The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133

https://doi.org/10.1007/s00170-022-08966-y

ORIGINAL ARTICLE

A machine learning approach for the prediction of melting efficiency

in wire arc additive manufacturing
Germán O. Barrionuevo1 · Pedro M. Sequeira‑Almeida2 · Sergio Ríos2,3 · Jorge A. Ramos‑Grez1 · Stewart W. Williams2

Received: 18 November 2021 / Accepted: 21 February 2022 / Published online: 2 March 2022
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
Wire arc additive manufacturing (WAAM) appears as one of the most promising technologies due to its capacity to process
all types of materials used in welding, its high production rate, and capacity to process large geometries of particular inter-
est in the aeronautical industry. Since this technology is still under investigation, it is important to determine the efficiency
of the process; in this sense, the melting efficiency stands out not only as a parameter of interest in energy terms but also as
a measure of the stability of the process. For calculating melting efficiency, it is necessary to use tailored colorimeters or
apply models requiring specific dimensions that involve destructive testing. For this reason, in the development of this work,
the melting efficiency is evaluated through machine learning algorithms. Processing parameters such as wire diameter, wire
feed speed, travel speed, and net power are used to determine melting efficiency. In addition, a simplified analytical model
was developed to compare the results. The average melting efficiency analytically calculated was 44.56 ± 5.48%, while the
predicted value reaches a comparable value of 44.32 ± 4.79% obtained with the Gaussian process regressor, which shows
the highest accuracy. Moreover, the known relationship with travel speed was verified.

Keywords Melting efficiency · Wire arc additive manufacturing · Cold metal transfer · Machine learning

1 Introduction (powder bed fusion, PBF), and the second, where the mate-
rial is deposited directly in either powder or wires, called
Additive manufacturing (AM) has evolved substantially dur- directed energy deposition (DED) [3]. Wire arc additive
ing the last decades. It has transited from being a rapid pro- manufacturing (WAAM) is a DED technology that com-
totyping technology, mainly focused on polymer processing bines an electric arc as the heat source and a metal wire as
[1], to a technology capable of producing final functional feedstock, typically employing standard welding equipment.
parts and processing almost any type of material [2]. Within WAAM technology has many advantages compared to PBF
metal AM, two major technologies are distinguished, one processes. It has a higher deposition rate of up to 8 kg/h
where there is a bed of metallic powder and the content is [4], it does not always require a highly inert atmosphere
fused through a heat source, e.g., laser or electron beam [5], and it is possible to manufacture large and medium-
complex metal components [6] and process any alloy used
in welding [7].
* Jorge A. Ramos‑Grez In welding, it is common practice to use the heat input
[email protected]
to control process characteristics such as the cooling rate
Germán O. Barrionuevo and temperature gradient. The heat input is calculated as the
[email protected]
ratio of arc power to travel speed times the arc efficiency.
1
Department of Mechanical and Metallurgical Engineering, The arc efficiency considers the losses from the arc to the
School of Engineering, Pontificia Universidad Católica de environment. However, as Fuerschbach and Knorovsky [8]
Chile, Av. Vicuña Mackenna 4860, Macul, Santiago, Chile have pointed out, the heat input is insufficient to estimate the
2
Welding Engineering and Laser Processing Centre, Cranfield amount of metal that is actually melted. Conduction losses
University, University Way, Cranfield, Bedfordshire, UK through the material account for more than half the absorbed
3
University of Magallanes, Región de Magallanes y de la energy which goes into the material [9]. Moreover, the lat-
Antártica Chilena, Manuel Bulnes 01855, Punta Arenas, ter gets worse at low travel speed. These losses give origin
Chile

13
Vol.:(0123456789)

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

3124 The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133

to the melting efficiency, as the ratio between the minimum

necessary energy to melt the fusion zone and the energy
which is effectively absorbed by the material. It has been
shown that the melting efficiency can be related to the width
of the heat-affected zone [10]. Therefore, it is relevant not
only as any efficiency obviously is, however also as a param-
eter possibly affecting the stability of the process and the
component’s properties.

