 7) the random forest model and the PLS model
performance is similar. 59% to 60% of octrees are predicted within a metre of the actual value and 33–37%
within half a metre. Based on the analytical assessment and using the minimum criteria set, 51% are excellently
predicted for the random forest model compared to 49% for the PLS, 27% compared to 22% are well predicted
(good prediction), 14% compared to 11% have only the OB and UB location predicted adequately (limited
prediction) and 8% compared to 19% are insufficient. The statistical criteria used for categorising the predictions
were presented in Section 2.4.4.

Table 7. 7: Summary of the performance metrics for the various Westwood models

Performance metrics PLS Random forest LDA

Classification accuracy (ACC) 91% 91% 92%

Positive OB predictions (POP) 99% 99% 92%

True OB detection (TOD) 92% 92% 99%

True UB detection (TUD) 04% 23% 0%

Matthews correlation coefficient (MCC) 0.008 0.06 0.004

% <0.5 m 37% 33% –

% <1 m 59% 60% –

% <2 m 85% 84% –

% >2 m 15% 16% –

% >3 m 9% 9% –

% >4 m 4% 4% –

RMSE 1.58 1.60 –

199
 Table 7. 8: Summary of the models’ performance at determining the OB and UB locations and magnitude distribution within a face for
the Westwood models

3D prediction
PLS RF LDA
representation
Excellent 49 51 –
Good 22 27 –
Limited 11 14 76
Insufficient 19 8 24

The PLS and random forest model present two different approaches at predicting stope geometry. While PLS is
a linear approach, random forest is a non-linear approach. The inclusion of a probabilistic approach with the
random forest model allowed the determination of where large OB would occur more often than the PLS.
However, in both models, only about 25% of the HW and FW are well predicted (good prediction). The variables’
impact on the models’ prediction varies between the two models (Table 7. 9). The PLS model gives more
importance to the undercut, ERF and UCS contrast. This is seen in the predictions as the projected distance
increases towards the middle and around the drives. For the random forest model, the undercut, ERF and dip
are important variables for making a prediction, followed by the RQD and distance to fault. The dip is a non-
linear variable that is used to distinguish the different faces. Since the models generate beneficial predictions,
the variables’ importance is used to interpret, to a certain degree, which variables seem to influence more the
stope OB and UB. In this case, the stope geometry (face distinction), how the drifts cut the stope faces and the
rock mass quality are the most important variables for determining the stope OB and UB.
Table 7. 9: Variables ordered by importance in the PLS and random forest model for Westwood

200
Given the lower insufficient case with the random forest model and higher percentage of excellent and good
predictions, the following section will further explore the results of the model, giving visual examples and
discussing the capabilities of predicting the geometry, probabilities, and limitations of the approach. The results
for the PLS and LDA models are presented in Appendix F.

7.2.1. Random forest

The goal of the random forest model is to see if a tree-based approach should be used for predicting stope
geometry at an octree level. The model used 500 trees, picks two variables at random for each node and triples
the weight for octrees with a projected distance over 2 m and under 0 m to balance out the larger and smaller
distances when picking them at random for each tree. These parameters were decided after preliminary testings
of the model, generating the best predictions. The weighting thresholds were determined from the distribution of
the projected distance presented in Section 5.2. These differ from the two other case studies as the stope OB
and UB are different. The predictions generated for the stopes in group 3 are plotted in a 3D view (Figure 7. 12
and Figure 7. 13) showing a HW and FW view.

Figure 7. 12: 3D view of the HW predictions versus observations for the Westwood random forest model

201
 Figure 7. 13: 3D view of the FW predictions versus observations for the Westwood random forest model

The statistical metrics were presented in Table 7. 7. The octrees are correctly predicted as OB or UB 91% of the
time (ACC). This means 91% of the design surface is correctly predicted as OB or UB. The MCC value is 0.06,
meaning the classification is random (MCC = 0). The model performs better at predicting if an octree is
overbroken (92%) compared to underbroken (23%). Table 7. 10 presents the model’s performance at correctly
classifying OB and UB according to the rounded projected distance for the random forest model. This allows the
comparison of the model’s OB and UB classification performance according to the projected distance that is
measured with the real CMS. The distance is rounded to the lowest unit, creating eight classes. The model has
more difficulty classifying UB (2% to 4%) than OB (97% to 100%). This is explained by the mine site’s approach
to their stope as they do not wish to leave ore behind; therefore, they generate little UB to calibrate the model
on. The percentage of octrees within the projected distance class that is predicted as the same class is also
shown. Octrees with a projected distance between 0 m and 2 m are better predicted (45% to 62%) than the other
class (0–8%). This indicates that the locations where large OB occurs, and OB in general, will be identified as
OB.

202
 Table 7. 10: Westwood random forest model performance at correctly predicting OB and UB grouped per projected distance

Given the stope OB and UB situation for the mine site, where there is more OB over 2 m than UB, the models’
capabilities at determining if there will be OB over 2 m was also tested. Table 7. 11 presents the model’s
performance at correctly classifying over or under 2 m OB according to the rounded projected distance for the
random forest model. Octrees with a projected distance under 2 m performed well (over 90% classified correctly)
while octrees with a projected distance over 2 m do not perform as well (9% to 26%), indicating that only a fifth
of the octrees with OB over 2 m would be identified correctly. This, however, does not consider the prediction
interval, only the median case. Given the underprediction of large OB, the selection of variables used is not fully
characterising the stoping environment needed for capturing the large OB with a higher rate.

203
 Table 7. 11: Westwood random forest model’s performance at correctly predicting over or under 2 m OB according to the rounded
projected distance

The predictions generated for group 3 are plotted against the observations in Figure 7. 14. The points are
coloured according to their confusion matrix classification. Linear regressions are overlaid, with the blue line
representing the ideal linear fit for predictions, and the red line representing the best linear fit that is obtained
from the predictions. The linear regressions help to show that the data does have a linear trend, but the slope is
smaller than what is expected. The RMSE is 1.60 meaning the average error is around 1.60 m. In terms of
prediction error, 33% of the octrees are predicted within half a metre of the actual projected distance, 60% under
1 m and 84% under 2 m. This means that when looking at the predicted stope surface, it is assumed that just
over half of the surface is within a metre of its actual position and the majority within 2 m. Knowing that the
projected distance has a range of 8 m (-2 m to 6 m), the model allows to narrow down the expected value range
to 1 m to 2 m most of the time, up to 4 m for the majority. Given that the median is under a 1 m error, the higher
average error is due to localised areas of the stope where OB and UB are underpredicted.

204
Figure 7. 14: Predictions versus observations for the Westwood random forest model. The points are coloured by their confusion matrix
classification. The ideal and best linear fit of the data are overlaid

The prediction error varies depending on the observed projected distance. From Figure 7. 15, it can be seen
that octrees with a projected distance between 0 m and almost 3 m tend to have smaller errors than larger
distances. This is due in part to the fact that there are less octrees with a projected distance over 3 m and under
0 m. Since the predictions from random forest model are the result of averaging, they are less sensitive to
extreme values and will underpredict the larger projected distances. However, the standard error is used to
identify the probability for areas in the stope where large OB or UB could occur.

Figure 7. 15: Prediction absolute error versus the projected distance for the Westwood random forest model

To better visualise this statistical information, an example of an excellently predicted stope is presented for which
the predictions match the statistical results presented in Table 7. 7. Figure 7. 16 presents two views of the stope,

205
showing the actual value, the predictions, and the prediction error side-by-side. The following conclusions are
drawn from this figure:
• The model predicts overall OB for the HW and FW, with larger OB towards the middle of the face.
• The side wall will break close to even.
• Larger OB is expected in the HW than the other faces.
• When compared with the actual CMS, a similar trend is observed overall for the whole stope with the
difference being in the magnitude of OB and UB.
• The top of the FW is overpredicted, expecting large OB (over 2 m) but in reality, it is under 2 m.
• As we can see from the prediction error, the predictions tend to underpredict the OB.

Figure 7. 16: Westwood stope random forest model example: observed projected distances of the octrees on the left, the predicted
distances in the middle, and the absolute prediction error on the right

From the standard error and predicted value, the prediction interval is calculated, within which the projected
distance is most likely. Figure 7. 17 shows the observed projected distance and the 60% prediction interval. The
upper bound represents a probability that 80% of the projected distance will be smaller than this value and the
lower bound represents a probability that 80% of the projected distance will be larger than this value. These
limits highlight the possibilities of large OB and UB. Large OB is expected in the HW and FW. UB is expected in
the side walls. The HW geometry is excellently predicted with the median case; however, the upper boundary is
a better match for what actually happened and shows the strong possibility of larger OB than what is predicted
with the median case. Octrees that fall within the interval are also highlighted, which is the majority for this stope.

206
 Figure 7. 17: Observed projected distances for the octrees of the Westwood stope: the lower and upper bound of a 60% prediction
interval calculated from the random forest model and the octrees identified in or out of the prediction interval

The quality of the predictions varies from stope to stope. Figure 7. 18 gives an example of inadequate predictions
where OB is expected for an UB region of the FW and the HW OB is underpredicted. The probabilities do not
capture the large OB in the HW or the FW performance.

207
Figure 7. 18: An example of an insufficient prediction with the Westwood random forest model. The observed and predicted projected
distances can be seen at the top, while the absolute prediction error and the lower and upper bound of the 60% prediction interval can
be seen at the bottom with the octrees identified in or out of the prediction interval.

The calculation of a prediction interval enables the development a probabilistic approach to stope design. For
the probabilistic approach to be beneficial, the projected distance needs to be captured within the prediction
interval. The percentage of octrees expected to fall within the interval will depend on the level of confidence. For
example, for a 60% prediction interval, around 60% of the octrees should have their actual projected distance
fall within the interval. The ability to capture the projected distance within the predicted interval varies from stope
to stope. For this prediction set, three stopes out of the 10 have 60% or more of their octrees’ projected distances
fall within the 60% probability interval (the worst stope has only 23% of the octrees fall within the 60% interval)
and the same three out of the 10 stopes have 95% or more octrees fall within the 95% probability interval (the
worst stope has only 46% of the octrees fall within the 95% interval). Part of this result can be attributed to the
fact that only 8% of the training database are octrees with a projected distance over 3 m, while 21% of the
prediction database is over 3 m.

208
From the standard error and predicted value of an octree, the PDF is calculated. This step enables the
determination of the likelihood of observing a projected distance inside the prediction interval for a given octree.
The likelihood fluctuates given the selected percentile of the prediction interval. The maximum likelihood for an
octree is obtained with the initial predicted distance (median case) and reflects the confidence that can be
attributed to observing the predicted distance for a selected octree. This should be reflected in the prediction
error, where octrees with a larger likelihood value for their predicted distance have a smaller prediction error.
Therefore, it is used to assess the quality of the model. To analyse the link between the likelihood and prediction
error, Figure 7. 19 shows the cumulative prediction error for all the octrees (considering only the median cases),
grouped by binned likelihood values (values rounded to the first decimal). These groups are arbitrarily
determined, but are used to distinguish the spread in the likelihood values. The prediction error decreases when
the likelihood value increases, indicating that the model’s confidence in the prediction reflects the expected
margins of error. Small likelihood values (under 0.3) for octrees often occur when some of the trees predict large
OB or UB (over 2 m). This would mean large OB or UB has been observed previously for octrees with similar
parametric values and, therefore, the probability of large OB or UB is higher for these octrees. The likelihood
groups merge into one line for prediction errors over 3 m. These errors are obtained for large OB and UB
distances and can be attributed to selected variables not fully capturing the cause for these large OB and UB
distances, and that some critical variables are missing.

Figure 7. 19: The cumulative distribution chart of the absolute prediction error, grouped by the rounded likelihood of observing the
predicted distance for the Westwood random forest model

If the octrees with errors over 3 m are selected and plotted in the 3D view, two observations are made. First,
only four of the 10 stopes have errors of over 3 m and are near one another (Figure 7. 20). Second, these errors
are a result of underprediction of the OB for these stopes. These four stopes’ HW have a projected average of
over 5 m which, is higher than all the other stopes. The HW OB is structurally controlled, breaking on structures

209
that are further away from the face and even located behind other structures (Figure 7. 22). This kind of
geological complexity is not well captured with the current fault variables, which only consider the closest fault.

Figure 7. 20: 3D view of the Westwood stopes (circled in blue) with large prediction errors (over 4 m)

Figure 7. 21: Side view of the Westwood analysed stope with predictions that are inadequate, as well as adjacent stopes and faults

Overall, the random forest model is the recommended statistical model for generating predictions based on the
three models presented for this mine site. The random forest model allowed the generation of excellent
predictions of the stope geometry for 51% of the cases, and generated minimal cases that predict insufficiently
the stope geometry (10%). An additional 27% of the cases allowed the prediction of the stope geometry, but
with a slightly higher prediction error (good prediction). For the predictions that are not adequate, two of the 10
stopes have the HW and FW not adequately predicted, and seven of the 10 HW.

7.3. Applicability of the proposed methodology

The objective of this thesis is to develop a methodology for operational mines to predict stope geometry to
understand the spatial distribution and magnitude of the OB and UB generated across stope surfaces and utilise
this knowledge during the final steps of stope design to predict stope mined geometry to optimise stope
performance. To assess if the proposed methodology meets the objective of this thesis, the methodology was

210
applied at three different sites, and the prediction results of the models were characterised and quantified. The
model performance for the three mine sites is summarised in Appendix G. Following the model performance
assessment presented in this chapter, it is determined that the proposed methodology (using a random forest
model) allows the prediction of the stope mined geometry. To further support this statement, a catalogue of
predicted stopes that allowed the determination of the mined geometry is presented in Appendix H.

Overall, the results showed that the proposed methodology is a flexible and stable approach that provides key
insight into the expected stope geometry when establishing the design shape during the stope design, and closer
to the blast date when defining the drill rings, providing key insight for engineers to plan and optimise their stope.
This thesis was able to show that by using a limited quantity of stopes (40 to 50 stopes), a quality predictive
model can be built. Increasing the database could help minimise prediction errors; although, there is a limit to
the benefit linked to the quantity of additional stopes used in the model. The methodology enables engineers to
go beyond the restraints of predicting per OB or UB. It is well suited for the stoping data and will provide a path
for engineers to improve their stope prediction at the design stage and use it for the purpose of optimisation.
More importantly, it constitutes a major advancement from the stope design tools currently available.

7.3.1. Improvement on other prediction methods

The proposed prediction methodology addresses various limitations associated to current prediction tools. The
three main improvements are the resolution (per face versus per octree), the variables considered and the
inclusion of UB. To put it into perspective, a face example is used (Figure 7. 22). The prediction for the HW goes
from a qualitative measure of OB (the HW is stable using the Stability Chart) to a quantitative measure of ELOS
(0.8 m using an ELOS chart or site-specific chart) to a spatial depiction of the OB and UB magnitude across the
HW surface (proposed approach). The predicted projected distance of the HW suggests localised OB with up to
2.5 m OB in the middle of the face, which is three times larger than the suggested ELOS distance. The ELOS
chart and Stability Chart do not allow the capture of this kind of information. The ability to calculate a prediction
interval is also an improvement on the other methods which do not typically include a probabilistic approach.
The model’s confidence in the prediction (using the standard error) is determined, as well as the chances and
locations of critical OB or UB, allowing the assessment of the possible scenarios. This means that the
predictions, using the proposed methodology, allows to assess the most likely expected geometry of the stope,
which in turn gives engineers the necessary information for understanding how their stope will behave and
explore solutions for optimising the stope geometry, whether it is for the whole stope or only a specific area of
the stope. The control measures on the OB and UB will depend on their causes, but the inclusion of operational
variables, such as blasting, offers the possibility to improve and work towards the optimal design. If the OB or
UB volumes can’t be minimised, the knowledge obtained from the predictions can still be used to reduce the

211
negative impact from the OB and UB through the extraction process. The octree resolution of the data also
makes it easier to integrate economic analysis (discussed in the next chapter) which will complete the stope
performance assessment for optimisation. All these key features enabled by this proposed methodology cannot
be done with other methods.

Figure 7. 22 : Face example of the different predictions using different prediction tools for the HW

The increased resolution will reduce the impact of a prediction error, from mispredicting the correct category in
the Stability Chart (stable instead of unstable for example, which can be a difference of several OB percentages),
to mispredicting the ELOS (0.5 m or more error, which translate into 5% or more of OB depending on the size
of the face), to a localised projected distance error using the proposed method. The mean error for the proposed
method is less than 1 m in the majority of cases that predict the geometry excellently, indicating a volume error
(OB and UB is considered) of less than 5% for the example in Figure 7. 22.

7.3.2. Site usage of the prediction method

The proposed methodology was developed closely with the mine sites in the case studies. This enabled the
collection of feedback and testing by the engineers to validate the practical implementation of the proposed
methodology during stope design.

In order for the predictions to be used during the design stage, the prediction model needs to be reliable (i.e. a
consistent approach and to be successful at predicting the stope geometry) when predicting the mined geometry
and easy to use. The predictions’ success was discussed previously in Section 7.2. In terms of consistency
when building and predicting, the random forest model uses multiple trees for generating the predictions which
minimises the variance caused by the model. This allows a stable approach to generating the predictions from
one model to another. Once the model’s parameters are set for a mine site (number of trees, number of variables
selected per node etc.), little parameter manipulation for the following models would be necessary (would depend

212
on whether the number of variables used changes significantly from one area to another or the correlation
structure).

Random forest is a simple model to build and quick to use (less than five minutes for running predictions),
allowing for site personnel to do it on their own. The predictive model can be generated for different sectors and
types of stope, and updated through time. Stopes can be added to the model, or removed if they are outdated,
making sure the stopes in the model are relevant to the active stoping area. These predictions can be visualised
in a 3D view with the design shape and adjacent stopes and structures for analysing and making decisions. The
probabilities can be overlaid as well. Since a prediction per point is generated, a full predicted CMS is created.
The predicted values are used to create probability shells. This is done for one stope or multiple stopes at once
if a certain area or period of mining needs to be analysed. It is also easy to rerun predictions if variables are
changed in the perspective of optimising the expected stope performance.

There is, however, a basic understanding of random forest recommended beforehand. Selecting the data that
goes into the model is the main subjective factor that will impact the model’s variables and the quality of the
predictions. Modifying the number of selected variables impacts the decision process at each node since only X
number of variables are selected for making the splitting decision. Careful filtering of the octrees for each stope
should be done (for example, excluding octrees that represent drives or backfill) to ensure the quality and
representation of the data in the model. Except for the undercut variable, all the variables are automatically
calculated and do not need any manipulation from the engineer during data processing, thus limiting the
subjectivity in the method. Users have to be careful as it is an empirical approach, and significant changes when
modifying the input values can result in extrapolation of the data when predicting.

