ERM 2005

Morgantown, W.V.
SPE Paper # 98012

Reservoir Characterization Using

Intelligent Seismic Inversion

Emre Artun, WVU

Shahab D. Mohaghegh, WVU
Jaime Toro, WVU
Tom Wilson, WVU
Alejandro Sanchez, Anadarko

September 15, 2005

motivation > Reservoir Modeling Workflow
Exploration: A structural model of the reservoir can
Seismic Surveys be attained.

Exploration Some data can be obtained from wells

( i.e. well logs, cores, well tests … )
Drilling

Geostatistical variogram models can be developed

Reservoir with the available data to interpolate / extrapolate
Characterization available well data to the entire field.

Flow in that 3D reservoir can be modeled with

Reservoir commercial reservoir simulators to predict
Simulation reservoir performance.

Field
Development
 motivation > Reservoir Characterization

- Issues about the data and problems regarding data analysis

must be considered carefully in reservoir characterization.

- Geostatistical models become insufficient in dealing with

issues like uncertainty, large variety of scales, immense size
of data, etc.

- As an alternate; our industry has realized the power of soft

computing tools, which are capable of dealing with
uncertainty, imprecision, and partial truth.
 motivation > Reservoir Characterization

Integrating all different types

Ten-feet
of data in an accurate and
high-resolution reservoir model
SEISMIC

One of inches

WELL LOGS

Fraction of inches

CORES
 motivation > Reservoir Characterization

- Due to its low resolution, seismic data is used only to attain

a structural view of the reservoir.

- However, its 3D coverage over a large area attracts engineers

to merge it more detailed characterization studies.

SEISMIC LOGS

- Inverse modeling of reservoir properties from the seismic

data is known as seismic inversion.
 Statement of the Problem

1. Does a relationship exist between seismic data and

reservoir characteristics, beyond the structural relationship?

2. If such a relationship exists, can it be extracted through the

use of soft computing tools, such as artificial neural
networks?

3. How that tool should be designed to develop the most

reliable correlation models?
i.e. neural network algorithm, number and type of seismic attributes
that should be included... etc.
 Previous Work

Chawathe et. al (1997)

neural network neural network

Surface Cross-well Gamma ray
seismic seismic logs

Reeves et. al (2002)

- In this study; vertical seismic profile (VSP) is incorporated

into the study as the intermediate scale instead of cross-well
seismic.
neural network neural network
Surface
seismic VSP Well logs
 Vertical Seismic Profile (VSP)

- Signal receivers are located in the borehole instead of

surface, both down-going and up-going signals are received.

Well Source
surface

rock layer boundary

Receivers VSP resolution ≈

(Geophones)
2 * Surface seismic resolution
 Statement of the Problem

- Using artificial neural networks is proposed to find a

desirable correlation between well logs and seismic data.
Generalized regression neural network (GRNN) algorithm is
used.

- Vertical seismic profile (VSP) is incorporated into the study

as the intermediate scale data.

- Another unique feature of this study was to develop and

work on a synthetic model, before dealing with real data.
 Two-step Correlation Methodology

Two steps of correlation

1) Correlation of surface seismic with VSP
2) Correlation of VSP with well logs

Step 1 Step 2

Surface Well
VSP
Seismic Logs

Low frequency Medium frequency High frequency

 Case 1
Synthetic Model
 Description of the Model

- The model represents the Pennsylvanian stratigraphy of the

Buffalo Valley Field in New Mexico, including the gas-
producing Atoka and Morrow formations.

- The geological complexity increases with depth;

0.8 – 1.124 sec. (6,600 – 9,000 ft) interval has been used.

- Surface seismic and VSP responses have been computed

through a synthetic seismic line of 100 traces.
 Description of the Model

A synthetic seismic line with 100 traces, having 3 wells @ traces 20, 50, and 80.