1.1 Efficiency in metal additive manufacturing

processes

Saving energy is essential in any manufacturing process. In

that sense, to accurately determine the global efficiency of
a metal AM process, it is necessary to differentiate between Fig. 1  Cross-section micrograph of a single layer deposited by
source efficiency ( 𝜂s ) and melting efficiency ( 𝜂m ). Source WAAM
efficiency is described as the heat transferred to the work-
piece divided by the total heat generated by the heat source
and depends on the interaction between the heat source and WAAM in carbon steel; W represents the bead width, H is
workpiece; it is described by Eq. (1): the bead height, Pd is the penetration depth, θ is the contact
Enet angle, and the darker zone is the heat-affected zone (HAZ)
𝜂s = (1) [13]. Melting efficiency can be estimated after the process
Ei
has finished, focusing on the energy used in melting the base
where Ei expressed in [J ­mm−1] is the energy or heat input plate, given by the penetration area (A1), and the energy con-
per unit length and can be calculated as the ratio between sumed in melting the deposited material (A2).
power (P) in watts and the travel speed v [mm ­s−1]; Enet is the Melting efficiency is defined as the reference value of
net energy input per unit length. In arc sources, the power heat required to cross the solid–liquid phase change barrier.
is calculated by the product between voltage and current; In real terms, the melt pool reaches a higher temperature,
therefore, the arc efficiency is given by Eq. (2): but the liquidus temperature is taken as a reference. The
𝜂m can be determined as the ratio between the sum of the
Enet v ⋅ Enet heat required to melt the base plate and the wire by the net
𝜂s = = (2)
P V ⋅I energy input.
v
𝛾 ⋅ Aw
For gas metal arc welding (GMAW) and submerged 𝜂m = (3)
arc welding (SAW), 𝜂s typically follows between 0.8 and Enet
0.9, while for non-consumable electrode processes such as
­ m−3] represents the melting enthalpy, a
In Eq. (3), 𝛾 [J m
gas tungsten arc welding (GTAW) and plasma arc welding
quantity representing the change in enthalpy from inter-pass
(PAW), the arc efficiency falls to a value of around 0.5 [11].
to liquidus temperature, Aw ­[mm2] is the cross-section area
Thus, net energy is obtained from Eq. (2).
given by the sum between A1 and A2. Since the net energy is
On the other hand, the melting efficiency denotes the frac-
known, 𝜂m can be expressed by Eq. (4):
tion of the net energy that in fact melts the material, while
the rest is lost through heat conduction in the workpiece. 𝛾 ⋅ v ⋅ (A1 + A2 )
The 𝜂m depends on both the heat source and the material
𝜂m = (4)
𝜂s ⋅ P
properties. In welding, 𝜂m is defined as the amount of heat
required to melt the weld deposit per unit length divided by Equations 5 and 6 show predicted melting efficiencies for
the net energy input [12]. Similarly, for WAAM, the melt- 2D and 3D heat conduction by Wells [14] and Okada [15],
ing efficiency can be calculated using the same definition. respectively, where 𝛼 represents the thermal diffusivity of
Therefore, WAAM technology will depend on the wire the material ­[mm2 ­s−1], and w the wall width [mm].
material, bead geometry, substrate properties, and pro- 1
cessing parameters. For instance, Fig. 1 shows the cross- 𝜂m = (
8𝛼
)
(5)
section and main dimensions of a bead-on-plate deposited by 5vw
+2

