International Journal of Fatigue 142 (2021) 105941

Contents lists available at ScienceDirect

International Journal of Fatigue

journal homepage: www.elsevier.com/locate/ijfatigue

Machine learning based fatigue life prediction with effects of additive T

manufacturing process parameters for printed SS 316L
Zhixin Zhana,b, Hua Lic,
⁎

a
School of Aeronautic Science and Engineering, Beihang University, Beijing 100191, People's Republic of China
b
Shen Zhen Institute of Beihang University, Shenzhen 518000, People's Republic of China
c
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798, Singapore

ARTICLE INFO ABSTRACT

Keywords: In aerospace engineering, many additive manufacturing (AM) metal parts subject to fatigue loadings, resulting in
Additive manufacturing their fatigue failure. Therefore, it is essential to develop an advanced approach for fatigue issues. Although some
Fatigue life prediction theoretical methods are used for fatigue analysis of AM metal parts, their implementations are time-consuming.
Machine learning models Furthermore, these methods cannot directly consider the effects of AM parameters. In this study, a platform is
Continuum damage mechanics
developed for a data-driven analysis of continuum damage mechanics (CDM)-based fatigue life prediction of AM
Stainless steel 316L
stainless steel (SS) 316L, in which the effects of AM process parameters (including laser power P, scan speed v,
hatch space h, powder layer thickness t) are considered. Here, three typical ML models: an artificial neural
network (ANN), a random forest (RF), and a support vector machine (SVM), are trained effectively by a database
produced by the CDM technique, and then further comparisons are made between the predicted results and
published experimental data to verify the proposed platform. Finally, detailed parametric studies using the ML
models are conducted to investigate some of the significant characteristics.

1. Introduction on the fatigue performance. In addition, the low-and high-cycle fatigue

resistance [16], notch fatigue behavior [17], the fatigue life and mi-
As is well known, additive manufacturing (AM) has many more crostructure [18] of AM Ti6Al4V were studied, and the relationship
advantages than the conventional fabrication process [1–4], particu- between microstructure, material cyclic properties, and component
larly for the manufacture of extremely complex geometric parts or performance were investigated [19]. In terms of AM AlSi10Mg parts,
customized products. In general, in aerospace engineering, numerous the defect-based modelling [20], fatigue life prediction [21], and fa-
AM metal parts frequently bear complex fatigue loadings, sometimes tigue strength assessment [22] were conducted. Moreover, the effects of
resulting in fatigue failure. Therefore, it is certainly essential to develop HIP treatment [23], thermal post-processing [24], and surface rough-
an advanced approach for fatigue life prediction. ness [25] on microstructure and fatigue properties of AM AlSi10Mg
The literature review reveals that the studies of fatigue issues are were elucidated. Studies were also carried out on the very-high-cycle
performed experimentally, theoretically and numerically for AM metal fatigue behavior [26], microstructural damage and fracture mechan-
parts. Three kinds of typical metal materials in aerospace engineering isms [27] of AM AlSi10Mg.
include SS 316L, Ti6Al4V and AlSi10Mg. Previously, for AM SS316L In general, several methods are used for fatigue life prediction, and
parts, the effects of AM processing parameters were carried out on the each has its own benefits and drawbacks. For instance, experiment-
fatigue and fracture behavior of laser powder bed fusion SS 316L [5], based statistical techniques require enormous amounts of fatigue data,
and the high cycle fatigue and ratcheting interaction [6] and some and sometimes this is problematic. The critical plane approach [28,29],
predictive models were studied for the fatigue property of SS 316L [7]. which is based on a semi-empirical equation, is usually used for the life
Furthermore, the influence of microstructure [8], porosity, surface prediction of components bearing fatigue loadings. However, it lacks
roughness [9], post-treatments [10] and heat treatment [11] were the precise physical relevance of the damage parameters. Fracture
discussed on the fatigue behavior of AM SS 316L. Regarding AM mechanics [30,31] are extensively used for the prediction of crack
Ti6Al4V parts, the influence of manufacturing defects [12], surface propagation in structures with small initial defects. However, it is not
roughness [13], HIP [14] and heat treatments [15] were investigated useful in analyzing crack initiation problems. In addition, the

⁎
Corresponding author.
E-mail address: [email protected] (H. Li).

https://doi.org/10.1016/j.ijfatigue.2020.105941
Received 15 July 2020; Received in revised form 29 August 2020; Accepted 5 September 2020
Available online 10 September 2020
0142-1123/ © 2020 Elsevier Ltd. All rights reserved.
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

continuum damage mechanics (CDM) technique [32–34], which char- solved numerically to predict the fatigue lives with different AM
acterizes the material degradation process by damage variables, is parameters and fatigue loadings, and the predicted results are used for
employed extensively for fatigue life prediction in engineering appli- training the ML models. Subsequently, ANN, RF, and SVM models are
cations. Although the above techniques are helpful in the analysis of employed for fatigue life prediction, and the predicted fatigue lives and
fatigue in metals, their implementations are time-consuming. Further- experimental results are compared to verify this novel method. Detailed
more, these methods cannot directly account for the effects of AM parametric studies for different ML models are then conducted to in-
parameters, such as laser power, scan speed and powder layer thick- vestigate some of the significant characteristics.
ness. These parameters are strictly related to the AM fatigue behavior
[5–7,35–37]. Therefore, it is necessary to propose a new effective 2. CDM-based equations
method to address the current issues.
In addition, machine learning (ML) is an advanced technique to 2.1. Damage-coupled constitutive relationship
characterize a problem with numerous uncertainties. It is quite popular
in the prediction of complex mechanical problems, such as the simu- According to the CDM theory [66,67], for materials with isotropic
lation of electric-thermal–mechanical breakdown of polymer-based di- properties, the damage variable D can represent the stiffness degrada-
electrics [38], establishment of the structure–mechanical stability re- tion of the representative volume element (RVE), as defined by
lations of metal–organic frameworks [39], optimal design of new
E ED
functional and customizable composites with very limited training data D= ,
(1)
E
[40], prediction of the mechanical properties of concrete [41] and
composite materials [42], damage detection [43] and mode identifi- where E is the undamaged Young’s modulus and ED represents the
cation [44], and structural health monitoring [45]. Currently, artificial damaged one. D varies from 0 to 1 when ED decreases from E to 0.
neural networks (ANNs), random forests (RFs), and support vector For materials with a small deformation, the total strain, ij , is ex-
machines (SVMs) are commonly used in the prediction of fatigue da- pressed as the sum of the elastic strain, ije , and the plastic strain, ijp ,
mage and mechanical properties. As a non-linear model having nu- e p
ij = ij + ij (2)
merous interconnected neurons, the ANN technique is frequently used
for highly non-linear problems, such as the assessment of the remaining The elastic strain is obtained as follows:
fatigue life of in-service road bridge decks [46], fatigue loading analysis
1+
with the effect of mean stress [47], fatigue damage prediction of vortex- e
ij =
ij kk ij
,
induced vibration [48] and shot peened mild carbon steels [49]. In E 1 D E 1 D (3)
comparison, the RF approach is used for the issues of classification or where v represents the Poisson's ratio and ij is the stress component. p
regression and the RF model is typically trained by bagging and random represents the plastic strain rate,
variable selection. It has been employed to fatigue strength analysis of
steel alloys [50], fatigue life prediction of rubber under constant am- p f
= ,
plitude stress [51], modelling the uniaxial compressive strength of ij (4)
lightweight self-compacting concrete [52] and mechanical properties of where represents the plastic multiplier. Here, the non-linear kine-
cemented paste backfill [53]. For SVM method, it is mainly based on matic hardening law is adopted [68], and the von Mises yield function,
the statistical learning theory and the structural risk minimization f, can be obtained by
principle. It is often used to solve classification, forecasting, and re-
gression problems, such as the remaining useful life estimation of air- 3 Sij Sij
f= ( ij )( ij ) Q
craft engines [54], fatigue life prediction for vibration isolation rubber 2 1 D 1 D (5)
[55], analysis of compressive strength of high volume fly ash self- M
compacting concrete [56], and compaction property assessment of in which Sij represents the deviatoric stress and ij = (k )
ij represents
mixed gangue backfill materials [57]. Although several studies have the deviatoric back stress. The rate of the yield surface radius can be
k=1

