Additive Manufacturing 35 (2020) 101355

Contents lists available at ScienceDirect

Additive Manufacturing
journal homepage: www.elsevier.com/locate/addma

Residual stresses in wire-arc additive manufacturing – Hierarchy of T

influential variables
Q. Wua,b, T. Mukherjeeb, A. Dec, T. DebRoyb,*
a
School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081, China
b
Department of Materials Science and Engineering, The Pennsylvania State University, University Park, PA 16802, USA
c
Department of Mechanical Engineering, IIT Bombay, Powai, Mumbai 400076, India

A RT ICLE INFO ABSTRACT

Keywords: Residual stresses and distortion are common serious defects in wire-arc additive manufacturing. Commercial
Wire-arc additive manufacturing thermomechanical models are often used to understand how these defects form. However, no clear mitigation
Residual stresses strategy has evolved from previous research. Identification of the hierarchy of variables that influence residual
Machine learning stresses will help to uncover practical means of mitigating this difficulty. Here we use multiple machine learning
Neural network
algorithms and a mechanistic model to rank separately both easy to measure process parameters as well as
Random forest
Delamination
thermomechanical variables that affect the evolution of stresses. We analyze 243 sets of residual stress data for
three alloys using random forest and neural network algorithms to uncover the relative influences of the vari-
ables. Both these algorithms predict residual stresses with 97 % accuracy. More important, both algorithms
provide the same hierarchical influence of process variables on stresses. The substrate preheat temperature is the
most influential variable among the process variables. Among the thermomechanical variables, the following
variables are the most influential in decreasing order of importance: the gap between the solidus and preheat
temperatures, the product of elastic modulus and the coefficient of thermal expansion, molten pool volume,
substrate rigidity, and heat input.

1. Introduction However, the aforementioned results do not identify the most im-
portant variables that influence the evolution of residual stresses. The
Wire-arc additive manufacturing (WAAM) is widely used because of roles of individual variables on the evolution of stresses in a multi-
its high deposition rate and low cost [1–3]. However, shrinkage of large variable process are often masked by the impacts of other variables and
pools of liquids during solidification and repeated heating and cooling the complexity of the process. As a result, it is difficult to uncover the
result in high residual stresses [1] and defects such as delamination, hierarchy of the important variables.
warping, buckling and dimensional inaccuracy [4–7]. Many simulta- Data-driven machine learning techniques are often beneficial to
neously occurring physical processes affect the evolution of residual establish relations among different sets of variables when phenomen-
stresses in the components. The resulting complexity of the process ological relations among them are unavailable. Machine learning (ML)
precludes any straightforward determination of variables that could be methods have been used to predict stresses in the related field of
adjusted to reduce stresses and mitigate defects. welding [13,14]. Both neural networks and neuro-fuzzy system models
The effects of individual variables on the accumulation of residual have been used to predict residual stress distribution in pipe welds. The
stresses have been investigated in several variants of AM. For example, performances of the two machine learning models have also been
preheating of the substrates is known to reduce residual stresses and compared [13]. It was also shown that the artificial neural network and
distortions in WAAM [8]. A decrease in laser power was found to re- multi-objective optimization algorithms helped to reduce the residual
duce residual stresses in components fabricated by both directed energy stresses and distortion in welding [15]. The application of support
deposition (DED) [9] and powder bed fusion (PBF) [10]. Besides, re- vector regression and neuro-evolutionary computing were also used in
sidual stresses along the scanning direction are decreased by increasing welding [16,17] based on data calculated from the finite element
scanning speed during PBF [11]. Thicker substrates are thought to in- models. Probabilistic kernel machine models were tested on experi-
crease residual stresses at the substrate-deposit interface [12]. mentally-obtained data of axial and hoop residual stresses in two

⁎
Corresponding author.
E-mail address: [email protected] (T. DebRoy).

https://doi.org/10.1016/j.addma.2020.101355
Received 6 March 2020; Received in revised form 15 May 2020; Accepted 21 May 2020
Available online 30 May 2020
2214-8604/ © 2020 Elsevier B.V. All rights reserved.
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

