Arch Appl Mech (2021) 91:223–239

https://doi.org/10.1007/s00419-020-01765-5

O R I G I NA L

H. Ersen Balcıoğlu · Ahmet Çağdaş Seçkin

Comparison of machine learning methods and finite element

analysis on the fracture behavior of polymer composites

Received: 27 April 2020 / Accepted: 24 August 2020 / Published online: 9 September 2020
© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract In recent years, it became possible to use different methods for the analysis of mechanical systems
with the help of computers to learn like humans and by increasing their interaction with the world by observ-
ing autonomously. One of these mechanical analyzes is the fracture mechanics in which the behavior of the
laminated composites having a crack is examined. In this study, experimental methods, finite element analysis
(FEA) and machine learning algorithms (MLA) were used to analyze the fracture behavior of polymer compos-
ites in Mode I, Mode I/II and Mode II loading situations. For the experimental study, the fracture behaviors of
the laminated composites reinforced with pure glass, pure carbon and glass/carbon hybrid knitted fabrics were
tested with the help of Arcan test apparatus. In the finite element method, the linear elastic fracture behavior
at the crack tip was analyzed by using the J-integral method. In the field of MLA, there is no single learning
algorithm that provides good learning on all real-world problem data. Therefore, algorithm selection is done
experimentally so various machine algorithms were used in the study. The analysis result showed that the finite
element analysis and machine learning results were in good agreement with experimental measurements. This
study is particularly important for the comparison of machine learning techniques with FEA in regression
applications.

Keywords Fracture toughness · Knitting fabric · Pure and hybrid composite · Finite element analysis ·
Machine learning algorithms

1 Introduction

Also, to be lightness, fiber-reinforced composite materials have high tensile strength and elastic modulus in
the fiber direction, superior impact resistance, and high fatigue life. In the last three decades, the laminated
composites have gained immense popularity in different weight-sensitive engineering branches including
engineering applications. Laminated composite materials, which used especially in the aerospace sector, have
many application areas such as building, automotive, and sports equipment. Tough environmental conditions,
uneven stress distribution and structural damages arise from production cause unexpected mechanical failures.
One of the most common types of failure encountered during production, assembly or use in composite
structures is the cracking of the composite structure. For this reason, it is necessary to know the fracture
behavior and fracture damage mechanism of the laminated composites which used as a construction material.

H. E. Balcıoğlu (B)
Department of Mechanical Engineering, Usak University, Uşak, Turkey
E-mail: [email protected]

A. Ç. Seçkin
Department of Computer Engineering, Adnan Menderes University, Aydın, Turkey

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

224 H. E. Balcıoğlu, A. Ç. Seçkin

Many experimental methods have been developed to determine fracture behavior of the laminated composite
having crack and studies have been published using them.
Edge crack is a common type of damage in composite structure. Thusly, the evaluation of edge crack
fracture behavior under different loading modes (Mod I, Mode II, Mode III, or mixed modes) is one of the
basic stage in predicting the damage of laminated composite materials. Several researcher has conducted on
the Mode I, Mode II, and Mode I/II fracture behavior of composite materials. The most useful methods to
characterize fracture toughness are double cantilever beam (DCB), end notched flexure (ENF), compact tension
shear (CTS) and the mixed-mode bending (MMB) test methods [1–4]. In addition, finite element analyses were
performed to determine compliance and fracture energy of composites by modeling end-loaded split (ELS)
and single leg bending (SLB) specimens [5,6]. Also, some studies are available in the literature on the fracture
behavior of single edge crack (SEC), double edge crack (DEC) and central crack (CC) specimen for laminated
composites [7]. However, crack growth in the laminated composite is usually not a result of opening (Mode I)
or shearing (Mode II) loading, so it is important that the fracture toughness is known for mixed-mode (mode
I/II) loading. The Arcan test can be used to measure fracture toughness over a wide range of mixed-mode I/II
ratios including pure mode I and pure mode II [8]. Hasanpour and Choupani [9] have presented a study about
an application of fracture mechanics to determine mixed-mode fracture properties of rock using the modified
Arcan test specimen. Alizadeh and Soares [10] have studied mixed-mode inter-laminar fracture behavior of
glass/vinylester composites under different loading angles by using Arcan test fixtures. Gao et al. [11] have
investigated mixed-mode cracking behavior of cold recycled mixes that have single edge crack, by using Arcan
test configuration. Alfonso et al. [12] have used the modified Arcan test fixture to determine the strength of
bonded assemblies for automotive applications. Xavier et al. [13] have measured the nonlinear shear behavior
of bovine cortical bone by coupling the Arcan test with digital image correlation.
Every event encountered in nature is tried to be understood with the help of the laws of physics and
mathematics. Each event can be expressed largely with the help of algebraic, differential or integral equations.
With the development of computer technology, the behavior of structural systems under load can be modeled by
using mathematical equations or computer algorithms. Nowadays, finite element analysis (FEA) is a frequently
used computational approach to analyze engineering construction systems under various loads. Useful for
problems with complicated geometries, loadings, and material properties where analytical solutions cannot
be obtained. The FEA was introduced in 1950 in the aircraft and aerospace industry for stress analysis. Many
studies are using the finite element method in fracture mechanics problems. Vales et al. [14] have used the 2D
Arcan finite element model to characterize the adhesive of bonded assemblies under combined and dynamic
loading. Song et al. [15] have reviewed the development and the application of the scaled boundary finite
element method for fracture analysis. Choupani [16] have presented a study on the mixed-mode fracture
behavior in woven carbon fiber/polyetherimide composite material based on numerical analyses. Gruben et
al. [17] have studied ductile fracture toughness of dual-phase steel numerically. They used the Arcan fracture
test results to validate the accuracy of the numerical model. Meite et al. [18] have performed a finite element
analysis to determine the mixed-mode energy release rate and local stress field characterization for orthotropic
wood material. Majidi et al. [19] have employed extended finite element model to determine mixed-mode
fracture behavior of key-hole notched polystyrene specimens. Mladensky and Rizov [20] have investigated
nonlinear interlaminar fracture of polymer composite beam under mode II crack loading conditions using the
J -integral approach. LeBlanc and LePlante [21] have investigated mixed-mode I/II delamination growth in a
carbon/epoxy composite as experimental and numerical. They found that the results obtained from the FEA
converge the experimental results with an average error of 10.32%. Hadavinia et al. [22] have compared the
results of interlaminar fracture toughness of carbon fiber composites, which obtained from experimental and
FEA methods. They observed 8.33% error percentage between real and numerical methods.
Although the finite element method is an effective method for modeling the behavior of structures under
any loading, there are some difficulties in the application. Modeling of materials having complex geometry and
internal structure such as nonlinear geometry, nonlinear mechanical properties, and the functional graded sub-
structure system is time-consuming. Expensive finite element software and a computer having high computing
power are needed for the solution of the finite element of complex models. The mechanical strength values of
the material must be known to define the material properties of the finite element model. Therefore, extra tests
should be performed to determine the mechanical properties of the material. When analyzing with FEA, it is
necessary to create a separate model for each dimensional parameter examined and this process can be done
by an expert and it is exhausting. These necessities create a disadvantage in terms of time, cost, manpower,
and equipment.