7.4. Conclusion
Stope geometry was predicted at an octrees resolution as demonstrated in this chapter for two mine sites using
multivariate and machine learning models. PLS, PCA and random forest were tested for predicting the OB and
UB locations, and PLS and random forest were tested for predicting the mined geometry. Good quality results
from the random forest predictive model (the geometry of around 71% of the faces is excellently predicted) show
the power and possibilities of using multivariate analysis and machine learning to predict the stope geometry,
thus allowing to spatially visualise what the CMS shape could look like, providing key insight into the OB and UB
magnitude and location as well as developing a probabilistic approach. These results indicate that the set
objective of this thesis can be met through the proposed methodology. The approach is used for predicting the
stope geometry during the final steps of stope design and helping the engineers with planning and optimising
the stopes. It is adaptable to each mining context, easy to use and is a major advancement from the stope design

213
tools currently available. This chapter covered the statistical prediction models and overall approach. The next
and final step of the methodology is assessing these predictions in terms of OB, UB and grade to determine the
accepted stope performance and the need for optimisation. The next chapter presents, in the form of an article,
a case study of how predictions made at an octree resolution using a random forest model are assessed in a
geotechnical (OB, UB) and an economic (grade) way, and how they can be used and implemented on sites to
provide an integrated optimised solution for stope optimisation.

214
Chapter 8. Article 3 - Optimising stope design
through economic and geotechnic assessments of
predictions made at a meter scale resolution using
the sites' reconciled data
Benoît McFadyena, Martin Grenona, Kyle Woodwardb and Yves Potvinb

aDépartement de génie des mines, de la métallurgie et des matériaux, Faculté des sciences et de génie,
Pavillon Adrien-Pouliot, Université Laval, Québec, Canada;
bAustralian Centre for Geomechanics, The University of Western Australia, Perth, Australia

Keywords: stope design, stope reconciliation, overbreak, underbreak, prediction, economic.

To help mines understand how their stopes can behave and the stope performance impact, this thesis proposes
an approach for assessing and understanding the predictions made at an octree resolution. This is the last step
of the proposed methodology. The objective is to demonstrate that the predictions can be used for assessing
the expected OB and UB, as well as determining the economical performance of the stope, thus allowing a full
understanding of the potential impact of the stope design. This chapter presents the third and last scientific
article that is integrated in this thesis. The article discusses the economical and geotechnical assessment for
predictions made for the Prominent Hill mine site.

8.1. Résumé
Cet article présente comment la compréhension et la prédiction de la performance des chantiers à une résolution
de l'échelle du mètre, utilisant un modèle multivarié, peuvent être utilisées pour évaluer les performances
géotechniques (bris hors-profil) et économiques des chantiers attendus afin que l’ingénieur puisse planifier et
optimiser la conception et le minage du chantier. Ces prédictions estiment l’ampleur et l'emplacement des bris
hors-profil sur la surface de conception et sont utilisées pour prédire la géométrie minée. De plus, la résolution
des prédictions permettant d’incorporer les valeurs de teneur d’un modèle de blocs permet une estimation
économique de la performance des chantiers. Ces travaux constituent une étape importante vers un processus
de conception et de planification des chantiers qui minimise les problèmes de stabilité, la dilution et la perte de
minerai à travers le chantier. Étant donné que la conception visant une récupération plus élevée du minerai
augmente le potentiel de dilution et vice versa, cette approche permet de maximiser la valeur réalisée lors de
l'exploitation du chantier en quantifiant ce compromis économique. Cela marque un progrès notable dans la
capacité de concevoir et de planifier le chantier optimal.

215
 8.2. Abstract
The final geometry of open stopes often diverges from the design, creating unintentional volumes of overbreak
and underbreak. Tools, such as the Stability Chart, are used for predicting overbreak, but the limitations of such
tools regarding the resolution of the predictions and the exclusion of underbreak have constrained the benefits
of a predictive tool at the design stage. Furthermore, the economic aspects are not integrated when addressing
geomechanical stope design. This paper presents how understanding and predicting stope OB and UB at a
metre-scale resolution through a machine learning model can be used for assessing the geotechnical (overbreak
and underbreak) and economic stope performance. These predictions estimate the magnitude and location of
overbreak and underbreak across the design surface, and are used to resolve the expected geometry of the
mined void. The incorporation of a predicted mined void with grade values from a block model allows for a
comprehensive economic stope performance estimate. This work is a significant step towards a stope design
and planning process that minimises the stability issues, dilution and loss of ore across the stope. As designing
for higher ore recovery increases the potential for dilution and vice versa, this approach allows for the value
realised from mining a stope to be maximised by quantifying this inherent economic trade-off. The enhanced
resolution of both data and predictions has allowed for a more comprehensive evaluation of the predicted stope
performance and economic outcomes during the design phase. This marks a notable advancement in the ability
to design and plan the optimal stope.

8.3. Introduction
Open stoping is a popular mining method due to the high production levels that can be achieved (Potvin and
Hadjigeorgiou 2001). To maximise the profitability of the operation, one should maximise the volume of ore per
stope while respecting the designed geometry when mining it. However, what often occurs when mining is the
generation of volumes of overbreak (OB; rock or backfill material outside of the design boundaries) or underbreak
(UB; rock inside the designed boundaries left unrecovered). These unintended volumes can have a critical
impact on the operation and the profitability of the mine. OB can cause stability issues that affect adjacent stopes
and mining infrastructure, and delay the mucking process. It can also dilute the ore with waste material that can
cause non-optimal recovery at the mill. UB, on the other hand, can result in a loss of ore as it is left behind and
can’t be recovered. To minimise these OB and UB volumes, engineers will look to understand and predict them.

A popular predictive tool used in stope design is the Stability Chart (Mathews et al. 1981) or one of its subsequent
versions. This empirical tool looks at OB, enabling us to qualitatively predict whether a stope surface will be
stable, unstable or caving. The stability is assessed by calculating a stability number using the geomechanical
properties of the rock mass and calculating the hydraulic radius using the planned stope geometry. The tool is
used to make decisions on the planned geometry or the need for ground support. In the absence of stope

216
performance and operational data, the Stability Chart is a valuable tool at the feasibility study stage. Once mining
has started and extraction progresses, operational and stope performance data become available. However,
there is no efficient process to provide operational or quantitative stability performance data to the Stability Chart.
This severely limits the optimisation capability of the method. Furthermore, the Stability Chart does not consider
UB, which can have a major impact on the revenue realised from mining a stope or the profitability of a stope.
These limitations inherent to the Stability Chart decrease its utility when aiming to predict stope geometry and
optimise stope design.

Site-specific tools can be developed by using the previously mined stope data to identify variables that are the
root causes of OB and UB. This knowledge can then be used at the design stage. However, workshops
conducted in 2019 with mine sites in Canada and Australia have shown that mine sites produce limited
information from their reconciliation, even though a large quantity of data is being collected (Potvin et al. 2020).
This results in limited information and knowledge being fed back into the design of future stopes. Furthermore,
stope economics were not considered when addressing OB and UB from a stability perspective. Assessing stope
stability and economics in silos prevents a more efficient stope design optimisation.

The limitations to predicting stope geometry and the exclusion of economic data highlight the need in the mining
industry for developing a new empirical approach to stope design. An approach that can incorporate the different
variables critical to stope OB and UB may enable mines to understand and quantitatively predict stope
performance (OB, UB and grade) at the design stage to assess and optimise stope stability and profitability. This
novel approach can be achieved through statistical multivariate analysis (Guido and Grenon 2018,
McFadyen 2020) and stope reconciliation at a significantly finer resolution than on a per face basis (Woodward
et al. 2019).

This paper presents how to conduct an integrated economic and root cause assessment of the predicted OB
and UB and their associated probabilities obtained using multivariate statistics and data measured at an
approximately cubic metre resolution, working towards a more-optimal stope design. This new methodology
presents many novelties facilitated by the increased resolution of the data and the inclusion of a wide range of
variables impacting stope OB and UB. The data structure follows an octree representation and provides
georeferenced points along the stope design surface. This structure provides a fine spatial reference stope OB
and UB compared to a per face basis. A detailed economic reconciliation can be achieved by spatially mapping
a grade block model to the octree data structure built from predicted geometry, which integrates the financial
aspect of mining into the decision-making process of stope design. This research represents a significant leap
in the capability of mine sites to predict stope performance (OB, UB and profit) at the design stage. Economic-
based optimisation represents the final step toward a new stope design methodology which can be achieved if

217
the predictive model considers variables that can be adjusted during stope design and if economic data is
quantified. The following work details the approach for assessing and using the predictions in the stope design
step and presents the results of a case study where the novel approach was tested.

8.4. Background on stope optimisation

Optimal stope design can be achieved by building an understanding of the root causes of OB and UB, and
integrating this knowledge with economic data at the design stage so that engineers can utilise this key insight
to optimise or plan stope extraction and meet the geotechnical and economic goals set by mine management.
This knowledge is gained by analysing and understanding the mine site’s historic stope OB and UB. Some stope
OB and UB issues can be managed with changes that are site-wide or per zone, such as changing the mining
sequence for managing the induced stress. However, some stope OB and UB issues are localised and will vary
from stope to stope or within the stope. To assess the expected performance of a designed stope, sites turn to
predictive tools that will quantify the stope OB and UB. The predictions can then be assessed to determine the
paths for planning or optimising the stope. The key steps for achieving an optimal design are stope reconciliation,
the prediction of stope OB and UB, and the economic analysis.

8.4.1. Stope reconciliation

Stope OB and UB are quantified by reconciling the design geometry and the mined geometry (Figure 8. 1)
obtained with a CMS scan (Miller et al. 1992) or, as of late, a LiDAR mounted on a drone (Woolmer et al. 2020).
The most common levels of resolution are on a per stope and per face basis (Potvin et al. 2020). Stope OB and
UB are quantified as volumes or percentages, and, for faces only, as ELOS (Clark 1998) and ELLO.

218
 Figure 8. 1:Stope overbreak and underbreak with respect to a stoping mining method (McFadyen et al. 2023a)

During the reconciliation process, variables are quantified for the root cause assessment and reflect the general
understanding of rock mechanics that control OB and UB. The variables will vary according to the resolution and
the available mine site data. A limited number of variables is quantified on a per stope basis (e.g. design volume
and mining method). More variables can be quantified per face (Table 8. 1). However, these variables usually
represent an average value for the face.

219
 Table 8. 1: Variables quantified on a per face basis

Reference and example of

Variable Description
application

Geometrical:

Hydraulic radius Ratio between the area of a face and its Mathews et al. 1981
perimeter.

Dip Average angle with respect to the vertical axis Mathews et al. 1981
for a stope face (°).
McFadyen 2020

Operational:

Blasting standoff Average distance of the closest drill ring to the Clark 1998
face (m).
Potvin et al. 2016

McFadyen 2020

Blasting orientation Average angle between the boreholes and Clark 1998
the face for the closest ring (°).
Potvin et al. 2016

McFadyen 2020

Undercut Width of the drive exceeding the stope Clark 1998

boundary (m).
Wang 2004

Ground support Presence of cables on the stope face, Potvin 1988

quantified as bolt density, cable length and
unsupported span. Nickson 1992

Stand-up time Number of days between the final blast and Wang 2004
the final CMS.
Guido et al. 2017

McFadyen 2020

Geomechanical:

Classification of the Rock mass quality according to one of the Mathews et al. 1981
rock mass (Q’, RQD, standard classifications.
RMR etc.) McFadyen 2020

220
 Stress regime (σ1, σ2, Characterisation of the induced stress and Mathews et al. 1981
σ3 etc.) deformation around the stope.

Geological:

Presence of faults Position, angle and distance of a fault to a Suorineni 1998

face.

There is also a finer resolution that represents a new advancement in stope reconciliation where the OB and UB
are assessed on a per octree basis (Woodward et al. 2019). The three-dimensional space is recursively
subdivided into blocks (octrees) to a given size along the design surface, which is usually less than a metre
(Figure 8. 2). At this resolution, thousands of data points are generated along the design surface, exponentially
increasing the amount of data points compared to a per face basis.

Figure 8. 2: Illustration of the recursive process of octrees as well as the octrees defined along the design surface (modified from
McFadyen et al. 2021)

Stope OB and UB at an octree resolution are quantified for each octree by calculating the distance from the
design surface to the CMS surface in the direction normal to the design surface (Figure 8. 3). This distance is
referred to as the projected distance. Projected distances are classified as OB and UB if this measure is positive
or negative, respectively.

221
 Figure 8. 3: Illustration of octree stope OB and UB quantified by calculating the distance in a direction normal to the design surface
between the design surface and the CMS for each octree (modified from McFadyen et al. 2021)

Given the finer resolution of the octree data structure, geomechanical, geometrical, geological and operational
variables can be refined (Woodward et al. 2019). The resolution enables the spatial variation of values to be
captured along the design surface, addressing the loss of information caused by averaging values over a stope
face and increasing the level of detail an analysis can achieve.

8.4.2. Predicting stope OB and UB

Predictive tools are built on the understanding of the root causes of OB and UB, which is done through analysis
of the quantified variables. The identified critical variables feed into the charts or models which output a
qualitative or quantitative measure of OB or UB. These tools are empirical approaches and rely on the mining
experience as well as general concepts of rock mechanics for predicting stope OB and UB.

The predictive tools can be used at different stages of the design to determine the expected stability of the stopes
and plan or optimise. The first tool that was developed is the Stability Chart by Matthews et al. in 1981. It allows,
at the feasibility stage, the determination of the general OB and selection of stope dimensions, which can be
used to come up with an initial optimal geometry. The chart uses fixed variables (stability number and hydraulic
radius) to determine if the face OB will be stable, unstable or caving. As mentioned in the introduction, the chart
presents several limitations regarding the predictions; namely, the precision of predicted OB and exclusion of
UB, and the difficulty to include new factors such as drilling and blasting in the chart’s structure. Furthermore, a
bivariate approach is limited as variables can be cross-correlated and interdependent.

The shortcomings of the Stability Chart have led to the exploration of statistical models for predicting OB or UB
that offer flexibility for integrating other variables. These models can be used once mining has started as they
rely on the site’s previously mined stopes. Multiple multivariate and machine learning models have been tested
and used for classifying or quantifying OB on a per face basis. These models are listed below:

222
 • MLR (Wang 2004, Hughes 2011, Guido and Grenon 2018).

• Logistic regression and PCR (Guido and Grenon 2018).

• PLS (McFadyen 2020).

• Random forest (Qi et al. 2018).

• Artificial neural network (Adoko et al. 2022).

These models showed accurate predictions or classifications of OB on a per face basis and use additional critical
variables, thus making them more flexible and site-specific approaches. They can also be more precise than
most of the Stability Charts by predicting ELOS or ELLO in comparison to general stability state (stable, unstable,
caving). The statistical models quantify directly (correlation coefficients) or indirectly (models’ variable
importance) the impact of each variable on the predictions and can be used to assess the root causes of the
predicted stope OB and UB. The main limitation with these models is that separate models are needed for OB
and UB, and their implementation are on a per face basis, thus not enabling the assessment of stope OB and
UB at a finer resolution. Multivariate and machine learning models (PLS, LDA and random forest) at an octree
resolution can be applied to address these limitations (McFadyen et al. 2023b). The resolution of the data
enables the prediction of the magnitude and location of OB and UB in one model which allows a predicted mined
geometry to be constructed from these results.

While these models and charts provide a prediction for a set of input values, there can be uncertainties and
variability in the variables that go into these tools due to the rock mass sampling, the rock mass heterogeneity
or the data quality. To encompass these factors, a probabilistic approach can be applied to provide a prediction
interval and evaluate the probability of large OB or UB. This approach relies on quantifying the distribution of
each input variable to consider the variability. This approach is used by Diederichs and Kaiser 1996 with the
Stability Chart to quantify the stability interval and perform a risk analysis. It is also used by Idris and
Nordlund 2019 to consider the effect of the variability in the geomechanical variables for a numerical model and
determine the optimal stope dimensions. However, this probabilistic approach relies on the distribution of the
variable values for each observation to be known. It also does not consider the variance of the stope OB and
UB that is attached to each prediction when building the model or chart. For example, with the Stability Chart,
not all faces used for defining the unstable category are unstable. Some faces are stable or caving. This variance
impacts the likelihood of the prediction, but is rarely considered.

223
 8.4.3. Economic analysis
Stope profitability can be complex to assess as it will depend on different mining costs, rate of mining, amount
of metal in the stope, sale price, optimal smelter feedstock, ore recovery, fragmentation, dilution, loss of ore and
more. To simplify the stope design and calculate its profitability, mines will simplify part of the economic analysis
by determining the cut-off grade, and, when possible, the NSR.

The cut-off grade is the minimum grade of ore that is economically viable to mine. It is calculated, depending on
the metal, as a percentage or grams per tonne and can be expressed as dollars per tonne if multiple metals are
mined. The cut-off grade represents the limit between ore and waste. The calculation of the cut-off grade will
depend on different factors, such as recovery, sale price, mining costs, processing cost, capacity of the mill and
refinery, discount rate and so forth (Lane 1988). Some of these factors will vary spatially and through time, thus
the cut-off grade will also vary.

The NSR represents the estimated value of the ore, calculated in $/tonne, and is based on the sale price of the
metals minus the transport and treatment costs at the mill. The NSR considers all the metals being mined and
is calculated on a point basis along with the grades of all the metals in that location. The grade and NSR values
can be presented as a block model and used with the cut-off grade to determine the limits for the design of the
stopes (Wang and Webber 2012). Since OB and UB inevitability occur when mining a stope, the effect of dilution
and loss of ore must be considered to assess the profitability of the stope. The loss of ore represents the amount
of ore left in the stope. The dilution represents the amount of waste that is mixed with the ore (Equation 8.1,
Pakalnis 1986). Dilution can be planned (within the design) or unplanned (from OB) and will incur direct (blasting,
mucking, crushing etc.) and indirect (metal recoveries) costs (Villaescusa 2014). The level of tolerance to dilution
and loss of ore will depend on the grade and quantity of ore in the stope, OB and UB, and will vary from mine to
mine. For example, dilution for a mine with a disseminated orebody will still contain metals and will not influence
the overall grade to the same degree as a mine with a distinct separation between ore and host rock.

𝑊𝑎𝑠𝑡𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑

𝐷𝑖𝑙𝑢𝑡𝑖𝑜𝑛 =
𝑂𝑟𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑

Or [8.1]

𝑊𝑎𝑠𝑡𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑

𝐷𝑖𝑙𝑢𝑡𝑖𝑜𝑛 =
𝑂𝑟𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑 + 𝑤𝑎𝑠𝑡𝑒 𝑡𝑜𝑛𝑛𝑒𝑠 𝑚𝑖𝑛𝑒𝑑

Dilution is managed through strategic (mining method chosen, stope design) and tactical (cable bolt, blasting,
backfill, draw control) controls (Potvin and Wesseloo 2022). If dilution and ore loss can be predicted, it is possible
to be proactive with these management methods. Predicting the grade value at the design stage using current

224
design tools is limited as an average value (ELOS) or a general stability level (stable, unstable or caving) for a
stope surface is predicted while UB is excluded. Dilution and ore loss can be estimated if the ELOS or ELLO is
predicted, but only if there is a clear boundary between ore and waste, with the design surface falling on this
boundary. Therefore, the average prediction per face inhibits the quantification of the grade value of the OB and
UB as well as operational optimisation. Predictions at a metre-scale resolution would, however, enable the
quantification of the grade value. The next section will detail the methodology for predicting and assessing stope
performance at a metre-scale resolution.