stainless-steel pipes [18]. These examples indicate the effectiveness of 2.1. Thermomechanical model
applying machine learning methods to predict residual stresses in
welding. However, there is a scarcity of literature in predicting stresses A finite element method based thermomechanical model is used to
in WAAM using machine learning methods. What is needed and not calculate 3D, transient temperature and residual stresses distributions.
currently available is a data-driven analysis of residual stresses in The calculation procedures have been described in detail in our pre-
WAAM to evaluate the hierarchy of the most important factors that vious publication [4] and are not repeated here. Only a few salient
affect residual stresses. features of the model are described here. The solution domain for the
Here, for the first time, we analyze data on residual stresses during thermomechanical calculations consisting of the substrate and a single-
various conditions of WAAM using data-driven machine learning. Two layer and a single-hatch deposit is shown in Fig. 2. First, temperature
hundred and forty-three sets of data on residual stresses for single layer fields are computed by solving the heat conduction equation [22]
deposits of three alloys, IN 718, SS 316 and 800H are analyzed. In the where the heat input is applied as a double ellipsoidal heat source [4].
appendix, we justify in detail why we consider the results of single layer The heat losses by convection and radiation from the surfaces of the
deposits for machine learning. First, hierarchical influences of easy to deposit and substrate are applied as boundary conditions. Temperature-
measure WAAM variables such as arc power, scanning speed, substrate dependent thermophysical and mechanical properties of the alloys used
preheat temperature, and substrate thickness, as well as alloy proper- in the calculations are provided in the supplementary document. 3D
ties, e.g., yield stress, coefficient of thermal expansion, thermal diffu- distribution of residual stresses is calculated based on the transient
sivity and Young’s modulus on residual stresses, are investigated. temperature field [23]. During the deposition, the components are
However, these process parameters are distributed over a very large clamped at four corners to resist movement which is simulated by fixing
range where the effects of individual variables are concealed by the the nodes at the four corners on the top surface of the deposit as shown
complexity of the thermomechanical behavior of the system. There are in Fig. 2. When the components cool down to the room temperature, the
complex thermomechanical variables that are better representatives of clamps are released by deactivating the restriction of the nodes at the
the complex evolution mechanism of residual stresses. The use of sev- four corners. A commercial software, Abaqus, is used for these calcu-
eral complex variables that cannot be easily measured compared to the lations [24].
individual process variables is well-recognized in the fluid flow through
a pipe. The variables such as the diameter of the pipe, average velocity, 2.2. Generation of datasets
density and viscosity of the fluid can determine whether the flow in a
pipe is turbulent or laminar. However, the flow behavior is much better Residual stresses have three primary components along x, y and z
showed by the Reynolds number than the four individual variables. axes. Besides, all three components are spatially non-uniform.
Therefore, here we separately identify five complex thermomechanical Therefore, to figure out which component of residual stresses and at
variables that influence residual stresses and delamination. They in- what location should be considered to generate the database for ML, all
clude three thermal variables, the difference between solidus and pre- three components are compared. Fig. 3 compares the three components
heat temperature, liquid pool volume, heat input, and mechanical of residual stress distributions of an IN 718 part. For all three compo-
variables, substrate rigidity, and the product of elastic modulus and nents, very high tensile stress is observed near the substrate-deposit
coefficient of thermal expansion. interface. In Fig. 3 (d), all three components are plotted along line 1
The residual stresses are obtained using a well-tested commercial (see Fig. 2) which is near the substrate-deposit interface. The stress
thermomechanical model. A neural network and a random forest-based component along y-direction (scanning direction) which is also called
machine learning algorithms are selected to predict their hierarchical the longitudinal stress is the highest among the three components. This
influences of variables that affect residual stresses. Neural network al- is primarily because the deposit mainly shrinks along the scanning di-
gorithms have a strong ability to recognize the underlying complex rection during cooling [25,26]. From Fig. 3 (d), it can also be found that
relationships between input variables and the responses [19,20]. Also, a the longitudinal stress at the substrate-deposit interface is almost con-
random forest algorithm works well with different types of input vari- stant. Therefore, longitudinal stresses (y-direction) at the center of the
ables and is good at handling non-linear parameters efficiently [21]. substrate-deposit interface (Point ‘A’ in Fig. 2) are used to generate the
Although we investigate residual stresses in WAAM of three alloys, it is database for ML. This value of residual stress is also responsible for
equally applicable for other alloys and AM variants. delamination in WAAM [4,23].
The residual stresses in WAAM are influenced by process parameters
such as arc power, scanning speed, substrate preheat temperature and
2. Methods substrate thickness as well as alloy properties [1]. Therefore, these four
aforementioned process parameters are varied for three commonly used
The methodology for this research is shown schematically in Fig. 1. alloys, IN 718, SS 316 and 800H to generate the database for ML. For
Two types of datasets are used in machine learning as inputs. First, the each of the four process parameters, three different levels are selected
raw, unprocessed WAAM variables that are easy to measure and record within the commonly used range in WAAM. Therefore, there are 34 i.e.
during the experiments and simple alloy properties are used. Second, 81 cases for each of the three alloys which generate 243 data points as
the causative variables that represent complicated thermomechanical input for machine learning. The ranges of the process parameters and
behavior of the system are calculated using a well-tested, thermo- alloy properties are provided in Table 1. Causative variables calculated
mechanical model and used in machine learning. The residual stresses using the thermomechanical model corresponding to the 243 cases are
calculated by the thermomechanical model are used as the outputs of also used as inputs to machine learning. The ranges of the causative
machine learning. The details of the thermomechanical model, data variables are also provided in Table 1. The output variable of the ma-
generation method and machine learning (ML) algorithms are de- chine learning is longitudinal residual stress at the mid-length of the
scribed below. deposit near the substrate-deposit interface. For both the raw process

2
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. 1. Schematic representation of the methodology used in this research. The symbol β represents the coefficient of thermal expansion and E is the Young’s
Modulus.

variables and causative variables, 80 % of the total of 243 data points that provides the best accuracy. In this work, two hidden layers with 32
(i.e. 194 data points) are randomly selected for training and the re- nodes in each layer and the ReLU activation function [29] are found to
maining 20 % (i.e. 49 data points) are used for testing. The random be the optimum combination. RF algorithm creates multiple decision
selection process is repeated five times and the accuracy reported in this trees as subsets and combines the outputs of them, which can help to
work is the average accuracy of these five trials. reduce the overfitting of the dataset [30]. Also, RF is good for providing
a feature-based ranking of variables [31]. In this work, the ranking of
2.3. Machine learning algorithms variables is performed by calculating the feature importance of vari-
ables which is defined as the total decrease in node impurity averaged
Two machine learning algorithms, artificial neural networks (ANN) over all trees of the random forest [32]. Therefore, feature importance
and random forests (RF) are used in this work. For the ANN algorithm, is an indicator that can evaluate the relative importance of the input
the number of hidden layers, nodes in each hidden layer and the acti- variables on the desired outputs. The larger is the feature importance,
vation function are the critical factors [27]. The randomized search the more important is the variable to the output. An open-source Py-
method [28] is used to find the optimum combination of these factors thon package “Sci-kit learn” was used in this work to implement both

3
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

ANN and RF [33]. The software package, instruction manuals, libraries

and sample cases for “Sci-kit learn” are freely available on the website
https://scikit-learn.org/stable/.