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

Comparison of machine learning methods… 225

For two decades, artificial intelligence (AI) researchers have carried out some work to developed new
techniques and applied them to solve engineering problems. MLA is a sub-branch of (AI) that provides sys-
tems the ability to learn and improve from data without being explicitly programmed. MLA is used especially
for applications where a clear equation is not generated from the experimental data [23,24]. Lots of stud-
ies about MLA techniques and approaches are available in mechanical engineering. Nasiri et al. [25] have
presented a review article on the state of the art of AI methods such as artificial neural networks, Bayesian
network, genetic algorithm, fuzzy logic, and case-based reasoning, which are used in the field of fracture
mechanics, was surveyed. Wang et al. [26] have investigated the fatigue crack growth calculation method is
proposed based on machine learning algorithms such as extreme learning machine (ELM), radial basis function
network (RBFN) and genetic algorithms optimized backpropagation network (GABP). Rovinelli et al. [27]
have utilized Bayesian network and machine learning techniques to identify relevant micromechanical and
microstructural variables that influence the direction and rate of the fatigue crack propagation. Balcioglu et al.
[28] have employed six different artificial neural network algorithms such as Levenberg–Marquardt, BFGS
quasi-Newton method, Resilient backpropagation, Scaled conjugate gradient, One step secant, and Bayesian
regularization to predict the failure loads of bonded pultruded composite specimens. Atilla et al. [29] have
used the Levenberg–Marquardt backpropagation algorithm to predict buckling load and the natural frequency
of laminated composites.
When the studies that have been done in the literature were examined, it was seen that the finite element
method and artificial intelligence algorithms were used for the mechanical behavior of composite materials. In
this study, the fracture toughness of pure (glass/epoxy and carbon/epoxy) and intra-ply hybrid knitting (glass–
carbon/epoxy) fabric reinforced composite plates have been investigated, by using experimental, numeric
and machine learning algorithms. In the experimental study, fracture toughness of pure and hybrid laminated
composites were investigated under different loading angle such as 0◦ (Mode I), 30◦ , 45◦ , 60◦ (Mode I/II),
and 90◦ (Mode II ). Besides, the effects of different knitting pattern width (50 mm, 25 mm, and 12.5 mm) on
fracture toughness were observed experimentally during the hybridization process. In the finite element studies,
fracture toughness values of composites were analyzed using the J-integral method. Also, the fracture behavior
of stratified composites was modeled by using machine learning algorithms such as k near neighbors (KNN),
decision tree (DT), random forest (RF), and support vector machine (SVM).

2 Materials and methods

In this study, reinforcement fabrics, which have a 1x1 rib knitting structure, were knitted by using a V-bed
semi-automatic knitting machine as pure and hybrid fabrics. Hybrid fabrics, which were knitted by using 50%
carbon and 50% glass fiber and having widths of 12.5 mm, 25 mm, and 50 mm, were used for hybrid structures.
Knitted fabric laminated composite materials having 4 laminae were produced by hand lay-up methods. Two-
component matrix material (DTE1000 resin and DTS 1105 hardener) was obtained from DURATEK companies
in Turkey. Epoxy matrices material were prepared by mixing resin and hardener 74/26 in weight, respectively.
After all, laminas were saturated with epoxy resin, semi-product laminated composites were cured at 100◦ C
under the pressure of 8 MPa for 100 min, by using temperature–time–pressure controlled hydraulic press.
Since the Arcan test specimen is in a butterfly form, it has notches on both sides. As the butterfly form was
created by using a milling cutter with a 4 mm diameter. After cutting, crack having an average 4 mm length
that parallel to course direction was created on the Arcan test sample by using a jigsaw, which had 0.6 mm
diameter. To investigate the effects of the crack location on the fracture behavior of hybrid composite samples,
the crack was created separately on the glass and carbon side (Fig. 1).
Fracture tests were performed by using U-Test brand mark universal tensile test machine with a capacity
of 50 kN at a 1 mm/min deformation rate. In the case of Mode I loading, the load applied tries to open the
crack tip, causing damage to the sample. In Arcan fracture tests, Mode I loading was applied as parallel to wale
direction. Also, in the case of Mode II, the applied load shear the crack tip, causing damage. Similarly, Mode
II loading was applied as parallel to the course direction. Thus, the stress concentration under load occurs at
the crack end and the crack advances, finally the sample fractured and is damaged. In some loading situations,
both Mode I and Mode II effects appear at the crack tip simultaneously. The fracture toughness obtained in
this case is called mixed-mode Mode I/II (Fig. 2).
The theory of linear elastic fracture mechanics (LEFM) has been used to the investigate of fracture behaviors
in fiber reinforced composite materials. In this study, the Arcan test fixtures which allow opening mode (Mode
I), shearing mode (Mode II), and mixed-mode (Mode I/II), was used for fracture test of specimens. The opening

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

226 H. E. Balcıoğlu, A. Ç. Seçkin

Fig. 1 Pure and hybrid cracked Arcan test specimens

Fig. 2 The components of load at fracture (PC ) for the Mode I/II in the Arcan test specimen

mode occurs when the loading angle, α is 0◦ , shearing mode occurs when α is 90◦ and when loading angle
between 0◦ and 90◦ , mixed-mode occurs. The values of fracture toughness for the opening mode (mode I),
and shearing mode (mode II) were found using the following formulae. The critical stress intensity factor or
fracture toughness in the crack tip is denoted by KIC and KIIC , in mode I and mode II loading angle, respectively
[30].
√
Pc πa  a 
KIC = fI (1)
wt w

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

Comparison of machine learning methods… 227

Table 1 Mechanical properties of pure knitting fabric reinforced composite specimens

Material type Mechanical properties (MPa)

Ew Tw Ec Tc Gwc Swc Scw Cw Cc
Glass/Epoxy 24105.84 127.82 19621.19 117.80 4160.16 40.33 51.94 98.61 81.94
Carbon/Epoxy 40437.71 200.91 29891.17 156.36 5112.89 36.21 46.23 125.80 99.27

√ a
Pc πa
KI IC = fI I (2)
wt w
a
where Pc is load at fracture, a crack length, w the specimen
 a  width, t athe
 specimen thickness, and f w is the
geometrical factor. The geometric factor formulae f I w and f I I w used to calculate the KI and KII were
provided below Eqs. (3–4) [16];
a  a 4  a 3  a 2 a
fI = 182.12 − 293.81 + 187.87 − 51.492 + 6.1137 (3)
w w w w w
a  a 4  a 3  a 2 a
fI I = −18.622 + 36.753 − 25.182 + 7.759 + 0.0944 (4)
w w w w w
To calculate the mixed-mode fracture toughness of composite material using the Arcan test, the load at fracture,
(PC ), at which the crack initiates, was measured. Using the equations available in the literature [1,31,32], the
mixed-mode loading effective fracture toughness, Keff , is calculated as:

(K C )e f f = K I2 + K I2I (5)

where KI and KII are modes I and mode II components of the fracture toughness in each specific loading angle.
√ √
Pc cosα πa  a  Pc sinα πa a
KI = fI KI I = fI I (6)
wt w wt w

