Curved and Layered Structures 2024; 11: 20240004

Research Article

Najib Zemed*, Kaoutar Mouzoun, Hanane Moulay Abdelali, Toufik Cherradi, Azzeddine Bouyahyaoui

Reliability and sensitivity assessment of

laminated composite plates with high-
dimensional uncertainty variables using active
learning-based ensemble metamodels
https://doi.org/10.1515/cls-2024-0004 marine engineering, civil engineering, and renewable energy,
received December 23, 2023; accepted March 18, 2024 due to their outstanding performance and lightweight prop-
Abstract: Laminated composite materials play a crucial erties. Their ability to reduce structural weight while pre-
role in various engineering applications due to their excep- serving strength in comparison to conventional metallic
tional mechanical properties. This article explores the materials is a key driver for their widespread use.
analysis of reliability and the local sensitivity of failure Moreover, laminated composites demonstrate superior resis-
probability associated with laminated composite materials. tance to both fatigue and corrosion when compared to
It delves into the impact of uncertainties on the performance metals, making them well-suited for enduring use in
of these materials and presents an active learning-based relia- demanding environments.
bility methodology. This methodology combines Monte Carlo To investigate the behavior of composite materials,
simulation with a metamodel derived from the combination various modeling theories are employed. Among these,
of three metamodels: artificial neural networks, support the classical Love–Kirchhoff theory, the oldest, neglects
vector regression, and Kriging. The article illustrates this the effects of transverse deformation to predict the beha-
methodology through two practical applications on laminated vior of thin plates [1,2]. Subsequently, the first-order shear
composite plates and compares its performance with other deformation theory (FSDT) was developed [3], accounting
reliability methods. This approach offers valuable insights to for variations in transverse shear throughout the plate
enhance the analysis of reliability, strengthen the design pro- thickness. Furthermore, higher-order shear deformation
cess, and facilitate decision-making while fully considering theories [4,5] were formulated, utilizing increased depen-
uncertainties related to material properties. dence on the thickness coordinate to describe displace-
ment fields in n-planes. Moreover, research efforts have
Keywords: composite material reliability, reliability sensi- been made to enhance the modeling of laminate behavior.
tivity, artificial neural network, support vector regression, For instance, Tornabene et al. [6] devised an innovative
Kriging, Monte Carlo simulation, active learning layer-wise approach employing the generalized differen-
tial quadrature (GDQ) method for dynamic analysis of
anisotropic shells. They demonstrated the accuracy of
1 Introduction their method through numerical examples, aligning
with classical 3D finite element predictions in terms of
Laminated composite plates find extensive applications frequencies and mode shapes. In addition, in another
across various sectors, including aerospace, automotive, study, Tornabene et al. also proposed [7] an equivalent
single-layer formulation for linear static analysis of shell
structures. They derived fundamental relations from the

* Corresponding author: Najib Zemed, Mohammed V University in stationary configuration of total potential energy and
Rabat, Mohammadia School of Engineering, Civil Engineering and addressed them numerically using GDQ. To illustrate
Construction laboratory, Avenue Ibn Sina B.P 765, Agdal, Rabat 10090, their method, they presented several examples demon-
Morocco, e-mail: [email protected]
strating its effectiveness. Furthermore, Brischetto [8] ana-
Kaoutar Mouzoun, Hanane Moulay Abdelali, Toufik Cherradi,
Azzeddine Bouyahyaoui: Mohammed V University in Rabat,
lyzed the effects of hygrothermal loading on the bending
Mohammadia School of Engineering, Civil Engineering and Construction behavior of multilayered composite plates. Using refined
laboratory, Avenue Ibn Sina B.P 765, Agdal, Rabat 10090, Morocco two-dimensional models within the framework of Carrera’s

Open Access. © 2024 the author(s), published by De Gruyter. This work is licensed under the Creative Commons Attribution 4.0 International License.
2  Najib Zemed et al.

unified formulation. Also, Civalek [9] proposed a method of metamodels, there are response surface methods (RSMs)
discrete singular convolution to obtain frequencies and [24–27], artificial neural networks (ANNs) [28–30], radial
buckling loads of composite plates. By utilizing geometric basis functions [31–33], Kriging (also known as Gaussian
transformation, he then compared the obtained results process) [34,35], support vector machine (SVM), and sup-
with those of other numerical methods. However, con- port vector regression (SVR) [36,37].
ducting reliability and sensitivity analyses on these plates Over the past few decades, extensive research has
under uncertain conditions can be exceptionally challen- been conducted on the failure and sensitivity of composite
ging, especially when dealing with numerous high-dimensional materials. For example, Huh [26] employed the stochastic
uncertainty variables. Ensuring the structural integrity and per- finite element method (SFEM) to assess the reliability of
formance of these materials necessitates a meticulous task relia- angled plies, comparing results with other reliability methods
bility and sensitivity assessment of laminated composite plates such as the β-method. The LSF used in this context was based on
within a high-dimensional uncertainty framework. the Tsai-Wu, Hoffman, and Tsai-Hill criteria. Similarly, Onkar
The reliability of composite materials fundamentally et al. [38] utilized SFEM for reliability analysis, employing the
relies on the use of a limit state function (LSF). This func- Tsai-Wu and Hoffman criteria as failure criteria for orthotropic
tion plays a crucial role by establishing a relationship plates with random material properties and random loads.
between the state of the structure and the applied loads. Furthermore, recent research on composite reliability has fre-
Reliability analysis is typically performed using two main quently employed metamodels to reduce computational costs.
approaches: simulation methods and moment methods. For instance, Lopes et al. [39] employed an ANN to replace the
Moment methods are employed to estimate a reliability LSF based on the Tsai-Wu failure criterion, demonstrating sig-
index, often defined as the distance from the origin to nificant computational efficiency. Dey et al. [40] focused on
the most probable failure point (MPFP). This index is cal- quantifying thermal uncertainty in the frequency responses of
culated using algorithms such as the first-order reliability laminated composite plates, using a surrogate model called high-
method (FORM) or the second-order reliability method dimensional model representation to propagate thermal, ply-
(SORM) [10,11], which rely on linear and nonlinear approx- level, and material uncertainties in frequency responses. Chen
imations based on Taylor series expansion of the LSF and Jia [41] investigated interlaminar stress analysis and LSF
around the MPFP to estimate the reliability index. While approximation methods for assessing the reliability of composite
these techniques are fast and efficient, they may lack pre- structures, discussing the use of surrogate methods such as SVM,
cision for highly nonlinear problems. To overcome this RSM, and ANN to approximate the true implicit LSF. Momeni
limitation, Monte Carlo simulations (MCS) have become Badeleh et al. [42] have introduced an advanced mesh-free finite
the preferred and widely used method in various fields volume approach, utilizing it to actively manage the vibrations
[12]. MCS involves generating random samples from prob- of a temperature-dependent piezoelectric laminated composite
ability distributions to model uncertain parameters, fol- plate based on the principles of first-order shear deformation.
lowed by performing repeated simulations to estimate The findings indicated a reduction in the reliability of the com-
the reliability of a system or process. Although this method posite plate with rising temperatures. Martinez et al. [43] con-
is highly effective, it can be costly when the probability of ducted a reliability analysis on a smart laminated composite
failure is very low. To address this issue, several specia- plate comprising a graphite/epoxy cross-ply substrate with a
lized sampling techniques have been developed based on piezoelectric fiber-reinforced composite actuator layer under
MCS, such as importance sampling (IS) [13–17], subset static electrical and mechanical loads. In their study, they devel-
simulation [18,19], directional simulation [20,21], and line oped a finite element (FE) model in COMSOL Multiphysics,
sampling [22,23]. While these approaches require fewer coupled with an ANN and integrated with MCS as well as first-
simulations than the MCS, they can still be computationally and second-order reliability methods (FORM/SORM). Haeri and
intensive, especially for complex models that demand sig- Fadaee [44] proposed an efficient and accurate reliability
nificant evaluation time. analysis approach for laminated composites called AKM-
For this reason, since the 1990s, several simulation MCS, employing an advanced Kriging model to approximate
techniques based on metamodels rather than the original the mechanical structure and applying the Tsai-Wu criterion
model have been developed to reduce the number of to define LSF. This approach demonstrates high computa-
necessary evaluations of the LSF. The goal of these techni- tional efficiency and accuracy. Mathew et al. [45] introduced
ques is to create a substitute model Y(X) that provides an an innovative approach that combines ANN with the SORM
explicit function to approximate the original performance and IS to accurately estimate failure probabilities and sen-
function. This substitute model is then used in place of the sitivities of variable stiffness composite laminate plates
original function in reliability methods. Among these while considering multiple sources of uncertainty. The
 Reliability and sensitivity assessment of laminated composite plates  3