8.5. Methodology
In this article, a new stope design approach is proposed that uses reconciliation data and prediction at a fine
resolution to predict and assess stope performance (OB, UB and grade). This approach enables one to work
towards a method to optimise stope design. This article will demonstrate this new stope design methodology
using a case study. The prediction assessment process is described in four steps (Figure 8. 4).

Step 1: Predict Step 2: Build the Step 3: Stope Step 4: Root

the stope probabilistic performance cause
geometry surfaces assessment (OB, assessment
• Predict the stope • Calculate the UB and grade) • Identify critical
geometry using the standard error • Assess predicted OB locations of OB and
critical variables per • Determine the and UB and their UB
octree and a machine prediction interval grade value • Identify the critical
learning model • Build the predicted • Use a probabilistic factors
surfaces approach to assess
the possible range of
stope geometries

Figure 8. 4: The proposed four steps for assessing the predicted stope performance as part of the stope design methodology

8.5.1. Octree data

The data used in this new stope design approach is quantified using an octree data structure at a metre-scale
resolution. The two main reasons are the increased spatial information and parametric possibilities enabled by

225
this resolution. Since information is quantified on a point basis along the design surface and at a metre-scale
resolution, the stope OB and UB spatial variation across the stope design is characterised. An example is given
in Figure 8. 5 where the variation from OB in the left to the UB on the right of the HW is characterised through
thousands of georeferenced octrees.

Figure 8. 5: An example of how stope OB and UB can spatially vary along a stope surface (McFadyen et al. 2021)

Variables are quantified for root cause analysis for each octree location. Given the resolution, geometrical,
operational, geomechanical and geological variables can be quantified, enabling an understanding of why there
is significant OB on the left of the HW and UB on the right in Figure 8. 5. The increased resolution also enables
one to quantify variables that couldn’t be quantified before at a per face resolution; therefore, offering a wider
range of variables as well as the possibility to develop a predictive model for optimisation. The list of variables
considered in the new stope design approach have been presented in McFadyen et al. 2023a. The variables
are quantified using the scripts developed in the mXrap (Harris and Wesseloo 2015) stope reconciliation and
analysis app (Woodward and Harris 2019).

8.5.2. Predictive model

A machine learning model is used to predict stope geometry. Given the use of octree data, the magnitude of OB
or UB (projected distance) is predicted on a point basis enabling the prediction of the expected mined geometry
(Figure 8. 6). The statistical model used for predicting the stope geometry is random forest (Breiman 2001). The
statistical approach is selected based on the work done by McFadyen et al. 2023b, showing strong prediction
performance at a metre-scale resolution. Random forest is a tree-based non-linear approach that can consider
a complex dataset and predict with high accuracy. Random forest enables the construction of an empirical

226
predictive model using the quantified critical variables as well as being adaptable to each mine site. The
predictive models are built using the Ranger script (Wright and Ziegler 2017) in the R software (R Core
Team 2021). The predicted geometry is built using the Poisson surface reconstruction process (Kazhdan et al.
2006) in CloudCompare (Girardeau-Montaut 2022).

Figure 8. 6: Example of a CMS (left) and the predicted mined geometry (right) using a random forest model

Probabilistic stope geometry prediction

The use of a random forest predictive model enables to building a probabilistic design approach. A prediction
obtained using random forest represents the average predicted value from all the trees used in the model (Figure
8. 7). The standard error is calculated from these trees using the infinitesimal jackknife approach (Wager et al.
2014 by omitting observations used and rebuilding the trees to determine the new average predicted value. This
process is repeated, and the variance in the average predicted value is calculated. The standard error is then
used to calculate the probability density function (PDF) and the cumulative density function (CDF) for the
predicted octree, allowing to determine a prediction interval within which the projected distance is expected.
Different geometries can be outputted using the different percentiles of the prediction interval (Figure 8. 8). The
probability of observing a specific projected distance can be determined using the PDF, and the probabilities of
exceeding a given projected distance can be determined using the CDF.

227
 Figure 8. 7: Random forest diagram for generating a prediction

Figure 8. 8: Example of different prediction geometries for a stope design shape (black) according to probabilities. The blue, green and
red surfaces are the 20th, 50th and 80th percentile cases of the prediction interval, respectively

228
 8.5.3. Stope performance assessment
Once the predictions are obtained, the predicted projected distances are analysed to understand where and to
what probability OB and UB will occur in the stope, along with the associated economic impact (revenue in
regard to a cut-off grade). The stope performance assessment is detailed as follows.

Geometry assessment
The octree resolution of the predictions enables the assessment of the areas where OB and UB are expected,
and using the standard error calculated for each predicted octree, the probability at each octree location of
exceeding a given tolerated threshold of OB or UB. The tolerance for OB and UB will vary from one mine site to
another as it depends on the geological and mining environment, which dictates how the OB and UB impact the
profitability and stability of the stope and subsequent stopes. The predicted OB and UB are assessed by first
plotting the octrees in a 3D view. The octrees are coloured using markers, showing the predicted distance or the
probability of exceeding a given threshold of OB or UB. This allows one to determine the areas of concern and
the probability for critical OB or UB. The percentage of the design surface that is possibly exceeding a given
threshold is calculated and analysed by plotting a cumulative chart of the proportion of octrees exceeding the
threshold at each percentile of the CDF.

The CDF is also used to analyse the prediction interval for the final mined geometry. For example, three different
geometries (20th, 50th and 80th percentile of the CDF) are generated by calculating the projected distance of the
octrees for these percentiles, giving a prediction interval. This is presented for one octree in Figure 8. 9. Since
it’s a cumulative function, each percentile represents a probability of observing a projected distance smaller or
equal to the distance calculated for that percentile. Therefore, if at the 20th percentile mark, the projected distance
for an octree is 1.2 m, there is a 20% chance the actual distance will be smaller and a 80% chance it will be
larger. For the example in Figure 8. 9, there is a 60% probability (prediction interval between 20% and 80%) that
the projected distance for the given octree will be between 1.2 m and 2.8 m. The 50th percentile represents the
geometry generated by the model (mean case) and is used to determine where the OB and UB will most likely
occur. To evaluate worst-case scenarios for OB, percentiles over 50% are analysed. The same is done inversely
for UB by looking at probabilities under 50%. The different geometries generated using the probabilities are then
reconciled with the stope design to calculate the predicted OB and UB volumes and perform an economic
assessment.

229
 Figure 8. 9: Example of the predicted distance for an octree based on the CDF and percentile considered

Economic assessment
The NSR designates the value of the ore after milling and smelting, and is often used as the economic variable
for designing the stope as it encompasses all the contained metals. Therefore, the NSR value is used for the
present analysis. Figure 8. 10 shows an example block model with the NSR for each economic mineral as well
as the NSR given per block.

Figure 8. 10: Example of the grade block model imported around a stope displaying the NSR value per block

230
The block model data is spatially associated to the octree data structure to calculate the grade value in the OB,
UB and other volumes. The weighted average NSR value and the total metal in the octree block are calculated
from the grade blocks that intersect the octree block (Figure 8. 11). If backfill information is available, the grade
value for the octree which contains fill is adjusted. This information is assessed by plotting the data in 3D view,
identifying the areas of low value or high value, and comparing with the probabilities of OB and UB. Since each
octree is associated to one or more volume type (OB, UB, design, mined and recovered) the average and total
NSR is also summed up per face and per stope for the different volume types. This information is calculated for
the different probabilities to assess and compare the economic impact of the OB and UB on the profitability of
the stope.

Figure 8. 11: Example of the calculation of the NSR value per octree (left) and an example of the result per octree block of the NSR
value adjusted with backfill information (right)

The calculation of the grade value at an octree resolution also means that a cumulative distribution of the NSR
value for each tonne in the stope or volume type is calculated (Figure 8. 12). As a footnote, the total tonnage for
every category is indicated to show the relative importance of the cumulative volume. This chart is used to
compare the distribution of the grade, or in this case the NSR, in each volume (design, mined, OB and UB) and
determine if the predicted geometry follows the design NSR tonne distribution. It is also used to calculate the
number of tonnes under the cut-off grade for each volume type. The economic value of the OB and UB is also
assessed, evaluating the proportion of waste and ore to determine how the OB and UB impact the stope dilution
and loss of ore.

231
Figure 8. 12: Example of the cumulative distribution of the NSR value for each tonne based on the different categories of volume. Total
tonnes per volume are indicated as a footnote and the cut-off value with the black line

These tonnes are also queried using a selection tool on the graph and plotted in 3D view to identify their spatial
locations. Figure 8. 13 is an example of a selection of OB octrees plotted in 3D and coloured using their NSR
value. Octrees with an NSR value bellow cut-off are semitransparent to highlight the areas where dilution is
expected to occur.

232
 Figure 8. 13: 3D view of the NSR value for the OB octrees. Points with an NSR value below the selected cut-off grade are
semitransparent while points above the cut-off grade are opaque

The dilution and loss of ore are calculated using the cut-off grade and are used to quantify the profitability of the
stope. There are different ways of calculating the dilution based on the literature and from what is seen on mine
sites. The measures implemented are listed in Table 8. 2. These values are compared between the different
predicted geometries to assess the impact of OB and UB on dilution.

Table 8. 2: Different dilution and loss of ore variables

Variables Formula

Predicted dilution (Pakalnis 1986) 𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑

𝑂𝑟𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑

Planned dilution (Scoble and Moss 1994) 𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

Unplanned dilution (Scoble and Moss 1994) 𝑊𝑎𝑠𝑡𝑒 𝑖𝑛 𝑂𝐵

𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

Loss of ore (Villaescusa 2014) 𝑂𝑟𝑒 𝑖𝑛 𝑈𝐵

𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

Recovered ore 𝑂𝑟𝑒 𝑖𝑛 𝑚𝑖𝑛𝑒𝑑

𝑂𝑟𝑒 𝑖𝑛 𝑑𝑒𝑠𝑖𝑔𝑛

233
 8.5.4. Root cause assessment
Once the prediction for the stope is generated and the economic information is compiled, different geometries
are compared to assess the expected performance. If the predicted stope performance is satisfactory
(economically and geometrically) and the probabilities for critical OB or UB are within the mine site’s risk
tolerance, engineers can proceed with the final planning of the stope. If concerns about the stability and
profitability are flagged during the stope performance assessment, a root cause assessment is done to identify
the factors influencing the predicted dilution or loss of ore from the OB and UB and determine if possible
modifications to the design are practically viable.

The areas of concern in the stope are highlighted through the stope performance assessment. The critical
variables for each area are analysed to identify the primary causes of OB or UB by plotting the critical variable
versus the projected distance in bivariate charts. The octrees from the previously mine stopes are also plotted
to compare how the current octrees correlate to OB or UB. The analysis is facilitated by plotting the octrees in a
3D view to visualise the area of concern and the variable value for that area. The variable importance of the
model is also used to direct this assessment as it identifies the general importance of each variable for prediction
using the model.

8.6. Case study

Prominent Hill is an underground copper deposit with gold and silver located in South Australia (Figure 8. 14).
The mining method is transversal stopes to extract the ore using a primary–secondary sequence. Stope
dimensions vary depending on the sector but are typically 20 m wide by 20 m long or 30 m by 30 m and up to
90 m high. The orebody reaches a depth of approximately 1 km.

234
 Figure 8. 14: Location of Prominent Hill mine, South Australia, Australia

8.6.1. Data overview

The stope database is comprised of 73 stopes mined between 2018 and 2022, inclusive (Figure 8. 15).
Reconciliation on a per stope, per face and per octree basis was conducted for each stope. Octrees associated
to drives are excluded as they represent existing voids, and backfill faces are excluded as well, due to their
different strength and stress conditions. This is to ensure the quality and consistency of the analysis. Of the 73
stopes, 55 stopes are transversal stopes and 18 are longitudinal stopes.

235
 Figure 8. 15: 3D section view of the Prominent Hill stope database

Economic, geological and operational data was made available by the mine site. The approach for selecting the
critical variables is presented in McFadyen et al. 2023a. The stopes, as well as a selection of critical
geomechanical, geometrical and operational variables (Table 8. 3) were used to build a random forest model. A
stope is used for each example, displaying the octree variable values according to the accompanying legend.
Explanations in this table discuss the underlying influences that these variables aim to characterise. These
variables require the CMS, stope design geometry, drift surveys, drill rings and structural model to be available
for each stope. It is assumed and accepted that there is different levels of inherent uncertainty in the data arising
from the complexity of the underground geological and mining environment where the variation in the
geomechanical properties of the rock mass, as well as the variation between the mined and planned data, cannot
be fully captured. Furthermore, blasting is not an exact cutting tool compared to a tunnel-boring machine,
implying some uncertainty around the position of the final surface. The variability and uncertainty in the variables
is considered to a certain degree by incorporating a probabilistic approach.

236
Table 8. 3: Critical variables established through root cause analysis with a description (modified from Woodward et al. 2019 and
McFadyen et al. 2023a)

Variable Concept Interpretation Example

Blasthole Distance (coloured by Blast-induced

standoff metre) from an octree damage.
distance. block to the nearest
designed blasthole.
Characterises how
OB and UB are
affected by the Distribution of the distance for each octree
standoff distance to along the design surface (dots in the
the blastholes. image) to their nearest blasthole. Cold
colours mean the octrees are further away
from the blastholes and, generally, UB is
expected.

Blasthole Shortest angle Blast-induced

orientation. (coloured by degree) damage.
between an octree’s
design normal vector
and the blasthole. Characterises how
Convention: 90° = OB and UB is
toeing, 0° = parallel. affected by the
angle of the Distribution of the angle for each octree
blastholes to the along the design surface (dots in the
design surface. image) to their nearest blasthole. Warm
colours (90°) mean the blastholes are
toeing into the design surface.

237
Proxy for 𝐵𝑙𝑎𝑠𝑡 𝑒𝑛𝑒𝑟𝑔𝑦 Blast-induced
𝑛
blast energy. L damage.
≈∑
max( 0.5, 𝐷)2
0

where: Characterises how

L = length of hole. OB and UB are

affected by the Distribution of the energy proxy for each
D = distance to hole. density of blastholes octree along the design surface (dots in
and the drill pattern. the image). The energy calculation
Max. distance of
0.5 m to account for considers the density of blastholes around
stemming near collars the octree (inverse distance relationship)
(coloured by energy and the length of the blastholes to
index). determine the spatial distribution of the
blasting energy according to the blasthole
design. Warm colours mean more energy
is expected in that area due to the drill
design.

Dip of the The dip calculated Influence of gravity.

design from the normal vector
surface. associated with the
octree block (coloured Characterises how
by degree). OB and UB are
affected by the
Distribution of the dip for each octree
effect of gravity.
along the design surface (dots in the
image). Warm colours (90°) mean the
design face attached to the octree is
vertical.

238
Direction of The direction Influence of
the design calculated from the orientation with
surface. normal vector respect to stress
associated with the and rock mass
octree block (coloured properties.
by degree).

Characterises how Distribution of the direction for each octree

OB and UB are along the design surface (dots in the
affected by face image). The direction varies from 0° to
orientation. 360°.

Effective Calculated by Local stability

radius factor measuring, at regular conditions –
radial angle intervals, deformation of
(ERF).
the distance from a excavated span.
point located on the
selected surface of
the stope to the edge Characterise how
of that surface. OB and UB are
A maximum value is affected by the size
Distribution of the ERF for each octree
obtained towards the and geometry of the
along the design surface (dots in the
centre of the surface. faces.
image). Warm colours mean the octrees
are further away from the surface edges
and have higher stability concerns.

239
Undercut. The stope design as Local stress
well as the drifts are conditions.
‘wrapped’ in a convex
hull.
Loss of
Distance (coloured by
confinement.
metre) is found from
an octree block to the
nearest point on the
Characterise how
convex hull surface.
OB and UB are
affected by the
complexity of the
geometry and the Distribution of the undercut for each
cutting of the stope octree along the design surface (dots in
by the drives. the image). Warm colours indicate that,
due to the presence of the drives and/or
the shape of the stope, the octrees are
further away from the simplified shape that
In comparison with
wraps the stope and drives, and highlights
the stope geometry
areas where there can be a loss of
complexity variable,
confinement due to how the drives cut the
the undercut
stope.
considers the full
geometry of the void
and the impact the
drives have.

240
Distance to Distance to the Deconfinement and
drives. nearest drift in regard gravity effect on
to the octree’s design surface
position. areas near the
structures.

Damage induced
during development.

Distribution of the distance to the drives

Characterises how for each octree along the design surface
OB and UB are (dots in the image). Cold colours (0 m)
affected by the mean the octrees are on the drive or close
presence of drives. and warm colours indicate the octrees are
further away from the drives.

RQD. Value from the Characterisation of

nearest block of the how OB and UB are
geotechnical block affected by the rock
model mapped to the mass quality.
octree.

The block model is

built using borehole
data. The nearest
Distribution of the RQD for each octree
RQD value is mapped
along the design surface (dots in the
to the block with the
image). Warm colours will highlight areas
nearest distance
where there is poor rock mass quality.
varying between 0 m
and 30 m with 20%
within 5 m, 60% within

241
 10 m and 90% within
15 m.

The RQD is directly

logged from the core.

The list of variables presented in Table 3 is based on the root-cause analysis and the data made available by
the mine site. No geomechanical variables are included in this analysis, as no data other than an RQD block
model was available. Producing a geotechnical model at a 1m scale resolution for data currently available at
underground mining sites is challenging. If such data are incorporated into the analysis, great care must be taken
to ensure they are compatible with the other data. There are also currently no stress variables that have been
developed at an octree resolution.

8.6.2. Random forest model overview

For this case study, only the transversal stopes will be used. Of the 55 stopes, four stopes were excluded from
the model due to quality concerns (missing drill rings and stope tunnelling into adjacent backfill). A first model
was built to evaluate the prediction performance of the random forest model for this dataset. 75% of the stopes
were used to build the model and 25% for evaluating the model’s performance (Figure 8. 16). The model has an
ACC of 66% (proportion of octrees correctly classified) for classifying an octree as OB or UB and, using the
mean predicted case, has a projected distance error smaller than half a metre 33% of the time, smaller than a
metre 58% of the time and smaller than 2 m 83% of the time. This statistical performance matches what was
observed by McFadyen et al. 2023b at other mining sites and it is, therefore, determined that random forest can
be used for this dataset.