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133 3125

1 works have been developed to face these common problems,

𝜂m =
(6) where deposition strategy optimization, path planning, and
( √ )
10.4𝛼 2
1.35 1 + 1 + (vw)2 in situ monitoring are some strategies studied. Huang et al.
[28] analyzed the effect of depositing direction on the resid-
Accurate measurement of the arc and melting efficien-
ual stress and distortion. They found that alternating direc-
cies requires calorimetric methods to determine the amount
tion of deposition in every layer causes 25% less distortion.
of heat deposited in a workpiece. However, these methods
Xiong et al. [7] developed a closed-loop control system for
typically do not consider heat losses between the start of
controlling the layer width. The controller was useful for
the process and the beginning of calorimetric measurement
layer width ranging from 6 to 9 mm. Nguyen et al. [29]
[16]. Other methods use thermocouples to determine thermal
applied machine learning to generate optimal tool paths,
cycles, but just at one point of the workpiece near the HAZ.
avoiding welding defects and uneven build-up. Ding et al.
Therefore, an alternative to assess 𝜂m consists of applying
[30] developed an automated manufacturing system apply-
analytical or numerical models [17]. Pepe et al. [18] deter-
ing artificial neural network (ANN) for path planning and
mined process efficiency of GMAW and cold metal transfer
void-free deposition with high geometrical accuracy. Ríos
(CMT); the average process efficiency was around 85% by
et al. [31] developed an analytical model to predict the power
averaging the results of a liquid nitrogen calorimeter and an
consumption for pulsed TIG and plasma deposition, and the
insulated box calorimeter. Mezrag et al. [16] determined the
model was validated for wall width less than 12 mm in Ti-
efficiency of CMT by applying two methods: (1) an analyti-
6Al-4 V processing.
cal heat transfer model coupled with thermocouple measure-
ML algorithms appear as a powerful tool to increase our
ments and (2) numerical heat transfer simulation. The values
understanding of metal AM. Li et al. [32] employed ANN
obtained vary between 0.78 and 0.92. CMT minimizes the
to enhance a bead-overlapping model, preventing defects
heat input by actively limiting the arc current during deposi-
inside the fabricated samples. Deng et al. [33] compared
tion [19]. Since the AM process involves all heat transfer
boosting algorithms and ANN algorithms for bead geometry
phenomena and besides the heat source moves continuously,
prediction; they trained the ML algorithms using average
the Navier–Stokes equations are also involved, making their
current, deposition rate, and inter-pass temperature as inputs
computation highly complex and time-consuming. One alter-
and predicted the bead height and width. In the same line,
native to solve these problems is the application of machine
Barrionuevo et al. [34] compared several ML regressors to
learning (ML) algorithms. Supervised ML algorithms map a
predict layer height and wall width of Ti-6Al-4 V processed
function from known input–output pairs and are mainly used
by plasma WAAM. Xia et al. [35] applied ML to predict
for classification and regression [20]. Regression is employed
the surface roughness and improve the surface integrity of
to estimate the relationships between a dependent variable
deposited layers by WAAM. Ikeuchi et al. [36] developed
and one or more independent variables [21]. Thus, it appears
a model based on ANN for track profile modeling in cold
as an appropriate mathematical tool to build predictive mod-
spray AM. ML algorithms bring new opportunities to opti-
els of the melting efficiency in AM as a function of process
mize advanced manufacturing systems capable of modeling
parameters.
high nonlinear relations and learning from available data.
With the motivation to determine the melting efficiency
1.2 Alloys, parameters, and machine learning of the WAAM technology, copper-coated, G3Si1 solid wire
in metal additive manufacturing low-alloyed carbon-manganese steel was deposited by cold
metal transfer technology. Then, three of the most powerful
WAAM emerges as a powerful technology whose primary ML algorithms were trained to find a possible relationship
applicability has been established in the aerospace, aero- between processing parameters and melting efficiency. A
nautical, automobile, and marine industries [22]. The capa- feature-importance analysis was developed to identify which
bility of fabricating large-scale metal components at rela- parameter dominates the melting efficiency.
tively low prices and a low buy-to-fly ratio, compared to
conventional manufacturing and PBF processes, makes it
suitable for these industrial sectors [23]. In WAAM, the heat 2 Material and methods
source power falls in a range from 1 to 5 [kW], a travel speed
between 5 and 50 [m ­s−1], and a deposition rate from 100 to 2.1 Experimental setup
1000 ­[cm3 ­h−1] [24].
Process parameters must be chosen carefully to avoid Mild steel solid wire consumable electrodes G3Si1 (AWS:
problems such as spatter, excessive distortion [25], high A5.18ER70S-6) were deposited on carbon steel plate A572
residual stresses [26], and poor surface finish [27]. Several Gr50 (ASTM A1011/1011 M) by a modified GMAW variant