results demonstrate a high level of accuracy in reliability these materials. It relies on the analysis of stresses, loads,
estimates and sensitivity studies. Finally, Zhou et al. [46] and the study of interactions within the composite structure,
proposed an adaptive Kriging-based approach for the relia- while seeking to anticipate potential failure modes. Several
bility and sensitivity analysis of composite structures with failure criteria have been developed to assess the strength of
uncertainties, using Kriging to approximate the structural composites in various loading scenarios. These criteria are
response. This method is applied to composite beams, plates, generally categorized into two main groups:
and random structures, yielding accurate results and effi- – Simple limit failure theories, such as the maximum
cient reliability evaluation. stress criterion,
Furthermore, to ensure the reliability of composites, it – Interaction theories, such as the Tsai-Hill and Tsai-Wu
is essential to investigate the contributions of uncertain criteria.
variables to the behavior of the composite material. This
allows for a more precise adjustment of the parameters
that have the most significant impact on the probability of 2.1 Maximum stress criterion
failure, a process commonly achieved through the theory of
failure probability sensitivity analysis. There are two main The maximum stress criterion is one of the primary
types of sensitivity analysis [47]: local reliability sensitivity failure criteria used to assess the strength of composite
analysis and global reliability sensitivity analysis. Local materials. This criterion is based on the concept that
reliability sensitivity analysis focuses on quantifying the failure occurs when a material exceeds its maximum
local effects of distribution parameters of random input stress-carrying capacity in one of its directions. In other
variables on the failure probability [48]. Its objective is to words, if any of the principal stresses applied to a com-
rank these distribution parameters based on their impact on posite material exceeds its corresponding strength,
the failure probability. On the other hand, global reliability failure occurs.
sensitivity analysis assesses the contributions of uncertain- To apply this criterion, the principal stresses computed
ties present in the input variables to the failure probability within the structure are compared with the material’s
[49]. It ranks these sources of uncertainty based on their maximum strengths in each direction. Failure happens
impact on the failure probability. when one of the following inequalities is not satisfied:
This article introduces an approach to analyze the
− Xc > σ1 > Xt , − Yc > σ2 > Yt , − Zc > σ3 > Zt
reliability and local sensitivity of the failure probability (1)
σ4 > R, σ5 > S , σ6 > T ,
of composite materials using an LSF that incorporates
the criteria of maximum stress, Tsai-Hill, and Tsai-Wu. where σ1, σ2, and σ3 represent the principal components of
The reliability methodology relies on an active learning normal stresses, and σ4, σ5, and σ6 represent the principal
method that combines MCS with a metamodel resulting components of shear stress, while Xt, Yt, Zt, Xc, Yc, and Zc
from a weighted combination of three metamodels: ANN, correspond to the ultimate tensile and compressive strengths
SVR, and Kriging. The weighting coefficients for these three of the lamina in the 1, 2, and 3 directions, and R, S, and T
metamodels are determined through a heuristic method denote the ultimate shear strengths of the lamina in the 2–3,
(Goel et al.). This approach provides valuable insights into 1–3, and 1–2 planes, respectively.
the behavior and performance of laminated composites, For a 2D in-plane stress state (σ3 = 0, σ4 = 0, σ5 = 0),
contributing to the enhancement of design and decision- which is often the case for laminates, the maximum stress
making processes while considering uncertainties related failure criterion is expressed as follows:
to material properties. Two applications on flat composite
− Xc > σ1 > Xt , − Yc > σ2 > Yt , σ6 > T . (2)
laminates were used to test this method and to demonstrate
its effectiveness.