3. Results and discussions

3.1. Validation of the thermomechanical model

Since the reliability of the results from machine learning depends on

the accuracy of the residual stresses calculated using the thermo-
mechanical model, the model outputs are rigorously tested using in-
dependent experimental data. Fig. 4 shows fair agreements between the
calculated residual stresses using the thermomechanical model and the
corresponding experimentally measured [34] stresses for a single track
Fig. 2. Schematic representation of the solution domain of the thermo-
WAAM deposit of carbon steel (S235JR). The reasons for the slight
mechanical model. Line 1 represents the substrate-deposit interface at the mid-
mismatch between the computed residual stresses and the measured
width of the deposit. The point ‘A’ is at the mid-length, mid-width of the de-
posit. Deposit and substrate dimensions are indicated in the figure. Substrate
stress values may be due to the measurement error in the hole drilling
thickness is a variable input as indicated in Table 1. Scanning direction is along method and the assumptions made in the thermomechanical model.
the positive Y-axis. The agreement gives us confidence that the residual stresses calculated
using the thermomechanical model can be used for machine learning.

Fig. 3. Residual stresses of the component

shown in Fig. 2 along (a) x-direction, σxx, (b) y-
direction, σyy, and (c) z-direction, σzz, for IN
718 when the deposit cools down to room
temperature and the clamps are released. (d)
Comparison of the three components of re-
sidual stresses along line 1. The simulation was
done for 1450 W arc power, 4.5 mm/s scan-
ning speed, 293 K substrate preheat tempera-
ture and 5 mm substrate thickness.

Table 1
Ranges of raw and causative variables.
Raw variables Causative variables

Variable name Range Variable name Range

Power, W 1200, 1450, 1700 Pool volume, mm3 2.34∼301.79

Scanning speed, mm/s 3, 4.5, 6 Substrate rigidity, ×105 N mm2 2.98∼85.64
Substrate preheat temperature, K 293, 450, 600 Difference between solidus and preheat temperatures, K 933∼1365
Substrate thickness, mm 5, 10, 15 Heat input, J/mm 200∼566.67
Yield stress at room temperature, MPa 577∼937 Stress generated per unit temperature change, MPa /K 2.66∼4.19
Expansion coefficient, ×10−5 K-1 1.31∼2.19
Thermal diffusivity, mm2/s 2.50∼3.72
Young’s modulus, MPa 191∼203

4
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

process conditions and alloys generate a large dataset that is used to

train and test ANN and RF as described in Section 2. In practice, two
metrics are used to evaluate the performance of ANN and RF models,
mean absolute error (MAE) [19] and correlation coefficient (R2) score
[35]. MAE represents the mean absolute difference between the pre-
dicted stress values by ML and the stress values calculated by the
thermomechanical model. A small MAE is desirable for a good ML
model. The R2 score quantifies the deviation of data from their mean
value and a score of 1 indicates that all predicted stress values by ML
exactly match with that calculated using the thermomechanical model.
Fig. 6 shows the comparison between predicted residual stresses by
both ANN and RF and the calculated residual stresses by the thermo-
mechanical model. The figures show that all data points for training and
testing datasets of both ML algorithms are close to the diagonal line,
indicating the predicted stress values by ML agree well with the cal-
culated stress values by the model. These validation results indicate that
both ANN and RF can be used for predicting the hierarchical influence
of variables with good accuracy. ANN performs slightly better than RF
with better performance metrics values (lower MAE and higher R2
score). This is mainly because ANN is good at solving complicated
multi-scale multi-physics mathematical models and capturing the
complex relationship between inputs and outputs [19].
Running the validated ANN to generate process maps of residual
stresses for different combinations of arc power, scanning speed, sub-
strate thickness and preheat temperature is a time-efficient way since it
precludes the necessity of time-consuming thermomechanical simula-
tions. Fig. 7 shows the process maps of longitudinal residual stresses for
different combinations of the four process parameters for three alloys.
Fig. 7 (a–c) show that high arc power and slow scanning speed increase
residual stresses for all three alloys, which is consistent with the trends
shown in Fig. 5 (a–b). Also, Fig. 7 (d–f) show that thicker substrate and
low preheat temperature are favorable conditions for accumulating
high residual stresses, which is also consistent with the trends shown in
Fig. 5 (c–d). Fig. 7 (g–i) also shows that high arc power and low preheat
temperature are the ideal conditions for high stress accumulation. In
addition, all three alloys exhibit the same trend of residual stresses for
the four process parameters. High residual stresses may cause thermo-
Fig. 4. Comparison between the experimentally measured [34] and numeri- mechanical defects such as delamination in WAAM components
cally computed (a) longitudinal (along y-direction, σyy) and (b) transverse [1,2,36]. Delamination is likely to occur when the longitudinal residual
(along x-direction, σxx) residual stress profiles along the line AB which is at the stress value is higher than the room temperature yield strength and
mid-length of a single track carbon steel deposit. The line AB is on the top close to the ultimate tensile strength [1]. Therefore, it is assumed that a
surface of the substrate. The processing conditions for these simulations are component is prone to delamination if the longitudinal residual stress is
available in the corresponding literature [34]. Error bars for the experimental higher than the average value of yield strength and ultimate tensile
data are not reported in the original paper [34]. strength of the alloy at room temperature. The regions of the process
maps in Fig. 7 where delamination is likely to occur are indicated in
3.2. Hierarchy of raw process variables
orange as the background color. It is evident that for the same range of
process conditions, the three alloys have different regions of suscept-
As described in Section 2.2, a database for ML is generated by
ibility to delamination because of the differences in their thermo-
varying four WAAM variables, arc power, scanning speed, substrate
physical and mechanical properties.
thickness and preheat temperature, because these four parameters
The process maps in Fig. 7 obtained using ANN can be used to
significantly affect the residual stresses. For example, Figs. 5 (a), (b), (c)
evaluate the relative sensitivity of the four process parameters on re-
and (d) show that the longitudinal residual stresses during WAAM of an
sidual stresses. Fig. 8 shows the process maps for IN 718, as an example,
IN 718 component vary significantly with arc power, scanning speed,
where the four variables are normalized to the same scale of 0–1 for
substrate thickness and preheat temperature, respectively. Both in-
comparison. Normalization is done by using the min-max normalization
creases in arc power and decreases in scanning speed result in a larger
method [37] where a variable V is normalized to V* as, V* = (V - Vmin)
molten pool that shrinks more during solidification [1]. The shrinkage
/ (Vmax - Vmin). Vmin and Vmax are the minimum and the maximum
generates high tensile stress near the substrate-deposit interface. The
values of the variable, respectively. By comparing the slopes of the
thicker substrate has more rigidity that can restrict the deformation of
contours for two variables in each plot in Fig. 8, the relative sensitivity
the component and accumulate high residual stresses as shown in Fig. 5
of the variables to residual stresses is evaluated. For example, in Fig. 8
(c). Residual stresses usually develop during cooling until the entire
(b), when the normalized preheat temperature increases from 0.06 to
component cools down to a constant temperature [4]. With higher
0.17, the residual stress decreases from 690 MPa to 660 MPa. However,
preheat temperature, the difference between the solidus temperature of
to achieve the same 30 MPa decrease in residual stress, the normalized
alloy and the preheat temperature is smaller and, the residual stresses
arc power needs to decrease from 0.53 to 0.3. Therefore, substrate
are also smaller. Thus, residual stresses decrease with an increase in
preheat temperature is more sensitive to residual stresses than arc
preheat temperature as shown in Fig. 5 (d).
power. Similarly, by comparing the relative sensitivities of variables in
The aforementioned variations in residual stresses for different
all four plots in Fig. 8, it can be found that the residual stresses are the