3 Finite element analysis

In this scope of finite element analysis, ANSYS finite element package program was used to model and solve
the damage behavior of the cracked laminated composite structure. Table 1 shows the mechanical strengths
of pure and hybrid fabric reinforced lamşnated composites under tensile, compression and shear loads. These
values were used to define the strength values of numerical models that created for finite element analysis.
EW and TW show the elasite modulus and tensile strength values in the wale direction, while EC and TC
show the elasticity and tensile strength in the course direction, respectively. Shear modulus values of laminated
composites was expressed GWC . CW and CC were expressed the wale direction compressive strength and course
direction compressive strength, respectively. SWC and SCW represents the in-plane shear strength in the wale
dand course direction, respectively [33].
Cracked numerical modes were sized based on the dimensions of the experimental samples. The Quadri-
lateral 8-Nodes plane elements, which have two degrees of freedom per node, were used to create the finite
element mesh structure. The mesh refinement was applied around the crack tip to minimize the computational
error level. For this reason, a smaller mesh structure was formed around the crack tip, the size of the elements
being 0.2 mm. The mesh structure of the geometric model contains about 29600 nodes. Figure 3 presents the
mesh structure of the 2D geometric model of the Arcan test apparatus and the cracked butterfly specimen.
The J-integral method was employed for numerical analyses. According to linear elastic fracture mechanics,
the J integral method is a way of calculating the resistance of the material against its crack propagation under
plane strain. This indicates that J Integral and fracture toughness are related to each other. J-Integral is a
frequently used method for numerically characterizing the fracture behavior of polymer composites [34,36,37].
The function of J-integral for thin plate (2D geometric model) were defined by Eq. 7. It assumes that the crack
lies in the global Cartesian x(u) – y(v) plane.
  
∂u x ∂u y
J= W dy − tx + ty ds (7)
  ∂ x ∂y

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

228 H. E. Balcıoğlu, A. Ç. Seçkin

Fig. 3 2D FEA model of Arcan test specimen a before and b after numerical analysis

where  is a contour around the crack tip in a coordinate system, W is stress field energy density, u the
displacement vector, ds is the length of an element along the contour, tx the traction vector along the global
x-axis is xn x + σx y n y ; ty the traction vector along the global y-axis is σ y n y + σx y n x ; and σ is component
stress [7]. When calculating the J integral value from paths at the crack tip, it should be noted that the integral
line starts from the bottom of the crack and it goes in the clockwise direction and ends at the upper side of the
crack.
Owing to its dual roles in describing the crack tip fracture mechanics behaviors and in measuring fracture
resistance, J integral value has become a notable symbol and parameter of linear elastic fracture mechanics.
This phenomenon can be used to infer an equivalent fracture toughness value from J integral measurements.
In Eq. (8), the relationship between J integral and fracture toughness is mathematically formulated [37]:
√
(K C )Num = J.E (8)

E value in the Eq. (8) refers to the modulus of elasticity of the material.

4 Machine learning

Machine Learning is a sub-branch of artificial intelligence and is a study area that includes automatic calculation
procedures based on logical or binary operations that learn from a series of samples [38–40]. In the field of MLA,
there is no single learning algorithm that provides good learning on all real-world problem data. Therefore,
algorithm selection is done experimentally [23]. The learning models developed vary according to the training
data. Machine learning hosting many learning techniques such as K Nearest Neighbors (KNN), Decision Trees
(DT), Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), etc. and each
technique has its advantages and disadvantages. This raises the choice of the best model. Moreover, the fact
that there is no high-performance algorithm for all data sets is known as the “no free lunch” theorem [41,42].
Therefore, different algorithms will be used in the study. In this section, the basic operations which are shown in
Fig. 4 to be used during the study are briefly explained. The main purpose of this article is to compare the results
of the machine learning model of the fatigue tests of various composite materials and the results of the finite
element analysis (FEA) model. Machine learning is divided into three basic types of operations: classification,
regression, and clustering. Classification is the learning of data previously categorically labeled. Clustering is
a technique involving the grouping of unlabeled data. Regression is a method in which learning is performed
if numerical outputs are given in response to input values. As the purpose of the study is an interpolation,
regression, which is a branch of supervised learning, will be used as a machine learning technique. The
algorithms used in the study are KNN, DT, RF and SVM algorithms. The codes used for this paper were
prepared using the Python programming language. The scikit-learn library was used in a machine learning

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

Comparison of machine learning methods… 229

Training Data Set Test Data Set New Data

Machine Learning Performance

Selected Model
Algorithms Evaluation

Prediction

Fig. 4 Machine learning steps

Table 2 Experimental data set and their properties

Feature Id Properties Range/Values Type

X1 Crack location 0: Glass Categorical
1: Carbon
X2 Pattern Width 1: Pure composite Categorical
2: Hybrid with 50 mm pattern width
4: Hybrid with 25 mm pattern width
8: Hybrid with 25 mm pattern width
X3 Loading angle 0◦ 30◦ 45◦ 60◦ 90◦ Numeric
X4 Crack length 3.03–4.24 mm Numeric
X5 Sample thickness 2.51–4.27 mm Numeric
Y1 Fracture toughness Numeric

process which is shown in Fig. 4 steps [43]. The machine learning process was carried out on a laptop computer
with an i7-9750H processor, 8GB ram and 64 bit Windows 10 operating system.

4.1 Data set and properties

In this study, the effect of crack position (glass and carbon), pattern type (pure and hybrid), loading angle,
sample thickness and the effect of crack length on fracture toughness were investigated experimentally. The
data set consists of 5 features and 1 target value. In Table 2, X1 indicates the location of the crack. X2 refers
to the knitting pattern width of the reinforcing fabric. Number 1 indicates that the material is pure composite.
Numbers of 2, 4 and 8 denote hybrid composite samples having a pattern width of 50 mm, 25 mm, and 12.5 mm
respectively. The feature of X3 refers to the loading angle applied for crack onset. X4 and X5 correspond to
the length of the crack and the thickness of the sample, respectively. Y1 is the target parameter calculated by
the machine learning algorithm and represents the fracture toughness value. These attributes and the target
value are summarized in Table 2. The data set contains data from 200 experiments.

4.2 Machine learning algorithms

4.2.1 K nearest neighbors

The K Nearest Neighbors (KNN) algorithm is a machine learning algorithm, which works on the basis of
the values of the nearest K neighbor. The KNN algorithm is a nonparametric method for classification and
regression duties [44]. It was first applied in the news articles classification problem [45]. When performing
learning process with the KNN algorithm, the distance between the individual data in the examined data set
is first calculated. This length calculation is performed with the Euclidian, Manhattan or Hamming distance
function. Then the mean value of the nearest K neighbors is calculated for each data set The K value is the
only hyperparameter of the KNN algorithm. If the K is too low, the boundaries flicker and overfit situation
occurs, while if the K value is too high, the separation boundaries become smoother and under-fit situation
occurs. The disadvantage of the KNN algorithm is the distance calculation process, because it increases the
processing load with increasing data [23,40,46].