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

3126 The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133

Table 1  Experimental variables employed for the fabrication of mild 2.3 Machine learning algorithms
steel by cold metal transfer
Factor Symbol Min Max The core of the ML algorithms is the data. Therefore, Fig. 3
presents the statistical distribution of each variable and the
Wire diameter (mm) WD 0.8 1.2
histograms on the diagonal. The input parameters to train the
Wire feed speed (mm/s) WFS 28 222.5
algorithms were the wire diameter (WD), wire feed speed
Travel speed (mm/s) TS 1.66 25
(WFS), travel speed (TS), and nominal power (P). The out-
Power (kW) P 0.65 4.03
put parameter is the melting efficiency (ME) (Fig. 4). To
accurately predict melting efficiency in WAAM processes,
based on a controlled dip transfer mode mechanism (Fronius three types of the most potent ML regressors were employed,
CMT GmbH). All deposits were conducted on a 5-axis Trio which are detailed in the following sections.
Motion rig coupled with the Robacta Drive CMT welding
torch (Lincoln Power Wave) [13]. Seventy-five single layer 2.3.1 Gaussian process regressors
deposits were fabricated, where the wire diameter, wire feed
speed, and travel speed were controlled while bead width, bead Gaussian process regressor (GPR) is a stochastic method
height, and energy were quantified. Furthermore, the cross- based on statistical learning and Bayesian theory that meas-
sectional area was determined by optical micrographs. The ures the similarity between points using a kernel function to
range of processing parameters is listed in Table 1. predict the value for an unseen point from the training data
[37]. GPR works well on small datasets and can provide
2.2 Melting efficiency evaluation uncertainty measurements on the predictions [38].

For assessing melting efficiency, Eq. (4) was applied, the 2.3.2 Extreme gradient boosting regressor
source efficiency was held constant ( 𝜂s = 0.9 ). In cases
where it is difficult to assess the cross-section area, it is pos- Extreme gradient boosting regressor (XGBR) is an ensemble
sible to replace Aw with the product between layer height boosting method, which predicts the desired outcome based
(H) and wall width (W), as they show a high coefficient of on a forward stage-wise fashion [33]; it allows the optimiza-
correlation (R2 = 0.9797) (Fig. 2). Moreover, a coefficient tion of arbitrary differentiable loss functions [38]. It uses a
(𝜁 = 0.87) is added to Eq. (4) in order to improve the model regularized model formalization to control overfitting, which
accuracy and increase its precision. Therefore, the melting gives it better performance.
efficiency can be calculated using Eq. (7):
2.3.3 Multi‑layer perceptron
𝜁.𝛾 ⋅ v ⋅ H ⋅ W
𝜂m = (7)
𝜂s ⋅ P Multi-layer perceptron (MLP), commonly known as artifi-
cial neural network (ANN), is based on the perceptron as an

Fig. 2  Relationship between the

cross-section area (Aw) and the
product between layer height
(H) and wall width (W)

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133 3127

Fig. 3  Scatter matrix to show a possible correlation between process parameters and melting efficiency

operational unit trained using backpropagation; therefore, it uses and attempt to find the optimal values that maximize the
the square error as a loss function output set of continuous val- coefficient of determination (R2). The ranges of explored
ues. MLP determines a function y = f (x;𝜃) and learns the value hyperparameters are listed in Table 2.
of the parameters 𝜃 that result in the best approximation [39]. The dataset assembled with the experimental data was
GPR, XGBR, and MLP were all trained and tested. randomized and then split into training (80%) and testing
The dataset was assembled from the experimental results (20%) portions. Before initiating the training process, the
(Fig. 3). In order to achieve the highest accuracy, the dataset was scaled using zero mean and unit variance. The
hyperparameters of each ML algorithm were tuned by ML algorithms employed to assess the melting efficiency
applying random search optimization. This procedure were executed in Google Colaboratory (Colab) environ-
automatically examines the hyperparameter search space ment using Scikit-learn and XGBoost libraries.

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

3128 The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133

Fig. 4  Machine learning regres-

sors for predicting melting
efficiency

2.4 Evaluation of the precision of machine learning into the data, and the model can improve the usefulness of a
algorithms predictive model on the estimations achieved.

K-fold cross-validation (CV) was employed to avoid over-

fitting during the training process [40]. In this validation 3 Results and discussion
technique, k represents the number of parts in which the data
is divided; k-1 folds are used for training, and the remain- 3.1 Bead geometry evaluation
ing fold is used to test the model [38]; in this work, 5 k CV
was utilized. According to Barrionuevo et al. [41], ML algo- Sequeira-Almeida [13] determined layer geometry through
rithms’ accuracy evaluation is more straightforward when AxioVision image analysis software utilizing built-in
applying several metrics or a combination of them. They measurement functions. The obtained results are shown in
applied a custom metric in the form of Eq. (8): Fig. 5. The bead width (W) varies from 2.52 to 8.83 mm;
√ the bead height (H) presents a compact range from 1.61
(8) to 3.62 mm.
)2
1 − R2 + MSE + (MAE)2
(
IM =