5
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. 5. Distribution of residual stress along y-

direction (scanning direction) i.e. longitudinal
residual stress along line 1 (indicated in Fig. 2)
for IN 718 for different (a) arc power, (b)
scanning speed, (c) substrate thickness and (d)
substrate preheat temperature. When one
variable is varied, other variables are kept
constant at the middle values of their range.
For example, for figure (a), scanning speed of
4.5 mm/s, substrate thickness of 10 mm and
substrate preheat temperature of 450 K are
used.

Fig. 6. Comparison of output stress values: (a)

training dataset and (b) testing dataset calcu-
lated by thermomechanical model with corre-
sponding stress values predicted by neural
network model with raw variables in WAAM,
(c) training dataset and (d) testing dataset
calculated by thermomechanical model with
corresponding stress values predicted by
random forest model with raw variables in
WAAM. The diagonal lines in each plot re-
present calculated stress values match exactly
with the predicted values by ML. The training
data and testing data comprise 194 and 49
datasets, respectively. MAE in the plot re-
presents the mean absolute error of predicted
stress values by ML and the stresses values
calculated by thermomechanical model. R2
score is a statistical measure of how close the
data are to the diagonal line. R2 score is always
between 0 and 1. In general, the closer the R2
score is to 1, the better the model fits the data.

6
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. 7. Process maps showing longitudinal residual stress (MPa) contours computed using the ANN model for WAAM of (a) IN 718, (b) SS 316 and (c) 800H for
different combinations of arc power and scanning speed, for (d) IN 718, (e) SS 316 and (f) 800H for different combination of substrate thickness and substrate preheat
temperature, for (g) IN 718, (h) SS 316 and (i) 800H for different combinations of arc power and substrate preheat temperature, for (j) IN 718, (k) SS 316 and (l)
800H for different combinations of scanning speed and substrate preheat temperature. For a particular process map, when two variables are varied, other parameters
are kept constant at the middle values of their range. The processing conditions that may result in delamination are indicated by orange color in each process maps.
The delamination criteria used in this work for IN 718, SS 316 and 800H is 604.5 MPa, 430.5 MPa, 561.5 MPa, respectively.

most sensitive to the substrate preheat temperature, followed by the arc obtained by analyzing their impact on residual stresses from the process
power and substrate thickness. Although both arc power and scanning maps in Fig. 8, can also be achieved by using RF. In RF, variables are
speed equally contribute to the heat input to the system, the scanning ranked based on feature importance that evaluates the importance of
speed has a lower influence on residual stresses than arc power. This is input variables in controlling the output [32]. Fig. 9 shows that the
primarily because the effect of low heat input on residual stresses at a ranking of the four process parameters identified by RF is the same as
fast scanning speed is masked by the low exposure time due to rapid what is obtained from the ANN in Fig. 8. In addition, four alloy prop-
scanning. Therefore, the effect of heat input on residual stresses is erties e.g. yield stress, Young’s modulus, coefficient of thermal expan-
largely controlled by the variations in arc power which make arc power sion/contraction and thermal diffusivity are also found to be important
more important than scanning speed. in controlling residual stresses in WAAM.
The hierarchical importance of the process variables that are

7
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. 8. Process maps showing longitudinal re-

sidual stress (MPa) contours computed using
the ANN model for WAAM of IN 718 for dif-
ferent combinations of (a) arc power and
scanning speed, (b) arc power and substrate
preheat temperature, (c) substrate thickness
and substrate preheat temperature and (d)
scanning speed and substrate preheat tem-
perature. For a particular process map, when
two variables are varied, other parameters are
kept constant at the middle values of their
range.