been conducted to investigate fatigue via ML models trained by an represented as

experimental database, to our best understanding of the three ML
models, no studies have been performed to compute the fatigue damage Q = l (O Q), (6)
of AM parts by CDM-based ML models. where l and O are material parameters. The rate of plastic strain can be
The model material for this research is SS 316L, which is a tough, obtained by
ductile, and high corrosion resistant austenitic stainless steel [58], and
it was developed for use in nuclear reactors thirty years ago in Europe p 3 Sij /(1 D) ij
= ,
[59]. It is known that the SS 316L is widely used in engineering ap- 21 D (Sij /(1 D) ij )eq (7)
plications such as aerospace [60], construction, automobile [61], 2 p p
marine equipment, and biomedical industries [62], due to its reason- p= 3 ij ij
= 1 D
, and , (8)
able cost, ease of fabrication, biocompatibility, and strong mechanical (k ) 2 p (k )
= (1 D)( Tk ij p ),
strength [58]. Furthermore, the current AM technologies (such as the ij
3
ij k (9)
selective laser melting [63], laser metal deposition [64] and laser
in which p represents the accumulated plastic strain rate, and Tk and
powder bed fusion [65]) could provide wide platforms for the better k
are parameters.
and further uses of the SS 316L. In this study, a platform is developed
for the data-driven analysis of the CDM-based fatigue damage analysis
of AM stainless steel 316L by considering several significant AM para- 2.2. Fatigue damage evolution laws
meters. These include laser power P, scan speed v, hatch space h,
powder layer thickness t, and fatigue loadings such as maximum stress For uniaxial fatigue loadings, the fatigue damage evolution equation
σmax and stress ratio R. As the first step in the data-driven analysis of is in the form of [69]:
fatigue life prediction, CDM-based equations are derived for damage- dD + 1] a
coupled constitutive relationship and damage evolution laws. Fol- D= = [1 (1 D) ,
dN m (1 n m )(1 D) (10)
lowing this, the calibration for material parameters in the theoretical
equations is presented. Second, the above CDM-based equations are in which N represents the number of cycles, a represents the stress

2
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

amplitude, and m is the mean stress. , , and m are material para- models, and the following are the steps involved:
meters. (1) Parameter initialization in established models. For example, the
For multiaxial fatigue loadings, the fatigue damage evolution damage variable and fatigue life are set as zero.
equation [70] is expressed as (2) Fatigue damage computation. To reduce the computation time, the
dD AII jump-in-cycle approach is employed for the computational stress and
strain fields. Subsequently, the damage increment with the fatigue
D= = [1 (1 D) + 1] [ ]
dN m (1 n H , m )(1 D) (11)
loading cycles, ΔN, can be computed by the formula,
in which AII is the amplitude of the octahedral shear stress, expressed as
AII
1 3 1/2 D (k + 1) = [1 (1 D) + 1] · · N.
AII = (Sij,max Sij,min ) 2 , m (1 n H , m )(1 D) (17)
2 2 (12)
Thus, the total fatigue damage regarding the loading cycles is ac-
where Sij represents the deviatoric stresses, and the mean stress is
H ,m
quired as follows:
in the form of
D (k + 1) = D (k ) + D (k + 1) . (18)
1
= [max(tr ( )) + min(tr ( ))]
H ,m
6 (13) (3) Fatigue crack initiation judgment. When the total fatigue damage
where tr ( ) = 11 + 22 + 33. 11, 22 , and 33 represent the principal within materials reaches 1, the fatigue crack is initiated, and the fatigue
stresses. life is obtained. Otherwise, the material properties are recomputed by
Given the conditions of constant fatigue loadings, the number of the following formula:
cycles, N, can be computed by integrating Eqs. (10) and (11) from E (k + 1) = E (k ) (1 D(k + 1) ). (19)
D = 0 to D = 1.
For uniaxial fatigue loadings, Subsequently, a reanalysis of the stress, strain and damage fields is
conducted until the determination criterion of crack initiation is at-
m a tained.
N=
(1 )(1 + ) (1 n m) (14)
3.3. Fatigue data acquisition for AM stainless steel 316L parts
For multiaxial fatigue loadings,

m AII To realize a relatively accurate ML model, it is extremely important

N= to train the ML model comprehensively with sufficient data. In this
(1 )(1 + ) (1 n H , m) (15)
study, the database for the training is composed of six input parameters
and one output parameter. The inputs include AM process parameters
3. Fatigue data acquisition for AM stainless steel 316L parts by (laser power P, scan speed v, hatch space h, and powder layer thickness
CDM–based numerical computation t) and fatigue loadings (maximum stress σmax and stress ratio R). The
output is the predicted fatigue life, N. The database consists of more
Stainless steel 316L, which possesses excellent fatigue properties, is than thousand sets of data to train the ML model for AM stainless steel
a commonly used metal in engineering applications ranging from body 316L parts.
implants to heat exchangers. This section presents acquisition of the
fatigue data for AM stainless steel 316L parts. 4. Machine learning models