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

230 H. E. Balcıoğlu, A. Ç. Seçkin

Table 3 SVM kernel functions

Kernel function Formulation

 
Linear K xi , x j  = (xiT x j )
Polynomial K xi , x j = (xi x j )d
  2
Radial Basis Function (RBF) K xi , x j  = exp(−γ xi − x j ), γ > 0
Sigmoid K xi , x j = tanh (kxi x j − δ)

4.2.2 Decision tree

Decision Tree (DT) is an algorithm commonly used in statistical learning and data mining. DT uses simple “if-
then-else” decision rules and is used for both classification and regression in supervised learning [40,46,47].
The DT has three basic steps. In the first step, the most meaningful feature is placed as the first node. In the
second step, the data set is divided into subsets according to this node. Subsets should be formed so that each
subset contains data with same value for a characteristic. In the third step, repeat steps one and two until you
have found the last nodes in all branches. The DT builds classification or regression models in the form of a
tree structure. It splits a data set into smaller subsets, while at the same time an associated decision tree is built
incrementally. The result of the algorithm is like a root with decision leafs. DT can process both categorical
and numerical data [48].

4.2.3 Random forest

The Random Forest (RF) method is designed in the form of a forest of DT. With the RF algorithm, classification
or regression trees can be created and clustering can be performed. Each DT in the forest is formed by selecting
the sample from the original data set using the bootstrap technique and selecting the random number of all
variables in each decision node. The RF algorithm consists of four steps. First, n features are randomly selected
from a total of m features. For the second step, you calculate the node d from the n features using the best
distribution point. In the third step, you check whether the number of final nodes that the target has reached if
it does not go to step 1 otherwise it goes on to the next step. For the last step, build forest by repeating steps
one to three for n times [40,46,49–51].

4.2.4 Support vector machine

Support Vector Machine (SVM) algorithm introduced by Vapnik, and based on statistical learning theory [52].
SVM was originally developed for binary classification problems. While SVM was initially only used for
classification, it was later extended to linear and nonlinear regression [53]. The goal of SVM is to find the
hyperplane that separates the classes from each other and which is the furthest away from both classes. In
cases that cannot be separated linearly, the data is moved to another space of a higher dimension, and the
classification is performed in this space. While the classification processes aim to separate the data using
the generated vectors, the regression, in some ways does the opposite and aims to create a hyperplane by
identifying the support vectors so that they contain as much data as possible [54]. The SVM is mainly split in
two according to the linear separability of the data set. The nonlinear SVM decision function is given in Eq. 9.

N
f (x) = sign ai yi (x) (xi ) + b (9)
i=1

In cases where the data cannot be separated linearly, the data is moved to a higher-dimensional space and kernel
functions are used for resolution. The transformations can be performed using kernel functions expressed as
K(xi , xj ) = (x) (xi ) instead of the (x) (xi ) scalar product given in Eq. 8. Nonlinear transformations can be
performed, and the data can be separated in the high dimension thanks to the Kernel Functions. Support Vector
Machine Learning (SVML), Support Vector Machine Polynomial (SVMPOLY), Support Vector Machine
Radial Basis Function (SVMRBF), and Support Vector Machine Sigmoid (SVMSIGMOID) are the most
commonly used Kernel functions and they were listed in Table 3 [40,46,54,55].

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

Comparison of machine learning methods… 231

Table 4 Performance metrics for regression

Metric Equation Explanation

 
n
Root-mean-square error (RMSE) RMSE = 1
n j=1 (y j − ŷ j )2 It represents the sample standard devia-
tion of the differences between predicted
n   values ( ŷ j ) and observed values (y j )
Mean Average Error (MAE) MAE = n1  y j − ŷ j  MAE is the average of the absolute differ-
j=1
ence between predicted values ( ŷ j ) and
observed values (y j )
| y j − ŷ j |
Error Percentage (EP) EP = yj ∗ 100 EP is the percentage of the absolute dif-
ference between predicted values ( ŷ j ) and
n
observed values (y j ) according to y j
(y j − ŷ j )2
R Squared (R2 ) R2 = 1 − j=1
n The numerator is Mean Square Error and
j=1 (y j − ȳ)2
the denominator is the variance in y values

4.3 Regression model evaluation and performance metrics

There are various model evaluation techniques but some of the most well-known are percentage split and cross-
validation. It is essential to use a train and a test data set in the evaluation processes. The split by percentage is
the most basic method. In this method, all the data is split subsets as train and tests by percentage definition.
The training subset is used for the learning process and the test subset is used for performance evaluation.
However, the evaluation results may not be reliable, because of the reasons such as not having the similar
distribution when selecting the train and test subsets in the data set, uneven distribution of outliers and so on.
Therefore, the K Fold cross-validation method was developed. In this method, train and test data are integrated
and turned into a single data set. All data is divided into K equally sized subsets. The K value, also known
as number of folding, is determined by the user. Then, learning and testing are performed for each of the K
sub-sets; here one of the subset will be used for test process, and the others will be used in train process. As
a result, performance metrics are obtained for K times for each sub-set. The mean of the performance metrics
is considered as the overall performance metric. K-Fold cross-validation method is known to produce more
reliable results than other methods. However, since learning and testing for each subset is done separately for
all subsets, the total time is longer than the other methods [23].
The main criteria used for performance evaluation and model selection are called metrics. The most com-
monly used regression metrics are root-mean-square error (RMSE), mean average error (MAE), and coefficient
of determination (R2 ). The RMSE, MAE, and R2 metrics used in performance evaluation in regression process
are listed in Table 4 [23,40,56,57].

4.4 Hyper parameter tuning

Hyperparameters are coefficients that are not learned with Machine Learning Algorithms. These parameters
affect the algorithm performance so it should be optimized or well-chosen. The process of searching hyper-
parameters is called tuning. There are two kinds of the searching method as Grid Search (GS), Parameter
Optimization (PO). With the GS, the results of the performance are generated for all values in the user-
defined parameter range. Then the best hyper-parameters are selected according to these results. The GS is an
exhaustive process; therefore, parameter search operations are performed using various optimization methods
like the Genetic Algorithm in the PO [58–60]. After the hyperparameter searches the optimum values and used
in the algorithms are presented in Table 5.

5 Results and discussion

In this study, fracture behaviors of knitted fabric reinforced laminated composites under different loading
conditions were modeled using FEA and ML algorithms. The effects of fracture location, material type and
loading angle on fracture toughness of composite material were investigated in FEA and MLA analyzes. The
fracture toughness of the pure and hybrid laminated composite materials according to their crack locations
were given in Table 6–9. Due to the lots of experimental, FEA and MLA analysis results, the fracture toughness

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

232 H. E. Balcıoğlu, A. Ç. Seçkin

Table 5 Hyper parameter tuning results

Algorithm Parameter Search range Step Value

KNN Number of neighbors 2–10 1 5
DT Maximum depth 2–10 1 4
RF Maximum depth 2–10 1 2
SVM Cost (C) 1–10 1 7
Epsilon (ε) [10−4 , 10−3 10−2 , 10−1 ] 1 10−4

Table 6 Fracture toughness of carbon/epoxy pure composites having a crack at carbon side