multiple simultaneously occurring physical processes [1]. Therefore,

complex variables that represent both thermal and mechanical beha-
viors of the system during WAAM can be better representatives of the
evolution of residual stresses compared to the raw process variables. In
WAAM, residual stresses develop depending on the transient variation
of temperature field [4]. Fig. 10 (a) shows the temperature and corre-
sponding stress variations with time at point A (in Fig. 2) for an IN 718
deposit. As the arc heat source approaches point A, the temperature at
that location increases. The expansion of the material at that location
due to a rise in temperature is restricted by the surrounding cold ma-
terial which results in compressive stress at point A. When the heat
source moves away from the location, the temperature at point A de-
creases. The stresses primarily develop during the cooling from the
solidus to preheat temperature (Fig. 10 (a)) as the complete build is
expected to cool down uniformly from the preheat to room temperature
without evolution of any significant additional stresses. Therefore, the
difference between solidus and preheat temperature is an important
Fig. 9. Feature importance of the four WAAM parameters and alloy properties factor controlling the residual stresses. Fig. 10 (b) shows that the
calculated by the random forest model. Tpreheat is substrate preheat temperature, longitudinal residual stress at point A (in Fig. 2) for three alloys in-
P is arc power, d is substrate thickness and v is scanning speed. σ, E, β and α creases with the difference between solidus and preheat temperature. In
refer to yield stress at room temperature, Young’s modulus, expansion coeffi- addition, it should be noted that for the same temperature difference,
cient and thermal diffusivity at room temperature, respectively. The corre- the residual stresses vary significantly for different alloys. Mechanical
sponding feature importance values are indicated below the variables in the
properties of alloys such as co-efficient of thermal expansion / con-
figure. The feature importance of the variables is calculated every time for the
traction and Young’s modulus and therefore their product which is
five-times randomly selected 194 datasets from 243 datasets. The final feature
importance values of variables are averaged by the results from the five times, known as the stress developed per unit temperature change is also an
and the standard deviation is shown as an error bar. Process parameters win- important factor. Fig. 10 (c) shows that the three alloys accumulate
dows for which the feature importance of different variables are calculated in different stresses values due to the differences in their mechanical
this figure are given in Table 1. properties.
Transient temperature variation in Fig. 10 (a) which is responsible
for the evolution of residual stresses, is primarily governed by the heat
3.3. Hierarchy of thermomechanical variables
input per unit length of the deposit (arc power/scanning speed) [38].
High heat input results in both large molten pool and high peak tem-
The hierarchical importance shown in Figs. 7–9 is for raw process
parameters in WAAM. However, the evolution of residual stresses in perature, both of which are favorable for the accumulation of more
residual stresses as shown in Fig. 10 (d). In addition to the thermal
WAAM is a complex thermomechanical phenomenon that involves

8
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. 10. (a) Temperature and longitudinal

stress variations with time at point A (in Fig. 2)
for an IN 718 deposit. All processing para-
meters are selected as the mid values of their
ranges in Table 1. Effects of (b) difference be-
tween solidus and preheat temperatures, (c)
stress generated per unit temperature change,
(d) heat input, (e) substrate rigidity and (f)
pool volume on longitudinal stresses at point A
(in Fig. 2) for three alloys. Pool volumes are
calculated at the mid length of the deposit.
Stresses are calculated with substrate preheat
temperature of 293 K, 450 K and 600 K in-
dividually for each material in (b). Stresses are
calculated with arc power of 1200 W, 1450 W
and 1700 W individually for each material in
(d) and (f). Stresses are calculated with sub-
strate thickness of 5 mm, 10 mm and 15 mm
individually for each material in (e).

effect, the evolution of residual stresses in WAAM also depends on the and the calculated residual stresses by the thermomechanical model.
mechanical behavior of the system such as rigidity of the substrate. The The figures show that all data points for training and testing datasets of
rigidity of the substrate is defined by the product of Young’s modulus to both ML algorithms are close to the diagonal line, indicating the pre-
the moment of inertia [1]. The rigid substrate has more capability to dicted stress values by ML agree well with the calculated stress values
restrict the shrinkage during the solidification and cooling that results by the model. This agreement also verifies the effectiveness of the se-
in high residual stresses as shown in Fig. 10 (e). Although most of the lection of the five thermomechanical variables, based on which the
stresses develop during cooling from solidus to preheat temperature as residual stresses can be predicted accurately.
shown in Fig. 10 (a), localized solidification shrinkage of the molten Unlike raw process parameters, for which both ANN and RF are
pool also partially contributes to the development of residual stresses. used for predicting hierarchical influence, only RF is used to rank the
Therefore, the volume of the molten pool that can quantitatively in- five complex thermomechanical variables. Since the five complex
dicate the solidification shrinkage is considered as a contributing factor thermomechanical variables are inherently connected, it is not possible
to the residual stresses. Fig. 10 (f) shows that a large molten pool to vary two variables independently while other variables are kept
shrinks more during solidification and results in high residual stresses. constant. For example, an increase in arc power results in higher heat
From the aforementioned discussions, it is evident that the five input per unit length of the deposit as well as larger pool volume both of
thermomechanical variables, the difference between solidus and pre- which are causative variables responsible for residual stress accumu-
heat temperatures, the stress generated per unit temperature change, lation. Therefore, the ranking of variables by analyzing the sensitivity
heat input, substrate rigidity, and pool volume are important for re- of residual stresses to these variables based on process maps generated
sidual stresses in WAAM. These variables are calculated for different by ANN is not applicable to evaluate the hierarchical influence of five
process conditions and alloys to generate a large dataset that is used to thermomechanical variables.
train and test ANN and RF. Two metrices, mean absolute error (MAE) Fig. 12 shows that the difference between solidus and preheat
[19] and correlation coefficient (R2) score [35] are used to evaluate the temperature is the most influential factor for residual stresses. This is
performance of ANN and RF models, same as Fig. 6. Fig. 11 shows the primarily because most of the stresses develop during the cooling from
comparison between predicted residual stresses by both ANN and RF solidus to preheat temperature as explained in Fig. 10 (a). Stress