242
 Figure 8. 16: Prominent Hill first model and prediction database

After evaluating the random forest model performance for this dataset, a final model was built from 40 of the
stopes in the database that were mined before the stope used for this case study (Figure 8. 17). Octrees were
excluded (drives, muckpile, backfill and floor) to ensure the quality of the model.

Figure 8. 17: Stopes used for the Prominent Hill model database

243
 8.7. Results
Stope A is used as an example to demonstrate the stope’s performance and root cause assessment (Figure 8.
18a). The octree data structure was built at a 1 m resolution and six faces were delimitated (Figure 8. 18b). The
critical variables presented in Table 8. 3 are computed for each octree. Octrees representing drives, as well as
the floor, were filtered out for the predictions. These octrees will be attributed a prediction of 0 m after the
predictions are generated to maintain the point density across the surface for building the predicted mined
surface.

(a) (b)
Figure 8. 18: 3D view of Prominent Hill stope A; a) The design geometry with the drives; b) The octree data structure coloured per face

8.7.1. Geometry analysis

The predictions were generated for stope A and surfaces were generated for three different geometries based
on the CDF percentile: the 20th, 50th and 80th percentile (Figure 8. 19). The 20% and 80% cases were chosen
arbitrarily for this example and could vary for a mine site based on their preferences.

244
 20th percentile 50th percentile 80th percentile

Figure 8. 19: The three predicted geometries (20th, 50th and 80th percentiles of the prediction interval) for Prominent Hill stope A

Per stope analysis

The predicted surface built from these cases is first used to determine the global stability of the stope by
quantifying OB and UB on a per stope or per face basis (Table 8. 4). The volume difference, recovered volume
and compliance to design are calculated according to Equations 8.2 to 8.4. The volume difference and
compliance to design are based on McFadyen 2020 and Woodward et al. 2019. For comparison, the mined
results obtained from the CMS are also included. The prediction interval gives a prediction volume with a volume
difference with the design between -23% and +16%. For the predictions, the OB could vary between 316 m³ and
5 995 m³ and the UB could vary between 7 847 m³ and 913 m³. Overall, the prediction interval encompasses
the mined volumes with the 50th percentile being the scenario that matches best the mined volume with a 3%
difference.
Table 8. 4: Prominent Hill predicted per stope summary for the geometries’ 20th, 50th and 80th percentiles of the prediction interval

20% 50% 80% Mined

Designed volume 32 212 32 212
(m³)
Predicted volume 24 680 30 451 37 295 29 626
(m³)
Volume difference -23 -5 16 -8
(%)
OB (m³) 316 1 706 5 995 1 817
UB (m³) 7 847 3 467 913 4 400
Recovered (m³) 24 365 28 745 31 299 27 812
Compliance to 75 85 82 82
design (%)

245
 (𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑−𝑑𝑒𝑠𝑖𝑔𝑛)
𝑉𝑜𝑙𝑢𝑚𝑒 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑐𝑒 = 𝑑𝑒𝑠𝑖𝑔𝑛
∗ 100 [8.2]
𝑅𝑒𝑐𝑜𝑣𝑒𝑟𝑒𝑑 = 𝐷𝑒𝑠𝑖𝑔𝑛 − 𝑈𝐵 [8.3]
𝐷𝑒𝑠𝑖𝑔𝑛−𝑈𝐵
𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒 𝑡𝑜 𝑑𝑒𝑠𝑖𝑔𝑛 = ∗ 100 [8.4]
𝐷𝑒𝑠𝑖𝑔𝑛+𝑂𝐵

Per octree analysis

The stope stability can then be assessed at a finer resolution, looking at the projected distance per octree (Figure
8. 20). Analysing the results:
• The mean case predicts OB for the majority of the HW and FW with UB occurring more in the corners
and on the left side of the sidewalls.
• Potential large OB (over 3 m) is seen in the FW along the middle.
• The side walls present probable large UB (over 3 m).
20th percentile 50th percentile 80th percentile

Figure 8. 20: The three predicted geometries per octree (20th, 50th and 80th percentiles of the prediction interval) for Prominent Hill
stope A

Probabilistic analysis
The probabilities are also used to determine the chance of exceeding a given threshold. For example, in Figure
8. 21, a 0 m (separation of OB and UB) and 2 m OB threshold are used. If we look at the area around the drives
in the HW (circled in black), there is a 60% to 70% chance of observing OB, and a 20% chance that it will exceed

246
2 m. A cumulative distribution chart can be used to determine the percentage of octrees (equivalent to the portion
of the surface) below or above the threshold for each probabilistic scenario. For example, the portion of the
surface that would be overbroken (below the curve) for the mean case (probability of 50) is 33%, meaning there
is a 50% chance of observing 33% or more OB for the surface. This surface area drops to 6% for OB over 2 m
for the same probability.
0 m threshold 2 m threshold

Figure 8. 21: 3D view of the probability of exceeding a 0 m or 2 m threshold, and a cumulative distribution chart of the percentage of
octrees above the threshold value for the probability geometry (CDF percentile) for Prominent Hill stope A

8.7.2. Economic analysis

Following the same process as the stability analysis, the economic analysis is first done on a per stope or per
face basis to assess the global economic performance of the stope. With the geological block model, an NSR
value is attributed to each octree and, for this example, a fictitious but plausible cut-off grade of $50/tonne is
used.

247
Per stope analysis
The dilution and loss of ore summary for the predictions and mined geometry of stope A is presented in Table
8. 5. The geometry that best matches the design and the mined geometry of the three is the median case with
the main difference being the predicted dilution (30% compared to a mined dilution of 19%). The predicted
dilution is 12% higher than the planned dilution with only 11% loss of ore and a predicted recovery of ore of 86%.
The predicted dilution interval varies between 22% and 33% (over-predicting the actual dilution) with the
unplanned dilution varying between 1% and 8%. The loss of ore varies between 25% and 3%, and the expected
ore recovery is expected to vary between 73% and 104%, encompassing the amount actually mined.
Table 8. 5: Dilution and loss of ore summary for the three different geometries of the prediction interval and the mined geometry for
Prominent Hill stope A

Measure 20% 50% 80% Mined Equation

Wastedesign
Planned dilution (%) 18 18 ∗ 100
Oredesign
Wastepredicted
Predicted dilution (%) 22 30 33 19 ∗ 100
Orepredicted
WasteOB
Unplanned dilution (%) 1 3 8 1 Orepredicted
∗ 100
OreUB
Loss of ore in UB (%) 25 11 3 13 ∗ 100
Oredesign
Orepredicted
Predicted recovered ore (%) 73 86 104 92 ∗ 100
Oredesign

The impact of the dilution and loss of ore on stope profit can also be computed as the NSR value is available.
The NSR value per stope and per face are shown in Table 8. 6 and Table 8. 7. Profit is calculated according to
Equation 8.5.
𝑃𝑟𝑜𝑓𝑖𝑡 = (Average NSR per tonne – cut-off) * Tonnes [8.5]
If mining goes according to the design, the stope should generate a profit of $8 420 969 with an average NSR
of $123.1/tonne. The model predicts a profit between $6 776 073 and $8 828 682 with an average NSR between
$127 and $116/tonne according to the 20th and 80th percentile with the mean predicted profit being $7 790 844
at $121/tonne. The loss of profit can be associated to the dilution in the FW where the average NSR is around
$44–46/tonne, which is below the $50 cut-off, and the loss of ore in the UB which could vary between $169 031
and $1 647 529. However, UB in the FW and HW only represent 18% to 20% of the loss of profit, indicating that
the majority could be recuperated with the adjacent stopes. Extra ore can be recuperated in the HW as the
average NSR in its OB varies between $74 and $97/tonne.

After mining, the stope generated a profit of $7 996 580 at $125/tonne which closely matches the 50th percentile.
$748,025 is lost in the UB with 25% coming from the FW and HW, reducing the amount of ore that can be

248
recuperated from the adjacent stopes. The OB and UB in the FW is closer to the 20th percentile with dilution in
the FW as predicted.
Table 8. 6: Per stope summary of the NSR values per volume type for the predictions and mined geometry of Prominent Hill stope A.
Profit is calculated according to Equation 8.5

20th percentile 50th percentile 80th percentile Mined

Average NSR 123.1

($/tonne)
Design Total ($) 14 182 410

Profit ($) 8 420 969

Average NSR 127 121 116 125

($/tonne)
Predicted
Total ($) 11 199 266 13 245 603 15 504 464 13 327 633

Profit ($) 6 776 073 7 790 844 8 828 682 7 996 580

Average NSR 52 63 77 103

($/tonne)
OB
Total ($) 61 391 393 416 1 653 413 671 663

Profit ($) 2 633 83 108 576 743 345 613

Average NSR 109 108 102 99

($/tonne)
UB
Total ($) 3 044 536 1 330 223 331 359 1 511 315

Profit ($) -1 647 529 -713 233 -169 031 -748 025

249
Table 8. 7: Per face summary of the NSR values for the FW and HW OB and UB for the predictions and mined geometry of Prominent
Hill stope A. Profit is calculated according to Equation 8.5

20th percentile 50th percentile 80th percentile Mined

Average NSR 44 46 45 49
($/tonne)
OB
FW Total ($) 45 435 215 672 479 536 50 629

Profit ($) -6 196 -18 296 -52 927 -1 033

Average NSR 97 77 74 79
($/tonne)
OB
HW Total ($) 1 081 45 373 348 666 165 490

Profit ($) 521 16 013 111 800 60 749

Average NSR 86 88 87 85
($/tonne)
UB
FW Total ($) 323 502 119 098 33 690 344 233

Profit ($) -134 320 -51 057 -14 238 -141 743

Average NSR 85 83 81 77
($/tonne)
UB
HW Total ($) 550 279 178 723 44 117 127 529

Profit ($) -224 670 -70 446 -16 904 -44 718

The quantification of the NSR per octree also allows one to quantify the distribution of the NSR value per tonne
for the design, mined, OB and UB volumes (Figure 8. 22).

There are 115 229 tonnes in the design. The NSR distribution for the 20th and 50th percentiles closely follows
the design. The 80th percentile has a higher proportion of tonnes below $100/tonne than the design. The
predicted total tonnes varies between 88 464 and 133 516, meaning fewer tonnes will most likely be extracted
than planned. The UB follows a similar distribution for the three percentiles with 90% to 95% being over the cut-
off limit but has less value than the tonnes in the predicted geometry. Waste in the OB varies between 40% and
65%. The total tonnes of OB represent, however, only a fraction of the predicted mined tonnes.

250
 20th percentile 50th percentile 80th percentile

Figure 8. 22: Cumulative distribution of the tonnes in each volume type based on the NSR value for the three different geometries (20th,
50th and 80th percentiles of the prediction interval) of Prominent Hill stope A. The mined volume represents the predicted mined

Per octree analysis

With the stope economic summary considered, the analysis is refined on an octree basis to identify where the
costs and profits are located on the stope. The octrees associated with the different volumes (design, predicted,
OB and UB) are shown in Figure 8. 23 for each predicted geometry and are coloured by NSR value. Octrees
with values under $50/tonne are semitransparent. Looking at the design, dilution concerns are in the left side of
the FW and part of the top of the HW. Loss of ore can be a concern for the rest of the stope, although the higher-
grade ore is mostly located in the side walls. This ore could be recuperated when mining the adjacent stopes for
the side walls. Looking at the other volumes and the different geometries, there is an 80% probability to generate
dilution in the FW (see OB for the 20th percentile) and leave ore in the corners of the faces (see UB for the 80th
percentile).

251
 20th percentile 50th percentile 80th percentile
Design

Predicted

OB

UB

Figure 8. 23: NSR value per octree and per volume for the three different geometries (20th, 50th and 80th percentiles of the prediction
interval) of Prominent Hill stope A. A cut-off value of $50/tonne is set. Octrees below the cut-off value are semitransparent

8.7.3. Root cause analysis

Given the predictions of the stope, the site’s risk and OB or UB tolerance, and the site’s economic goals (no ore
left behind or minimise dilution, for example), the engineer can determine if the potential outcome of the stope
does not meet the objectives of the stope, the respect of the design geometry and the profitability of the stope.
If the outcome is judged to be not meeting the design requirements, a root cause analysis can be done to
determine the key variables controlling the predicted performances. For stope A, if the goal is to minimise

252
dilution, dilution in the FW would be the main concern as the average NSR value is below the cut-off value. If
the loss of ore is the main concern, UB in the other faces could result in potentially millions of dollars’ worth of
ore being left behind, while OB would imply limited dilution and some ore recovery. When considering the
economics in isolation, it would be reasonable to aim to minimise UB and tolerate some additional OB. Additional
factors considered by the engineer may also impact to what degree this trade-off is made; for example, ore
fragmentation and the impact on the stability of adjacent stopes to be mined in the future.

Interpretation of the model

The first step in the root cause analysis, as described in Section 8.6, is to consider the importance of the variable
calculated by the model. When building the model, a value is attributed by the model to each variable based on
its importance in predicting stope performance. The order of variable importance is presented in Figure 8. 24.
From this order, it is interpreted that performance is determined by separating the different faces (dip and
orientation) as the side walls, HW, FW and crown perform differently. After which, the position of the drives
(undercut) and the distribution of the blasting energy will impact where the OB and UB will occur. Going down
the list, the size of the faces, the angle of the boreholes to the face, the gravity effect caused by the intersection
of the drives and the rock mass quality will decrease in impact.

Figure 8. 24: Variables ordered by variable importance calculated from the Prominent Hill random forest model

Visual analysis
Given the general interpretation of stope performance by the model, the analysis is extended, looking at the
stope in the 3D view and colouring the octrees with the different variables (Figure 8. 25) to determine, for specific
areas, which variable is critical. As determined before, if dilution were the main concern, one would want to limit
OB in the FW. Looking at Figure 8. 25, the following observation can be made:

253
 • The FW presents high energy (Figure 8. 25d) at the top and bottom of the face and a relaxation zone
due to undercut (Figure 8. 25j).
• There is also a lower RQD area (50% to 70%, Figure 8. 25k), compared to the rest of stope, at the
bottom of the face.
Potential optimisation could involve changing the charging against the FW and/or potentially some ground
support; although, in this case, the impact of ground support has not been analysed. The undercut issue would
remain.
If the loss of ore were the main concern, UB would need to be minimised in the other faces. Looking at Figure
8. 25, the following observations can be made:
• UB is more pronounced in the corners (Figure 8. 25a).
• UB in the corners is mostly associated to the drilling pattern (Figure 8. 25c and Figure 8. 25d).
Potential optimisation could be achieved by changing the drilling pattern.
This root cause analysis only considered the variables used in the model. Other factors could also have an
impact on stope performance and should not be excluded.

254
 Figure 8. 25: Multiple 3D views of the design octrees of Prominent Hill stope A coloured with the different critical variables

8.7.4. Exploring other designs

The model structure enables one to modify certain design variables and rerun the model to explore how stope
performance could fluctuate to work towards an optimal design that matches the mine operator’s goals. For this
case study, the design geometry or the drill rings could be altered as they are considered in the model, and
certain aspects could be modified in the long-term or short-term design. Stope A is used as an example again.
Given the disseminated ore in the HW and the possibility to recuperate the ore in the corners with the adjacent
stopes, the goal of the exercise will be to minimise dilution in the FW. Using the information from the root cause
analysis and the model, blast rings against the FW will be omitted to evaluate how the FW performance could
change. Figure 8. 26 shows the original blast design and the alternative blast design, as well as the variation of
the blasting energy proxy for the FW, which diminishes.

255
 Original design Alternative design

Figure 8. 26: Original and alternative blast designs proposed for optimisation analysis of Prominent Hill stope A

The alternative predictions are presented in Figure 8. 27. The probability of OB over 2 m diminished for the
majority with UB becoming more probable. Waste can be seen to increase in the UB while decreasing in the
OB.

256
Figure 8. 27: Original and alternative projected distance and economic predictions for optimisation analysis of Prominent Hill stope A

257
Table 8. 8 provides a comparison of the overall and FW predicted performance for the original and alternative
designs. The profit intervals are similar for the two scenarios, with the alternative design predicting a slightly
smaller profit for the 20th and 50th percentile. However, the average NSR increase for the alternative case,
reducing the interval from $11/tonne ($127 to $116/tonne for the original scenario) to $9/tonne ($130 to
$121/tonne for the alternative scenario).

The predicted OB volume is also reduced from a possible 6 000 m³ to 4 000 m³, increasing the predicted average
NSR in the FW OB from below the cut-off ($44 to $46/tonne) to over the cut-off ($53 to $60/tonne). The predicted
dilution interval drops by 9% from 22% to 33% to 13% to 24%. The average NSR in the FW UB drops from $86
to $88/tonne to $77 to $81/tonne, indicating more waste would be left behind. However, the predicted UB
increases by up to 1 500 m³ from the original scenario, and the loss of ore interval increases by 2%.

This alternative economic performance that implies omitting the drill ring close to the FW would reduce the
dilution overall while maintaining the economic performance of the stope, all while mining between 6% and 8%
less rock than the original predicted scenario. This would reduce the operation time dedicated to the stope and
fill material. Overall, the economic risk from dilution would be minimised as the predicted dilution interval drops
and the average NSR per tonne increases.

258
 Table 8. 8: Comparison of the predicted performance between the original and alternative design for Prominent Hill stope A

Design 20th percentile 50th percentile 80th percentile

Total value Original 11 199 266 13 245 603 15 504 464
($) Alternative 10 702 314 12 763 733 15 061 394
Original 6 776 073 7 790 844 8 828 682
Profit ($)
Alternative 6 560 348 7 687 677 8 812 804
Predicted Original 127 121 116
Average NSR
Alternative 130 126 121
($/tonne)
Predicted Original 73 86 104
recovered
Alternative 72 88 105
ore (%)
Predicted Original 22 30 33
dilution (%) Alternative 13 19 24
Original 316 1 706 5 995
OB (m³)
Alternative 32 498 3 982
FW avg NSR Original 44 46 45
($/tonne) in
Alternative 53 60 54
OB
Loss of ore in Original 25 11 3
UB (%) Alternative 28 13 4
Original 7 847 3 467 913
UB (m³)
Alternative 9 355 4 374 1 286
FW avg NSR Original 86 88 87
($/tonne) in
Alternative 81 77 80
UB
Compliance Original 75 85 82
to design (%) Alternative 71 85 85
Volume Original -23 -5 16
difference
Alternative -29 -12 8
(%)

This optimisation exercise was quick to complete (approximately 15 minutes for one stope) and showed that
limiting the energy proxy in the FW would help reduce the dilution and cost of waste processing from the FW.
Caution needs to be taken when optimising through this approach as it is an empirical approach and not
analytical. Therefore, proposing a new design that does not resemble the ones used in the model would mislead
the engineer in their assessment of the predicted stope performance. A good understanding of how the model
works and the mine site environment is needed. However, it remains that this proposed approach offers a new
innovative and powerful tool for predicting stope performance at the design stage, and a pathway for facilitating
optimisation at the design stage.