Loading angle(◦ ) EXP (MPamm0.5 ) FEA (MPamm0.5 ) Machine Learning Algorithms (MPamm0.5 )
DT KNN RF SVML SVMPOLY SVMRBF SVMSIGMOID
0 664.64 623.63 648.70 562.55 598.18 631.22 564.51 580.97 587.44
30 537.41 564.88 492.86 445.71 470.73 490.20 510.29 499.49 337.92
45 509.09 468.68 435.30 353.73 437.43 419.94 451.75 440.07 399.36
60 394.49 299.55 346.37 281.74 390.37 342.64 372.26 353.79 429.78
90 177.15 189.78 148.86 161.30 160.58 147.03 150.43 168.60 447.92

Table 7 Fracture toughness of glass–carbon/epoxy hybrid composites with 50 mm pattern width having a crack at carbon side

Loading angle(◦ ) EXP (MPamm0.5 ) FEA (MPamm0.5 ) Machine learning algorithms (MPamm0.5 )
DT KNN RF SVML DT SVMRBF SVMSIGMOID
0 599.85 524.25 648.70 585.46 598.18 651.08 564.51 581.73 587.65
30 483.91 492.00 492.86 459.06 470.73 493.78 510.29 500.35 338.88
45 429.60 484.21 435.30 394.61 437.43 433.72 451.75 440.84 400.05
60 367.56 406.19 346.37 331.10 390.37 360.88 372.26 354.24 429.98
90 148.98 160.65 148.86 141.90 155.13 188.50 150.43 168.06 448.02

Table 8 Fracture toughness of glass/epoxy pure composites having a crack at glass side

Loading angle(◦ ) EXP (MPamm0.5 ) FEA (MPamm0.5 ) Machine learning algorithms (MPamm0.5 )
DT KNN RF SVML DT SVMRBF SVMSIGMOID
0 562.55 505.55 648.70 664.64 598.18 651.20 564.51 580.77 587.31
30 445.71 420.74 492.86 537.41 470.73 504.21 510.08 499.07 337.76
45 353.73 353.33 435.30 509.09 437.43 430.72 451.67 439.59 399.13
60 281.74 226.06 346.37 394.49 390.37 357.22 371.24 353.31 429.55
90 161.30 111.83 148.86 174.26 159.90 210.23 149.93 168.38 447.51

values of each composite structure were not given. Only one table of the fracture toughness of glass, carbon
and hybrid composite materials were given according to their fracture locations, to make comparison and
evaluation. When the experimental results were examined, maximum fracture toughness values for pure and
hybrid composite structures were obtained at 0◦ loading angle. On the other hand, for each pure and hybrid
composite structure, the minimum fracture toughness value was obtained at a 90◦ loading angle. As the loading
angle is increased from 0 to 90, the effect of Mode I decreases and the effect of Mode II increases. For this
reason, the fracture toughness value is high at low loading angles, and when the loading angle increases, the
fracture toughness value decreases. When examined the fracture behavior in terms of reinforcing materials, it
was seen that carbon fiber reinforced composites have superior toughness values over glass one. The addition
of carbon fiber to the glass fiber in hybrid structures increased the fracture strength of the structure. The
knitting structure is formed by the sequencing of each row successively. The energy of the crack progression
was absorbed in the location where the glass and the carbon row overlap. This had led to an increase in the
fracture energy required for the onset of the crack. Consequently, fracture toughness increased by decreasing
the pattern width or increasing the number of patterns.
As can be seen from Table 6–9, FEA and MLA have shown different performances in the convergence
of experimental fracture toughness results. The fracture toughness values were analyzed for different loading

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

Comparison of machine learning methods… 233

Table 9 Fracture toughness of glass–carbon/epoxy hybrid composites with 50 mm pattern width having a crack at glass side

Loading angle(◦ ) EXP (MPamm0.5 ) FEA (MPamm0.5 ) Machine Learning Algorithms (MPamm0.5 )
DT KNN RF SVML DT SVMRBF SVMSIGMOID
0 585.46 499.28 648.70 599.85 598.18 663.68 564.53 581.54 587.48
30 459.06 451.36 492.86 483.91 470.73 516.69 510.35 499.97 338.42
45 394.61 386.25 435.30 429.60 437.43 443.20 451.98 440.38 399.64
60 331.10 338.96 346.37 367.56 390.37 369.70 372.51 353.77 429.81
90 141.90 127.37 148.86 148.98 154.64 222.71 149.69 167.83 447.90

Table 10 Error performance metrics of results according to Experimental and FEA

Convergence methods Experimental Finite element method

RMSE MAE EP R2 RMSE MAE EP R2
FEA 38.90 29.68 8.01 0.94 – −− – –
DT 36.36 28.95 6.93 0.95 73.97 58.36 15.99 0.79
KNN 56.86 38.29 8.50 0.87 77.53 61.96 17.12 0.73
RF 37.65 27.30 7.25 0.94 72.62 57.98 16.75 0.74
SVML 55.63 48.77 15.40 0.89 66.06 54.92 15.29 0.79
SVMPOLY 41.09 28.98 7.57 0.93 52.77 43.23 11.73 0.85
SVMRBF 35.49 26.48 7.45 0.95 54.14 46.75 12.58 0.85
SVMSIGMOID 155.85 121.95 51.38 0.09 173.78 149.39 46.26 0.13

angles (from 0◦ to 90◦ with 5◦ intervals) that were not examined experimentally using FEA and MLA. The
performance of the fracture toughness values, which were obtained from the analysis results was determined in
two different categories. The experimental fracture toughness values, which were obtained from loading angles
such as 0◦ , 30◦ , 45◦ , 60◦ , and 90◦ , were based on the performance of fracture toughness. Fracture toughness
values obtained from FEA were based on the error performance metrics of fracture toughness values in other
angles. RMSE, MAE, EP, and R2 obtained from the comparison of both experimental and FEA methods were
given in Table 10.
RMSE, MAE, EP, and R2 metrics are used when comparing algorithms with each other. The explanations
and equations of these metrics are presented in Table 2. Algorithms with lower RMSE, MAE and EP values
than others are considered successful. The R2 metric consists of values between 0 and 1. When the R2 close
to 1, the better the predictions made by that algorithm. Each of these metrics can be used alone to determine
which algorithm is best.
When the experimental results are analyzed in terms of RMSE, the three most successful algorithms
are SVMRBF (35.49), DT (36.36) and RF (37.65), respectively. If the experimental results are compared
according to the MAE, the three most successful algorithms are SVMRBF (26.48), RF (27.30) and DT (28.95),
respectively. When the experimental results for the EP metric are examined, it is seen that the three most
successful algorithms are DT (6.93), RF (7.25) and SVMRBF (7.45) respectively. When the experimental
results for R2 are examined, the three most successful algorithms are SVMRBF (0.95), DT (0.95) and RF
(0.94), respectively. The comparison of the MLA results versus the finite element method was performed
to see how the algorithms behave for intermediate values where the experimental results were not measured.
According to the experimental results, R2 values obtained by MLA are higher than FEA. In this section, the three
most successful algorithms for comparison in terms of RMSE, MAE, and EP are SVMPOLY (52.77, 43.23,
11.73), SVMRBF (54.14, 46.75, 12.58) and SVML (66.06, 54.92, 15.29) respectively. When the algorithms
are compared in terms of R2 , the three most successful algorithms are SVMPOLY(0.85), SVMRBF (0.85) and
SVML (0.79) respectively. After the experimental and FEA comparisons of the algorithms, it is seen that DT
and RF algorithms perform poor regression in the intermediate value estimation without experimental data.
The fact that SVMRBF is in the top three positions in all comparisons shows that this algorithm performs
better regression than the others. In both comparisons, SVMSIGMOID was the worst-performing algorithm
in terms of all metrics.
In Table 11, the fracture toughness values of pure carbon/epoxy composites were given for loading angles,
which were not examined experimentally, and obtained by convergence methods. Fracture toughness values
for other types of composite materials were not given numerically.