Trace 20

Trace 50
( VSP well )

Trace 80
 Available Data

1. Density and acoustic velocity distributions.

2. Surface seismic and VSP responses in the form of the

following seismic attributes:

- Amplitude
- Average energy
- Envelope
- Frequency
- Hilbert transform
- Paraphase
- Phase
Seismic Amplitude Distribution
 Case 1 – Synthetic Model

Step 1 Correlation of surface seismic with VSP

Step 2 Correlation of VSP with well logs

Case 1 _ Step 1 (Surface seismic VSP)

32 57
ce ce
Tra T ra
 Case 1 _ Step 1( Surface seismic VSP)

Neural network design:

Inputs Output

Time
+ neural network Single VSP
7 surface attribute
seismic
attributes
Case 1 _ Step 1 (Surface seismic VSP)
Correlation Map

Step 1 Step 2

Surface Seven
Sevenseparate
VSP separateprediction models
modelshave
Well
prediction have
Seismic Logs
been
beendeveloped
developedfor
forseven
sevenVSP
VSPattributes
attributes
Model found with
withthe
thedata
dataof
oftraces
traces32
32++57.
57.

Now,
Now,let’s
let’sapply
applythese
thesemodels
modelstotothe
theother
other
traces
tracestotohave
havethe
thepredicted
predicteddistributions.
distributions.

Surface Virtual Virtual

Seismic VSP Well Logs
Case 1 _ Step 1 (Surface seismic VSP)

FREQUENCY

Actual Predicted
Case 1 _ Step 1 (Surface seismic VSP)

PHASE

Actual Predicted
Case 1 _ Step 1 (Surface seismic VSP)

HILBERT TRANSFORM

Actual Predicted
Case 1 _ Step 1 (Surface seismic VSP)

ENVELOPE

Actual Predicted
Correlation Map

Step 1 Step 2

Surface Well
VSP
Seismic Logs

Model found

Step
Step11--ACCOMPLISHED
ACCOMPLISHED!..
!..

Surface Virtual Virtual

Seismic VSP Well Logs
 Case 1 – Synthetic Model

Step 1 Correlation of surface seismic with VSP

Step 2 Correlation of VSP with well logs

 Case 1 _ Step 2 ( VSP Well Logs )

- Density log has been selected as the target log, and data
of t-50 have been used in building network models.

- Instead of using actual values, the problem was converted

to a classification problem, because of observable
averaged values of density log of t-50.
 Case 1 _ Step 2 ( VSP Well Logs )

Class 1
Class 1
ρ ≈ 1.9 g/cc
Class 2

Class 2
ρ ≈ 2.3 g/cc

Class 3
Class 3
ρ ≈ 2.65 g/cc
 Case 1 _ Step 2 ( VSP Well Logs )

Neural network design:

Inputs Outputs

Time Three
+ neural network
Classes of
7 VSP Density
attributes
 Case 1 _ Step 2 ( VSP Well Logs )

r2 = 0.82
Class 1
ρ ≈ 1.9 g/cc

Class 2
ρ ≈ 2.3 g/cc

Class 3
ρ ≈ 2.65 g/cc
 Case 1 _ Step 2 ( VSP Well Logs )

Class 4 Class 1 Class 2

ρ ≈ 2.09 g/cc
Class 3

Class 4
r2 = 0.94
 Correlation Map

Step 1 Step 2

The
Theprediction model
modelfor
Surface
prediction fordensity
densityhas
hasbeen
been
VSP Well
Seismic Logs
developed
developedwith
withthe
thedata
dataofoftrace
trace50.
50.
Model found Model found
Now,
Now,we
wecan
cangenerate
generatethethecross-sectional
cross-sectional
density
densitydistribution.
distribution.

Surface Virtual Virtual

Seismic VSP Well Logs
Case 1 _ Step 2 ( VSP Well Logs )

DENSITY

Actual Predicted
Case 1 _ Step 2 ( VSP Well Logs )
Correlation Map

Step 1 Step 2

Surface Well
VSP
Seismic Logs

Model found Model found

Step
Step22--ACCOMPLISHED
ACCOMPLISHED!..
!..