3.1. Calibration for material parameters in CDM-based equations In this section, the three typical ML models (ANN, RF, and SVM)
used for the fatigue life prediction of AM stainless steel 316L parts are
Here, the calibration of the material parameters is briefly presented, presented in detail.
for which some experimental results [5-7] are employed. First, the
material parameters in the elastic–plastic constitutive equations can be 4.1. Artificial neural network model
calibrated from uniaxial tensile test data. In this study, the non-linear
kinematic hardening characteristic of AM metallic materials is con- It is well known that an ANN is an ML algorithm inspired by the
sidered, and three sets of back stresses are employed in the stress–strain brain neural network [71], and the accuracy of the predicted results by
relationship below. an ANN depends strongly on the training process. In this study, the
3 multi-layer perceptron (MLP) artificial neural network is employed. The
Tk
= y + (1 e k p), schematic diagram of an MLP network is plotted in Fig. 1. The optimal
k=1 k (16) architecture for achieving better results at a lower computational cost is
where y represents the yield stress. The least-squares approach is em- presented in Table 1. The network is composed of three types of layers
ployed for the calibration of the parameters. (input, hidden, and output layers), and each type of layer has a few
Second, the four material parameters ( , , m , n ) are calibrated in neurons. The behavior of an individual neuron i could be represented as
the fatigue damage evolution equations in Eqs. (14) and (15). Para- follows [46]:
meters and n and the term, m /[(1 )(1 + )], can be determined
yi = fi ( wij x j ti ) (20)
using the fatigue data of smooth specimens with a stress ratio of
R = 0.1. Similarly, the least-squares approach is preferred. Subse- The inputs x j are multiplied by the weights wij . When the sum of the
quently, a CDM-based finite element (FE) approach is used to calibrate weighed inputs wij xj are higher than a threshold ti , the output of the
the other two parameters, and m , numerically. neuron yi is computed by using an activation function fi . The activation
function is a mathematical “gate” between the input feeding the current
3.2. Computational approach for fatigue life prediction neuron and its output going to the next layer. In this study, the sigmoid
1
function is employed as the activation function, f (x ) = 1 + e x . This
Here the computational approach is briefly demonstrated for the life function could ensure that all values passing to the next neuron lie in
prediction of AM metal parts subjected to fatigue loadings. The core of the range between zero and one. The initial weights and the thresholds
this approach lies in the implementations of CDM-based theoretical are set to random values and are subsequently revised by the gradient

3
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 1. Schematic of an MLP neural network with three types of layers (input, hidden, and output layers), in which xk represents the laser power P, scan speed v, hatch
space h, and powder layer thickness t, maximum stress σmax and stress ratio R, and y represents the predicted fatigue life.

Table 1 initialized.
Optimized architecture of the MLP neural network. (2) The data for the training are collected and pre-processed. The
Network Number
data pre-processing is the process of cleaning and transforming raw
data prior to processing and analysis. In this work, the data pre-pro-
Neurons of the input layer 6 cessing includes standardizing data formats, enriching source data, and
Hidden layers 3 removing outliers. It is the first and crucial step while creating an ar-
Neurons in the 1st hidden layer 20
tificial neural network model. Subsequently, the data are input into the
Neurons in the 2nd hidden layer 10
Neurons in the 3rd hidden layer 5 MLP neural network to acquire the outputs.
Neurons in the output layer 1 (3) For a predicted output, the error of the output ei (o) means the
error between the actual output yi (o) and the predicted value yi pre (o) ,
that is, ei (o) = yi (o) yi pre (o). And the global error e (o) is defined as the
descent method in the training process to minimize the output errors. error of all the predicted outputs, which is computed according to the
In this study, the commonly used back-propagation algorithm and following equation
gradient-descent optimization are employed to train the MLP neural n n
network effectively. The process [72,73] is summarized as follows: 1 1
e (o ) = ei (o) 2 = (yi (o) yi pre (o)) 2
(1) All the weights and thresholds in the MLP neural network are 2 i=1
2 i=1 (21)

Fig. 2. Diagram of the RF regression structure with bootstrap samples and trees, in which Snk is the training samples selected by bootstrap aggregating algorithm and
Ypre represents the estimated value.

4
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 3. Structure of the SVM model with three types of layers, in which xn
represents the inputs (laser power P, scan speed v, hatch space h, and powder
layer thickness t, maximum stress σmax and stress ratio R), K(x, xn) represents
the kernel function and f(x) represents the predicted fatigue life.

Fig. 6. Variations in the predicted results by the SVM against the experimental
data [5-7] for AM stainless steel 316L under fatigue loads ranging from
145 MPa to 660 MPa and stress ratio R of 0.1, in which seven sets of data
beyond the three-error band, and the coefficient of determination R2 is 0.482
and the normalized mean-squared error MSE is 1.

Fig. 4. Variations in the predicted results by the ANN against the experimental
data [5-7] for AM stainless steel 316L under fatigue loads ranging from
145 MPa to 660 MPa and stress ratio R of 0.1, in which two sets of data are
beyond the three-error band, and the coefficient of determination R2 is 0.602
and the normalized mean-squared error MSE is 0.769.

Fig. 7. Comparison of the predicted performances (coefficient of determination

R2 and mean-squared error MSE) of different ML (ANN, RF and SVM) models
for AM stainless steel 316L.

Table 2
Predicted performances of the ANN models with different numbers of hidden
layers.
ANN Number of neurons Metrics

1st layer 2nd layer 3rd layer 4th layer R2 MSE

Model 1 20 0.57 1
Model 2 20 10 0.596 0.939
Model 3 20 10 5 0.602 0.921
Model 4 20 10 10 5 0.59 0.954

where yi (o) represents the actual output or the exact result of the
training samples and yi pre (o) is the predicted result by the trained MLP
Fig. 5. Variations in the predicted results by the RF against the experimental neural network in the training procedure.
data [5-7] for AM stainless steel 316L under fatigue loads ranging from
(4) The values of the weights and thresholds are revised based on
145 MPa to 660 MPa and stress ratio R of 0.1, in which all the 70 sets of data are
the gradient descent method using the following equations:
located within the three-error band, and the coefficient of determination R2 is
0.942 and the normalized mean-squared error MSE is 0.112. e (o )
wij (o + 1) = wij (o) lr
wij (o) (22)

5
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 8. Variations in the predicted life by the different ANN models against the experimental data [5–7] for AM stainless steel 316L: (a) Model 1 with one hidden
layer, (b) Model 2 with two hidden layers, (c) Model 3 with three hidden layers, (d) Model 4 with four hidden layers. It is observed that only two sets of the predicted
data by Model 3 are outside the three-error band, which are better than that by the other ANN models.

Table 3 An RF generates hundreds of decision trees, and the average value

Predicted performance of the ANN models with different numbers of neurons. of the outputs from all decision trees is computed as the final predicted
ANN Number of neurons Metrics
result. A typical decision tree is composed of leaf nodes and decision
nodes, which are used to estimate the input data according to the test
1st layer 2nd layer 3rd layer R2 MSE functions. Specifically, the core of the training process lies in the con-
struction of uncorrelated decision trees.
Model 1 10 10 5 0.56 1
Model 2 15 10 5 0.581 0.956
The bootstrap aggregating algorithm [75] is frequently employed
Model 3 20 10 5 0.602 0.901 for training the RF regression model. The process for randomly col-
Model 4 30 10 5 0.599 0.912 lecting samples is called bootstrap. The training set is
Sn = {(X1, Y1), , (Xn , Yn)} , where X is the input vector with m variables
X = {x1, x2, , xm} , and Y is the output scalar. The input data are op-
e (o ) erated to split at every node, and the function, f (X , Sn ) , is established
ti (o + 1) = ti (o) lr
ti (o) (23) at the end of the training stage. The bootstrap aggregating algorithm
selects some samples (Sn1, , Snk ) and constructs k prediction
where lr is the learning rate associated with the convergence and sta-
trees f (X , Sn1), , f (X , Snk ) , and thus, k outputs
bility of the training procedure , and the value of lr is 0.01.
Y1pre = f (X , Sn1), , Ykpre = f (X , Snk ) are obtained. Subsequently, the
(5) If the global error of the predicted results exceeds the allowed
estimated value, Y pre , is computed by the following formula [76]:
value, the previous steps are repeated, and the weights and thresholds
are re-adjusted. The process repeats until the error is within a reason- k k
1 1
able range from 0.001 to 0.01. Y pre = Yipre = f (X , Snk )
k i=1
k i=1 (24)