where R2 is the coefficient of determination, MSE is the 3.2 Analytical melting efficiency results
mean squared error, and MAE is the mean absolute error.
This index of merit (IM) evaluates the accuracy of the According to the methodology explained in Sect. 2.2, Eq. (7)
predictions; as the magnitude of the index approaches zero, shows good agreement with the conventional model depicted
the maximum predicting accuracy is achieved. This metric in Eq. (4). A 𝜂m of 44.56 ± 5.48% was obtained using Eq. (4),
is beneficial when all metrics present similar results. while 𝜂m = 44.47 ± 6.47% is reported when applying Eq. (7).
Once identified which algorithm presents the higher Therefore, the use of the height and width of the WAAM
accuracy (lower IM), feature importance analysis (FIA) deposit appears as a good alternative to replace the pen-
was employed. FIA assigns a score to input features based etration area (A1), and the energy consumed in melting the
on how useful they are at predicting a target variable ( 𝜂m ). deposited material (A2). The obtained results are in agree-
Moreover, FIA provides scores that help us obtain insight ment with the results reported by DuPont and Marder [42].
Furthermore, it is common to report the melting efficiency
as a function of the travel speed; thus, Fig. 6 shows that the
melting efficiency increases with increasing travel speed.
Table 2  Hyperparameter ranges employed in random search optimi- Equations (4), (5), and (7) were employed to assess the melt-
zation
ing efficiency in WAAM analytically.
ML algorithm Hyperparameter ranges

GPR Noise level α: [0.001, 0.1, 1, 10] 3.3 Machine learning melting efficiency results
Number of optimizers: [5, 10, 20]
XGBR Number of estimators: [100, 1000, 10000] The obtained hyperparameters that maximize the accuracy
Maximum depth: [5, 10, 20] by random search optimization are reported in Table 3.
Learning rate: [0.1, 0.01, 1e-3] Gaussian process regressor (GPR), extreme gradi-
MLP Hidden layers: [5, 8, 10] ent boosting regressor (XGBR), and multi-layer per-
Activation function: [“relu,” “tanh”]
Solver: [“lbfgs,” “sgd,” “adam”] ceptron (MLP) algorithms were applied to predict the

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133 3129

Fig. 5  Boxplots of measured

bead geometry: height (H) and
width (W)

Fig. 6  Melting efficiency as a

function of travel speed for the
cold metal transfer process

melting efficiency during WAAM. All algorithms show and MLP present similar metrics, but MLP presents a
similar predictions, GPR reports a melting efficiency higher index of merit, therefore, lower accuracy.
of 44.32 ± 4.79%, XGBR 44.69 ± 4.62%, and MLP
44.12 ± 4.80%. The dispersion of these results is due in part
to the range of travel speed involved in the CMT experi- Table 3  Hyperparameters that provide the highest accuracy of each
ments. The obtained metrics during the cross-validation ML algorithm
and testing are reported in Table 4. ML algorithm Hyperparameters
Cross-validation provides a practical evaluation of the
GPR Noise level α = 0.001
models’ capability to predict new data and face common Number of optimizers = 40
problems as underfitting or overfitting. For the melting XGBR Number of estimators = 100
efficiency evaluation, all the examined algorithms pro- Maximum depth = 10
vide similar metrics. Nonetheless, GPR achieves a higher Learning rate = 0.1
coefficient of determination, which explains how well the MLP Hidden layers = 5
model replicates the observed results and presents the Activation function = relu
Solver = lbfgs
same value for MAE with the XGBR algorithm. XGBR

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

3130 The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133

Table 4  Metrics to assess the ML algorithm Cross-validation Testing

accuracy of the ML algorithms
2
employed to predict melting R RMSE MAE IM R2 RMSE MAE IM
efficiency in CMT
GPR 0.8678 0.0004 0.0153 0.1345 0.9190 0.0146 0.0113 0.0830
XGBR 0.8453 0.0005 0.0153 0.1570 0.8667 0.0187 0.0154 0.1355
MLP 0.8172 0.0005 0.0173 0.1851 0.8864 0.0187 0.0143 0.1160