Surface Virtual Virtual

Seismic VSP Well Logs
 Case 2
Real Case
The Buffalo Valley Field
The Buffalo Valley Field, New Mexico
 Available Data

- Paper logs from around 40 wells within a 3D seismic

survey area have been digitized.

- Only one well had a VSP survey, i.e. it’s the only well to
build network models.

- Seismic data were loaned by WesternGeco; a total of 27

seismic attributes were available.
Map of Wells and Seismic Survey Area

VSP well
Seismic Amplitude Distribution

Well #1 Well #2 Well #3 Well #4 Well #5

( VSP well )
 Case 2 – Real Case: The B.Valley Field

Step 1 Correlation of surface seismic with VSP

Step 2 Correlation of VSP with well logs

Case 2 _ Step 1 (Surface seismic VSP)
Case 2 _ Step 1 (Surface seismic VSP)
Correlation Map

Step 1 Step 2

Surface Well
VSP
Seismic Logs

Model found

Surface Virtual Virtual

Seismic VSP Well Logs
 Case 2 – Real Case: The B. Valley Field

Step 1 Correlation of surface seismic with VSP

Step 2 Correlation of VSP with well logs

 Case 2 _ Step 2 ( VSP Well Logs )

- After a quality check of available logs, gamma ray and

neutron porosity logs were selected as target logs,
considering their availability, and quality.
 Case 2 _ Step 2 ( VSP Well Logs )

- Data from all available wells were used in developing the

neural network models.

- A ‘Key Performance Indicators’ (KPI) study was conducted

to see influences of each seismic attribute on the target
log.
 Key Performance Indicators (KPI)

Intelligent Reservoir Characterization and Analysis (IRCA) software:

- Most influent attributes were selected due to large number

of available attributes.
 Gamma Ray Log

Well #1 Well #2 Well #3 Well #4 Well #5

r = 0.76 r = 0.86 r = 0.81 r = 0.90 r = 0.90
Gamma Ray Log
Neutron Porosity Log
Well #1 Well #2
r = 0.98 r = 0.97
Neutron Porosity Log
Correlation Map

Step 1 Step 2

Surface Well
VSP
Seismic Logs

Model found Model found

Step
Step22--ACCOMPLISHED
ACCOMPLISHED!..
!..

Surface Virtual Virtual

Seismic VSP Well Logs
 Conclusions

- The proposed two-scale-step, intelligent seismic inversion

methodology has been successfully developed on a
synthetic model. The same methodology has then been
applied to real data of the Buffalo Valley Field in New
Mexico.

- Density logs for the synthetic model, and gamma ray logs
for the field data have been produced from seismic data.
 Conclusions

- The complex and non-linear relationships have been

extracted with the power of artificial neural networks with
both classification and prediction.

- A novel approach has been presented to solve an

important data integration problem in reservoir
characterization.

- The same methodology can be applied to a 3D seismic

block to obtain 3D distributions of reservoir properties.
 ERM 2005
Morgantown, W.V.
SPE Paper # 98012

Reservoir Characterization Using

Intelligent Seismic Inversion

Acknowledgements
Acknowledgements

--This
Thisstudy
studywas
wassupported
supportedbybythe
theU.S.
U.S.Department
Departmentof ofEnergy.
Energy.Help
Helpand
andsupport
support
of
ofMr.
Mr.Thomas Mroz(project
ThomasMroz (projectmanager)
manager)isisappreciated.
appreciated.

--Seismic
Seismicdata
datawere
wereused
usedwith
withthe
thecourtesy
courtesyof
ofWesternGeco.
WesternGeco.

--Mrs.
Mrs.Janaina
JanainaPereira’s
Pereira’shelp
helpinindigitizing
digitizingwell
welllogs
logsisisalso
alsoappreciated.
appreciated.

September 15, 2005