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

234 H. E. Balcıoğlu, A. Ç. Seçkin

Table 11 Fracture toughness of pure carbon/epoxy laminated composites

Loading EXP FEA Machine learning algorithms (MPamm0.5 )

angle(◦ ) (MPamm0.5 ) (MPamm0.5 )
DT KNN RF SVML SVMPOLY SVMRBF SVMSIGMOID
0 664.64 623.63 648.70 562.55 598.18 608.38 564.51 580.78 587.29
5 610.86 648.70 562.55 598.18 583.88 561.29 579.52 424.10
10 597.54 648.70 562.55 598.18 559.38 555.75 570.84 322.93
15 593.12 648.70 515.88 598.18 534.88 547.87 556.39 286.83
20 574.84 492.86 445.71 470.73 510.38 537.67 538.33 290.63
25 560.38 492.86 445.71 470.73 485.89 525.15 518.75 311.87
30 537.41 564.88 492.86 445.71 470.73 461.39 510.29 499.11 337.63
35 516.13 492.86 445.71 470.73 436.89 493.11 479.87 361.90
40 491.35 435.30 353.73 437.43 412.39 473.59 460.49 382.52
45 509.09 468.68 435.30 353.73 437.43 387.89 451.75 439.61 399.11
50 421.94 435.30 353.73 437.43 363.39 427.58 415.57 412.04
55 379.30 346.37 281.74 390.37 338.90 401.09 386.99 421.93
60 394.49 299.55 346.37 281.74 390.37 314.40 372.26 353.37 429.39
65 270.65 346.37 281.74 390.37 289.90 341.11 315.51 434.98
70 267.60 346.37 281.74 390.37 265.40 307.63 275.63 439.14
75 232.10 346.37 202.35 390.37 240.90 271.82 237.15 442.23
80 201.00 148.86 161.30 156.91 216.40 233.69 204.09 444.52
85 185.71 148.86 161.30 156.91 191.90 193.22 180.25 446.21
90 177.15 189.78 148.86 161.30 156.91 167.41 150.43 168.35 447.46

Fracture toughness behaviors according to loading angle were given in Fig. 5 graphically. To be clarity
of curves, three MLA algorithms having the best performance were given in Fig. 5. The experimental data
set obtained from the specified loading angles was used for the MLA algorithms. Although the performance
values of the fracture toughness values, which produced by the DT, KNN and RF algorithms for the angles in
the test, are good, the performance value is bad for the angles that non-experimental. Such that, the estimation
error in these angles caused the algorithm to show a zigzag behavior. As DT, KNN, RF and SVMSIGMOID
algorithms are designed primarily for classification purposes, it is seen that it produces poor results in regression
application.
Table 12 presents the error percentage and convergence coefficients in the studies, where the fracture
toughness values of the composite materials were calculated by using FEA and MLA methods. When the error
percentages in the convergence to the experimental value for the fracture toughness obtained from the finite
element method were examined, it can be seen that there are error amounts varying up to 10% in the literature.
In the current study, the error percentage for the fracture toughness value obtained from the finite element
method is 8.01%. Therefore, the finite element model, which was created to simulate the fracture behavior of
laminated composites, is consistent with the literature. On the other hand, considering the studies in which
the fracture behavior of composite materials were examined using numerical methods such FEA and MLA,
the minimum convergence error was 7.48%. Besides, a comparative study, which the FEA and MLA methods
were used together to estimate the fracture toughness of laminated composites, was not found. In some studies,
convergence between experimental and predicted values was expressed in terms of the correlation coefficient.
Accordingly, the maximum coefficient of correlation obtained for estimating fracture toughness is R2 =0.94.
In this respect, the error percentage and correlation coefficient values, which belong to estimated fracture
toughness values obtained by using the MLA method, are consistent with the literature.

6 Conclusion

In this study, the fracture toughness of the pure and hybrid knitted fabric reinforced laminated composites
in different loading angles was investigated using experimental, FEA and MLA methods. The general results
obtained in this study, where the true fracture toughness values were calculated by FEA and estimated by MLA
are as follows:
• Experimentally, fracture toughness value was directly affected by loading angle, pattern width, and type
of material. Thus, the maximum fracture toughness value was obtained from carbon/epoxy composites
and the minimum one was obtained from glass/epoxy composites. Also, the loading angle increased from

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

Comparison of machine learning methods… 235

Fig. 5 Changing of fracture toughness values that obtained from experimental, FEA and MLA, according to the loading angle

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

236 H. E. Balcıoğlu, A. Ç. Seçkin

Fig. 5 continued

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

Comparison of machine learning methods… 237

Table 12 Comparison of error rates of fracture toughness values obtained using FEA and MLA methods

Citation FEA methods MLA methods

Error percentage (%) Error percentage (%) R2
[21] 10.32 – –
[22] 8.33 – –
[26] – 7.48 –
[61] – – 0.94
Current study 8.01 7.45 0.95

0◦ to 90◦ , it was shown that the fracture toughness value decreased. Also, the value of fracture toughness
increased as the pattern width value decreased.
• The J-Integral method used with FEA was seen to be successful for both experimental and non-experimental
loading angles.
• MLA methods are particularly preferred in applications that are difficult to extract equations or modeling
of experimental structure. However, as we can see from the performance values obtained in this study,
good performance results have been obtained in an application having the equation.
• In fracture toughness analysis, it was shown that MLA can be used in fracture mechanics problems in terms
of computer computational methods when compared to FEA method.
• According to the performance metrics obtained from the SVMRBF algorithm, it has been shown that the
fracture behavior of laminated composite materials can be used for studies. However, according to the
performance value of the SVMSIGMOID algorithm, it is seen that it is not suitable for such applications.
DT, KNN, RF, and SVMSIGMOID algorithms are designed primarily for classification purposes. It is seen
that these algorithms produce poor results in regression application.
• DT, KNN and RF algorithms produce good results for experimental data. However, it was observed that
a natural behavior curve could not be obtained for the angles which were not examined experimentally.
Therefore, in the preparation of the data set for these algorithms in the fracture mechanics studies, the
loading angles should be chosen close to each other and, if possible, to address all values in the examination
range.