259
 8.8. Discussion and summary
The proposed methodology for predicting stope performance using machine learning (random forest) and data
at an octree resolution enabled a probabilistic prediction of a stope’s expected geometry to be developed. The
calculation of the standard error of each prediction allows one to evaluate the possibility of large OB or UB
occurring (prediction interval) and, given the resolution of the predictions, to know where it would occur. Contrary
to a single prediction, this approach allows one to determine the probability of OB or UB exceeding an acceptable
threshold and allows the engineer to integrate the economic impact in a stope performance analysis.

The octree data structure enables one to incorporate the grade block model for assessing the economic impact
of the OB and UB. The inclusion of the economic assessment also allows one to evaluate the impact the OB
and UB will have on dilution and loss of ore and, using the different geometries, how the increase and decrease
in OB and UB impact stope profitability. The economic assessment only considers the NSR value and the cut-
off limit. Therefore, only dilution, loss of ore, total value and average value can be determined. The impact of
mucking time, fragmentation and how the waste mixes with the ore is not considered in this analysis.

Existing methods for predicting stope performance only give a general stability assessment (stable, unstable or
caving) or an ELOS or ELLO value for each face depending on the method used. The predictions don’t allow
one to know where the OB and UB would occur in the face. They also limit the economic assessment. This
proposed methodology presents a big step forward in how the predictive tools will impact and improve future
stope performance. It provides a unique path towards integrating economic and stability assessment of stope
performance – all of which are at a metre-scale resolution, thus allowing to move away from averaged values.
While only a single case study is presented in this article, this methodology is meant to be flexible and applicable
to various mine sites. The type of deposit and the critical factors impacting the mine site stope performance can
affect the methodology's performance.

The knowledge obtained from the predicted stope performance assessment can be used by the engineer to
assess the expected performance of the stope and design, and consequently provide paths to optimise the
stope. Possible optimised scenarios can be generated using the model due to the flexibility of the model to
incorporate any variables, such as the operational variables calculated (drill and blast). Possible optimisation
will, however, depend on the critical factors identified through the root cause analysis and if they are used in the
model. Current optimisation through the model can remain limited, as some critical variables could not be
considered in the models at this stage but will improve with time as more variables are quantified on an octree
basis, thus improving the model at the same time. This approach enables one to rerun the predictions after
modifying certain variables to evaluate how they would impact the expected mined geometry.

260
 8.9. Conclusion
This article presented, through a case study, a new stope design approach that uses geometrical,
geomechanical and operational variables, as well as reconciliation data, measured at a metre-scale resolution,
to predict a stope’s OB and UB locations and magnitude on the design. Through the machine learning model
(random forest), a standard error was calculated for each prediction, allowing the generation of different possible
mined geometries and determine the probability of large OB and UB based on the mine site’s tolerance. The
economic performance was assessed by mapping the grade block model to the octree data structure to
determine the NSR value, the dilution and the loss of ore for the stope. This assessment was then used to
determine the root causes of the OB and UB based on the variables used in the model which can be used by
the engineer to determine the need and possible ways to optimise the design.

This new methodology works towards a full integration of an economic and stability analysis from predictions at
the design stage by enabling engineers to extract more information and precision from the predictions, and build
an understanding of the controlling factors of OB and UB and how they impact the profitability of a stope. This
knowledge is built from the site’s data and enables engineers to integrate operational data to facilitate the
optimisation of stope design and improve performance.

This is a significant leap forward in open stope design technology when compared to the qualitative stability
output of the stability graph or the very crude estimation of ELOS and ELLO.

8.10. Chapter conclusion

This chapter presented the economical and geotechnical approach for assessing predictions made at an octree
resolution through a case study. The proposed approach enabled, through a probabilistic approach, an
understanding of the expected magnitude of OB and UB and their spatial distribution while associating an
economic value to the OB and UB. This allows mine sites to get complete picture of the expected stope
performance and offers them the possibility to reassess their design, providing a path to work towards an optimal
stope design. This is the last step of the proposed approach for helping operational mines to understand and
predict their stope performance and shows that the methodology allows to work towards an optimal stope design,
all during the late stage of the stope design step. The conclusion of this thesis follows.

261
Conclusion
Summary
Optimising stope design can add significant value to open stoping mines. OB and UB have economic and
operational consequences that can influence the profitability and safety of a mine. To help limit these volumes
of OB or UB, empirical tools were created in the 1980s to predict OB. Although these tools have evolved with
time, they are limited by the variables considered, the resolution of the predictions (per face) and the exclusion
of UB. It is, therefore, essential to push the envelope by developing a new prediction method that can integrate
these limitations.

The main objective of this thesis is to develop a methodology that can be used during the final steps of the
design stage for understanding and predicting stope geometry at an octree resolution based on a multivariate
and probabilistic approach. This was achieved by building a stope database that integrated geometrical,
geomechanical, geological and operational variables. The database is then used to determine the critical
variables impacting stope OB and UB, and build the predictive approach. The octree resolution of the data also
enables one to integrate economic data, allowing engineers to assess the economic performance of the
predicted stope and work towards a methodology that helps optimise the stability and profitability of the stope.

Several researchers have already worked on the subject of open stope design and OB and UB analysis.
Chapter 1 provided an overview of the empirical prediction tools currently available and their limitations. The
chapter also covered the variables impacting OB and UB, the statistical methods used for assessing stope OB
and UB, and economic analysis.

Chapter 2 covered the proposed methodology for understanding and predicting stope geometry. The approach
is divided into four main steps: data collection, root cause analysis, predicting the stope geometry and the
stability and economic assessment of the predicted stope geometry. The mXrap software was used to quantify
the different variables, build the stope database and automate some of the analyses. Several variables make
up the database. The choices were justified and presented. Different statistical analysis and prediction models
were used. These were detailed and justified as well.

To test the proposed approach, data from three mine sites were collected. These mines sites were briefly
described in Chapter 3. The mines are Dugald River (MMG) and Prominent Hill (BHP) in Australia, and
Westwood (IAMGOLD) in Canada. These sites were selected for their different mining and geological conditions.

262
The chapter covered the mining process and geological properties that are necessary for assessing stope
performance.

The first results presented are those from the multivariate stope OB and UB analysis conducted for the Dugald
River mine site. These results were incorporated into an article that is presented in Chapter 4. OB and UB were
assessed using PCA and PLS and georeferenced data measured at an approximately cubic metre resolution
(octrees). Results showed that OB is observed in areas of the stope where there is a large blasting energy proxy
and large undercut, and towards the middle of the face (based on ERF). UB tends to be observed near the
border of the faces and where there is large blasting standoff distance and toeing of the rings into the wall.

The multivariate stope OB and UB analysis for the other two mine sites was presented in Chapter 5 as well as
a summary for all three mine sites. Overall, the root cause assessment at an octree resolution showed that the
face geometry (angle, shape and size), the undercutting and overcutting from the drives, the blasting pattern
and the presence of faults were key factors impacting OB and UB and where they occur on the stope surface.
The octree resolution of the data and the use of multivariate analysis have enabled the depiction of the variation
of stope OB and UB along the design surface, going beyond the simple per face analysis.

Results from the multivariate and machine learning predictions conducted for the Dugald River mine site were
incorporated into an article that is presented in Chapter 6. This article presented how stope OB and UB are
measured and georeferenced using octrees at an approximately cubic metre resolution and predicted using
multivariate and machine learning models (PLS, LDA and random forest). These predictions were used to build
the expected geometry of the open stope. Results show that the stope geometry is predicted excellently with a
high success rate using a random forest model (76% of the faces).

Prediction results and model performance from the two other sites were presented in Chapter 7 and used to
make a statement regarding the ability of the proposed methodology to meet the objective set in this thesis.
Overall, the random forest model performed better at predicting the stope geometry, enabling the prediction of
71% (on average) of the faces for the transversal stopes and 51% for the longitudinal stopes. Different criteria
were used for the different stope types to consider the impact the magnitude of OB or UB will have on different
stope sizes. In addition, the probabilistic approach enabled the estimation of the maximum OB and UB that can
occur and its location. After assessing the model performance, it was stated that the proposed methodology
enabled the prediction of the expected mined geometry. The octree resolution of the data and the use of
multivariate analysis and machine learning has enabled the prediction of the distribution of the stope OB and
UB along the design surface, going well beyond the simple qualitative per stope face prediction provided by a

263
traditional Stability Chart approach. The practical impact the prediction approach has on a mine site’s stope
design was also discussed.

The final results presented cover the stability and economic assessment of the predictions generated for the
Prominent Hill mine site. These results were incorporated into an article that is presented in Chapter 8. This
article presents how understanding and predicting stope OB and UB at a metre-scale resolution through a
machine learning model is used for assessing the geotechnical (OB and UB) and economic stope performance.
These predictions estimate the magnitude and location of OB and UB across the design surface, and are used
to resolve the expected geometry of the mined void. The incorporation of a predicted mined void with grade
values from a block model allows for a comprehensive economic stope performance estimate.

The work carried out and the results obtained demonstrate that the proposed methodology achieves the
objectives of this thesis. Critical variables impacting open stope OB and UB were identified using statistical
analysis. The expected mined geometry was predicted successfully using the random forest model. The metrics
set for quantifying the model performance showed the stope geometry can be predicted excellently (71% on
average). This work is a significant step towards a stope design and planning process that minimises the stability
issues, dilution and loss of ore across the stope. As designing for higher ore recovery increases the potential for
dilution and vice versa, this approach allows for the value realised from mining a stope to be maximised by
quantifying this inherent economic trade-off. The enhanced level of detail of both data and predictions has
allowed for a more comprehensive evaluation of the predicted stope performance and economic outcomes
during the design phase. This marks a notable advancement in the ability to design and plan the optimal stope.

Limitations of the thesis

While the proposed methodology allows to improve our comprehension of stope OB and UB and predict the
stope geometry, the work carried out in this thesis has certain limitations. These limitations are associated to the
database and the predictive models that were built:

• The octree data is sensitive to the survey quality. The data processing has shown that the design
geometry will sometimes have internal faces from shapes being merged or irregular edges. These
affect the projected distance direction and magnitude.
• The octree resolution is smaller than the geotechnical block model and is, therefore, sensitive to how
the geotechnical data is interpreted. Data collection varies for different areas of the mine site and so
the density or amount of data available per stope is not necessarily suitable for the finer octree
resolution. Great caution must be taken if geotechnical block model data want to be used at this stage
as it is not necessarily meant to be used for an even finer resolution. Furthermore, the interpretation of

264
 directional data such as RQD within a block model is still challenging and is the topic of ongoing
research.
• Stope reconciliation at a per octree resolution is currently only available with the mXrap software. While
this is not a limitation of the thesis, it is a limitation of how the proposed methodology can be applied at
mine sites. mXrap is not used at every mine site and, therefore, limits the accessibility of the prediction
approach.
• Variables characterising stress, ground support, mining sequence, seismicity were not a part of the
stope database. These factors are known to have an impact on stope performance and can be critical
to understanding stope performance. However, there are currently no variables at an octree resolution
characterising these factors, therefore, their impact on OB, UB and the model’s performance could not
be assessed. Stress is indirectly considered as the stope shapes have been determined using the
stability chart and indirect measures of stress such as undercut and ERF are used, but there are no
direct measures. Further research is needed to take them into account. The database used in this
thesis covers a large number of variables, but it is not complete. This is also represented in the
predictive models which can be improved. These factors are also presented in the next section.
• Variables characterising backfill were also not a part of the stope database. Therefore, the model can
only predict faces mined against rock, which limits the proportion of the stope geometry that can be
predicted for stopes mined later in the mining sequence.
• The model capacity at predicting the geometry for longitudinal stopes was limited (based on arbitrary
criteria). While improvement can be made with data for the mine site, there is a limit to the maximum
accuracy (minimal prediction error) the random forest model can have across a whole stope. This is in
part due to the variance in the data and the precision of the measures.
• The PLS models were built using only one component. To better evaluate the predictive performance
of the PLS models, additional components should have been explored when optimising the model
parameters.
• The proposed methodology is an empirical approach. Therefore, it is limited to the data used for building
the model. If extrapolation is to be made (variable values outside of the used range or changes in the
correlation structure between the variables), the predictions should be used with caution.

Recommendations for future work

To follow-up and build on the work carried out as part of this thesis, some recommendations for future work are
formulated and listed below. These recommendations aim to better characterise and analyse the performance
of open stope mining. They also aim to improve the predictive models developed.

265
• Continue to expand the octree database by integrating additional geomechanical, seismic and
operational variables. Factors such as stress, seismicity and ground support were not considered in
these analyses.
• Link numerical analysis outputs to the octree data structure to incorporate an assessment of the mining-
induced stress regime at the stope various mining stages.
• Improve the creation process of the geotechnical block model and resolution to consider the inherent
variability and directional nature of geotechnical data allow to properly integrate the block models at an
octree resolution.
• Further develop the integration of the impact of mining sequence into the models. The adjacent mined
stope’s OB and UB will have an impact on the stope OB and UB, but this is not considered in this thesis.
• Use the drilled blasting pattern, which considers deviations, for analysing the blasting impact. The
designed drill rings were used in this thesis. The actual drilled rings could help improve the predictions.
The use of the charging length instead of the whole drilled length could also improve the estimate of
the blasting energy distribution.
• Revise how the major structural variables are calculated to better capture more complex settings where
multiple structures are present.
• Build a backfill database for predicting stope OB and UB against backfilled walls.
• Explore the inclusion of a spatial variable in the model as the position between octrees is not considered
and could help improve the predictions.
• Continue to update the model’s stope database with the more recently mined stopes to ensure the
quality of the predictions.
• Explore tools for warning site users when there are differences in the data correlation structure between
the selected stope for predictions and the stopes used for building the predictive model in order to
determine if they are extrapolating.
• Explore other statistical methods for predicting the stope geometry (for example neural network) as not
all statistical methods were explored. It is possible other methods will perform better than random forest.

266
References
Adoko, A. C., F. Saadaari, D. Mireku-Gyimah and A. Imashev. 2022. A Feasibility Study on The Implementation
of Neural Network Classifiers for Open Stope Design. Geotechnical and Geological Engineering. 40:677-696.
[Link]

Barton, N., R. Lien and J. Lunde. 1974. Engineering classification of rock masses for the design of tunnel
support. Journal of the international society for rock mechanics. 6:189-236.

Bewick, R. P. and P. K. Kaiser. 2009. Numerical Assessment of Factor B in Mathews’ Method for Open Stope
Design. In. 3rd CANUS Rock Mechanics Symposium; Toronto, Canada.

Bieniawski, Z. T. 1973. Engineering classification of jointed rock masses. South African Institute of Civil
Engineers. 15(12).

Breiman, L. 2001. Random Forest. Machine Learning. 45(1):5-32. [Link]

Capes, G. W. 2009. Open stope hangingwall design based on general and detailed data collection in rock
masses with unfavourable hangingwall conditions [Doctor of Philosophy]. Saskatoon: Saskatchewan. 300.

Castro, L. A. M., R. P. Bewick and T. G. Carter. 2012. An overview of numerical modelling applied to deep
mining. In Luis Ribeiro e Sousa, Eurípedes Vargas Jr., M.M. Fernandes and R. Azevedo. Innovative Numerical
Modelling in Geomechanics. London: CRC Press; p. 474.

Cepuritis, P. M. 2011. The use of scale independent shape measures to assess open stope performance.
International Journal of Rock Mechanics and Mining Sciences. 48(7):1188-1195.
[Link]

Chambers, J. M., W. S. Cleveland, B. Kleiner and P. A. Tukey. 1983. Graphical method for data analysis New
York: Chapman and Hall/CRC. 410.

Clark, L. and R. Pakalnis. 1997. An empirical design approach for estimating unplanned dilution from open stope
hangingwalls and footwalls. In. 99th Annual AGM–CIM conference; Vancouver, BC: Canadian Institute of Mining,
Metallurgy and Petroleum.

Clark, L. 1998. Minimizing dilution in open stope mining with a focus on stope design and narrow vein longhole
blasting [master’s thesis]. Vancouver, Canada: University of British Columbia. 316.

Deere, D. U., A. J. Hendron, P. F.D. and E. J. Cording. 1967. Design of surface and near-surface construction
in rock. In: C. Fairhurst. Failure and Breakage of Rock; New York: Society of Mining Engineers of AIME. p. 237-
302.

Deere, D. U. and D. W. Deere. 1988. The rock quality designation (RQD) index in practice. In: L. Kirkaldie. Rock
classification systems for engineering purposes; Cincinnati, OH: ASTM international. p. 91-101.

Deswik. 2016. [Link] [Software]. 2016.2. Deswik Software Solutions Pty Ltd.

Diederichs, M. S. and P. K. Kaiser. 1996. Rock instability and risk analyses in open stope mine design. Canadian
Geotechnical Journal. 33(3):431-439. [Link]

267
Diederichs, M. S. and P. K. Kaiser. 1999. Tensile strength and abutment relaxation as failure control mechanisms
in underground excavations. International Journal of Rock Mechanics and Mining Sciences. 36(1):28.
[Link]

Dunn, K. 2023a. Principal Component Analysis. Kevin Dunn. Retrieved April 25, 2023, from
[Link]

Dunn, K. 2023b. Projection to Latent Structures. Kevin Dunn. Retrieved April 25, 2023, from
[Link]

Fisher, R. A. 1936. The Use of Multiple Measurements in Taxonomic Problems. Annals of Eugenics. 7(2):179-
188. [Link]

Girardeau-Montaut, D. 2022. CloudCompare [Software]. 2.11.3.

Grieco, N. and R. Dimitrakopoulos. 2007. Managing grade risk in stope design optimisation: probabilistic
mathematical programming model and application in sublevel stoping. Mining Technology. 116(2):49-57.
[Link]

Guido, S., M. Grenon and P. Germain. 2017. Stope performance assessment at the Goldcorp Eleonore mine
using bivariate analysis. In: T. S. A. I. o. M. a. Metallurgy. AfriRock 2017 Rock Mechanics for Africa; 2-7 October;
Cape Town, South Africa. p. 1107-1124.

Guido, S. and M. Grenon. 2018. Contributions to Geomechanical Stope Optimization at the Goldcorp Eleonore
Mine Using Statistical Analysis. In. 10th Asian Rock Mechanics Symposium; October 2018; Singapore: ISRM.

Hadjigeorgiou, J., J. G. Leclair and Y. Potvin. 1995. An update of the stability graph method for open stope
design. In. 97e CIM-AGM; Halifax, NS: CIM.

Harris, P. and J. Wesseloo. 2015. mXrap software. Version 5. T. A. C. f. Geomechanics. Perth, WA: The
University of Western Australia.

Hassell, R., E. Villaescusa and J. R. Player. 2015. Stope Blast Vibration Analysis at Dugald River Underground
Mine. In. 11th International Symposium on Rock Fragmentation by Blasting; Sydney, NSW: AusIMM. p. 16.