2.2 Tsai-Hill criterion

2 Failure theory The Tsai-Hill failure criterion is based on the concept of

distortion energy [50]. It is an adaptation of the Von Mises
Failure theories play a central role in assessing the robust- criterion, initially designed for isotropic materials, but it
ness of structures. In the context of composite materials, has been specifically tailored for anisotropic materials
this fundamental discipline aims to understand the mechan- like composites. This failure criterion assesses the strength
isms and conditions leading to the rupture or degradation of of a composite material by considering the intricate
4  Najib Zemed et al.

interactions between stresses and strains that occur in an 3 Reliability and probability of
anisotropic material.
The failure condition according to the Tsai-Hill cri-
failure sensitivity theory
terion is defined as follows:
3.1 Failure probability
CritTsai − Hill
σ1 2 σ2 2 σ3 2 1 1 1
= ⎛ ⎞ + ⎛ ⎞ + ⎛ ⎞ − ⎛ 2 + 2 − 2 ⎞σ1σ2 The failure probability of structures is defined through the
⎝X⎠ ⎝Y ⎠ ⎝Z⎠ ⎝X Y Z ⎠
LSF G(x). This probability is defined as the percentage of
1 1 1 (3)
− ⎛ 2 + 2 − 2 ⎞σ2σ3 the region in which the LSF G(x) is less than zero. It is
⎝Y Z X ⎠
calculated using the following integral:
1 1 1 σ4 2 σ5 2 σ6 2
− ⎛ 2 + 2 − 2 ⎞σ1σ3 + ⎛ ⎞ + ⎛ ⎞ + ⎛ ⎞ < 1,
⎝R⎠ ⎝S⎠ ⎝T ⎠
⎝Z X Y ⎠ Pf = ∫ fx dx , (8)
G (x ) ≤ 0
where the values of X, Y, and Z represent Xt, Yt, Zt or Xc, Yc,
and Zc depending on the sign of σ1, σ2, and σ3. where fx represents the joint density of the input random
For a 2D in-plane stress state (σ3 = 0, σ4 = 0, σ5 = 0), the variables.
Tsai-Hill failure criterion is expressed as follows: The failure probability can be expressed as follows:
2 2 2
σ1 σ2 σ1σ2 σ6 ∫
CritTsai − Hill = ⎛ ⎞ + ⎛ ⎞ − ⎛ 2 ⎞ + ⎛ ⎞ < 1. (4) Pf = IDf (x )fx dx = E fx[IDf (x )]Ω , (9)
⎝X⎠ ⎝Y ⎠ ⎝ X ⎠ ⎝T ⎠ Ω

where E represents the expectation function with respect to fx ,

and IDf is the indicator function, which is defined as follows:
2.3 Tsai-Wu criterion ⎧ 1 si G (x ) ≤ 0
IDf (xk ) = ⎨
⎩0 si G (x ) ≥ 0.
The Tsai-Wu criterion is a widely used failure model for
assessing the safety of composite materials. It is based on The integral in Eq. (8) is complex to solve directly. For this
Beltrami’s theory of failure by total strain energy [51]. reason, the MCS is employed to estimate this integral. This
According to this criterion, a lamina or composite material method involves generating a population, denoted as P, which
is considered to fail if the following inequality is not is then used to simulate the system N times. The number of
satisfied: failure events is counted, and therefore, the probability of
failure can be estimated by the following expression:
CritTsai − Wu = Fiσi + Fijσiσj < 1, (5)
N
IDf (xi )
where Pf = E fx[IDf (x)]Ω ≈ Pˆf = ∑ . (10)
i=1 N
1 1 1 1 1 1
F1 = − ; F2 = − ; F3 = − ; Furthermore, to obtain a reliable degree of precision for
XT XC YT YC ZT ZC
this result, it is necessary to calculate the coefficient of varia-
1 1 1
F11 = ; F22 = ; F33 = , (6) tion (CoV) of P̂f , defined by the following relationship:
XTXC YTYC Z TZC
1 1 1 1 − Pˆf
F44 = ; F55 = ; F33 = , CoV = . (11)
R2 S2 T2 NPˆf
1 1 1 1 In most cases, a precision level of CoV < 0.05 is com-
F12 = − ; F13 = − ;
2 XTXCYTYC 2 XTXCZ TZC monly employed, as indicated by researchers [52,53].
1 1
F23 = − .
2 YTYCZ TZC
3.2 Local sensitivity of failure probability
For a 2D in-plane stress state (σ3 = 0, σ4 = 0, σ5 = 0), the
Tsai-Wu failure criterion is expressed as follows: The primary objective of local sensitivity analysis of failure
2
CritTsai − Wu = F1σ1 + F2σ2 + F11σ1 + F22σ2 + F66σ6 2 2 probability is to measure the local impacts of the distribu-
(7) tion parameters of input random variables on the failure
+ 2F12σ1σ2 < 1.
probability. It has been employed in the context of
 Reliability and sensitivity assessment of laminated composite plates  5

probability by the researcher [48]. To achieve this, the set of metamodels designed to simulate the LSF G(X). These
gradient of the failure probability with respect to these metamodels are subsequently utilized to assess the prob-
variables is evaluated: ability of failure and its sensitivity, replacing the LSF itself.
∂Pf This approach is typically characterized by five key features:
∂fx
∂θi
∫
= IDf (x )
∂θi
dx , (12) 1) The formulation of an LSF specifically adapted to com-
Ω posite laminates based on failure criteria.
where θi represents the parameter of the probability dis- 2) The construction of an average and robust metamodel,
tribution function of the ith random variable. which combines three distinct metamodels weighted by
Furthermore, to ensure a fair comparison of the sen- coefficients.
sitivities of all random variables in the studied system, it is 3) The use of an active learning method through a pro-
necessary to standardize these sensitivities with respect to posed learning function enables the intelligent selection
the probability of failure, making them dimensionless. A of samples to be added to the experimental plan.
normalization method proposed by Wu and Mohanty [54] 4) The evaluation of a stopping criterion tailored to this approach
is expressed as follows: ensures proper convergence of results towards the exact
5) The estimation of the probability of failure and the local
∂Pf /Pf σi ∂fx sensitivity of the probability of failure using MCS, in con-
S θi =
∂θi /σi
∫
= IDf (x )
Pf ∂θi
dx , (13)
Ω junction with the constructed average metamodel.