References

1. Jamali, J., Mourad, A.-H.I., Fan, Y., Wood, J.T.: Through-thickness fracture behavior of unidirectional glass fibers/epoxy
composites under various in-plane loading using the CTS test. Eng. Fract. Mech. 156, 83–95 (2016). https://doi.org/10.1016/
j.engfracmech.2016.01.016
2. Bin Mohamed Rehan, M.S., Rousseau, J., Fontaine, S., Gong, X.J.: Experimental study of the influence of ply orientation
on DCB mode-I delamination behavior by using multidirectional fully isotropic carbon/epoxy laminates. Compos. Struct.
161, 1–7 (2017). https://doi.org/10.1016/j.compstruct.2016.11.036
3. Srivastava, V.K., Gries, T., Veit, D., Quadflieg, T., Mohr, B., Kolloch, M.: Effect of nanomaterial on mode I and mode II
interlaminar fracture toughness of woven carbon fabric reinforced polymer composites. Eng. Fract. Mech. 180, 73–86 (2017).
https://doi.org/10.1016/j.engfracmech.2017.05.030
4. Tiber, B., Balcıoğlu, H.E.: Flexural and fracture behavior of natural fiber knitted fabric reinforced composites. Polym.
Compos. (2017). https://doi.org/10.1002/pc.24635
5. Szekrényes, A., Uj, J.: Beam and finite element analysis of quasi-unidirectional composite SLB and ELS specimens. Compos.
Sci. Technol. 64, 2393–2406 (2004). https://doi.org/10.1016/j.compscitech.2004.05.002
6. Davidson, B.D., Bansal, A., Bing, Q., Sun, X.: Geometrically nonlinear determination of energy release rate and mode ratio
in single leg bending tests. J. Reinf. Plast. Compos. 28, 1881–1901 (2009). https://doi.org/10.1177/0731684408089235
7. Arasan, Ş., Aktaş, M., Balcıoğlu, H.E.: Fracture toughness of woven glass and carbon reinforced hybrid and non-hybrid
composite plates. Polym. Compos. (2016). https://doi.org/10.1002/pc.23999
8. Arcan, M., Hashin, Z., Voloshin, A.: A method to produce uniform plane-stress states with applications to fiber-reinforced
materials: a specially designed specimen yields material properties under pure shear or uniform plane-stress conditions. Exp.
Mech. 18, 141–146 (1978). https://doi.org/10.1007/BF02324146
9. Hasanpour, R., Choupani, N.: Rock fracture characterization using the modified Arcan test specimen. Int. J. Rock Mech.
Min. Sci. 46, 346–354 (2009). https://doi.org/10.1016/j.ijrmms.2008.07.004
10. Alizadeh, F., Guedes Soares, C.: Experimental and numerical investigation of the fracture toughness of Glass/Vinylester
composite laminates. Eur. J. Mech. - A Solids. 73, 204–211 (2019). https://doi.org/10.1016/j.euromechsol.2018.08.003

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

238 H. E. Balcıoğlu, A. Ç. Seçkin

11. Gao, L., Ni, F., Braham, A., Luo, H.: Mixed-Mode cracking behavior of cold recycled mixes with emulsion using Arcan
configuration. Constr. Build. Mater. 55, 415–422 (2014). https://doi.org/10.1016/j.conbuildmat.2013.12.079
12. Alfonso, L., Badulescu, C., Carrere, N.: Use of the modified Arcan fixture to study the strength of bonded assemblies for
automotive applications. Int. J. Adhes. Adhes. 80, 104–114 (2018). https://doi.org/10.1016/j.ijadhadh.2017.09.014
13. Xavier, J., Morais, J., Pereira, F.: Non-linear shear behaviour of bovine cortical bone by coupling the Arcan test with digital
image correlation. Opt. Lasers Eng. 110, 462–470 (2018). https://doi.org/10.1016/j.optlaseng.2018.07.004
14. Valès, B., Marguet, S., Créac’hcadec, R., Sohier, L., Ferrero, J.-F., Navarro, P.: Experimental & numerical study of the
tensile/compression-shear Arcan test under dynamic loading. Int. J. Adhes. Adhes. 78, 135–147 (2017). https://doi.org/10.
1016/j.ijadhadh.2017.06.010
15. Song, C., Ooi, E.T., Natarajan, S.: A review of the scaled boundary finite element method for two-dimensional linear elastic
fracture mechanics. Eng. Fract. Mech. 187, 45–73 (2018). https://doi.org/10.1016/j.engfracmech.2017.10.016
16. Choupani, N.: Experimental and numerical investigation of the mixed-mode delamination in Arcan laminated specimens.
Mater. Sci. Eng. A. 478, 229–242 (2008). https://doi.org/10.1016/j.msea.2007.05.103
17. Gruben, G., Fagerholt, E., Hopperstad, O.S., Børvik, T., Langseth, M.: Numerical simulation of ductile fracture in modified
Arcan test. Procedia Mater. Sci. 3, 661–666 (2014). https://doi.org/10.1016/j.mspro.2014.06.109
18. Méité, M., Dubois, F., Pop, O., Absi, J.: Mixed mode fracture properties characterization for wood by digital images correlation
and finite element method coupling. Eng. Fract. Mech. 105, 86–100 (2013). https://doi.org/10.1016/j.engfracmech.2013.01.
008
19. Majidi, H.R., Ayatollahi, M.R., Torabi, A.R.: On the use of the extended finite element and incremental methods in brit-
tle fracture assessment of key-hole notched polystyrene specimens under mixed mode I/II loading with negative mode I
contributions. Arch. Appl. Mech. 88, 587–612 (2018). https://doi.org/10.1007/s00419-017-1329-7
20. Mladensky, A., Rizov, V.: Analytical investigation of nonlinear interlaminar fracture in trilayered polymer composite beam
under mode II crack loading conditions using the J -integral approach. Arch. Appl. Mech. 83, 1637–1658 (2013). https://
doi.org/10.1007/s00419-013-0770-5
21. LeBlanc, L.R., LaPlante, G.: Experimental investigation and finite element modeling of mixed-mode delamination in a
moisture-exposed carbon/epoxy composite. Compos. Part Appl. Sci. Manuf. 81, 202–213 (2016). https://doi.org/10.1016/j.
compositesa.2015.11.017
22. Hadavinia, H., Ghasemnejad, H.: Effects of mode-I and mode-II interlaminar fracture toughness on the energy absorption
of CFRP twill/weave composite box sections. Compos. Struct. 89(2), 303–314 (2009)
23. Alpaydin, E.: Introduction to Machine Learning. MIT press, Cambridge (2009)
24. Radivojac, P, White, M.: Machine Learning Handbook (2019). https://marthawhite.github.io/mlcourse/notes.pdf
25. Nasiri, S., Khosravani, M.R., Weinberg, K.: Fracture mechanics and mechanical fault detection by artificial intelligence
methods: a review. Eng. Fail. Anal. 81, 270–293 (2017). https://doi.org/10.1016/j.engfailanal.2017.07.011
26. Wang, Hongxun, Zhang, Weifang, Sun, Fuqiang, Zhang, Wei: A comparison study of machine learning based algorithms for
fatigue crack growth calculation. Materials 10, 543 (2017). https://doi.org/10.3390/ma10050543
27. Rovinelli, A., Sangid, M.D., Proudhon, H., Ludwig, W.: Using machine learning and a data-driven approach to identify the
small fatigue crack driving force in polycrystalline materials. Npj Comput. Mater. 4, 1–10 (2018). https://doi.org/10.1038/
s41524-018-0094-7
28. Balcıoğlu, H.E., Seçkin, A.Ç., Aktaş, M.: Failure load prediction of adhesively bonded pultruded composites using artificial
neural network. J. Compos. Mater. 50, 3267–3281 (2016). https://doi.org/10.1177/0021998315617998
29. Atilla, D., Sencan, C., Goren Kiral, B., Kiral, Z.: Free vibration and buckling analyses of laminated composite plates with
cutout. Arch. Appl. Mech. (2020). https://doi.org/10.1007/s00419-020-01730-2
30. Balcıoğlu, H.E., Yalçın, D.: The determination of fracture characterization of knitted fabric reinforced composites using
Arcan test. Fibers Polym. 21, 849–863 (2020). https://doi.org/10.1007/s12221-020-9619-z
31. Jamali, J., Fan, Y., Wood, J.T.: The mixed-mode fracture behavior of epoxy by the compact tension shear test. Int. J. Adhes.
Adhes. 63, 79–86 (2015). https://doi.org/10.1016/j.ijadhadh.2015.08.006
32. Richard, H.A.: Some theoretıcal and experımental aspects of mıxed mode fractures. In: Fract. 84, Elsevier, 1984: pp. 3337–
3344. https://doi.org/10.1016/B978-1-4832-8440-8.50358-6
33. Ersen Balcioglu, H., Baytan Ozmen, H.: The Fracture Behavior of Pure and Hybrid Intraply Knitted Fabric-Reinforced
Polymer Composites. In: Fract. Mech. Appl. Work. Title, IntechOpen, 2019. https://doi.org/10.5772/intechopen.89478
34. Arash, B., Park, H.S., Rabczuk, T.: Coarse-grained model of the J-integral of carbon nanotube reinforced polymer composites.
Carbon 96, 1084–1092 (2016). https://doi.org/10.1016/j.carbon.2015.10.058
35. Shahani, A.R., Abolfathitabar, R., Shooshtar, H.: On the validity of LEFM methods to investigate the fracture behavior of
angle-ply laminates. Compos. Part B Eng. 160, 249–253 (2019). https://doi.org/10.1016/j.compositesb.2018.10.013
36. Matthews, T., Ali, M., Paris, A.J.: Finite element analysis for large displacement J-integral test method for mode I interlaminar
fracture in composite materials. Finite Elem. Anal. Des. 83, 43–48 (2014). https://doi.org/10.1016/j.finel.2014.01.006
37. Gonzáles, G.L.G., González, J.A.O., Castro, J.T.P., Freire, J.L.F.: A J-integral approach using digital image correlation for
evaluating stress intensity factors in fatigue cracks with closure effects. Theor. Appl. Fract. Mech. 90, 14–21 (2017). https://
doi.org/10.1016/j.tafmec.2017.02.008
38. Kotsiantis, S.B., Zaharakis, I.D., Pintelas, P.E.: Machine learning: a review of classification and combining techniques. Artif.
Intell. Rev. 26, 159–190 (2006)
39. Coşkun, A.: Yapay zeka optimizasyon teknikleri: literatür değerlendirmesi. Doğu Anadolu Bölgesi Araştırmaları. 5(2),
142–146 (2007)
40. Seçkin, A.Ç., Coşkun, A.: Hierarchical fusion of machine learning algorithms in indoor positioning and localization. Appl.
Sci. 9, 3665 (2019)
41. Wolpert, D.H.: The lack of a priori distinctions between learning algorithms. Neural Comput. 8, 1341–1390 (1996)
42. Wolpert, D.H., Macready, W.G.: No free lunch theorems for optimization. IEEE Trans. Evol. Comput. 1, 67–82 (1997)
43. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R.,
Dubourg, V.: Scikit-learn: machine learning in python. J. Mach. Learn. Res. 12, 2825–2830 (2011)