4.2. Random forest model where Yipre is the output of the ith tree.
Furthermore, three important parameters are adjusted: number of
As a typical supervised ML algorithm, an RF combines a non-linear trees ntree , number of variables n var at each node, and maximum depth
statistical method with an ensemble learning strategy [74]. Currently, of the tree ndep . It is known that every node (excluding a root) in a tree is
the RF has not been used for the fatigue life prediction of AM stainless connected by a directed edge from exactly one other node, and each
steel 316L parts. In addition, comparisons of RFs with other ML models node can be connected to the arbitrary number of nodes. The depth of
have not been performed regarding this issue. tree is defined as the number of edges from the root node to the leaf

6
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 9. Variations in the predicted life by the different ANN models against the experimental data [5–7] for AM stainless steel 316L: (a) Model 1 with ten neurons, (b)
Model 2 with fifteen neurons, (c) Model 3 with twenty neurons, (d) Model 4 with thirty neurons. It is observed that the predicted accuracy increases as the number of
neurons increases, and there is no apparent difference when the number of neurons is over 20.

Table 4
Architectures for the ANN models with different number of neurons.
ANN Number of neurons ANN Number of neurons

1st layer 2nd layer 3rd layer 1st layer 2st layer 3st layer

Model 1 20 10 5 Model 6 20 10 5
Model 2 20 15 5 Model 7 20 10 10
Model 3 20 20 5 Model 8 20 10 15
Model 4 20 25 5 Model 9 20 10 20
Model 5 20 30 5 Model 10 20 10 25

node of the tree. The deeper the depth of tree, the more splits it has and no further splits, the regression tree is considered full grown.
it captures more information about the data. Generally, the predicted (3) Data prediction is conducted for the new inputs by the trained RF
accuracy improves with an increase in the number of trees, ntree . n var is a model.
sensitive parameter of the RF model, and it is found that a better pre-
dicted performance is achieved if n var is equal to one-third of the
4.3. Support vector machine model
number of input variables. In this study, ntree = 100, n var = 2 ,
andndep = 20 are employed to attain a good balance between the pre-
For the SVM regression problem, the training set is composed of the
dicted accuracy and computational cost. The diagram of the RF re-
input vector, X = {x1, , xn} , and the output vector, Y = {y1, , yn } . In
gression structure is shown in Fig. 2, and the process for the RF re-
this study, X is a six‐dimension variable including laser power P, scan
gression [77] is summarized as follows:
speed v, hatch space h, powder layer thickness t, maximum stress σmax
and stress ratio R. Y is a dependent variable that represents the fatigue
(1) Bootstrap sample extraction is performed from the training set.
life of the AM stainless steel 316L part. The relationship is established
(2) Regression trees are grown with the extracted bootstrap samples by
by an SVM between the output and input vectors using the following
the following operations. At each node, a suitable split is chosen
formula:
from the randomly sampled input variables, which are the adjust-
ment parameters for the RF regression algorithm. When there are f (x ) = · (x ) + b (25)

7
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

in which kp represents the kernel parameter, and the value of kp is 0.1.

The structure of SVM is plotted in Fig. 3.

5. Results and discussions

5.1. Predicted fatigue lives of AM stainless steel 316L parts by novel method

Stainless steel 316L is a commonly used metal material in aero-

nautics engineering, and the fatigue damage behavior have been in-
vestigated by many studies. Based on the proposed novel method, the
fatigue lives of AM 316L parts are predicted and then compared with
the experimental data from open publications. Fatigue experiments [5-
7] were performed on an MTS machine, in which the stress ratio, R, was
equal to 0.1, and the maximum stress, σmax, ranged from 145 MPa to
660 MPa. All the predicted fatigue lives against the experimental data
are shown in Figs. 4–6. A total of 70 sets of data are computed, and
most of the computed results agree well with the experimental data. It
is also found that all the predicted results by the RF model are located
within the three-error band. However, two sets of data by the ANN
model and seven sets of data by the SVM model are beyond the three-
Fig. 10. The predicted metric versus ANN models with different numbers of
error band. Therefore, it is effective to employ a novel method to the
neurons.
life prediction of AM stainless steel 316L parts in engineering applica-
tions.
Table 5 Furthermore, two metrics: coefficient of determination R2 and
Variable descriptions of the RF models. mean-squared error MSE, are used to evaluate the predicted perfor-
Number of training data 107, 215, 430, 644, 1074 mance of the ML models. The R2 and MSE are formulated as follows:
Number of trees 10, 20, 50, 100, 200 n
Number of sensitive parameters 1, 2, 3, 4, 6 (yi yi pre )2
Depth of trees 2, 5, 8, 10, 20 i=1
R2 (y , y pre ) =1 n and
(yi ymean ) 2
i=1 (30)
where f (x ) represents the regression function, is the weight coeffi-
cient, b is the bias term, and (x ) is the non-linear function. 1
n
MSE (y, y pre ) = (yi yi pre )2 ,
To evaluate the coefficients, and b, the SVM is formulated as a n (31)
i=1
constrained optimization problem [78],
where yi is the ith experimental data, yi is the ith predicted value, and
pre
n
min
1 2 +C + ymean represents the mean value of the experimental data. The best value
i
for R2 is 1.0, and for MSE, the smaller the value, the better the predicted
i
2 i=1
result. The computed MSE values are normalized between 0 and 1 to
yi · (x i ) b + i make the results easily understandable. Fig. 7 shows the plot of the
subject to , (26)in which C is the penalty
· (x i ) + b yi + i predicted performances of the different ML models, and it is observed
i 0, i 0 that the RF model performs best in the fatigue life prediction of AM
parameter, and the value of C is 100. i and i are relaxation factors. is stainless steel 316L parts.
the error precision, and the value of is 0.01. A Lagrangian function is
then employed to transform Eq. (26) into a dual optimization problem 5.2. Parametric studies of novel method
[79],
n n
This section describes the parametric studies performed on the
1 different ML models to further investigate the fatigue lives of AM
max ( i i )( j j )· K (x i , x j )
2 i=1 j=1 stainless steel 316L parts.
n n
( i + i )+ yi ( i i ) 5.2.1. Parametric studies for life prediction using the ANN model
i=1 i=1 In this subsection, the effects of the number of hidden layers and
n
neurons on the predicted fatigue lives of AM stainless steel 316L parts
( i i )=0 are discussed in detail.
subject to , (27)where K (x i , xj ) = (x i ) (xj ) re-
i=1 First, the sensitivity of the ANN model is studied for the variation in
[0, C ] i, i the number of hidden layers ranging from 1 to 4. Concurrently, 20
presents the kernel function, and i and i are the Lagrangian multi- neurons are set for the first hidden layer. The detailed architectures for
pliers. By solving the above dual optimization problem, Eq. (25) is the different ANN models are summarized in Table 2. It is also noted
expressed as follows: that all ANN models are trained by the same database. Fig. 8 shows the
n plot of the predicted results against the experimental data, where al-
f (x ) = ( i i )· K (x , x i) + b. most all the predicted results from the different ANN models lie in the
i=1 (28) three-error band. It is also observed that only two sets of the predicted
Owing to the excellent non-linear property in a high-dimension data by the ANN model with three hidden layers are found outside the
space, the radial basis function is regularly used as follows: three-error band. They are better than the results predicted by the other
ANN models. Furthermore, the predicted performances for the different
(29)
K (x , x i) = exp ( kp x i xj 2 ),
ANN models as listed in Table 2. It is also verified that the ANN model