Hastie, T., R. Tibshirani and J. Friedman. 2017. The Elements of Statistical Learning: Data Mining, Inference,
and Prediction New York: Springer.

Henning, J. G. and H. S. Mitri. 2007. Numerical modelling of ore dilution in blasthole stoping. International Journal
of Rock Mechanics and Mining Sciences. 44(5):692-703. [Link]

Hudson, J. A. and J. P. Harrison. 1997. Engineering rock mechanics: an introduction to the principles Oxford:
Elsevier Science Ltd.

Hughes, R. 2011. Factors influencing overbreak in narrow vein longitudinal retreat mining [master’s thesis].
Montreal, QC: McGill University. 143.

Hutchinson, D. J. and M. S. Diederichs. 1996. Cablebolting in Underground Mines Richmond, BC: BiTech. 406.

Idris, M. A., D. Salang and E. Nordlund. 2011. Probabilistic analysis of open stope stability using numerical
modelling. Int. J. Mining and Mineral Engineering. 3(3):26. DOI: 10.1504/IJMME.2011.043849.

268
Idris, M. A. and E. Nordlund. 2019. Probabilistic-Based Stope Design Methodology for Complex Ore Body with
Rock Mass Property Variability. Journal of Mining Science. 55:743-750.

Kaiser, H. F. 1960. The Application of Electronic Computers to Factor Analysis. Educational and Psychological
Measurement. 20(1):141-151. [Link]

Kazhdan, M., M. Bolitho and H. Hoppe. 2006. Poisson Surface Reconstruction. In: A. S. Konrad Polthier.
Eurographics Symposium on Geometry Processing; Cagliari, Sardinia, Italy: ACM International Conference
Proceeding Series. p. 10.

Krige, D. G. 1951. A statistical approach to some basic mine valuation problems on the Witwatersrand. Journal
of the Southern African Institute of Mining and Metallurgy,. 52(6):119-139.

Lane, K. 1988. The Economic Definition of Ore: Cut-Off Grades in Theory and Practice Brisbane, Australia:
COMET Strategy Pty Ltd. 147.

Le Roux, P. J. 2015. Measurement and prediction of dilution in a gold mine operating with open stoping mining
methods [Doctor of Philosophy]. Johannesburg: Witwatersrand. 467.

Le, S., J. Josse and F. Husson. 2008. FactoMineR: An R Package for Multivariate Analysis. Journal of Statistical
Software. 25(1):1-18. [Link]

Martel, M. and K. Tremblay. 2023. Competency contrast modelling Using XRF Data to Identify Areas of High
Seismic Risk. CIM.1-9.

Matheron, G. 1963. Principles of geostatistics. Economic Geology. 58:1246-1266.

Mathews, K. E., E. Hoek, D. C. Wyllie and S. B. V. Stewart. 1981. Prediction of stable excavation spans for
mining at depths below 1000 meters in hard rock. Vancouver, BC: CANMET.

Matthews, B. W. 1975. Comparison of the predicted and observed secondary structure of T4 phage lysozyme.
Biochimica et Biophysica Acta (BBA) - Protein Structure. 405(2):442-451. [Link]
2795(75)90109-9.

Mawdesley, C., R. Trueman and W. J. Whiten. 2001. Extending the Mathews stability graph for open-stope
design. Institution of Mining and Metallurgy, Section A: Mining Technology.A27-A39.

McCulloch, W. and W. Pitts. 1943. A Logical Calculus of Ideas Immanent in Nervous Activity. Bulletin of
Mathematical Biophysics. 5(4):115-133.

McFadyen, B., M. Grenon, K. Woodward and P. Morissette. 2019a. Preliminary assessment of primary stope
performance for a seismically active mine. In. 53rd US Rock Mechanics/Geomechanics Symposium; 23-26 June;
New York: ARMA. p. 11.

McFadyen, B., K. Woodward and Y. Potvin. 2019b. Stope Reconciliation Report Dugald River mine. Perth, WA:
The Australian Centre for Geomechanics.

McFadyen, B. 2020. Détermination de facteurs contrôlant la stabilité des chantiers souterrains pour une mine
sujette à la sismicité [master’s thesis]. Quebec, QC: Université Laval. 146.

269
McFadyen, B., K. Woodward and Y. Potvin. 2020. A new stope reconciliation approach. In: J. Wesseloo.
Proceedings of the Second International Conference on Underground Mining Technology; Perth, WA: Australian
Centre for Geomechanics. p. 16.

McFadyen, B., K. Woodward and Y. Potvin. 2021. Open stope design; beyond the Stability Graph. Underground
Operators Conference 2021; Perth, Australia: AUSIMM. p. 13.

McFadyen, B., M. Grenon, K. Woodward and Y. Potvin. 2023a. Assessing stope performance using
georeferenced octrees and multivariate analysis. Mining Technology. 132(2):132-152.
[Link]

McFadyen, B., M. Grenon, K. Woodward and Y. Potvin. 2023b. Predicting open stope performance at an octree
resolution using multivariate models. Journal of the Southern African Institute of Mining and Metallurgy,.
123(6):309-320. [Link]

Mevik, B.-H. and R. Wehrens. 2007. The pls Package: Principal Component and Partial Least Squares
Regression in R. Journal of Statistical Software. 18(2):1-23.

Miller, F., Y. Potvin and D. Jacob. 1992. Laser measurement of open stope dilution. CIM bulletin. 85:96-102.

Milne, D. 1997. Underground design and deformation based on surface geometry [Doctor of Philosophy].
Vancouver, BC: University of British Columbia. 271.

Milne, D., R. Pakalnis, D. Grant and J. Sharma. 2004. Interpreting hanging wall deformation in mines.
International Journal of Rock Mechanics and Mining Sciences. 41(7):1139-1151.
[Link]

Nickson, S. D. 1992. Cable support guidelines for underground hard rock mine operations [master’s thesis].
Vancouver, BC: University of British Columbia. 223.

Pakalnis, R. T. 1986. Empirical stope design at the Ruttan mine, Sherritt Gordon mine LTD. [Doctor of
Philosophy]. Vancouver, BC: University of British Columbia. 290.

Pearson, K. 1901. LIII. On Lines and Planes of Closest Fit to Systems of Points in Space. The London,
Edinburgh, and Dublin Philosophical Magazine and Journal of Science. 2(11):559.

Potvin, Y. 1988. Empirical open stope design in Canada [Doctor of Philosophy]. Vancouver, BC: University of
British Columbia. 372.

Potvin, Y. and J. Hadjigeorgiou. 2001. The stability graph method for open-stope design. In W. A. Hustrulid and
R. L. Bullock. Underground mining methods, engineering fundamentals and international case studies. Littleton,
CO: Society for mining, metallurgy, and exploration, Inc.; p. 513-520.

Potvin, Y., D. Grant, G. Mungur, J. Wesseloo and Y. Kim. 2016. Practical Stope Reconciliation in Large-scale
Operations Part 2, Olympic Dam, South Australia. Seventh international conference & exhibition on mass mining;
9-11 Mai; Sydney,NSW: AUSIMM. p. 501-509.

Potvin, Y. 2020. Stope design. Perth, Australia. Australian Centre for Geomechanics. 5.

270
Potvin, Y., K. Woodward, B. McFadyen, I. Thin and D. Grant. 2020. Benchmarking of Stope Design and
Reconciliation Practices. In: J. Wesseloo. Proceedings of the Second International Conference on Underground
Mining Technology Perth, WA: Australian Centre for Geomechanics. p. 9.

Potvin, Y. and J. Wesseloo. 2022. Geotechnical design for underground metalliferous mines course: Stope
design. Perth, WA: Australian Centre for Geomechanics.

Pyle, D. 1999. Data Preparation for Data Mining San Francisco, CA: Morgan Kaufmann Publishers.

Qi, C., A. Fourie, X. Du and X. Tang. 2018. Prediction of open stope hangingwall stability using random forests.
Natural Hazards. 92(2):1179-1197. [Link]

R Core Team. 2021. R: A language and environment for statistical computing [Software]. R. F. f. S. Computing.
Vienna, Austria: R.

Sainsbury, B., D. Sainsbury and A. Vakili. 2015. Discrete analysis of open stope stability. In: Y. Potvin.
Underground Design Methods 2015; Perth, WA: Australian Centre for Geomechanics. p. 79-94.

Scoble, M. J. and A. Moss. 1994. Dilution in underground bulk mining: implications for production management.
Geological Society, London, Special Publications. 79(1):95-108.
[Link]

Sewnun, D., W. Joughin, M. Wanless and P. Mpunzi. 2019. The creation and application of a geotechnical block
model for an underground mining project. In: J. Hadjigeorgiou and M. Hudyma. Ground Support 2019:
Proceedings of the Ninth International Symposium on Ground Support in Mining and Underground Construction;
23-25 October 2019; Sudbury, ON: Australian Centre for Geomechanics. p. 479-492.

Shapka-Fels, T. and D. Elmo. 2022. Numerical Modelling Challenges in Rock Engineering with Special
Consideration of Open Pit to Underground Mine Interaction. geosciences. 12(199):1-23.
[Link]

Singh, B. and R. K. Goel. 1999. Rock Mass Rating (RMR). In. Rock mass classification: A practical approach in
civil engineering. Oxford: Elsevier; p. 282.

Steffen, O. K. H. 1997. Planning of open pit mines on a risk basis. The Journal of The South African Institute of
Mining and Metallurgy. 97(2):47-56.

Stewart, S. B. V. and W. W. Forysth. 1995. The Matthews Method for Open Stope Design. CIM bulletin.
88(992):45-53.

Student. 1908. The probable error of mean. Biometrika. 6(1):1-25.

Suorineni, F. T. 1998. Effects of faults and stress on open stope design [Doctor of Philosophy]. Waterloo, ON:
University of Waterloo. 370.

Suorineni, F. T., D. D. Tannant, P. K. Kaiser and M. B. Dusseault. 2001a. Incorporation of a Fault Factor into
the Stability Graph Method: Kidd Mine Case Studies. Mineral Resources Engineering. 10(1):3-37.
[Link]

271
Suorineni, F. T., P. K. Kaiser and D. D. Tannant. 2001b. Likelihood statistic for interpretation of the stability graph
for open stope design. International Journal of Rock Mechanics and Mining Sciences. 38(5):735-744.
[Link]

Suorineni, F. T. 2010. The stability graph after three decades in use: Experiences and the way forward.
International Journal of Mining, Reclamation and Environment. 24(4):307-339.
[Link]

Tholana, T. and C. Musingwini. 2022. A Probabilistic Block Economic Value Calculation Method for Use in Stope
Designs under Uncertainty. Minerals. 12(4):1-15. [Link]

Tremblay, V. 2020. Analyse discriminante. Université Laval: Véronique Tremblay.

Truman, R., P. Mikula, C. Mawdesley and N. Harries. 2000. Experience in Australia with the application of the
Mathews' method for open stope design. CIM bulletin. 93(1036).

Vallejos, J. A., A. Delonca, J. Fuenzalida and L. Burgos. 2016. Statistical analysis of the stability number
adjustment factors and implications for underground mine design. International Journal of Rock Mechanics and
Mining Sciences. 87:104-112. [Link]

Venables, W. N. and B. D. Ripley. 2002. Modern Applied Statistics with S. New York: Springer.

Villaescusa, E. 2014. Geotechnical Design for Sublevel Open Stoping Boca Raton, FL: CRC Press.

Wager, S., T. Hastie and B. Efron. 2014. Confidence Intervals for Random Forests: The Jackknife and the
Infinitesimal Jackknife. Journal of Machine Learning Research. 15(1):1625-1651.

Wang, H. and T. Webber. 2012. Practical semiautomatic stope design and cutoff grade calculation method.
Mining Engineering. 64(10):85-91.

Wang, J. 2004. Influence of stress, undercutting, blasting and time on open stope stability and dilution [Doctor
of Philosophy]. Saskatoon: Saskatchewan, SK. 303.

Wang, J., D. Milne, L. Wegner and M. Reeves. 2007. Numerical evaluation of the effects of stress and excavation
surface geometry on the zone of relaxation around open stope hanging walls. International Journal of Rock
Mechanics and Mining Sciences. 44(2):289-298. [Link]

Wold, H. 1966. Estimation of principal components and related models by iterative least squares. Multivariate
Analysis.391-420.

Woodward, K., B. McFadyen, Y. Potvin and J. Wesseloo. 2019. Probabilistic Stope Design. MRIWA: Australian
Centre for Geomechanics.

Woodward, K. and P. Harris. 2019. Stope Reconciliation and analysis. Version 2. T. A. C. f. Geomechanics.
mXrap Software app. Perth, WA: The University of Western Australia.

Woolmer, D., E. Jones, J. Taylor, C. Baylis and D. Kewe. 2020. Use of Drone based Lidar technology at Olympic
Dam Mine and initial Technical Applications. In: F. B. R. Castro and K. Suzuki. MassMin 2020: Proceedings of
the Eighth International Conference & Exhibition on Mass Mining; 9-11 December; Online: University of Chile.
p. 565-582.

272
Wright, M. N. and A. Ziegler. 2017. ranger: A Fast Implementation of Random Forests for High Dimensional
Data in C++ and R. Journal of Statistical Software. 77(1):1-17. [Link]

273
Appendix A: Dugald River PCA analysis charts

a b
Figure A. 1: Component 1 versus component 2 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 2: Component 1 versus component 3 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

274
 a b
Figure A. 3: Component 1 versus component 4 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 4: Component 1 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

275
 a b
Figure A. 5: Component 1 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 6: Component 2 versus component 3 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

276
 a b
Figure A. 7: Component 2 versus component 4 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 8: Component 2 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

277
 a b
Figure A. 9: Component 2 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 10: Component 3 versus component 4 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

278
 a b
Figure A. 11: Component 3 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 12: Component 3 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

279
 a b
Figure A. 13: Component 4 versus component 5 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure A. 14: Component 4 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

280
 a b
Figure A. 15: Component 5 versus component 6 PCA charts for Dugald River. a: loading chart; b: score chart coloured by stope OB or
UB

281
Appendix B: Prominent Hill PCA analysis charts

a b
Figure B. 1: Component 1 versus component 2 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 2: Component 1 versus component 3 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

282
 a b
Figure B. 3: Component 1 versus component 4 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 4: Component 1 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

283
 a b
Figure B. 5: Component 1 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 6: Component 2 versus component 3 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

284
 a b
Figure B. 7: Component 2 versus component 4 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 8: Component 2 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

285
 a b
Figure B. 9: Component 2 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 10: Component 3 versus component 4 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

286
 a b
Figure B. 11: Component 3 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 12: Component 3 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

287
 a b
Figure B. 13: Component 4 versus component 5 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

a b
Figure B. 14: Component 4 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

288
 a b
Figure B. 15: Component 5 versus component 6 PCA charts for Prominent Hill. a: loading chart; b: score chart coloured by stope OB or
UB

289
Appendix C: Westwood PCA analysis charts

a b c
Figure C. 1: Component 1 versus component 2 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

a b c
Figure C. 2: Component 1 versus component 3 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

a b c
Figure C. 3: Component 1 versus component 4 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

290
 a b c
Figure C. 4: Component 1 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

a b c
Figure C. 5: Component 1 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

a b c
Figure C. 6: Component 1 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

291
 a b c
Figure C. 7: Component 1 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

a b c
Figure C. 8: Component 2 versus component 3 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

a b c
Figure C. 9: Component 2 versus component 4 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c: score
chart coloured by binned projected distance

292
 a b c
Figure C. 10: Component 2 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 11: Component 2 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 12: Component 2 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

293
 a b c
Figure C. 13: Component 2 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 14: Component 3 versus component 4 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 15: Component 3 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

294
 a b c
Figure C. 16: Component 3 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 17: Component 3 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 18: Component 3 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

295
 a b c
Figure C. 19: Component 4 versus component 5 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 20: Component 4 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 21: Component 4 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 22: Component 4 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

296
 a b c
Figure C. 23: Component 5 versus component 6 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 24: Component 5 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 25: Component 5 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 26: Component 6 versus component 7 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

297
 a b c
Figure C. 27: Component 6 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

a b c
Figure C. 28: Component 7 versus component 8 PCA charts for Westwood. a: loading chart; b: score chart coloured by OB-UB; c:
score chart coloured by binned projected distance

298
Appendix D: Prediction results for the Dugald River
PLS and LDA models
The 49 transversal stopes were separated into three groups chronologically. The first group contained 24 stopes,
the second 13 and the last 12. The models were built using data from the first group which consisted of 200 000
points, representing half the database. Since the final testing of the model is with the final group, only the results
for group 3 are presented. The stopes in group 3 are shown in Figure D. 1.

Figure D. 1: Stopes in group 3 used for predicting and testing the Dugald River models

Partial least squares model

The first multivariate model tested is the PLS. The goal is to see if a linear model that successfully predicted OB
on a per face basis (McFadyen 2020) could be applied at an octree resolution. Only one output component is
calculated since there is only one performance variable (projected distance). The variables are initially centred
and scaled since they have different units. The predictions generated for the stopes in group 3 are plotted against
the observations in Figure D. 2. The points have been coloured according to their confusion matrix classification.
Linear regressions have been overlaid, with the blue line representing the ideal linear fit for predictions and the
red line representing the best linear fit that is obtained from the predictions. The linear regressions show that the
data does have a linear trend, but the slope is shallower than the perfect prediction.

299
 Figure D. 2: Predictions versus observations for the Dugald River PLS model. The points are coloured by their confusion matrix
classification. The ideal and best linear fit of the data are overlaid

The statistical metrics are presented in Table D. 1. The octrees were correctly predicted as OB or UB 65% of
the time. This means 65% of the design surface would be correctly depicted as OB or UB. The MCC value is
0.25, meaning the classification is not perfect (MCC = 1), but not random either (MCC = 0). The model performs
better at predicting if an octree is UB (70%) compared to OB (50%). The RMSE is 1.31, meaning the average
error is around 1.31 m. 34% of the octrees are predicted within half a metre of the actual projected distance,
60% under a metre and 86% under 2 m. This means that when looking at the predicted stope surface, it is
assumed that just over half of the surface is within 1 m of its actual position and the majority within 2 m. Knowing
that the projected distance has a range of 10 m (-4 m to 6 m), the model allows to narrow down the expected
value range to 2 m most of the time, up to 4 m for the majority. Given that the median is under a 1 m error, the
higher average error is due to localised areas of the stope where OB and UB are underpredicted.