where σi represents the standard deviation of the prob-

The key characteristics of this approach are presented below.
ability distribution function of the ith random variable.
This sensitivity can be estimated using the MCS method:
σ ∂fx ∂fx ⎤
σi ⎡
S θi =
i
∫
Pf Ω
IDf (x ) f dx = E IDf (x )
fx ∂θi x Pf ⎣⎢ fx ∂θi ⎦⎥ Ω 4.1 LSF incorporating failure criteria
(14)
σi N IDf (x ) ∂fx In the context of reliability analysis theory, failure is always
≈ Sˆθi = ∑ .
Pf i = 1 N fx ∂θi described using an LSF, denoted as G(X). This function estab-
lishes a relationship between the limit state of the structure
If the input variables are independent, the joint den-
n and various factors, including applied loads, material proper-
sity can be decomposed into fx = ∏i = 1fxi . Thus, this sensi-
ties, and other relevant parameters. In this research, to ensure
tivity can be expressed as follows:
the reliability and safety of composites, and since none of the
N
σi IDf (x ) ∂fxi failure criteria guarantee the most conservative results in all
Sˆθi = ∑ . (15)
Pf i = 1 N fxi∂θi conditions according to Martinez and Bishay [55], it is assumed
that failure occurs when one of the three failure criteria is
According to Wu [48], two types of sensitivity can be
satisfied. To this end, an LSF based on the failure criteria pre-
defined, the first one being sensitivity of the probability of
sented in the previous section has been defined as follows:
failure with respect to the mean μ, and the second one with
respect to the standard deviation σ: ⎧ ⎡⎛ σ1 ⎞ σ2 σ3 σ4 σ5 σ6 ⎤ ⎫
⎪ 1 − i =max ; ⎛ ⎞ ; ⎛ ⎞ ; ⎛ ⎞; ⎛ ⎞; ⎛ ⎞ ⎪
N 1, … , 6 ⎢⎣⎝ X ⎠T / C ⎝ Y ⎠T / C ⎝ Z ⎠T / C ⎝ R ⎠ ⎝ S ⎠ ⎝ T ⎠⎥⎦
∂Pf /Pf σi IDf (x ) ∂fxi G (X ) = min ⎨ ⎬, (18)
S μi = ≈ Sˆμi = ∑ , (16) ⎪ 1 − CritTsai − Wu(X ) ⎪
∂μi /σi Pf i = 1 N fxi∂μi
⎩ 1 − CritTsai − Hill(X ) ⎭
N
∂Pf /Pf σi IDf (x ) ∂fxi where X = {x1, x2,..., xn} represents the vector of input
Sσi = ≈ Sˆσi = ∑ . (17)
∂σi /σi Pf i = 1 N fxi∂σi random variables, which typically include loads and mate-
rial properties, and the failure occurs when the LSF has a
negative value, G(X) ≤ 0.

4 Reliability and sensitivity analysis

method based on an ensemble of 4.2 Construction of a metamodel combining
metamodels three metamodels

The proposed reliability and sensitivity approach is based To construct a consolidated average metamodel from mul-
on an ensemble of metamodels, involving the creation of a tiple metamodels, the commonly used method is the
6  Najib Zemed et al.

metamodel ensemble approach, which has been recently experiments (DoEs) except the pair (x(k), y(k)), and NDoE is
incorporated into structural reliability analysis by the total number of DoE, and α and β are two parameters
researchers [56,57]. This approach involves training mul- that need to be defined. In this article, the same parameter
tiple metamodels using the same sample points and then values as those used by Goel et al. [58] have been applied,
combining them to obtain a more accurate prediction than α = 0.05 and β = −1.
what each individual model could provide separately. The
combination of these metamodels is achieved by deter-
mining the weights associated with each of them.
In the literature, there are two primary strategies for 4.3 UEM learning function
determining these weights. First, there is global weighting,
where the weights wi are constant for each metamodel ŷi. The accuracy of determining the failure probability largely
This includes methods such as weight calculation based on depends on the precise classification of points near the
root-mean-square error (RMSE) [58], Bayesian model aver- system’s state limit, where Y(x) = 0. To ensure this preci-
aging [59], or determining weights using an optimization sion, Echard et al. [53] implemented the U-learning func-
problem [42]. Second, there is local weighting, where the tion, which was developed in the context of the AK-MCS
weights depend on input values X. An example of this is algorithm. This function identifies points that may poten-
variance-based local weighting [60]. tially cross the predicted separator of the metamodel
In this article, the method of weight calculation based on Ŷ(x) = 0 incorrectly.
RMSE [58] is employed. This method suggests the calculation In this article, an adaptation of this LSF has been intro-
of the average metamodel using the following formula: duced. It involves the introduction of a minimum distance
M between the studied point and the points already present
Yˆens(x ) = ∑ wi(x )yˆi (x ), (19) in the design of experiments. This adaptation aims to
i=1 achieve a balanced distribution of all samples in the space,
where Ŷens(x) represents the global prediction of the EM at preventing excessive point concentration in a single area
given x, ŷi represents the output of the ith meta-model at and ensuring a uniform distribution. This learning func-
the specific point x, M stands for the total number of sur- tion is defined as follows.
rogates employed in the EM, and wi corresponds to the ∣ Yˆens(x )∣
weight assigned to the ith meta-model. UEM(x ) = , (25)
VYˆens(x ) d min(x )2
Moreover, this approach enables the detection of regions
where substantial prediction errors may occur by computing d min (x ) = min ‖x − xd‖2 , (26)
xd ∈ DoE
the variance VŶens(x ) of the EM, defined as follows:
M where Ŷens(x) represents the global prediction of the EM
VYˆens(x ) = ∑ wi(x )( yˆi (x ) − Yˆens(x ))². (20) at given x, VŶens(x ) is the variance of the metamodel in x,
i=1 and d min (x ) is the minimal distance, calculated using the
The weights are computed using a heuristic formulation Euclidean norm, between the point x and the nearest point
that employs a “leave-one-out” cross-validation strategy to in the DoE.
determine the RMSE. The weight calculation is as follows:
wi*
wi = M , (21)
∑i = 1wi* 4.4 Stopping criterion for the algorithm
wi* = (Ei + αE̅ ) β , (22)
The stopping criterion adopted in this method is based on
M
1 the evolution of the failure probability values over the last
E̅ = ∑ Ei , (23)
five iterations and is proposed as follows:
M i=1

1
NDoE ⎛ P fi − k − 1 − P fi − k ⎞
Ei = RSME = ∑ (Y (k ) − Yˆ (k )) , (24) εmax = max ⎜
P fi − k
⎟ < εstop, for i ≥ 5, (27)
NDoE k=1
k = 0,1,2,3,5 ⎝ ⎠

where Y(k) represents the real response at a given point x(k), where i represents the value of the last iteration and P fi − k − 1
Ŷ(k) represents the predicted response obtained from the and P fi − k represent the failure probabilities for iterations
ith surrogate model trained using all points of the design of (i−k−1) and (i−k), respectively.
 Reliability and sensitivity assessment of laminated composite plates  7

4.5 Calculation of failure probability and its nonlinearity in the output of a neuron, and in this study,
sensitivity we use the sigmoid function as activation function, which
is defined as follows:
The UEM learning function is computed for all points in the 1
population. The point with the lowest value is then incor- f (x ) = . (30)
1 + e −x
porated into the DoE. The best next point, denoted as X*, is
At the level of each neuron, a weighted set of inputs
then determined by:
from the previous layer is summed, and the activation
X * = argmin(UEM(x )) . (28) function is applied to this sum as follows:
x∈S