8
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 11. Variations in the predicted accuracy with different parameters of the RF models for AM stainless steel 316L: values of metrics versus (a) the number of
training data, (b) the number of trees, (c) the number of sensitive parameters, (d) the depth of trees.

with three hidden layers is the optimal architecture. The computed MSE The best value for R2 is 1.0, and for MSE, the smaller the value, the
values are normalized between 0 and 1 to make the results easily un- better the predicted result. The computed MSE values are normalized
derstandable. between 0 and 1 to make the results easily understandable. It is seen
Second, the sensitivities of the ANN models to the variation in that the R2 increases and MSE decreases as the number of neurons in
neuron number are studied. Concurrently, three hidden layers are the second or in the third hidden layer increases. For the ANN models
proposed for all the models. Typically, the predicted accuracy is more with different number of neurons in the second hidden layer, there is no
dependent on the number of neurons in the first hidden layer than those apparent difference in the predicted performances when the number of
in the other layers. Thus, the number of neurons in the first hidden layer neurons is greater that 20. In terms of the ANN models with different
ranges from 10 to 30, and the number of neurons in the second and number of neurons in the third hidden layer, the greater the number of
third hidden layers remain unchanged. The detailed architectures for neurons, the higher the predicted accuracy, and there is no apparent
different ANN models are summarized in Table 3, and the same data- difference when the number of neurons is greater than 15.
base is employed to train all the ANN models. Fig. 9 shows the plot of Based on the aforementioned discussions, some points are con-
the predicted fatigue lives versus the experimental data, and it is shown cluded for fatigue life prediction by the ANN model. First, there is no
that most of the results predicted by the ANN models are consistent apparent difference in the predicted accuracy if the number of hidden
with the experimental data. Furthermore, the predicted performance is layers exceeds 2. Second, the greater the number of neurons in the first
listed in Table 3, according to which the predicted accuracy increases as hidden layer, the higher the predicted accuracy. Third, the designed
the number of neurons increases, and there is no apparent difference values of the number of hidden layers and the number of neurons in the
when the number of neurons is over 20. first hidden layer should be over 2 and 20, respectively, to effectively
Third, the sensitivities of the ANN models to the variation in the improve the predicted accuracy of the ANN models.
number of neurons on other hidden layers (except first layer) are stu-
died. Concurrently, three hidden layers are proposed for all the models, 5.2.2. Parametric studies for life prediction using the RF model
and the number of neurons in the first hidden layer remain unchanged. This sub-section discusses the life prediction of AM stainless steel
Here, the number of neurons in the second hidden layer ranges from 10 316L conducted by different RF models, where two important features
to 30, and in the third hidden layer ranges from 5 to 25. The detailed were investigated in detail. They include: (a) variations in the predicted
architectures for different ANN models are summarized in Table 4, and accuracy against number of training data ndata , number of trees ntree ,
the same database is employed to train all the ANN models. The pre- number of sensitive parameters n var , and depth of trees ndep , (b) varia-
dicted performances versus different ANN models are plotted in Fig. 10. tions in the predicted fatigue lives against ndata , ntree , n var , and ndep . All

9
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 12. Variations in the predicted life with different parameters of the RF models for AM stainless steel 316L: (a) the predicted fatigue life versus the number of
training data, (b) the predicted fatigue life versus the number of trees, (c) the predicted fatigue life versus the number of sensitive parameters, (d) the predicted
fatigue life versus the depth of trees.

the variable descriptions of the RF models are listed in Table 5. fatigue loadings (σmax = 444 MPa and R = 0.1), the curves of the
Fig. 11 shows a plot of the variations in the predicted accuracy with predicted fatigue lives under different parameters of the RF models are
different variables. The computed MSE values are normalized between plotted in Fig. 12. For the given AM parameters and fatigue loadings,
0 and 1 to make the results easily understandable. First, the values of the experimental fatigue life is 377,050, which is also plotted in Fig. 12.
the metrics against the number of training data, ndata , are shown in It is seen that the minimum error between the predicted fatigue life and
Fig. 11(a), in which R2 increases and MSE decreases as the number of experimental data is attained if the ndata = 430, ntree = 20, n var = 2 ,
training data, ndata , increases. In general, the predicted accuracy de- and ndep = 10 are employed. It is noted that the predicted fatigue life is
creases as the number of training data, ndata , decreases. As such, it is short when the number of training data, ndata , is small, as shown in
necessary to have a sufficiently large number of training data, ndata , Fig. 12(a). Furthermore, a similar changing law for the predicted fa-
such as more than 500, to achieve highly accurate predicted results. tigue life versus the number of trees, ntree , is shown in Fig. 12(b), and
Second, the values of the metrics against the number of trees, ntree , are the converged value is obtained when the number of trees, ntree , is over
plotted in Fig. 11(b) when ndata , n var , and ndep remain constant. It is 50. Moreover, the curve of the predicted fatigue life versus the number
found that R2 changes slightly, and the MSE decreases by approxi- of sensitive parameters, n var , is shown in Fig. 12(c), in which all the
mately 20% when the number of trees, ntree , increases from 10 to 100, predicted fatigue lives are located between 415,000 and 430,000. Thus,
and thus, ntree = 100 , is suggested. Third, Fig. 11(c) demonstrates the the number of sensitive parameters, n var , has little influence on the
values of the metrics versus the number of sensitive parameters, n var , predicted results. Finally, the effect of the depth of trees, ndep , on the
when ndata , ntree , and ndep remain unchanged. It is observed that R2 and predicted fatigue life is presented in Fig. 12(d), according to the pre-
the MSE can achieve the convergence if the number of sensitive para- dicted fatigue life decreases slightly as the depth of trees, ndep , increases
meters, n var , is over 2. Finally, the values of the metrics versus the depth when the depth is greater than 2.
of trees, ntree , are plotted in Fig. 11(d), where ndata , ntree , and n var remain It is concluded that both the number of training data, ndata , and the
constant. It is observed that the changes of the R2 and MSE will be large, depth of trees, ndep , significantly influence the predicted accuracy for
if the depth of trees, ntree, is smaller than 8. Furthermore, the predicted the AM stainless steel 316L parts. The values of ndata 600 and
accuracy could be higher if the ntree is greater than 8. Therefore, it is ndep 5are recommended to achieve accurate predicted results. In ad-
important to make the depth of trees ntree larger than 8 to accurately dition, the predicted fatigue life is slightly influenced by the number of
predict the fatigue lives of AM stainless steel 316L parts. sensitive parameters, n var , and the predicted accuracy of the RF model is
To further understand the predicted fatigue life with the given AM much higher when the number of trees ntree is over 50.
parameters (P = 195 W, v = 1084 mm/s, h = 90 µm, t = 20 µm) and