Table D. 1: Results of the performance metrics of the Dugald River PLS model. These metrics do not consider the octrees with a
projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy (ACC) 65%
Positive OB predictions (POP) 56%
True OB detection (TOD) 50%
True UB detection (TUD) 70%
Matthews correlation coefficient (MCC) 0.25
% <0.5 m 34%
% <1 m 60%
% <2 m 86%
%> 2 m 14%
%> 3 m 3%
%> 4 m 0.4%
RMSE 1.31

300
The prediction error varies depending on the observed projected distance. From Figure D. 3, it can be seen that
octrees with a projected distance between -1.5 m and 1.5 m (distances between the dotted lines) tend to have
smaller errors than larger OB and UB distances. This is due, in part, to the fact that there are fewer octrees with
a projected distance exceeding -1.5 m and 1.5 m for the model to fit. There is one set of octrees that have a
larger error for a small projected distance (circled in red). These octrees come from one particular face.

Figure D. 3: Prediction absolute error versus the projected distance for the Dugald River PLS model

To better visualise this statistical information, a stope example is presented (Figure D. 4) for which the predictions
fit the statistics discussed in this section. Two views of the stope are presented, showing the actual values side-
by-side, the predictions, and the prediction error. The face with no octrees is a backfill face. The following
conclusions are drawn from this figure:
• The model predicts the bottom of the HW will be underbroken and the top part near the drift will be
overbroken.
• The side wall will be overbroken for the most part.
• The bottom of the footwall will be overbroken and the top part will be underbroken. UB will be observed
in most of the corners.
• When compared with the actual CMS, similar trends are observed.
• The main difference between the predictions and the CMS are the absolute magnitude of OB and UB.
Looking at the predictions error, the predictions tend to underpredict the OB and UB.

301
Figure D. 4: Dugald River stope PLS model example: observed projected distances of the octrees on the left, the predicted distances in
the middle, and the absolute prediction error on the right. The face with no octree is a backfill face

The quality of the predictions varies from stope to stope and even from face to face. The predictions generated
for the stopes in group 3 are plotted against the observations in 3D view (Figure D. 5 and Figure D. 6). It can be
seen that, although there are variations between observed and predicted CMS, the general trends of OB and
UB are captured and the predicted geometries are realistic for some of these stopes. Overall, with the PLS
model, the per face analytical assessment (presented in Section 2.4.4) enabled the determination that, looking
at the geometry of the 48 faces, 59% are excellently predicted, 14% are well predicted (good prediction), 4%
showed limited results (limited prediction) and 22% are insufficient. The statistical criterion used for determining
if a geometry is predicted excellently is 50% of the octrees with a prediction error within 1 m and 80% within 2
m. These values were selected arbitrarily based on the size of the stope and the impact 1 m OB or UB would
have on the economic performance. The statistical criterion for the OB and UB location (limited prediction) is
70% of the octree correctly predicted as OB or UB.

302
Figure D. 5: 3D view of the HW predictions versus observations for the Dugald River PLS model

303
 Figure D. 6: 3D view of the FW predictions versus observations for the Dugald River PLS model

Figure D. 7 gives an example where the predictions did not allow the determination of the location and magnitude
of OB and UB (insufficient prediction). In this case, UB was observed in the top HW while the model expected
OB. From discussions with mine site personnel, OB was expected due to the presence of a nearby fault in the
HW. Poor ground conditions were expected, which would have caused the rock to break towards the fault. The
different outcome indicates information about the fault and its influence was inaccurate in this instance, which
can occur because part of the structural model is inferred. The model also observed OB in the side wall, but in
fact, it was underbroken. The prediction errors depict a lot of areas of this stope with errors larger than 2.5 m.

304
 Figure D. 7: An example stope of an insufficient prediction with the Dugald River PLS model. The observed, predicted projected
distances and the absolute prediction error can be seen. Pale blue octrees on the crest of the HW have a distance of 0 m and were not
part of the predictions. The face with no octrees is a backfill face

Linear discriminant analysis model

The second model tested is the LDA model. The LDA model is a classification method, and as such only predicts
whether the octree will be OB or UB. If the model proved to be more accurate than the other models, it could be
used to establish where OB and UB would occur, and be used as an additional variable for the other models.

The statistical metrics for the predictions generated for group 3 are presented in Table D. 2. The octrees are
correctly predicted as OB or UB 66% of the time. This means 66% of the design surface would be correctly
depicted as OB or UB. The MCC value is 0.27, meaning the classification is not perfect (MCC = 1), but not
random either (MCC = 0). The model performs better at predicting if an octree is underbroken (74%) compared
to OB (57%).
Table D. 2: Results for the performance metrics of the Dugald River LDA model. These metrics do not consider the octrees with a
projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy (ACC) 66%
Positive OB predictions (POP) 53%
True OB detection (TOD) 57%
True UB detection (TUD) 74%
Matthews correlation coefficient
0.27
(MCC)

305
To better visualise this statistical information, a stope example is presented for which the predictions are similar
to the statistical results presented in Table D. 2. Figure D. 8 presents two views of the stope which compares
observed and predicted values along with error. The following conclusions are drawn from this figure:

• The model predicts that the bottom of the HW will be underbroken and the top part near the drift will be
overbroken.

• The side wall will be overbroken for the most part.

• The bottom of the footwall will be overbroken and the top part will be underbroken. UB will be observed
in most of the corners.

• When compared with the actual CMS, similar trends are observed with OB being more present in the
HW.

• As we can see from the prediction error, part of the FW and HW would be misled by the predictions if
interpreted as is.

Figure D. 8: Stope A observed OB and UB of the octrees on the left, the predicted OB and UB in the middle, and the prediction error on
the right using the LDA model. The face with no octrees is a backfill face

The quality of the predictions varies from stope to stope. The predictions generated for the stopes in group 3 are
plotted against the observations in 3D view (Figure D. 9 and Figure D. 10). Overall, with the LDA model, the per
face analytical assessment (presented in Section 2.4.4) enabled the determination that, of the 49 faces, 59%

306
showed limited results (OB-UB location predicted correctly) and 41% are insufficient. The statistical criterion for
the OB and UB location (limited prediction) is 70% of the octrees correctly predicted as OB or UB.

Figure D. 9: 3D view of the HW predictions versus observations for the Dugald River LDA model

307
 Figure D. 10: 3D view of the FW predictions versus observations for the Dugald River LDA model

Figure D. 11 gives an example where the predictions did not allow the determination of where OB and UB will
occur (insufficient prediction). OB is predicted in the side wall and the top of the HW while they are mostly UB.
The FW is predicted as mostly UB while the bottom half is OB.

308
Figure D. 11: An example of an insufficient prediction with the LDA model. The observed and predicted OB and UB can be seen as well
as the prediction error. Pale blue octrees on the crest of the HW were not part of the predictions. The face with no octrees is a backfill
face

309
Appendix E: Prediction results for the Prominent
Hill PLS and LDA models
A predictive model is built and tested for the 51 transversal stopes presented in Section 3.2.1. For this case
study, the variable inputs for the statistical models are the same as for the previous case study presented in
Chapter 6. Therefore, the first two groups are used for building the model (38 stopes) and the remaining 13
stopes for testing the predictions (group 3, Figure E. 1). The stope selection follows the extraction sequence in
a chronological order.

Figure E. 1: Prominent Hill stopes in group 3

Partial least squares model

The first multivariate model tested is the PLS. The goal is to see if a linear model that successfully predicted OB
on a per face basis (McFadyen 2020) could be applied at an octree resolution. Only one output component is
calculated since there is only one performance variable (projected distance). The variables are initially centred
and scaled since they have different units. The predictions generated for the stopes in group 3 are plotted against
the observations in Figure E. 2. The points have been coloured according to their confusion matrix classification.
Linear regressions have been overlaid, with the blue line representing the ideal linear fit for predictions and the
red line representing the best linear fit that is obtained from the predictions. The linear regressions show that the
data does have a linear trend, but the slope is shallower than the perfect prediction.

310
 Figure E. 2: Predictions versus observations for the Prominent Hill PLS model. The points are coloured by their confusion matrix
classification. The ideal and best linear fit of the data are overlaid

The statistical metrics are presented in Table E. 1. The octrees were correctly predicted as OB or UB 67% of
the time. This means 67% of the design surface would be correctly depicted as OB or UB. The MCC value is
0.39, meaning the classification is not perfect (MCC = 1), but not random either (MCC = 0). The model performs
better at predicting if an octree is OB (81%) compared to UB (73%). The RMSE is 1.45, meaning the average
error is around 1.45 m. 33% of the octrees are predicted within half a metre of the actual projected distance,
59% under a metre and 83% under 2 m. This means that when looking at the predicted stope surface, it is
assumed that just over half of the surface is within 1 m of its actual position and the majority within 2 m. Knowing
that the projected distance has a range of 10 m (-4 m to 6 m), the model allows one to narrow down the expected
value range to 2 m most of the time, and up to 4 m for the majority. Given that the median is under a 1 m error,
the higher average error is due to localised areas of the stope where OB and UB are underpredicted.

Table E. 1: Results for the performance metrics of the Prominent Hill PLS model. These metrics do not consider the octrees with a
projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy (ACC) 67%
Positive OB predictions (POP) 81%
True OB detection (TOD) 63%
True UB detection (TUD) 73%
Matthews correlation coefficient (MCC) 0.39
% <0.5 m 33%
% <1 m 59%
% <2 m 83%
%> 2 m 17%
%> 3 m 6%
%> 4 m 1%
RMSE 1.45

311
The prediction error varies depending on the observed projected distance. From Figure E. 3, it can be seen that
octrees with a projected distance between -1.5 m and 2.0 m (distances between the dotted lines) tend to have
smaller errors than larger OB and UB distances. This is due, in part, to the fact that there are less octrees with
a projected distance exceeding -1.5 m and 2.0 m for the model to fit. There is one set of octrees that have a
larger error for a small projected distance (circled in red). These octrees come from one particular face.

Figure E. 3: Prediction absolute error versus the projected distance for the Prominent Hill PLS model. Outliers circled in red

To better visualise this statistical information, a stope example is presented (Figure E. 4) for which the predictions
fit the statistics discussed in this section. Two views of the stope are presented, showing the actual values side-
by-side, the predictions, and the prediction error. The following conclusions are drawn from this figure:
• The model predicts OB around the drives, with the HW and FW mostly overbroken.
• The side wall will be overbroken in the middle of the face.
• UB will be mostly in the corners.
• When compared with the actual CMS, similar trends are observed except for the FW.
• The main difference between the predictions and the CMS are the absolute magnitude of OB and UB.
Looking at the prediction errors, the predictions tend to underpredict the OB and UB.

312
Figure E. 4: Prominent Hill stope PLS model example: observed projected distances of the octrees on the left, the predicted distances
in the middle, and the absolute prediction error. The face with no octrees is a backfill face

The quality of the predictions varies from stope to stope and even from face to face. The predictions generated
for the stopes in group 3 are plotted against the observations in a 3D view (Figure E. 5 and Figure E. 6). It can
be seen that, although there are variations between observed and predicted CMS, the general trends of OB and
UB are captured and the predicted geometries are realistic for some of these stopes. Overall, with the PLS
model, the per face analytical assessment (presented in Section 2.4.4) enabled the determination that, looking
at the geometry of the 48 faces, 54% are excellently predicted, 23% are well predicted (good prediction), 8%
showed limited results (limited prediction) and 15% are insufficient. The statistical criterion used for determining
if a geometry is predicted excellently is 50% of the octrees with a prediction error within 1 m and 80% within 2
m. These values were selected arbitrarily based on the size of the stope and the impact 1 m OB or UB would
have on the economic performance. The statistical criterion for the OB and UB location (limited prediction) is
70% of the octree correctly predicted as OB or UB. All insufficient predictions are FW (six out of 12) and the
majority of HW are well predicted (nine out of 12).

313
Figure E. 5: 3D view of the HW predictions versus observations for the Prominent Hill PLS model

314
 Figure E. 6: 3D view of the FW predictions versus observations for the Prominent Hill PLS model

Figure E. 7 gives an example where the predictions did not allow the determination of the location and magnitude
of OB and UB (insufficient prediction). In this case, large OB (over 2.5 m, coloured in red) was observed for most
of the HW while the model expected some OB, but only large OB in specific areas. The FW is mostly
underbroken, but the model predicts OB for the majority of the surface. From discussions with mine site
personnel, inclined FW do not generally get OB; however, the stopes used to train the model had multiple stopes
with FW OB, explaining the predictions. The FW results can be improved by updating the stope in the model,
removing older stopes that behaved differently. The prediction errors depict a large area of the stope with errors
larger than 2.5 m for this stope.

315
 Figure E. 7: Prominent Hill stope example where the predictions were not adequate with the PLS model. The observed, predicted
projected distances and the absolute prediction error can be seen. The faces with no octrees are backfill faces

Linear discriminant analysis

The second model tested is LDA. The LDA model is a classification method and, as such, only predicts whether
the octree will be OB or UB. If the model proved to be more accurate than the other models, it could be used to
establish where OB and UB would occur and be used as an additional variable for the other models.

The statistical metrics for the predictions generated for group 3 are presented in Table E. 2. The octrees are
correctly predicted as OB or UB 69% of the time. This means 69% of the design surface would be correctly
depicted as OB or UB. The MCC value is 0.27, meaning the classification is not perfect (MCC = 1), but not
random either (MCC = 0). The model performs well at predicting if an octree is underbroken (64%) compared to
OB (66%).

Table E. 2: Results for the performance metrics of the Prominent Hill LDA model. These metrics do not consider the octrees with a
projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy (ACC) 69%
Positive OB predictions (POP) 66%
True OB detection (TOD) 74%
True UB detection (TUD) 64%
Matthews correlation coefficient (MCC) 0.38

316
To better visualise this statistical information, a stope example is presented for which the predictions are similar
to the statistical results presented in Table E. 2. Figure E. 8 presents two views of the stope which compare
observed and predicted values along with the error. The following conclusions are drawn from this figure:
• The model predicts OB for the majority of the HW and FW and that is what is observed.
• UB will be mostly in the corners.
• The prediction errors are on the edges of the faces.

Figure E. 8: Prominent Hill stope LDA model example: observed OB and UB of the octrees on the left, the predicted OB and UB in the
middle and the prediction error on the right. The face with no octrees is a backfill face

As described in Section 7.1.1, the quality of the predictions varies from stope to stope. The predictions generated
for the stopes in group 3 are plotted against the observations in a 3D view (Figure E. 9 and Figure E. 10). Overall,
with the LDA model, the per face analytical assessment (presented in Section 2.4.4) enabled the determination
that, of the 48 faces, 63% showed limited results (limited prediction) and 38% are insufficient. The statistical
criterion for the OB and UB location (limited prediction) is 70% of the octree correctly predicted as OB or UB.
The insufficiently predicted faces are scattered between the different face types (HW, FW, side walls and crown).

317
Figure E. 9: 3D view of the HW predictions versus observations for the Prominent Hill LDA model

318
 Figure E. 10: 3D view of the FW predictions versus observations for the Prominent Hill LDA model

Figure E. 11 gives an example where the predictions did not allow the determination of where OB and UB will
occur (insufficient prediction). In this case, the side wall OB and UB were mostly insufficient predictions while
the FW OB and UB were correctly predicted. From discussions with mine site personnel, stopes adjacent to
paste do not have OB in the side walls due to stress shadowing, which is not accounted for in this model.

319
Figure E. 11: Prominent Hill example where the predictions were not adequate with the LDA model. The observed and predicted OB
and UB can be seen as well as the prediction error. The face with no octrees is a backfill face

320
Appendix F: Prediction results for the Westwood
PLS and LDA models
A predictive model was built and tested for the 40 longitudinal stopes of the WW28 area presented in Section
3.3.1. For this case study, the variable inputs for the statistical models are the same as for the previous case
studies. Therefore, the first two groups are used for building the model (30 stopes) and the remaining 10 stopes
for testing the predictions (group 3, Figure F. 1). The selection follows a chronological order.

Figure F. 1: Westwood stopes in group 3

Partial least squares model

The first multivariate model tested is PLS. The goal is to see if a linear model that successfully predicted OB on
a per face basis (McFadyen 2020) can be applied at an octree resolution. Only one output component is
calculated since there is only one performance variable (projected distance). The variables are initially centred
and scaled since they have different units. The predictions generated for group 3 are plotted against the
observations in Figure F. 2. The points are coloured according to their confusion matrix classification. Linear
regressions are overlaid, with the blue line representing the ideal linear fit for predictions and the red line
representing the best linear fit that is obtained from the predictions. The linear regressions show little linear trend
in the data since the slope is close to horizontal.

321
 Figure F. 2: Predictions versus observations for the Westwood PLS model. The points are coloured by their confusion matrix
classification. The ideal and best linear fit of the data are overlaid

The statistical metrics are presented in Table F. 1. The octrees are correctly predicted as OB or UB 91% of the
time (ACC). However, only 4% of the UB is correctly predicted as UB (TUD) while 99% of the OB is predicted
as OB (92% of the predicted OB is OB). The MCC value is 0.01, meaning the classification is random (MCC =
0). This is related to the unbalanced categories of OB and UB. The UB population in the dataset is small (14%),
which is due the mine’s goal of leaving no ore behind given the grade and size of the stopes. If we were instead
to look at the larger OB (over 2 m) which is more frequent for this mine site than the other two, the model doesn’t
perform any better (6% of the OB over 2 m correctly predicted). The RMSE is 1.58, meaning the average error
is around 1.58 m. 37% of the octrees are predicted within half a metre of the actual projected distance, 59%
under a metre and 85% under 2 m. This means that, when looking at the predicted stope surface, it is assumed
that just over half of the surface are within 1 m of its actual position and the majority within 2 m. Knowing that
the projected distance has a range of 8 m (-2 m to 6 m), the model allows one to narrow down the expected
value range to 2 m most of the time, up to 4 m for the majority. Given that the median is under a 1 m error, the
higher average error is due to localised areas with large OB that is underpredicted.

322
 Table F. 1: Results for the performance metrics of the Westwood PLS model. These metrics do not consider the octrees with a
projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy
91%
(ACC)
Positive OB predictions
99%
(POP)
True OB detection (TOD) 92%
True UB detection (TUD) 04%
Matthews correlation
0.008
coefficient (MCC)
% <0.5 m 37%
% <1 m 59%
% <2 m 85%
%> 2 m 15%
%> 3 m 9%
%> 4 m 4%
RMSE 1.58
The prediction error varies depending on the observed projected distance. From Figure F. 3, it can be seen that
octrees with a projected distance between 0.0 m and 2.4 m tend to have smaller errors than larger distances.
This is due, in part, to the fact that there are fewer octrees with a projected distance exceeding 0.0 m and 2.4 m
for the model to fit.

Figure F. 3: Prediction absolute error versus the projected distance for the Westwood PLS model

To better visualise this statistical information, a stope example is presented (Figure F. 4) for which the predictions
fit the statistics discussed in this section. Two views of the stope are presented, showing the actual values side-
by-side, the predictions, and the prediction error. The following conclusions are drawn from this figure:
• The model predicts OB for the majority of the HW and FW.
• The side wall will break close to even.
• Little UB is observed.

323
 • When compared with the actual CMS, similar trends are observed.
• The main difference between the predictions and the CMS are the absolute magnitude of OB and UB.
• Looking at the prediction errors, the predictions tend to underpredict the OB and UB.