At each iterative step, X* is identified and incorporated zil = ∑j wijlalj − 1 + bil and a il = f(zil ) , (31)
into the DoE. Subsequently, metamodels are refined using
where, wijl represents the weight of the connection between
this new DoE, and the failure probability is calculated at
neuron j in layer l − 1 and neuron i in layer l , a lj − 1 is the
each iteration according to the following expression:
activation of neuron j in layer l − 1, and bil the bias term in
N neuron i in layer l .
IYˆens(x ) ≤ 0(xi )
Pˆ f = ∑ .
Considering that the sigmoid function is sensitive in
i=1 N
the interval [−1,1], we proceed to normalize all inputs and
Then, the local sensitivity of the failure probability is outputs (targets) as described by Rafiq et al. [61]:
calculated according to the following expression: For inputs:
N
σi IYˆens(x ) ≤ 0(xi ) ∂fxi X − Xmin
Sˆθi = ∑ . (29) S=2× − 1, (32)
Pf i = 1 N fxi∂θi Xmax − Xmin

where S is the normalized value of the variable X and Xmin

and Xmax are the minimum and maximum values of the
variables.
5 Metamodels employed in the For targets:
T − Tmin
proposed method out = 2 ×
Tmax − Tmin
− 1, (33)

In this section, the theories of the metamodels used in the where out is the normalized value of the targets T, and Tmin
ensemble metamodel method are presented, namely, ANN, and Tmax are the minimum and maximum values, respec-
SVR, and Kriging. tively, of the targets.
In this study, an ANN was employed, featuring three
hidden layers, each comprising 20 nodes. To determine the
problem’s weight values, the Levenberg-Marquardt (LM)
5.1 ANN algorithm implemented in MATLAB was utilized to address
the nonlinear least squares problem.
ANNs, a subset of machine learning, form the cornerstone
of deep learning algorithms. They draw their name and
structure from the human brain, aiming to simulate how 5.2 SVR
the human brain processes information. These networks
consist of layers of nodes, including an input layer, one or SVR is an evolution of the SVM that expands its capabilities.
more hidden layers, and an output layer. Each node is It introduces the concept of an ε-insensitive tube, as illu-
interconnected with others and is assigned an associated strated in Figure 2. Within this tube, deviations from the
weight (Wi), as depicted in Figure 1. target output are allowed without any penalties, while out-
Each neuron transforms the input into an output using side the tube, penalties are applied to deviations. SVR is
an activation function, which has the role of introducing a fundamentally a machine learning algorithm [36] that
8  Najib Zemed et al.

Input layer Hidden layers Output layer

Wi Wi
Input 1

Input 2

Input 3

.
.
. . . .
. . .
Input n . . .

Figure 1: ANNs architecture.

relies on a loss function Lε called e-insensitive function. This approximation assumes that all inputs exist inside
This function is designed to handle cases where we are the ε-tube, but this is not always the case, so slack variables ξi,
less concerned with small deviations within the ε tube ξi* are introduced, and this optimization problem becomes:
but want to penalize larger deviations outside it, making n
1
SVR particularly useful for robust regression tasks. This Min ‖w‖2 + C ∑ (ξi + ξi*)²
2
loss function Lε is defined as follows: i=1

⎧ yi − b − 〈w , xi〉 ≤ ξi + ε (37)
⎪
⎧ 0
Lε ( y ) = max ⎨ (34) S.t⎨ − yi + b + 〈w , xi〉 ≤ ξi* + ε
⎩|y − f ( x )| − ε. ⎪
⎩ ξi* and ξi ≥ 0.

5.2.1 Linear case

In the linear case, f(x) is defined as follows:

Y (x ) = 〈w , x 〉 + b , (35)

where 〈.,.〉 is the inner product operator, w ∈ n is the

normal vector of the hyperplane defined by the regression
function, and b is the bias.
According to Vapnik [36,37], achieving a flat approxi-
mation linear function implies maximizing the margin,
and this objective can be accomplished by minimizing
the Euclidean norm of w, which means ‖w‖2 , then this
problem is formulated as a convex optimization problem:
1
Min ‖w‖2
2
(36)
⎧ yi − b − 〈w , xi〉 ≤ ε
S.t⎨
⎩− yi + b + 〈w , xi〉 ≤ ε ,
Figure 2: Curve of the SVR approximation function with the slack
where i = 1,…, N : Number of training points. variables.
 Reliability and sensitivity assessment of laminated composite plates  9

where the constant C > 0 plays a crucial role in determining function Φ is not needed, thanks to the kernel trick which
the balance between achieving a flat approximation (small w) allows us to use the kernel function K, defined as follows:
and the tolerance level for deviations exceeding ε.
K (x , x ′ ) = 〈φ(x ) , φ(x ′)〉. (44)
To solve the optimization problem with the linear con-
straints, the Lagrangian function is used: This function K represents the dot product between
n two projections. In this context, the Lagrangian optimiza-
∥w∥2
Min L = + C ∑ (ξi + ξi*)2 tion problem can be expressed as follows:
2 i=1
n 1 n
− ∑ αi(ξi + ε − yi + b + 〈w , xi〉) Max L = − ∑ (αi − αi*)K (xi , xj )(αj − αj*)
2 i,j= 1
i=1
n
(38) n n

− ∑ αi*(ξi* + ε + yi − b − 〈w , xi〉) − ε ∑ (αi + αi*) + ∑ yi (αi − αi*)

i=1 i=1 i=1 (45)
n n
⎧
− ∑ (ηi*ξi* + ηi ξi ), ⎪ ∑ (αi − αi*) = 0
i=1 S.t⎨ i=1
⎪
⎩ αi , αi* ∈ [0, C ].
where αi , αi*,ηi , ηi* are the Lagrange multipliers.
For optimality, according to the Lagrangian theory, it From Eq. (39), and considering the projection function
is necessary for the partial derivatives of L with respect to φ for the nonlinear case, the expression of the weight
each variable to vanish. vector is given as follows:
n
∂L n
=w− ∑ (αi − αi*)xi = 0, (39) w= ∑ (αi − αi*)φ(xi ). (46)
∂{ w } i=1
i=1
n
∂L To calculate the offset b, the Karush–Kuhn–Tucker
= ∑ (αi − αi*) = 0, (40)
∂{ b } i=1 condition [62] are utilized, These conditions state that the
∂L
n product between the constraints and the Lagrange multi-
= ∑ −αi − ηi + C = 0, (41) pliers must equal zero at the optimal solution. This implies:
∂{ξi } i=1
n ⎛ n ⎞
∂L b = ⎜yi − ∑ (αj − αj*)K (xj , xi ) − εup⎟ if
= ∑ −αi* − ηi* + C = 0. (42) (47)
∂{ξi*} i=1
⎝ j=1 ⎠

To solve the Lagrangian optimization in Eq. (38), it is αi ≠ 0 and αi ≠ C ,

preferable to solve it in its dual formulation, by substi- ⎛ n ⎞
tuting Eq. (39) into Eq. (42) in Eq. (38), yields the dual for- b = ⎜yi − ∑ (αj − αj*)K (xj , xi ) + εdown ⎟ if
⎝ ⎠ (48)
mulation, which depends only on αi: j=1

αi* ≠ 0 and αi* ≠ C .