10
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

Fig. 14. Variations in the predicted life by the SVM models for AM stainless
steel 316L with the (a) laser power, (b) powder layer thickness, and (c) max-
Fig. 13. Variations in the (a) predicted accuracy, (b) predicted life, and (c) imum stress.
relative error with the number of training data of the SVM models for AM
stainless steel 316L.
subject to different AM parameters and fatigue loadings. Here, R2 in-
creases and the MSE decreases rapidly as the number of training data,
5.2.3. Parametric studies for life prediction using the SVM model ndata , increases from 400 to 600. When the number, ndata , exceeds 600,
Using the SVM model, the fatigue lives of AM stainless steel 316L the variation in the predicted accuracy is small, and there is even a
parts are predicted to investigate two important aspects. They include slight decrease. Therefore, the increase in the number of training data,
(a) variations in the predicted accuracy and predicted fatigue life with ndata , does not imply an improvement in the predicted accuracy.
the number of training data, ndata , and (b) variations in the predicted Fig. 13(b) shows a plot of the variations in the predicted results against
fatigue life with the AM parameters and fatigue loadings. the number of training data, ndata , and Fig. 13(c) is a plot of the relative
Fig. 13(a) is a plot of the variation in the predicted accuracy with error between the predicted fatigue life and related experimental data
the number of training data, ndata , for the AM stainless steel 316L parts subject to the same AM parameters and fatigue loadings. It is noted that

11
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

the minimum relative error is obtained when the number of training elastoplastic response of AM SS 316L are not considered, which
data, ndata , is equal to 600. could also have important effects on the fatigue behavior. In the
To further understand the predicted fatigue life with different AM future work, the in-depth analysis on microstructures and heat
parameters and fatigue loadings, the variations in the predicted lives treatment for AM materials will be carried out. Furthermore, SS
with laser power P and powder layer thickness t are plotted in 316L is used as the model material in this study, and the established
Fig. 14(a–b). Fig. 14(c) shows a plot of the variation curves of the method will be employed to investigate the fatigue behavior of
predicted life under different maximum stresses. As illustrated in other commonly-used AM alloy materials in the area of aerospace
Fig. 14(a), the fatigue life increases slightly as the laser power, P, in- engineering, such as AM Ti6Al4V and AM AlSi10Mg.
creases from 100 W to 200 W, such that the effect of the laser power on
the improvement of fatigue life is not apparent. However, the increase Declaration of Competing Interest
in the powder layer thickness, t, from 20 µm to 80 µm leads to a sig-
nificant decrease in the fatigue life from 300,000 to 30,000, as shown in The authors declare that they have no known competing financial
Fig. 14(b). Furthermore, the predicted results are also highly sensitive interests or personal relationships that could have appeared to influ-
to the maximum stress for the AM stainless steel 316L parts, as shown in ence the work reported in this paper.
Fig. 14(c). Thus, it is important to design suitable service fatigue
loadings to ensure safety in engineering applications. Acknowledgements
Several conclusions are summarized for fatigue life prediction by
the SVM model. The authors sincerely acknowledge the support from Basic and
Applied Basic Research Foundation of Guangdong Province (No.
(1) The predicted accuracy is dependent on the number of training 2019A1515110334).
data, ndata . The decrease in the relative error becomes rapid when
the number of training data, ndata , increases from 400 to 600, but References
becomes insignificant when ndata is larger than 600.
(2) The powder layer thickness has a larger impact on the fatigue life [1] Mueller B. Additive manufacturing technologies–Rapid prototyping to direct digital
than the laser power. Specifically, it is more efficient for improving manufacturing. Assembly Automation 2012;32(2).
the fatigue performance by decreasing the powder layer thickness [2] Frazier WE. Metal additive manufacturing: a review. J Mater Eng Perform
2014;23(6):1917–28.
than that by modifying the laser power. [3] Gu DD, Meiners W, Wissenbach K, et al. Laser additive manufacturing of metallic
(3) The fatigue life of an AM stainless steel 316L part is extremely components: materials, processes and mechanisms. Int Mater Rev
sensitive to the maximum stress. 2012;57(3):133–64.
[4] Wycisk E, Solbach A, Siddique S, et al. Effects of defects in laser additive manu-
factured Ti-6Al-4V on fatigue properties. Physics Procedia 2014;56:371–8.
6. Conclusion [5] Zhang M, Sun CN, Zhang X, et al. Fatigue and fracture behaviour of laser powder
bed fusion stainless steel 316L: influence of processing parameters. Mater Sci Eng, A
2017;703:251–61.
In this study, a platform is developed for a data-driven analysis of [6] Zhang M, Sun CN, Zhang X, et al. High cycle fatigue and ratcheting interaction of
CDM-based fatigue life prediction of AM stainless steel 316L, and is laser powder bed fusion stainless steel 316L: fracture behaviour and stress-based
validated by comparing the predicted results with published experi- modelling. Int J Fatigue 2019;121:252–64.
[7] Zhang M, Sun CN, Zhang X, et al. Predictive models for fatigue property of laser
mental data. Moreover, the fatigue lives are studied in detail to identify
powder bed fusion stainless steel 316L. Mater Des 2018;145:42–54.
the important features of different ML models. Finally, parametric [8] Elangeswaran C, Cutolo A, Muralidharan GK, et al. Microstructural analysis and
studies of the novel method are conducted to analyze the variations in fatigue crack initiation modelling of additively manufactured 316L after different
the predicted accuracy and fatigue lives with different variables in the heat treatments. Mater Des 2020;108962.
[9] Solberg K, Guan S, Razavi SMJ, et al. Fatigue of additively manufactured 316L
ML models. Accordingly, several conclusions are drawn, which are stainless steel: the influence of porosity and surface roughness. Fatigue Fract Eng
summarized below. Mater Struct 2019;42(9):2043–52.
[10] Elangeswaran C, Cutolo A, Muralidharan GK, et al. Effect of post-treatments on the
fatigue behaviour of 316L stainless steel manufactured by laser powder bed fusion.
(1) Compared with the ANN and SVM models, the RF model having the Int J Fatigue 2019;123:31–9.
highest accuracy performs the best in the fatigue life prediction of [11] Fergani O, Bratli Wold A, Berto F, et al. Study of the effect of heat treatment on
AM stainless steel 316L parts. In addition, the results obtained by fatigue crack growth behaviour of 316L stainless steel produced by selective laser
melting. Fatigue Fract Eng Mater Struct 2018;41(5):1102–19.
the ANN model are better than those obtained by the SVM model. [12] Hu YN, Wu SC, Withers PJ, et al. The effect of manufacturing defects on the fatigue
(2) For the ANN model, the increase in the number of neurons in the life of selective laser melted Ti-6Al-4V structures. Mater Des 2020;108708.
first hidden layer increases the predicted accuracy. Furthermore, [13] Pegues J, Roach M, Williamson RS, et al. Surface roughness effects on the fatigue
strength of additively manufactured Ti-6Al-4V. Int J Fatigue 2018;116:543–52.
three hidden layers and twenty neurons in the first hidden layer are
[14] Masuo H, Tanaka Y, Morokoshi S, et al. Influence of defects, surface roughness and
recommended for excellent fatigue life prediction of AM stainless HIP on the fatigue strength of Ti-6Al-4V manufactured by additive manufacturing.
steel 316L parts. Int J Fatigue 2018;117:163–79.
[15] Yu H, Li F, Wang Z, et al. Fatigue performances of selective laser melted Ti-6Al-4V
(3) For the RF model, the number of training data, ndata , and depth of
alloy: Influence of surface finishing, hot isostatic pressing and heat treatments. Int J
trees ndep affect the predicted fatigue life. The values of ndata 600 Fatigue 2019;120:175–83.
and ndep 5are recommended to improve the predicted accuracy. [16] Benedetti M, Fontanari V, Bandini M, et al. Low-and high-cycle fatigue resistance of
In addition, the predicted accuracy becomes much higher if the Ti-6Al-4V ELI additively manufactured via selective laser melting: Mean stress and
defect sensitivity. Int J Fatigue 2018;107:96–109.
number of trees ntree exceeds 50. [17] Razavi SMJ, Ferro P, Berto F, et al. Fatigue strength of blunt V-notched specimens
(4) Regarding the SVM model, the relative error decreases rapidly with produced by selective laser melting of Ti-6Al-4V. Theor Appl Fract Mech
the increase in the number of training data, ndata , from 400 to 600. 2018;97:376–84.
[18] Denti L, Bassoli E, Gatto A, et al. Fatigue life and microstructure of additive man-
Furthermore, the fatigue life of an AM stainless steel 316L part is ufactured Ti6Al4V after different finishing processes. Mater Sci Eng, A
sensitive to the maximum stress and powder layer thickness. Thus, 2019;755:1–9.
it is efficient to decrease the thickness of powder layer to improve [19] Molaei R, Fatemi A, Sanaei N, et al. Fatigue of additive manufactured Ti-6Al-4V,
Part II: The relationship between microstructure, material cyclic properties, and
the fatigue performance. component performance. Int J Fatigue 2020;132:105363.
(5) Limitation in this study and the outlook for future research. In the [20] Romano S, Brückner-Foit A, Brandão A, et al. Fatigue properties of AlSi10Mg ob-
current work, the ML based approach is developed for fatigue life tained by additive manufacturing: Defect-based modelling and prediction of fatigue
strength. Eng Fract Mech 2018;187:165–89.
prediction of AM SS 316L. However, the potential effects of mi-
[21] Tang M, Pistorius PC. Fatigue life prediction for AlSi10Mg components produced by
crostructures and different heat treatment cycles on the