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

Comparison of machine learning methods… 239

44. Altman, N.S.: An introduction to kernel and nearest-neighbor nonparametric regression. Am. Stat. 46, 175–185 (1992).
https://doi.org/10.1080/00031305.1992.10475879
45. Masand, B., Linoff, G., Waltz, D.: Classifying news stories using memory based reasoning. In: Proceedings of the 15th
Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, ACM, (1992) pp.
59–65
46. Seçkin, M., Seçkin, A.Ç., Coşkun, A.: Production fault simulation and forecasting from time series data with machine learning
in glove textile industry. J. Eng. Fibers Fabr. 14, 155892501988346 (2019). https://doi.org/10.1177/1558925019883462
47. Quinlan, J.R.: Induction of decision trees. Mach. Learn. 1, 81–106 (1986)
48. Quinlan, J.R.: Simplifying decision trees. Int. J. Man-Mach. Stud. 27, 221–234 (1987)
49. Liaw, A., Wiener, M.: Classification and regression by randomForest. R News 2, 18–22 (2002)
50. Akman, M., Genç, Y., Ankarali, H.: Random forests yöntemi ve sağlık alanında bir uygulama. Turk. Klin. J. Biostat. 3, 36–48
(2011)
51. Breiman, L.: Random forests. Mach. Learn. 45, 5–32 (2001)
52. Boser, B.E., Guyon, I.M., Vapnik, V.N.: A training algorithm for optimal margin classifiers. In: Proceedings of the Fifth
Annual Workshop on Computational Learning Theory, ACM, (1992) pp. 144–152
53. Müller, K.-R., Smola, A.J., Rätsch, G., Schölkopf, B., Kohlmorgen, J., Vapnik, V.: Predicting time series with support vector
machines. In: International Conference on Artificial Neural Networks, Springer, (1997) pp. 999–1004
54. Smola, A.J., Schölkopf, B.: A tutorial on support vector regression. Stat. Comput. 14, 199–222 (2004)
55. Gunn, S.R.: Support vector machines for classification and regression. ISIS Tech. Rep. 14, 5–16 (1998)
56. Willmott, C.J.: Some comments on the evaluation of model performance. Bull. Am. Meteorol. Soc. 63, 1309–1313 (1982)
57. Moriasi, D.N., Arnold, J.G., Van Liew, M.W., Bingner, R.L., Harmel, R.D., Veith, T.L.: Model evaluation guidelines for
systematic quantification of accuracy in watershed simulations. Trans. ASABE. 50, 885–900 (2007)
58. Bergstra, J.S., Bardenet, R., Bengio, Y., Kégl, B.: Algorithms for hyper-parameter optimization. In: Advances in Neural
Information Processing Systems (2011) pp. 2546–2554
59. Coskun, A.: Optimization of a mini-golf game using the genetic algorithm. Elektron. Ir Elektrotechnika. 109, 97–100 (2011)
60. Bergstra, J., Bengio, Y.: Random search for hyper-parameter optimization. J. Mach. Learn. Res. 13, 281–305 (2012)
61. Gopinath, K.G.S., Pal, S., Tambe, P.: Prediction of hardness and fracture toughness in liquid-phase-sintered alumina system
using Gaussian process regression and minimax probability machine regression. Mater. Today. 5(5), 12223–12232 (2018)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional
affiliations.

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

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

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

[email protected]