1 n
Max L = − ∑ (αi − αi*)〈xi , xj〉(αj − αj*) Then, the approximation function is written as:
2 i,j= 1
n n n
+ ∑ yi (αi − αi*) − ε ∑ (αi + αi*) Yˆ (x ) = ∑ (αi − αi*)k (xi , x ) + b. (49)
i=1 i=1 (43) i=1
n
⎧
⎪ ∑ (αi − αi*) = 0
S.t⎨ i = 1
⎪
⎩ αi , αi* ∈ [0, C ]. 5.3 Kriging metamodel

The Kriging model, a nonlinear interpolation meta-model

5.2.2 Nonlinear case developed for Geostatistics by Matheron [63], also known
as Gaussian process, is a statistical interpolation method
For the nonlinear case, a nonlinear projector Φ is used, employed to predict or estimate unknown values within a
which has the objective of transforming the starting space given space by utilizing observations made at various
to a space of higher dimension than the starting one, points within that space. This approach is based on estab-
Φ : X → F. Also, the explicit equation of the projection lishing a stochastic random field that simulates the
10  Najib Zemed et al.

behavior of the limit function. Subsequently, the technique These two equations lead to:
known as the best linear unbiased predictor (BLUP) is uti-
βˆ = (F T R(θ )−1 F )−1F T R(θ )−1 F , (55)
lized to optimally estimate the value of this model at a
specific point. The Kriging function is written as a realiza- 1
σˆ 2 = (Y − 1β )−1R(θ )−1 (Y − 1β ) . (56)
tion of a random function as described by [64]. n

G (x ) = F (x )T β + z(x ) , (50) However, the two parameters in Eqs. (55) and (56)
depend on the correlation parameter θ, so it is first neces-
where F (x ) = [{ f1 (x )},…,{ fp (x )}]represents the basic func- sary to obtain it using maximum likelihood estimation:
tions and β = {β1,…, βp} is the vector of regression coeffi-
cients, and in this article, ordinary Kriging is chosen, which θˆ = arg minθ − log L(β , σ 2 ,θ ) . (57)
means that F (x ) is a vector of ones and the product
By expanding the optimization problem of the Eq. (57)
F (x )T β = β is a scalar.
and eliminating the constant terms, the likelihood function
Z(x) is a stationary Gaussian process that has an
is reduced to:
unknown form, zero mean and the following covariance
functions between two points of space:
θˆ = arg min θ⎛ det(R(θ ))n σ̂ 2⎞.
1
(58)
⎝ ⎠
cov(xi , xj ) = σ 2Rθ (xi ,xj ) , (51)
After the determination of the parameter θ, the best
where σ is the variance of the process and σ 2Rθ (xi ,xj ) is the
linear unbiased predictor (BLUP) is used to determine Gˆ (x )
autocorrelation function between the points xi and xj .
the estimation of the response G(x) and to determine σ Ĝ2(x )
In this type of Kriging, the Gaussian process is sta-
tionary, which means that the autocorrelation function R the variance of Gˆ (x ) at the point x.
depends only on the difference between the points and on Gˆ (x ) = β + r (x )T R(θ )−1 (Y − 1β ) , (59)
a set of hyperparameters θ ∈  +n , and this correlation
model can be formulated as follows: where r (x ) = {Rθˆ(x ,xi )}i = 1, … , n

N0 σ G2ˆ (x ) = σ 2(1 + u(x )t (1tR(θ )−1 1) , (60)

(l ) (l )
Rθ (xi ,xj ) = ∏ exp( − θl (xi − x j )2) , (52)
l=1 where u(x ) = 1tR(θ )−1 r (x ) − 1 and 1 is the vector of size n
and filled with 1.
where i, j = 1,., n, with n representing the number of
To build the Kriging model in this article and to solve
random variables, and N0 denoting the number of points
the global optimization problem of Eq. (58), which is com-
in the design of experiments.
The identification of the hyperparameters β , σ , and θ plex and cannot be solved analytically, an optimization
can be determined using the maximum likelihood method algorithm is developed in MATLAB to solve this problem.
defined as follows:
1 1
L= exp ⎡− 2 (Y − 1β )T R(θ )−1 (Y
(2πσ 2 )n ∣
1
R (θ )∣ 2
⎣ 2σ 6 Algorithm of the proposed
(53)
− 1β )⎤,
methodology
⎦
The methodology developed in this article for reliability
where R(θ ) = Rθ (xi ,xj ) is the matrix of correlation
analysis and sensitivity to reliability is organized into three
between each pair of points of the design of experiments,
main stages. The implementation details of this method are
and it is a symmetric correlation matrix and 1 is the vector
described below and illustrated in Figure 3.
of size n and filled with 1.
Then, since maximizing this likelihood is equivalent to
minimizing its opposite natural logarithm, the first-order
optimality conditions for this likelihood logarithm is used 6.1 Stage 1: convergence analysis and
to determine the two estimations of the parameters β numerical validation
and σ2.
This stage holds paramount significance within our meth-
⎧ ∇β − log L(β , σ 2 ,θ ) = 0
⎨ (54) odology, aiming to validate the FE model and determine
⎩ ∇σ 2 − log L(β , σ 2 ,θ ) = 0. the optimal mesh to use. The first substage involves
 Reliability and sensitivity assessment of laminated composite plates  11

Figure 3: Algorithm of the methodology used for calculating the reliability and sensitivity of composite materials.

validating the obtained results by comparing them with models based on different theories or concepts. Alternatively,
analytically calculated results. Subsequently, the second running it within a different finite element software is also a
substage aims to determine the optimal mesh size. It is well viable approach.
known that a finer mesh yields more precise results, albeit
often at the cost of longer computational time. Therefore, a
convergence study becomes necessary, employing various 6.2 Stage 2: reliability analysis
mesh sizes to ascertain the optimal size that provides both
adequate accuracy and reasonable computation time. The objective of this stage is to achieve a precise estimation
Moreover, in the absence of precise analytical results, of failure. It includes the following steps:
it is conceivable to assess the model’s validation and conver- 1. Generate the Monte Carlo population S using Latin
gence by comparing it to other established finite element hypercube sampling (LHS) technique.
12  Najib Zemed et al.