12
Z. Zhan and H. Li International Journal of Fatigue 142 (2021) 105941

selective laser melting. Int J Fatigue 2019;125:479–90. [52] Matin SS, Farahzadi L, Makaremi S, et al. by random forest. Appl Soft Comput
[22] Beretta S, Gargourimotlagh M, Foletti S, et al. Fatigue strength assessment of” as 2018;70:980–7.
built” AlSi10Mg manufactured by SLM with different build orientations. Int J [53] Qi C, Chen Q, Fourie A, et al. An intelligent modelling framework for mechanical
Fatigue 2020;105737. properties of cemented paste backfill. Miner Eng 2018;123:16–27.
[23] Schneller W, Leitner M, Springer S, et al. Effect of HIP treatment on microstructure [54] Chen Z, Cao S, Mao Z. Remaining useful life estimation of aircraft engines using a
and fatigue strength of selectively laser melted AlSi10Mg. J Manufact Mater Process modified similarity and supporting vector machine (SVM) approach. Energies
2019;3(1):16. 2018;11(1):28.
[24] Bagherifard S, Beretta N, Monti S, et al. On the fatigue strength enhancement of [55] Liu Q, Shi W, Chen Z. Fatigue life prediction for vibration isolation rubber based on
additive manufactured AlSi10Mg parts by mechanical and thermal post-processing. parameter-optimized support vector machine model. Fatigue Fract Eng Mater Struct
Mater Des 2018;145:28–41. 2019;42(3):710–8.
[25] du Plessis A, Beretta S. Killer notches: the effect of as-built surface roughness on [56] Azimi-Pour M, Eskandari-Naddaf H, Pakzad A. Linear and non-linear SVM predic-
fatigue failure in AlSi10Mg produced by laser powder bed fusion. Additive tion for fresh properties and compressive strength of high volume fly ash self-
Manufacturing, 2020: 101424. compacting concrete. Constr Build Mater 2020;230:117021.
[26] Qian G, Jian Z, Qian Y, et al. Very-high-cycle fatigue behavior of AlSi10Mg man- [57] Li B, Yan H, Zhang J, et al. Compaction property prediction of mixed gangue
ufactured by selective laser melting: Effect of build orientation and mean stress. Int backfill materials using hybrid intelligence models: A new approach. Constr Build
J Fatigue 2020;105696. Mater 2020;247:118633.
[27] Awd M, Siddique S, Walther F. Microstructural damage and fracture mechanisms of [58] Lodhi MJK, Deen KM, Greenlee-Wacker MC, et al. Additively manufactured 316L
selective laser melted Al-Si alloys under fatigue loading. Theor Appl Fract Mech stainless steel with improved corrosion resistance and biological response for bio-
2020;106:102483. medical applications. Addit Manuf 2019;27:8–19.
[28] Susmel L, Taylor D. A critical distance/plane method to estimate finite life of not- [59] Zhong Y, Rännar LE, Liu L, et al. Additive manufacturing of 316L stainless steel by
ched components under variable amplitude uniaxial/multiaxial fatigue loading. Int electron beam melting for nuclear fusion applications. J Nucl Mater
J Fatigue 2012;38:7–24. 2017;486:234–45.
[29] Walat K, Łagoda T. Lifetime of semi-ductile materials through the critical plane [60] Yakout M, Elbestawi MA, Veldhuis SC. A study of thermal expansion coefficients
approach. Int J Fatigue 2014;67:73–7. and microstructure during selective laser melting of Invar 36 and stainless steel
[30] Zerbst U, Ainsworth RA, Beier HT, et al. Review on fracture and crack propagation 316L. Addit Manuf 2018;24:405–18.
in weldments–A fracture mechanics perspective. Eng Fract Mech 2014;132:200–76. [61] Coelho LB, Kossman S, Mejias A, et al. Mechanical and corrosion characterization of
[31] Previati G, Kaliske M. Crack propagation in pneumatic tires: Continuum mechanics industrially treated 316L stainless steel surfaces. Surf Coat Technol
and fracture mechanics approaches. Int J Fatigue 2012;37:69–78. 2020;382:125175.
[32] Ferjaoui A, Yue T, Wahab MA, et al. Prediction of fretting fatigue crack initiation in [62] Kaliaraj GS, Vishwakarma V, Kirubaharan AMK. Biocompatible zirconia-coated 316
double lap bolted joint using Continuum Damage Mechanics. Int J Fatigue stainless steel with anticorrosive behavior for biomedical application. Ceram Int
2015;73:66–76. 2018;44(8):9780–6.
[33] Malcher L, Mamiya EN. An improved damage evolution law based on continuum [63] Kong D, Ni X, Dong C, et al. Bio-functional and anti-corrosive 3D printing 316L
damage mechanics and its dependence on both stress triaxiality and the third in- stainless steel fabricated by selective laser melting. Mater Des 2018;152:88–101.
variant. Int J Plast 2014;56:232–61. [64] Zhang Y, Zhang J. Modeling of solidification microstructure evolution in laser
[34] Amiri M, Arcari A, Airoldi L, et al. A continuum damage mechanics model for pit-to- powder bed fusion fabricated 316L stainless steel using combined computational
crack transition in AA2024-T3. Corros Sci 2015;98:678–87. fluid dynamics and cellular automata. Addit Manuf 2019;28:750–65.
[35] Read N, Wang W, Essa K, et al. Selective laser melting of AlSi10Mg alloy: Process [65] Eo DR, Park SH, Cho JW. Inclusion evolution in additive manufactured 316L
optimisation and mechanical properties development. Mater Des 2015;65:417–24. stainless steel by laser metal deposition process. Mater Des 2018;155:212–9.