GPR obtains the lowest IM and reports the highest R2 in Fig. 9. GPR and MLP algorithms converge to the maxi-
and the lowest error metrics about the set employed for mum melting efficiency value (≈ 51%); on the other hand,
testing evaluation. All the algorithms show good accuracy XGBR reports a melting efficiency near the average (44.8%).
for predicting melting efficiency in WAAM. Figure 7 shows Therefore, GPR and MLP capture the essence of the melting
the performance evaluation of each ML algorithm. The efficiency evaluation; they can reproduce the observed trend
training dataset is plotted in black dots and different colors in Fig. 6 and the model developed by Wells [14].
in the testing dataset by each algorithm. The predicted During the WAAM process planning, it is worth noting
melting efficiency corresponds to the vertical axes, while that although the melting efficiency increases with increasing
the horizontal axes show the measured melting efficiency. travel speed, there is a threshold where the heat input only
Figure 8 shows the feature-importance analysis, where melts the feedstock but does not melt the substrate, caus-
power (P) scores the highest value, 72.74%. Therefore, the ing delamination. Moreover, a roughness increase has been
nominal power represents the main factor for the prediction reported while increasing travel speed or decreasing power
model. Travel speed (TS) and wire feed speed (WFS) show due to a lower melting efficiency [25]. Experimentally, it has
feature importance of 11.56 and 10.47%, respectively. The wire been found that the power with which there is greater melting
diameter practically does not influence the model predictability. efficiency is around 2.6 kW. Thus, a linear energy density of
In order to assess the behavior of the ML algorithms, around 110 J/mm is recommended for processing carbon-
an average set of processing parameters was selected manganese steel. Furthermore, an inverse relationship has
(WD = 0.8 mm, WFS = 98 mm ­s−1, P = 2245 W); the melting been reported between travel speed, melt pool depth, and
efficiency evaluation is reported as a function of travel speed width, and a direct relationship with heat input [10, 43].

Fig. 7  Scatter plot for accuracy evaluation of the GPR performance in the melting efficiency prediction

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133 3131

Fig. 8  Feature importance analysis of the processing parameters in the melting efficiency prediction model