2. Select from S a number N0 (e.g., N0 = 12) of points X = stopping criterion rather than its original one to ensure a
{X1,…, X N0 }, and calculate the corresponding system fair and consistent comparison.
responses for the initial N0 points, Y = {Y1,…,YN0 }; Local sensitivity analysis was conducted to assess the
3. Build the three metamodels Kriging–ANN–SVR using {X, sensitivity of failure probability with respect to changes
Y} and determine the metamodel mean Yˆens(x ) and its in the mean and standard deviation of input variables.
variance VŶ ens(x ) This sensitivity was calculated using the reliability results
4. Estimate the failure probability Pf using the equation: obtained from the three methods employed to evaluate the
N laminate’s reliability. A positive sensitivity indicates that
IYˆens(x ) ≤ 0(xi )
Pf ≈ ∑ . (61) when the mean of a variable increases, the probability of
N
i=1 failure also increases, while a negative sensitivity suggests
5. Evaluate the stopping criterion in Eq. (46). If the condi- the opposite, namely, that an increase in the mean of a
tion is met, proceed to the next stage; otherwise, find the variable decreases the probability of failure.
best point X* from all points in S to add it to the DoE, by
minimizing the learning function UEM. Then calculate the
response Y(X*). Next, update the X = {X ∪ X*} ,
Y = {Y ∪ G (X*)} and return to step (3). 7.1 Techniques and applications utilized
6. Calculate the coefficient of variation of Pf using Eq. (11)
and check if CoV < 0.05 proceed to stage 3; otherwise, 7.1.1 Tools and software
expand the population by adding Snew to S, and return
to step (1). The simulations were conducted using the ANSYS APDL software
with the SHELL181 finite element, which provides six degrees of
freedom at each node, as illustrated in Figure 4. This software
was used to model the laminates. Furthermore, the necessary
6.3 Stage 3: reliability sensitivity analysis codes for performing reliability analysis and sensitivity analysis
were developed within the MATLAB environment. A coupling
Once the failure probability calculation algorithm is com- between the two software tools was established.
pleted, the sensitivity of the failure probability with respect
to a given parameter θi is estimated using the following
expression:
N
σi IYˆens(x ) ≤ 0(xi ) ∂fxi
Sˆθi = ∑ . (62)
Pf i = 1 N fxi∂θi

7 Numerical results and discussion

Numerical analysis was conducted on two graphite/epoxy
T300/5208 composite plates examined by [65]. The first com-
posite plate is characterized by an antisymmetric angle-ply
configuration [−45°, 45°, −45°, 45°], subjected to a uniformly
distributed transverse load. The second composite plate adopts
a symmetric cross-ply arrangement [0°, 90°, 90°, 0°], and it
experiences uniaxial loading on one side only.
Reliability analysis for these two applications was con-
ducted using three different methods to ensure the relia-
bility of the results and facilitate comparisons between
these methods. The three methods employed are the MCS
method, the proposed method in this article, and the well-
known AK-MCS method [53], which serves as a reference for
new approaches based on machine learning. It is worth Figure 4: Quadrilateral four-node shell element in ANSYS software
noting that, for the AK-MCS method, we utilized its proposed (SHELL181).
 Reliability and sensitivity assessment of laminated composite plates  13

This element is based on the FSDT, which assumes that 7.2 Convergence analysis and validation of
the transverse deformation through the thickness is con- the FE model
stant and also takes into account the transverse shear dis-
placement. The parameter settings and assumptions used A FE validation study was conducted to confirm the accu-
for this element in ANSYS are as follows: racy of the model. Furthermore, a convergence study was
– KEYOPT(1) = 0: This parameter is set to include both carried out by varying the mesh size (2 × 2), (4 × 4), (6 × 6),
membrane forces and moments in the calculation of (8 × 8), and (10 × 10) to determine the optimal mesh,
the element stiffness. allowing for precise analytical results while minimizing
– KEYOPT(3) = 0: It is set to zero to utilize reduced inte- resource utilization and computation time. Precise deflec-
gration, which reduces computational time while main- tion values for the plates were extracted from the research
taining acceptable accuracy. of Reddy and Pandey [65,66]. The properties of the epoxy
– KEYOPT(5) = 1: This option selects the standard formula- graphite material T300/5208 and the applied force for this
tion of the shell under the FSDT theory. validation are presented in Table 1.
– KEYOPT(8) = 3: By setting this parameter to 3, results are The results, presented in Table 2, indicate that the max-
computed at the top, bottom, and mid-plane of each imum deflection values obtained from numerical simula-
layer, allowing us to subsequently assess failure criteria tions of the plate are nearly equal to the exact solutions.
at three points for each layer, aiming to ensure more This confirms the validation of the FE-based model. Further-
reliable results. more, in the analysis of results convergence based on mesh
– KEYOPT(9) = 0: No subroutine is used to provide initial size, the choice was made to use the (8 × 8) mesh for all
thickness, meaning that thickness data is directly defined subsequent steps. This selection is justified by its ability to
within the model. provide sufficiently accurate results compared to the exact
– KEYOPT(10) = 0: By default, the normal stress is assumed solution while exhibiting minimal differences compared to
to be zero. the next mesh size, which is (10 × 10).
– KEYOPT(11) = 0: The standard orientation axes of the
element are used by default.

7.3 Example 1: Composite plate under

7.1.2 Transformation of stress to the local coordinate Uniformly distributed transverse load
systems of each laminate layer
The first example involves a laminated composite plate con-
In the context of evaluating failure criteria, it is imperative sisting of 4 layers with an antisymmetric angular arrange-
to analyze stresses and strains with respect to the material ment [−45°, 45°, −45°, 45°] made of graphite/epoxy T300/5208,
axes specific to each constituent layer of the laminate. as depicted in Figure 5. This laminate is simply supported
However, it is common for the intrinsic axes of the lami- at its edges and subjected to a uniformly distributed load over
nate not to align with the principal axes of individual its surface. Dimensional characteristics such as length, width,
layers. Since the stresses predicted by ANSYS are calculated and thickness, as well as the properties of the graphite/epoxy
with respect to the global axes (σxx, σyy, σzz, τyz, τxz, τxy), a material (elasticity modules in different directions, Poisson’s
transformation using MATLAB is performed to determine ratio in direction 12, maximum strengths in different direc-
the principal stresses (σ1, σ2, σ3, σ4, σ5, σ6) for each lami- tions), and the applied surface force are all considered as
nate. This transformation is carried out using the following random variables in this example. The probability distribu-
matrix: tions of these variables are presented in Table 3.