Figure F. 4: Westwood PLS model prediction example: observed projected distances of the octrees on the left, the predicted distances
in the middle, and the absolute prediction error on the right. The face with no octrees is a backfill face

The quality of the predictions varies from stope to stope and even from face to face. The predictions generated
for the stopes in group 3 are plotted against the observations in 3D view (Figure F. 5 and Figure F. 6). It can be
seen that, although there are variations between observed and predicted CMS, the general trends of OB and
UB are captured and the predicted geometries are realistic for some of these stopes. Overall, with the PLS
model, the per face analytical assessment (presented in Section 2.4.4) enabled the determination that, looking
at the geometry of the 37 faces, 49% are excellently predicted, 22% are well predicted (good prediction), 11%
showed limited results (limited prediction) and 19% are insufficient. The statistical criterion used for determining
if a geometry is predicted excellently is 50% of the octrees with a prediction error within 0.5 m and 80% within 1
m. These values were selected arbitrarily based on the size of the stope and the impact 1 m OB or UB would
have on the economic performance. The statistical criterion for the OB and UB locations (limited prediction) is
70% of the octree correctly predicted as OB or UB. For the insufficient predictions, five out of the 10 stopes have
the HW and FW predictions fail.

324
Figure F. 5: 3D view of the HW predictions versus observations for the Westwood PLS model

325
 Figure F. 6: 3D view of the FW predictions versus observations for the Westwood PLS model

Figure F. 7 gives an example of an insufficient prediction (prediction error drops to 19% under half a metre and
46% under a metre). In this case, large OB (over 2.5 m, coloured in red) is observed for most of the HW while
the model expected OB under 2 m with the maximum in the middle of the face. The FW is mostly close to an
even break, but the model predicts OB for the majority of the surface, similar to the HW. The prediction errors
depict 54% of the surface of the stope with errors larger than 1.5 m. This area of the mine is highly influenced
by the presence of the faults where the HW of the stopes in Figure F. 8 break on or near the fault surface. This
stope HW OB is smaller than the adjacent and previously mine stopes as the nearby fault is closer to the design
surface. In addition, the large OB observed in the surrounding stopes HW is not considered in the model and
most likely has an impact on the stress redistribution for this stope.

326
Figure F. 7: Westwood PLS prediction example where the predictions were not adequate. The observed, predicted projected distances
and the absolute prediction error can be seen. The faces with no octrees are backfill faces

Figure F. 8: Side view of the Westwood analysed stope with predictions that are inadequate as well as adjacent stopes and faults

Linear discriminant analysis

The second model tested is the LDA. The LDA model is a classification method and, as such, only predicts
whether the octree will be OB or UB. If the model proved to be more accurate than the other models, it can be
used to establish where OB and UB would occur and be used as an additional variable for the other models.

The statistical metrics for the predictions generated for group 3 are presented in Table F. 2. The octrees were
correctly predicted as OB or UB 92% of the time (ACC). This means 92% of the design surface would be
accurately depicted as OB or UB. The MCC value is 0.004, meaning the classification is random (MCC = 0). The
model does not perform well at predicting if an octree is underbroken (0%) compared to OB (92%).

327
 Table F. 2: Results for the performance metrics of the Westwood LDA model. These metrics do not consider the octrees with a
projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy (ACC) 92%
Positive OB predictions (POP) 92%
True OB detection (TOD) 99%
True UB detection (TUD) 0%
Matthews correlation
0.004
coefficient (MCC)

The LDA model performs poorly at separating OB and UB. This is in part due to the unbalanced categories.
There is a low percentage of UB (8%) in these longitudinal stopes and the rest of the database (below 15%) as
the stopes are designed to leave no ore behind, and therefore have no UB. This makes the UB prediction difficult.
The prediction errors are mainly located on the edges of the stope and side walls as that is where the UB is
found (Figure F. 9). This accounts for about 22% of the predicted faces. Overall, with this LDA model, the per
face visual assessment (presented in Section 2.4.4) enabled the determination that, of the 37 faces, 76% showed
limited results (limited prediction) and 24% are insufficient. The statistical criterion for the OB and UB location
(limited prediction) is 70% of the octrees are correctly predicted as OB-UB.

Figure F. 9: 3D view of the LDA prediction error for the predicted Westwood stopes

Given that this mine site experiences little UB since they do not wish to leave any ore behind, a 2 m OB threshold
is used to help determine areas where larger OB will occur. A second model was built, but this time with the
objective of separating smaller OB and UB (under 2 m) from larger OB (over 2 m) as around 20% of the database
is large OB and consists of a more critical scenario for the mine. The goal here is to see if the possible presence
of large OB can be easily identified with the model. Again, there is an imbalance in the categories.

328
The statistical metrics for the predictions generated for group 3 are presented in Table F. 3. The octrees were
correctly predicted 68% of the time (ACC). This means 68% of the design surface is correctly predicted as OB
over or under 2 m. The MCC value is 0.21, meaning the classification is not random (MCC = 0) but not perfect
either (MCC = 1). The model does not perform well at predicting if an octree is overbroken by more than 2 m
(18%) compared to under 2 m (95%). This model shows improvement at classifying the data compared to the
first LDA model, implying there is a stronger relationship with the OB than the UB.
Table F. 3: Results for the performance metrics of the second Westwood LDA model (over/under 2 m separation). These metrics do not
consider the octrees with a projected distance larger than -4 m or 6 m

Performance metrics Results

Classification accuracy (ACC) 68%
Positive OB predictions (POP) 65%
True OB detection (TOD) 95%
True UB detection (TUD) 18%
Matthews correlation
0.21
coefficient (MCC)

Overall, for this LDA model, 68% of the faces have the location of OB over/under 2 m predicted (limited
prediction) and 32% are insufficient. The statistical criterion for the over/under 2 m location is 70% of the octrees
are correctly predicted as over or under 2 m. The majority of the HW predictions are insufficient (eight out of 10)
while only a minority of the FW are insufficient (two out of 10). However, no faces with OB over 2 m are well
predicted (mostly in the HW).

329
Appendix G: Model prediction summary
The new proposed design approach enables one to predict stope OB and UB at a per stope (OB-UB volumes),
per face (ELOS and ELLO), and most importantly, an octree level (projected distance). This means the location
and magnitude of OB and UB are predicted, allowing one to visualise the expected geometry in 3D and calculate
volumes of OB and UB from it. This is a step change from current practice, which only predicts the ELOS per
face, not allowing engineers to know the true geometry, the distribution of the magnitude of OB, where it will
occur or the UB.

Statistical summary
A summary of the statistical performance for the three different models built for the three different mine sites is
presented in Table G. 1. In terms of classifying the OB and UB (ACC), the three statistical methods showed
similar performance when looking at the means. Random forest performed slightly better in general than PLS
(1% to 6% depending on the performance metric (introduced in Section 2.4.4)) and performed better than the
LDA with the exception of the TOD (proportion of OB predicted as OB). The Dugald River mine and Prominent
Hill mine showed similar performance where 65% to 69% of the octrees were classified correctly (ACC). In terms
of prediction error, random forest and PLS show almost the exact same performance on average. Around 35%
of the octrees are predicted within half a metre, 60% within a metre and 85% within 2 m. Overall, these models
showed consistent performance over the three mines sites (except for the WW OB-UB classification) implying
the stability and flexibility of the approach to different mine sites. These models enable mine sites to determine,
for the majority of stope surfaces, where the OB and UB are (65% of the time or more) and, for the PLS and
random forest models, the magnitude within 1 m to 2 m.

330
 Table G. 1: Summary of the models’ prediction performance for the three mine sites (DR = Dugald River, PH = Prominent Hill, WW =
Westwood). Performance metrics defined below

Variables summary
The statistical approaches are different for the random forest models (non-linear) and PLS models (linear). This
is seen when looking at the importance the models put on the different variables (Table G. 2). The top and
recurring variables for the PLS models are the ERF, undercut and blasting energy proxy, while the top and
recurring variables for the random forest models are the dip, undercut, direction and blasting (energy or
orientation). For the PLS models, the octree’s position in the face (ERF), the depth of undercut and the
distribution of the blasting energy will dictate the predicted OB and UB. For the random forest model, determining
which face the octree is in (using dip and direction), the depth of undercut and the blasting pattern will dictate
the predicted OB and UB.

331
 Table G. 2: Summary of the model variables’ importance for the three mine sites (DR = Dugald River, PH = Prominent Hill, WW =
Westwood)

Predicted performance summary

When building the prediction models, the spatial aspect is not considered, meaning the relative octree position
to each other is not accounted for. Therefore, to complete the assessment of the model performance, an
analytical 3D assessment was necessary to determine if the position and magnitude of OB and UB matches the
position magnitude measured from the CMS. This is to validate that the predictions can be used by the engineers
to determine the expected mined geometry. Different categories were used to classify the predictions (introduced
in Section 2.4.4):
• Excellent prediction:
o The prediction error criterion and shape criterion are met.
o The predictions identify the magnitude and distribution of the projected distance across the
face within a set threshold, allowing the prediction of the mined geometry.
o The prediction error threshold depends on the stope being analysed.
o These predictions can be used by the engineers to assess the expected stope geometry.
• Good prediction:
o The shape criterion is met, but not the prediction error criterion.
o The predictions identify the distribution of the projected distance across the face, allowing
engineers to determine the general shape of the mined stope. However, the magnitude of the
projected distance cannot be identified across the face within a set threshold.

332
 o The predictions allow engineers to determine where the larger OB and UB will occur, but not
its magnitude.
• Limited prediction:
o The OB-UB location criterion is met, but not the prediction error criterion and shape criterion.
o The predictions allow engineers to identify where OB and UB will occur, but do not identify the
magnitude or distribution of the projected distance for the majority of the surface.
• Insufficient prediction:
o None of the criteria are met.
o The predictions do not allow the determination of the distribution of the projected distance,
and where OB and UB will occur, and its magnitude for the majority of the surface.

A different prediction error criterion was used between the transversal and longitudinal stopes for determining if
the geometry is predicted excellently. The tolerance for error or the minimum prediction error that is required is
mine specific and will depend on the stope size, geometry and grade. For example, the impact a predicted 1 m
OB or UB will have on the estimated dilution or ore recovery varies significantly. Longitudinal stopes that are 2
m to 5 m wide with clear waste-ore boundaries will see dilution increase by 20% to 50% for every metre of OB.
Therefore, the tolerated error for the excellently predicted geometry should be small (under a metre) to allow the
mine site personnel to get an accurate estimate of their stope geometry. On the other hand, for low disseminated
grade using transversal stopes that are 20 m wide, dilution would be limited to only 5% or less for every metre
of OB. Therefore, the tolerated error for the excellently predicted geometry can be larger to allow the mine site
personnel to get an adequate estimate of stope geometry. For the transversal stopes (Dugald River and
Prominent Hill), at least 50% of the octrees needed to have a prediction error smaller than 1 m and 80% smaller
than 2 m, while for the longitudinal stopes (Westwood), 50% under half a metre and 80% under a metre. This
ensures the average prediction error is under 2 m for a transversal face and 1 m for a longitudinal face. It is
acceptable that the remaining 20% of octrees in both cases can exceed the 2 m or 1 m threshold. Given that at
least 80% of the surface is depicted within 1 m or 2 m, an adequate estimate of the surface geometry can be
interpreted. These criteria are based on an understanding of the mining practices but remain arbitrary. This
implies the threshold used for characterising the model performance can vary and improve according to the mine
site and their risk tolerance.

Table G. 3 gives a summary of the predicted geometry performance based on the per face assessment. Overall,
the random forest models perform better at generating excellent predictions of the geometry, capturing the shape
and magnitude of the face 71% of the time (on average) for a transversal stope and 51% of the time for a
longitudinal stope. In comparison, PLS models’ excellent predictions represent 57% (on average) and 49%,

333
respectively. 8% to 13% of the predicted faces failed to capture the geometry of the face for the random forest
models, compared to 15% to 22% for the PLS models. The remaining predicted faces at least allowed the
determination of where the OB and UB would occur or the general distribution of the projected distance (good
or limited predictions). In addition to these numbers, the random forest models enable a probabilistic approach.
This enables one to calculate a predicted interval and assess the maximum OB or UB that can be expected.
One of the limitations observed with the PLS model is the inability to predict the magnitude of the largest OB in
the face when the OB exceeded 2 m.
Table G. 3 : Summary of the models’ performance at determining the OB and UB location and magnitude distribution within a face for
the three mine sites (DR = Dugald River, PH = Prominent Hill, WW = Westwood)

The criteria for characterising the random forest model’s performance included a probabilistic criterion which
looked at the percentage of octrees with their projected distance within the prediction interval (70% of octrees
required for the 10th to 90th percentile interval, see Section 2.4.4). This criterion allows one to consider the
prediction interval when assessing the capabilities of the model to capture the observed projected distance and
the geometry. It is assumed here that its impact on the model performance is beneficial as it looks at a range of
predictions instead of a single value, thus increasing the chances of identifying the observed projected distance
using the model. However, if the prediction interval is large (for example, ±4 m), then it would not allow one to
narrow down the range for the expected projected distance and would leave a certain amount of uncertainty
even though the observed projected distance is within the range. Figure G. 1 shows the cumulative distribution
of the 10th to 90th percentile prediction interval (halved interval distance is used) for all the predicted octrees for
the three mine sites. For Prominent Hill and Westwood, 75% of the predicted octrees have a ±1 m to 1.5 m
interval indicating that the predictions do narrow down the expected projected distance to a 2 m to 3 m window.
The proportion of octrees for Dugald River decreases in their case to 45%.

334
 Figure G. 1 : Cumulative distribution of the 10th to 90 th percentile prediction interval (halved interval distance is used) for all the
predicted octrees for the three mine sites

The reduced model performance for the longitudinal stope is in part associated to the lower prediction error
tolerance that is imposed and, in other parts, to the geological context. Given the position error associated to
the surveys (CMS, drives, faults) and the rock mass sampling, which can be sparse, the inherent variability in
the dataset makes it difficult to predict, with a high success rate, the projected distance within 0.5 m as the
precision of these surveys is to the decimetre. If the prediction error threshold for the transversal stopes were
used instead (1 m and 2 m instead of 0.5 m and 1 m), 72% of the faces would have their geometry predicted
excellently. In terms of the geological context, all mispredicted faces were either HW or FW which are the key
faces that need to be predicted. Four out of the 10 HW have OB over 5 m for the majority of the faces. This
represents almost half of the insufficient faces. These OB are linked to the complex rock mass, where there was
unravelling to faults further away from the stope surface that were not captured by the fault variables. There are
no variables at an octree level currently that can properly capture these situations. The model for this mine site
does not capture all the critical factors. Therefore, for this mine site, further research is needed to capture all the
key variables and additional stopes in the database. For the longitudinal stopes, additional case studies are
recommended to validate the results as, so far, only a limited percentage of faces are predicted excellently
(51%). However, the probabilistic approach predicts the projected distance with a high success rate within a 1
m to 2 m error. The mean errors in these cases are all under 1 m, meaning that valuable information is obtained
72% of the time with minimal error. Also, the predictions can still provide key information for the engineer on the
chances and location of the large OB (over 2 m). 72% of the faces have their maximum and minimum projected
distance fall within the 5th and 95th percentile.

Prediction errors will occur. Analysis of the prediction error indicated causes such as data quality (abnormal
cases such as misfires or muckpiles, or misinformation such as fault position), variability in the geological and
mining environment or OB and UB causes not captured by the model (for example, unravelling). For the
insufficient cases caused by data quality, these cases are site-specific and not caused by the methodology.

335
These predictions are not necessarily wrong since the mined geometry diverged due to misinformation or
operational changes that are not accounted for and are not available at the design stage. It can also be a result
of mining plans changing over time, which can be considered by updating the database. The use of a probabilistic
approach enhances the capabilities of the prediction model as it enables one to determine the prediction interval,
narrowing down the expected range of the projected distance and assessing the chances for larger OB or UB.
For the insufficient cases caused by omitted variables, this indicates the model is not adapted for the specific
mining situation and would need to be modified to better account for specific mining conditions.

Overall, the assessment of the model performance (numbers presented in Table G. 1 and Table G. 3) show that
random forest performs better than the PLS models at predicting the mined geometry. Therefore, the random
forest models are used for determining the capabilities of the proposed approach at predicting the mined
geometry and meeting the objectives of this thesis.

336
Appendix H: Catalogue of predicted geometry that
meets the objective set in this thesis

Figure H. 1: Example no1 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 2: Example no2 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

337
Figure H. 3: Example no3 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 4: Example no4 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

338
Figure H. 5: Example no5 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 6: Example no6 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

339
Figure H. 7: Example no7 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 8: Example no8 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

340
 Figure H. 9: Example no9 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 10: Example no10 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

341
Figure H. 11: Example no11 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 12: Example no12 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

342
Figure H. 13: Example no13 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 14: Example no14 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

343
Figure H. 15: Example no15 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 16: Example no16 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

344
Figure H. 17: Example no17 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 18: Example no18 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

345
Figure H. 19: Example no19 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 20: Example no20 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

346
Figure H. 21: Example no21 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 22: Example no22 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

347
Figure H. 23: Example no23 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 24: Example no24 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

348
Figure H. 25: Example no25 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 26: Example no26 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

349
Figure H. 27: Example no27 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 28: Example no28 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

350
Figure H. 29: Example no29 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 30: Example no30 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

351
Figure H. 31: Example no31 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 32: Example no32 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

352
Figure H. 33: Example no33 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 34: Example no34 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

353
Figure H. 35: Example no35 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 36: Example no36 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

354
Figure H. 37: Example no37 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 38: Example no38 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

355
Figure H. 39: Example no39 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 40: Example no40 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

356
Figure H. 41: Example no41 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 42: Example no42 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

357
Figure H. 43: Example no43 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 44: Example no44 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

358
Figure H. 45: Example no45 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 46: Example no46 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

359
Figure H. 47: Example no47 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 48: Example no48 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

360
Figure H. 49: Example no49 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 50: Example no50 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

361
Figure H. 51: Example no51 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 52: Example no52 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

362
Figure H. 53: Example no53 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 54: Example no54 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

363
Figure H. 55: Example no55 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 56: Example no56 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

364
Figure H. 57: Example no57 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 58: Example no58 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

365
Figure H. 59: Example no59 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 60: Example no60 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

366
Figure H. 61: Example no61 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 62: Example no62 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

367
Figure H. 63: Example no63 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 64: Example no64 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

368
Figure H. 65: Example no65 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

Figure H. 66: Example no66 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

369
Figure H. 67: Example no67 of a predicted geometry that meets the objective set in this thesis. A side-by-side comparison between the
observation and predictions is shown for the projected distance and cumulative distribution of the projected distance. The projected
distance for the prediction interval is shown (10th and 90th percentile) as well as the prediction error

370