Fig. 9  Melting efficiency evalu-

ation as a function of travel
speed: comparison between
XGBR, GPR, and MLP

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

3132 The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133

4 Conclusions References

This work presents the evaluation of melting efficiency 1. Rinaldi M, Ghidini T, Cecchini F, Brandao A, Nanni F (2018)
Additive layer manufacturing of poly (ether ether ketone) via
in a WAAM process. The cold metal transfer process was FDM. Compos Part B Eng 145:162–172
assessed analytical and by applying machine learning algo- 2. Mycroft W, Katzman M, Tammas-Williams S, Hernandez-Nava
rithms. The main results can be summarized as follows: E, Panoutsos G, Todd I, Kadirkamanathan V (2020) A data-driven
approach for predicting printability in metal additive manufactur-
ing processes. J Intell Manuf
• The calculated melting efficiency results report a value
3. Li N, Huang S, Zhang G, Qin R, Liu W, Xiong H, Shi G, Blackburn
of 44.56 ± 5.48% for the cold metal transfer process J (2019) Progress in additive manufacturing on new materials: a
employing Eq. (4). For the proposed model in Eq. (7), review. J Mater Sci Technol 35(2):242–269
a melting efficiency of 44.47 ± 6.47% was obtained. 4. Karmuhilan M, Sood AK (2018) Intelligent process model
for bead geometry prediction in WAAM. Mater Today Proc
• By applying machine learning, it is possible to predict
5(11):24005–24013
the melting efficiency without using complex calorimetry 5. Li JLZ, Alkahari MR, Rosli NAB, Hasan R, Sudin MN, Ramli FR
measurement systems. This technique can adjust the pro- (2019) Review of wire arc additive manufacturing for 3d metal
cess parameters to ensure suitable adhesion between the printing. Int J Autom Technol 13(3):346–353
6. Liu J, Xu Y, Ge Y, Hou Z, Chen S (2020) Wire and arc additive
substrate and the deposited material avoiding delamination. manufacturing of metal components: a review of recent research
• The algorithm that better performs predicting melting developments. Int J Adv Manuf Technol 111(1–2):149–198
efficiency in CMT was the Gaussian process regres- 7. Xiong J, Yin Z, Zhang W (2016) Closed-loop control of variable
sor, with a predicted value of 44.32 ± 4.79%. It shows layer width for thin-walled parts in wire and arc additive manu-
facturing. J Mater Process Technol 233:100–106
the highest coefficient of determination, lowest mean 8. Fuerschbach PW, Knorovsky AG (1991) A study of melting effi-
squared error, and lowest absolute error during cross- ciency in plasma - Desconhecido.pdf. Weld Res Suppl 287–297
validation and testing procedures. The accuracy evalu- 9. Andani MT, Dehghani R, Karamooz-Ravari MR, Mirzaeifar R,
ation was validated through the lowest index of merit. Ni J (2018) A study on the effect of energy input on spatter parti-
cles creation during selective laser melting process. Addit Manuf
For these reasons, GPR is the algorithm recommended 20:33–43
for predicting melting efficiency in WAAM. 10. Fotovvati B, Wayne SF, Lewis G, Asadi E (2018) A review on melt-pool
• The factors that dominate the model prediction were deter- characteristics in laser welding of metals. Adv Mater Sci Eng 2018
mined through a feature-importance analysis. Nominal power 11. Stenbacka N, Choquet I, Hurtig K (2012) Review of arc efficiency
values for gas tungsten arc welding. IIW Comm. IV-XII-SG212,
represents over 72.7% of the model’s predictability, followed Intermed. Meet. BAM, Berlin, Ger. 18–20 April. 2012, pp 1–21
by travel speed (11.6%). These parameters make physical sense 12. American Welding Society (2001) Welding Handbook, vol 1
since the torch power to speed ratio represents the heat input. 13. Sequeira-Almeida PM (2012) Process control and development in
wire and arc additive manufacturing. Cranfield University, PhD
Dissertation
14. Wells AA (1952) Heat flow in welding. Weld J 263s-267s
Author contribution Germán O. Barrionuevo: conceptualization,
15. Okada A (1977) Application of melting efficiency in welding
methodology, data curation, writing–original draft preparation. P. M.
and its problems. Yosetsu Gakkai Shi/Journal Japan Weld Soc
Sequeira-Almeida: methodology, investigation, formal analysis. Sergio
46(2):53–61
Ríos: formal analysis, validation, supervision, writing–review and edit-
16. Mezrag B, Deschaux Beaume F, Rouquette S, Benachour M (2018)
ing. Jorge Ramos-Grez: supervision, writing–reviewing and editing.
Indirect approaches for estimating the efficiency of the cold metal
Stewart Williams: funding acquisition, writing–reviewing and editing.
transfer welding process. Sci Technol Weld Join 23(6):508–519
17. Cambon C, Rouquette S, Bendaoud I, Bordreuil C, Wimpory R,
Funding This work was supported by SENESCYT grant number Soulie F (2020) Thermo-mechanical simulation of overlaid layers
ARSEQ-BEC-000329–2017, the Research Center for Nanotechnol- made with wire + arc additive manufacturing and GMAW-cold
ogy and Advanced Materials (CIEN-UC), ANID FONDECYT grant metal transfer. Weld World 64(8):1427–1435
number 1201068 project, and by the WAAM-Mat research program of 18. Pepe N, Egerland S, Colegrove PA, Yapp D, Leonhartsberger
Cranfield University. A, Scotti A (2011) Measuring the process efficiency of con-
trolled gas metal arc welding processes. Sci Technol Weld Join
Data availability The dataset and the source code are available at 16(5):412–417
https://​github.​com/​Germa​nOmar/​Melti​ng/​blob/​master/​Melti​ngEff_​ 19. Selvi S, Vishvaksenan A, Rajasekar E (2018) Cold metal transfer
WAAM.​ipynb (CMT) technology - an overview. Def Technol 14(1):28–44
20. Gianey HK, Choudhary R (2018) Comprehensive review on
supervised machine learning algorithms. Proc. - 2017 Int. Conf.
Declarations Mach. Learn. Data Sci. MLDS 2017, vol 2018-Janua, pp 38–43
21. Kostopoulos G, Karlos S, Kotsiantis S, Ragos O (2018) Semi-
Ethics approval Not applicable. supervised regression: a recent review. J Intell Fuzzy Syst
35(2):1483–1500
Consent for publication All listed authors approve to publish. 22. Rodrigues TA, Duarte V, Miranda RM, Santos TG, Oliveira JP
(2019) Current status and perspectives on wire and arc additive
Conflict of interest The authors declare no competing interests. manufacturing (WAAM). Materials (Basel) 12(7)

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

The International Journal of Advanced Manufacturing Technology (2022) 120:3123–3133 3133