9
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. 11. Comparison of output stress values: (a)

training dataset and (b) testing dataset calcu-
lated by thermomechanical model with corre-
sponding stress values predicted by neural
network model with five complex thermo-
mechanical variables in WAAM, (c) training
dataset and (d) testing dataset calculated by
thermomechanical model with corresponding
stress values predicted by random forest model
with five complex thermomechanical variables
in WAAM. The diagonal lines in each plot re-
present calculated stress values match exactly
with the predicted values by ML. The training
data and testing data comprise 194 and 49
datasets, respectively. MAE in the plot re-
presents the mean absolute error of predicted
stress values by ML and the stresses values
calculated by thermomechanical model. R2
score is a statistical measure of how close the
data are to the diagonal line. R2 score is always
between 0 and 1. In general, the closer the R2
score is to 1, the better the model fits the data.

residual stresses depends more on the thermomechanical behavior of

the entire component compared to the localized solidification shrinkage
of the molten pool. Therefore, the pool volume is less influential on
residual stresses compared to the difference between solidus and pre-
heat temperature as shown in Fig. 12. Residual stresses depend on the
net heat content in the system, which contains both the heat input from
the arc source as well as the heat from the preheating. However, the
variable, heat input, represents the only heat available from the arc
source. In other words, for the same heat input, residual stresses can
vary largely for different preheat temperature. That makes the residual
stress least sensitive to heat input.
The identification of the five complex thermomechanical variables
and their hierarchical influence on residual stresses will substantially
advance the scientific understanding of the role of important variables
in WAAM. In addition, the hierarchical influence of the raw process
parameters will guide engineers to know which variables to adjust to
fabricate structurally sound and defect-free components.
Fig. 12. Feature importance of the causative parameters calculated by the
random forest model. The corresponding feature importance values are in- 4. Summary and conclusions
dicated below the variables. The feature importance of the variables is calcu-
lated every time for the five-times randomly selected 194 datasets from 243
We propose a unique combination of machine learning and a well-
datasets. The final feature importance values of variables are averaged by the
tested thermomechanical model to identify the most important vari-
results from the five times, and the standard deviation is shown as an error bar.
The symbol β represents the coefficient of thermal expansion and E is the ables and their hierarchical influence on residual stresses and delami-
Young’s Modulus. Process parameters windows for which the feature im- nation. Two hundred and forty-three data sets for three commonly used
portance of different variables are calculated in this figure are given in Table 1. alloys, IN 718, SS 316 and 800H are generated using a well-tested
thermomechanical model and are used in random forest and neural
network-based machine learning. Below are the specific findings:

generated per unit temperature change is the second most important

(1) Both random forest and neural network algorithms have been found to
variable primarily because the residual stresses can vary significantly
predict residual stresses and delamination in WAAM with more than 97
for different alloys for the same processing conditions. The evolution of

10
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

% accuracy. Besides, both algorithms provide the same hierarchical entire component and not only on the localized shrinkage of the
influence of WAAM process variables on stresses where preheat tem- molten pool volume. The pool volume was not as influential as the
perature is found to be the most important followed by arc power, other variables on residual stresses.
substrate thickness and scanning speed in controlling stresses. (5) Although the thermophysical properties of three alloys, IN 718, SS
(2) Using both neural network and random forest, we identified five 316 and 800H, vary within the same range of process parameters,
complex thermomechanical variables that represent the complex the variation of residual stresses shows a similar trend. However,
mechanisms of evolution of residual stresses better than the in- magnitudes of stresses are different for three alloys due to the dif-
dividual WAAM process variables. They include three thermal ferences in their thermophysical and mechanical properties. These
variables, the difference between solidus and preheat temperature, differences in stress values make their susceptibility to delamina-
liquid pool volume, heat input, and mechanical variables, substrate tion different under the same processing conditions.
rigidity, and the product of elastic modulus and coefficient of
thermal expansion. CRediT authorship contribution statement
(3) Among the five thermomechanical variables, the difference be-
tween the solidus and preheat temperature was found to be the Q. Wu: Methodology, Writing - original draft, Software, Validation.
most important parameter. This is mainly because residual stresses T. Mukherjee: Conceptualization, Software, Investigation, Writing -
originate during cooling from the solidus to preheat temperature. original draft. A. De: Investigation, Writing - review & editing. T.
As the temperature further decreases from the preheat to room DebRoy: Conceptualization, Supervision, Writing - review & editing.
temperature, the whole component cools down uniformly and no
significant additional stresses develop.
Declaration of Competing Interest
(4) The product of elastic modulus and the coefficient of thermal ex-
pansion that represents the stress developed per unit change in
temperature and volume of the molten pool are the second and None.
third most influential variables that affect residual stresses. For the
same range of cooling, residual stresses vary significantly de- Acknowledgement
pending on the material property, which makes residual stresses
highly sensitive to stress developed per unit change in temperature. Q. Wu acknowledges the support of the China Scholarship Council
Residual stresses depend on the thermomechanical behavior of the [grant number 201806030114].

Appendix A. Does the hierarchy depend on the absolute values of residual stresses?

In wire-arc additive manufacturing, generally, components are made by depositing multiple layers [4]. Residual stress values in the component
vary with the progress of the multi-layer deposition process [4]. Fig. A1 shows that the longitudinal residual stresses change with the deposition of
different layers along the build direction [39]. This is primarily due to the repeated thermal cycles experienced by the component during layer-by-
layer deposition [39]. Therefore, the residual stress values of a single layer may vary from that in a multi-layer deposit. It is important to evaluate if
the values of the first layer can be reliably used to determine the hierarchy of variables that affect the residual stresses. We do so using various
approaches.
First, it is well-established in the machine learning literature that the variables need to be non-dimentionalized [40–42] on a scale generally
between 0 and 1 for obtaining their hierarchy. This is because different variables have diverse order of magnitudes of values. In our case, all variables
including the residual stress are normalized on a scale between 0 and 1. Therefore, the hierarchy depends on the trends and not absolute values of the
residual stress which vary with the deposition of different layers.
Second, it can be shown mathematically that the hierarchy of the input variables does not depend on the magnitude of the output variable
(residual stresses). Consider the variation of the output variable residual stress (σ) as functions of two independent variables X and Y. For a particular
value of σ, if X has a higher influence on residual stress than Y, then the following expression is true:

/ >1
X Y (A1)
If the value of residual stress in a different layer is higher by Δσ, such that:
( ) ( )
= 0 and =0
X Y (A2)

Fig. A1. (a) A schematic diagram of the solu-

tion domain for the calculations showing the
line at the substrate deposit interface along
which the stresses are plotted [39]. Calcula-
tions are done for half of the build. This re-
duces the time necessary for the calculations
and the computer memory requirements for the
analysis. (b) Residual stresses [39] after de-
positing various layers along the line shown in
figure (a).

11
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

Fig. A2. Feature importance of the four WAAM parameters and alloy properties calculated by the random forest model using the new dataset explained in the
appendix. Tpreheat is substrate preheat temperature, P is arc power, d is substrate thickness and v is scanning speed. σ, E, β and α refer to yield stress at room
temperature, Young’s modulus, expansion coefficient and thermal diffusivity at room temperature, respectively. The corresponding feature importance values are
indicated below the variables in the figure. The feature importance of the variables is calculated every time for the five-times randomly selected 194 datasets from
243 datasets. The final feature importance values of variables are averaged by the results from the five times, and the standard deviation is shown as an error bar.
Process parameters windows for which the feature importance of different variables is calculated in this figure are given in Table 1.

The same hierarchy, i.e., X is more influential than Y would be valid as can be observed from the following expression.
( + ) ( + )
/ >1
X Y (A3)
Since Eq. (A3) is the same as Eq. (A1), the hierarchy of variables X and Y are unaffected by the absolute values but depends on the trends in
variations. It is well-known in the additive manufacturing literature [12] that the trends in the variations of residual stresses with variables such as
power, speed, substrate thickness, preheat temperature do not change with the number of layers deposited. Therefore, the hierarchy of variables
computed using the data of a single layer deposit is also valid for multi-layer deposits.
Third, we have repeated the calculations presented in this paper with two layers of deposition and evaluated the hierarchy using the same
procedure. In other words, we have calculated the residual stresses at the substrate deposit interface after depositing the second layer for the IN 718
deposit using 1450 W power, 4.5 mm/s scanning speed, 10 mm substrate thickness, and 450 K preheat temperature. The maximum stress value is
about 0.75 times the maximum value for the single layer deposit. Following the trend, we have made a new dataset with 243 data points and
calculated the hierarchy of raw variables and alloy properties which is shown in Fig. A2. The hierarchy of variables shown in the figure is the same as
what is provided in Fig. 9 for the single layer deposit. The hierarchy remained the same as that for a single layer deposit. In addition, the calculation
of hierarchy using the first layer deposition is computationally more efficient since the residual stress calculations are computationally intensive.

Appendix B. Supplementary data

Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.addma.2020.101355.

References during additive manufacturing, Scr. Mater. 127 (2017) 79–83.

[8] J. Xiong, Y. Lei, R. Li, Finite element analysis and experimental validation of
thermal behavior for thin-walled parts in GMAW-based additive manufacturing
[1] T. DebRoy, H.L. Wei, J.S. Zuback, T. Mukherjee, J.W. Elmer, J.O. Milewski, et al., with various substrate preheating temperatures, Appl. Therm. Eng. (2017) 126.
Additive manufacturing of metallic components - process, structure and properties, [9] M. Alimardani, E. Toyserkani, J.P. Huissoon, C.P. Paul, On the delamination and
Prog. Mater. Sci. 92 (2018) 112–224. crack formation in a thin wall fabricated using laser solid freeform fabrication
[2] T. DebRoy, T. Mukherjee, J.O. Milewski, J.W. Elmer, B. Ribic, J.J. Blecher, et al., process: an experimental–numerical investigation, Opt. Lasers Eng. 47 (2009)
Scientific, technological and economic issues in metal printing and their solutions, 1160–1168.
Nat. Mater. 18 (2019) 1026–1032. [10] J.L. Bartlett, X. Li, An overview of residual stresses in metal powder bed fusion,
[3] W. Ou, T. Mukherjee, G.L. Knapp, Y. Wei, T. DebRoy, Fusion zone geometries, Addit. Manuf. (2019).
cooling rates and solidification parameters during wire arc additive manufacturing, [11] Y. Liu, Y. Yang, D. Wang, A study on the residual stress during selective laser
Int. J. Heat Mass Transf. 127 (2018) 1084–1094. melting (SLM) of metallic powder, Int. J. Adv. Manuf. Technol. 87 (2016) 647–656.
[4] Q. Wu, T. Mukherjee, C. Liu, J. Lu, T. DebRoy, Residual stresses and distortion in [12] P. Mercelis, J.P. Kruth, Residual stresses in selective laser sintering and selective
the patterned printing of titanium and nickel alloys, Addit. Manuf. (2019) 100808. laser melting, Rapid Prototyp. J. 12 (2006) 254–265.
[5] J.D. Roehling, W.L. Smith, T.T. Roehling, B. Vrancken, G.M. Guss, J.T. McKeown, [13] J. Mathew, J. Griffin, M. Alamaniotis, S. Kanarachos, M. Fitzpatrick, Prediction of
et al., Reducing residual stress by selective large-area diode surface heating during welding residual stresses using machine learning: comparison between neural
laser powder bed fusion additive manufacturing, Addit. Manuf. 28 (2019) 228–235. networks and neuro-fuzzy systems, Appl. Soft Comput. 70 (2018) 131–146.
[6] C. Cunningham, J. Flynn, A. Shokrani, V. Dhokia, S. Newman, Invited review ar- [14] Y.D. Koo, K.H. Yoo, G.N. Man, Estimation of residual stress in welding of dissimilar
ticle: strategies and processes for high quality wire arc additive manufacturing, metals at nuclear power plants using cascaded support vector regression, Nucl. Eng.
Addit. Manuf. (2018). Technol. 49 (2017) S173857331630345X.
[7] T. Mukherjee, V. Manvatkar, A. De, T. Debroy, Mitigation of thermal distortion [15] N. Ansaripour, A. Heidari, S.A. Eftekhari, Multi-objective optimization of residual