[36] Fayazfar H, Salarian M, Rogalsky A, et al. A critical review of powder-based ad- [66] Darabi MK, Al-Rub RKA, Little DN. A continuum damage mechanics framework for
ditive manufacturing of ferrous alloys: Process parameters, microstructure and modeling micro-damage healing. Int J Solids Struct 2012;49(3–4):492–513.
mechanical properties. Mater Des 2018;144:98–128. [67] Ayoub G, Naït-Abdelaziz M, Zaïri F, et al. Fatigue life prediction of rubber-like
[37] Zhan Z, Li H, Lam KY. Development of a novel fatigue damage model with AM materials under multiaxial loading using a continuum damage mechanics approach:
effects for life prediction of commonly-used alloys in aerospace. Int J Mech Sci effects of two-blocks loading and R ratio. Mech Mater 2012;52:87–102.
2019;155:110–24. [68] Wali M, Chouchene H, Said LB, et al. One-equation integration algorithm of a
[38] Shen ZH, Wang JJ, Jiang JY, et al. Phase-field modeling and machine learning of generalized quadratic yield function with Chaboche non-linear isotropic/kinematic
electric-thermal-mechanical breakdown of polymer-based dielectrics. Nat Commun hardening. Int J Mech Sci 2015;92:223–32.
2019;10(1):1–10. [69] Zhan Z, Hu W, Meng Q, et al. Continuum damage mechanics-based approach to the
[39] Moghadam PZ, Rogge SMJ, Li A, et al. Structure-mechanical stability relations of fatigue life prediction for 7050–T7451 aluminum alloy with impact pit. Int J
metal-organic frameworks via machine learning. Matter 2019;1(1):219–34. Damage Mech 2016;25(7):943–66.
[40] Gu GX, Chen CT, Buehler MJ. De novo composite design based on machine learning [70] Shen F, Hu W, Meng Q. A damage mechanics approach to fretting fatigue life
algorithm. Extreme Mech Lett 2018;18:19–28. prediction with consideration of elastic–plastic damage model and wear. Tribol Int
[41] Chaabene WB, Flah M, Nehdi ML. Machine learning prediction of mechanical 2015;82:176–90.
properties of concrete: critical review. Constr Build Mater 2020;260:119889. [71] Gope D, Gope PC, Thakur A, et al. Application of artificial neural network for
[42] Chen CT, Gu GX. Machine learning for composite materials. MRS Commun predicting crack growth direction in multiple cracks geometry. Appl Soft Comput
2019;9(2):556–66. 2015;30:514–28.
[43] Regan T, Beale C, Inalpolat M. Wind turbine blade damage detection using su- [72] Artero-Guerrero JA, Pernas-Sánchez J, Martín-Montal J, et al. The influence of la-
pervised machine learning algorithms. J Vib Acoust 2017;139(6). minate stacking sequence on ballistic limit using a combined Experimental/FEM/
[44] Xu D, Liu PF, Li JG, et al. Damage mode identification of adhesive composite joints Artificial Neural Networks (ANN) methodology. Compos Struct 2018;183:299–308.
under hygrothermal environment using acoustic emission and machine learning. [73] Ahmad MW, Mourshed M, Rezgui Y. Trees vs Neurons: Comparison between
Compos Struct 2019;211:351–63. random forest and ANN for high-resolution prediction of building energy con-
[45] Bao Y, Chen Z, Wei S, et al. The state of the art of data science and engineering in sumption. Energy Build 2017;147:77–89.
structural health monitoring. Engineering 2019;5(2):234–42. [74] Belgiu M, Drăguţ L. Random forest in remote sensing: A review of applications and
[46] Fathalla E, Tanaka Y, Maekawa K. Remaining fatigue life assessment of in-service future directions. ISPRS J Photogramm Remote Sens 2016;114:24–31.
road bridge decks based upon artificial neural networks. Eng Struct [75] Zhang D, Zhou X, Leung SCH, et al. Vertical bagging decision trees model for credit
2018;171:602–16. scoring. Expert Syst Appl 2010;37(12):7838–43.
[47] Durodola JF, Ramachandra S, Gerguri S, et al. Artificial neural network for random [76] Rodriguez-Galiano V, Sanchez-Castillo M, Chica-Olmo M, et al. Machine learning
fatigue loading analysis including the effect of mean stress. Int J Fatigue predictive models for mineral prospectivity: An evaluation of neural networks,
2018;111:321–32. random forest, regression trees and support vector machines. Ore Geol Rev
[48] Wong EWC. A simplified method to predict fatigue damage of TTR subjected to 2015;71:804–18.
short-term VIV using artificial neural network. Adv Eng Softw 2018;126:100–9. [77] Hassan AR, Siuly S, Zhang Y. Epileptic seizure detection in EEG signals using tun-
[49] Maleki E, Unal O, Kashyzadeh KR. Fatigue behavior prediction and analysis of shot able-Q factor wavelet transform and bootstrap aggregating. Comput Methods
peened mild carbon steels. Int J Fatigue 2018;116:48–67. Programs Biomed 2016;137:247–59.
[50] Agrawal A, Choudhary A. An online tool for predicting fatigue strength of steel [78] Kang S, Cho S. Approximating support vector machine with artificial neural net-
alloys based on ensemble data mining. Int J Fatigue 2018;113:389–400. work for fast prediction. Expert Syst Appl 2014;41(10):4989–95.
[51] Liu Q, Shi W, Chen Z. Rubber fatigue life prediction using a random forest method [79] Lei C, Deng J, Cao K, et al. A comparison of random forest and support vector
and nonlinear cumulative fatigue damage model. J Appl Polym Sci machine approaches to predict coal spontaneous combustion in gob. Fuel
2020;137(14):48519. 2019;239:297–311.

13