23. Jin W, Zhang C, Jin S, Tian Y, Wellmann D, Liu W (2020) Wire 34. Barrionuevo GO, Ríos S, Williams SW, Ramos-Grez JA (2021)
arc additive manufacturing of stainless steels: a review. Appl Sci Comparative evaluation of machine learning regressors for the
24. DebRoy T, Mukherjee T, Wei HL, Elmer JW, Milewski JO (2021) layer geometry prediction in wire arc additive manufacturing.
Metallurgy, mechanistic models and machine learning in metal In 12th International Conference on Mechanical and Intelligent
printing. Nat Rev Mater 6(1):48–68 Manufacturing Technologies, pp 186–190
25. Dinovitzer M, Chen X, Laliberte J, Huang X, Frei H (2019) 35. Xia C, Pan Z, Polden J, Li H, Xu Y, Chen S (2021) Modelling and
Effect of wire and arc additive manufacturing (WAAM) process prediction of surface roughness in wire arc additive manufacturing
parameters on bead geometry and microstructure. Addit Manuf using machine learning. J Intell Manuf
26:138–146 36. Ikeuchi D, Vargas-Uscategui A, Wu X, King PC (2021) Data-
26. Dhinakaran V, Ajith J, Fahmidha AFY, Jagadeesha T, Sathish efficient neural network for track profile modelling in cold spray
T, Stalin B (2020) Wire arc additive manufacturing (WAAM) additive manufacturing. Appl Sci 11(4):1–12
process of nickel based superalloys-A review. Mater Today Proc 37. Chen Z, Wang B, Gorban AN (2020) Multivariate Gaussian and
21:920–925 Student-t process regression for multi-output prediction. Neural
27. Thapliyal S (2019) Challenges associated with the wire arc addi- Comput Appl 32(8):3005–3028
tive manufacturing (WAAM) of aluminum alloys. Mater Res 38. Baturynska I, Martinsen K (2020) Prediction of geometry devia-
Express 6(11) tions in additive manufactured parts: comparison of linear regres-
28. Huang J, Guan Z, Yu S, Yu X, Yuan W, Li N, Fan D (2020) A 3D sion with machine learning algorithms. J Intell Manuf
dynamic analysis of different depositing processes used in wire 39. Goodfellow I, Bengio Y, Courville A (2016) Deep learning adap-
arc additive manufacturing. Mater Today Commun 24:101255 tive computation and machine learning. vol 1
29. Nguyen L, Buhl J, Bambach M (2020) Continuous Eulerian tool 40. Prieditis A, Sapp S (2013) Lazy overfitting control. Lect Notes
path strategies for wire-arc additive manufacturing of rib-web Comput Sci (including Subser Lect Notes Artif Intell Lect Notes
structures with machine-learning-based adaptive void filling. Bioinformatics), vol 7988 LNAI, pp 481–491
Addit Manuf 35:101265 41. Barrionuevo G, Ramos-Grez J, Walczak M, Betancourt C (2021)
30. Ding D, Pan Z, Cuiuri D, Li H, Van Duin S, Larkin N (2016) Bead Comparative evaluation of supervised machine learning algo-
modelling and implementation of adaptive MAT path in wire and rithms in the prediction of the relative density of 316L stain-
arc additive manufacturing. Robot Comput Integr Manuf 39:32–42 less steel fabricated by selective laser melting. Int J Adv Manuf
31. Ríos S, Colegrove PA, Martina F, Williams SW (2018) Analyti- Technol
cal process model for wire + arc additive manufacturing. Addit 42. DuPont JN, Marder AR (1995) Thermal efficiency of arc welding
Manuf 21:651–657 processes. Weld. J (Miami, Fla) 74(12):406
32. Li Y, Sun Y, Han Q, Zhang G, Horváth I (2018) Enhanced beads 43. Obidigbo C, Tatman EP, Gockel J (2019) Processing parameter
overlapping model for wire and arc additive manufacturing of and transient effects on melt pool geometry in additive manufac-
multi-layer multi-bead metallic parts. J Mater Process Technol turing of Invar 36. Int J Adv Manuf Technol 104(5–8):3139–3146
252:838–848
33. Deng J, Xu Y, Zuo Z, Hou Z, Chen S (2019) Bead geometry Publisher's Note Springer Nature remains neutral with regard to
prediction for multi-layer and multi-bead wire and arc additive jurisdictional claims in published maps and institutional affiliations.
manufacturing based on XGBoost. 125–135

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]