12
Q. Wu, et al. Additive Manufacturing 35 (2020) 101355

stresses and distortion in submerged arc welding process using Genetic Algorithm of the fourteenth international conference on artificial intelligence and statistics
and Harmony search, Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. (2019) (2011) 315–323.
0954406219885977. [30] A. Liaw, M. Wiener, Classification and regression by randomForest, R news 2 (2002)
[16] J.E.R. Dhas, S. Kumanan, Evolutionary fuzzy SVR modeling of weld residual stress, 18–22.
Appl. Soft Comput. 42 (2016) 423–430. [31] L. Breiman, Random forests, Mach. Learn. 45 (2001) 5–32.
[17] J.E.R. Dhas, S. Kumanan, Neuro evolutionary model for weld residual stress pre- [32] M. Sandri, P. Zuccolotto, Analysis and correction of bias in total decrease in node
diction, Appl. Soft Comput. 14 (2014) 461–468. impurity measures for tree-based algorithms, Stat. Comput. 20 (2010) 393–407.
[18] M. Alamaniotis, J. Mathew, A. Chroneos, M.E. Fitzpatrick, L.H. Tsoukalas, [33] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, et al.,
Probabilistic kernel machines for predictive monitoring of weld residual stress in Scikit-learn: machine learning in Python, J. Mach. Learn. Res. 12 (2011)
energy systems, Eng. Appl. Artif. Intell. 71 (2018) 138–154. 2825–2830.
[19] X. Qi, G. Chen, Y. Li, G.X. Chen, C. Li, Applying neural-network-Based machine [34] H. Zhao, G. Zhang, Z. Yin, L. Wu, Three-dimensional finite element analysis of
learning to additive manufacturing: current AppliC.ations, challenges, and future thermal stress in single-pass multi-layer weld-based rapid prototyping, J. Mater.
perspectives, Engineering (2019). Process. Technol. 212 (2012) 276–285.
[20] Y. LeCun, Y. Bengio, G. Hinton, Deep learning. nature 521 (2015) 436. [35] P. Ahlgren, B. Jarneving, R. Rousseau, Requirements for a cocitation similarity
[21] V. Svetnik, A. Liaw, C. Tong, J.C. Culberson, R.P. Sheridan, B.P. Feuston, Random measure, with special reference to Pearson’s correlation coefficient, J. Am. Soc. Inf.
forest: a classification and regression tool for compound classification and QSAR Sci. Technol. 54 (2003) 550–560.
modeling, J. Chem. Inf. Comput. Sci. 43 (2003) 1947–1958. [36] T. Mukherjee, T. DebRoy, Printability of 316 stainless steel, Sci. Technol. Weld.
[22] K. Khan, A. De, Modelling of selective laser melting process with adaptive re- Join. (2019) 1–8.
meshing, Sci. Technol. Weld. Join. (2019) 1–10. [37] S. Patro, K.K. Sahu, Normalization: a preprocessing stage, arXiv preprint arXiv
[23] T. Mukherjee, W. Zhang, T. Debroy, An improved prediction of residual stresses and (2015) 150306462.
distortion in additive manufacturing, Comput. Mater. Sci. 126 (2017) 360–372. [38] T. Mukherjee, V. Manvatkar, A. De, T. Debroy, Dimensionless numbers in additive
[24] Abaqus Documentation, Version 6.14, Dassault Systems, 2015. manufacturing, J. Appl. Phys. 2017 (121) (2017).
[25] F. Martina, M. Roy, B. Szost, S. Terzi, P.A. Colegrove, S.W. Williams, et al., Residual [39] T. Mukherjee, J.S. Zuback, W. Zhang, T. DebRoy, Residual stresses and distortion in
stress of as-deposited and rolled wire+ arc additive manufacturing Ti–6Al–4V additively manufactured compositionally graded and dissimilar joints, Comput.
components, Mater. Sci. Technol. 32 (2016) 1439–1448. Mater. Sci. 143 (2018) 325–337.
[26] J. Ding, P. Colegrove, J. Mehnen, S. Ganguly, P.S. Almeida, F. Wang, et al., Thermo- [40] Y. Du, T. Mukherjee, P. Mitra, T. DebRoy, Machine learning based hierarchy of
mechanical analysis of Wire and Arc Additive Layer Manufacturing process on large causative variables for tool failure in friction stir welding, Acta Mater. 192 (2020)
multi-layer parts, Comput. Mater. Sci. 50 (2011) 3315–3322. 67–77.
[27] H.D. Beale, H.B. Demuth, M. Hagan, Neural network design, Pws, Boston (1996). [41] Y. Du, T. Mukherjee, T. DebRoy, Conditions for void formation in friction stir
[28] P. Liashchynskyi, P. Liashchynskyi, Grid search, random search, genetic algorithm: welding from machine learning, Npj Comput. Mater. 5 (2019) 1–8.
a big comparison for NAS, arXiv preprint arXiv 191206059 (2019). [42] O. Genc, A. Dag, A Bayesian network-based data analytical approach to predict
[29] X. Glorot, A. Bordes, Y. Bengio, Deep sparse rectifier neural networks, Proceedings velocity distribution in small streams, J. Hydroinformatics 18 (2016) 466–480.

13