⎛ cos2 θm sin2 θm 0 0 0 2 sin θm cos θm ⎞ ⎛ σxx ⎞

⎛ σ1 ⎞
⎜ sin2 θm cos2 θm 0 0 0 − 2 sin θm cos θm ⎟ ⎜ σyy ⎟
⎜ σ2 ⎟
σ ⎜ 0 0 0 0 0 1 ⎟ ⎜ σzz ⎟
⎜ 3⎟ = . (63)
⎜ ⎟ ⎜ τyz ⎟
⎜ σ4 ⎟ 0 0 0 cos θm − sin θm 0
σ ⎜ ⎟ ⎜⎜ τxz ⎟⎟
⎜ ⎟5 0 0 0 cos θm sin θm 0
⎝ σ6 ⎠ ⎜ ⎟
⎝ τxy ⎠
⎝ − sin θm cos θm sin θm cos θm 0 0 0 (cos2 θm − sin2 θm ) ⎠
14  Najib Zemed et al.

Table 1: Applied force and mechanical properties of epoxy graphite material T300/5208

Parameter Description Unit Value

E11 Elastic modulus in the 11 direction MPa 132,500

E22 Elastic modulus in the 22 direction MPa 10,800
v12 Poisson’s ratio in the 12 direction — 0.24
G12 Elastic modulus in the 12 direction MPa 5,700
G13 Elastic modulus in the 13 direction MPa 5,700
G23 Elastic modulus in the 23 direction MPa 3,400
L Length of the laminate m 0.2286
l Width of the laminate m 0.127
t Thickness of the laminate m 0.000127
XT Tensile strength of the laminate in the 11 direction MPa 1,515
Xc Tensile strength in the 22 and 33 directions MPa 1,697
YT & Z T Compression strength in the 22 and 33 directions MPa 43.8
Yc & Z c Compression strength in the 22 and 33 directions MPa 43.8
R Shear strength in the 23 direction MPa 67.6
S&T Shear strength in the 13 and 12 directions MPa 86.9
P Applied surface force MPa 0.000689476

Table 2: Results of the study on FE model validation and convergence

Laminate Solution found by FEM (mm) Exact solution (mm) [58] Error (%)

Mesh (2 × 2) (4 × 4) (6 × 6) (8 × 8) (10 × 10) — —

[−45°, 45°, −45°, 45°] 2.6740 2.8148 2.7713 2.7654 2.7634 2.7609 0.0877
[0°, 90°, 90°, 0°] 5.4153 5.8615 5.8451 5.8364 5.8323 5.8166 0.2699

7.3.1 Reliability analysis of the composite plate the AK-MCS method. It can be observed that the proposed
method rapidly converged to the exact solution starting
The reliability results for the laminate are presented in from the 86th iteration, and the algorithm did not require
Table 4. The exact probability was computed using the many simulations to meet the stopping criterion, as it
MCS method with 100,000 simulations. It is noteworthy reached it after only about 16 additional simulations. In
that the proposed method yielded results very close to the contrast, the AK-MCS method required a much higher
exact value, with an estimation error of only 2.27% com- number of simulations to meet the same stopping criterion.
pared to the MCS method. Furthermore, it is important to Figure 7 presents the evolution of weighting coeffi-
highlight that the proposed method required only 102 simu- cients associated with the three metamodels used in the
lations to achieve this probability, which is significantly proposed method. Overall, it can be observed that the Kri-
fewer than the simulations needed by the AK-MCS method. ging metamodel dominates in the initial and final segments
Figure 6 illustrates the evolution curve of the esti- of the curve, while in the central part of the curve, covering
mated failure probability using the proposed method and approximately one-third, the SVR metamodel takes prece-
dence at certain points. On the other hand, the ANNs meta-
model consistently maintains low weights in comparison to
the other two, except in 3 out of the 102 iterations where it
outperforms the other two metamodels.

7.3.2 Sensitivity analysis of the composite plate

Figures 8 and 9, respectively, depict the sensitivity of

failure probability regarding the mean and standard devia-
Figure 5: Composite plate under uniformly distributed transverse load. tion of variables. It is clearly observed that the thickness t
 Reliability and sensitivity assessment of laminated composite plates  15

Table 3: Random variable distribution of the antisymmetric angle-ply [−45°, 45°, −45°, 45°]

Variable Description Unit Mean value Standard deviation Distribution

E11 Elastic modulus in the 11 direction MPa 132,500 13,250 Normal

E22 Elastic modulus in the 22 direction MPa 10,800 1,080 Normal
v12 Poisson’s ratio in the 12 direction — 0.24 0.024 Normal
G12 Elastic modulus in the 12 direction MPa 5,700 570 Normal
G13 Elastic modulus in the 13 direction MPa 5,700 570 Normal
G23 Elastic modulus in the 23 direction MPa 3,400 340 Normal
L Length of the laminate m 0.2286 0.02286 Normal
l Width of the laminate m 0.1270 0.002 Normal
t Thickness of the laminate m 0.000127 0.0000127 Normal
XT Tensile strength of the laminate in the 11 direction MPa 1,515 151.5 Normal
XC Tensile strength in the 22 & 33 directions MPa 1,697 169.7 Normal
YT & Z T Compression strength in the 22 & 33 directions MPa 43.8 4.38 Normal
YC & Z C Compression strength in the 22 & 33 directions MPa 43.8 4.38 Normal
R Shear strength in the 23 direction MPa 67.6 6.76 Normal
S&T Shear strength in the 13 & 12 directions MPa 86.9 8.69 Normal
P Applied surface force MPa 0.0035 0.00035 Gumbel

of the laminate layers has a significant impact on the

Table 4: Results of reliability analysis for the antisymmetric angle-ply failure probability. Furthermore, given its positive sensi-
composite [−45°, 45°, −45°, 45°] tivity to the mean, it can be affirmed that it positively
contributes to the increase in the failure probability. Vari-
Method Number of calls Pf Error (%) ables such as Young’s moduli E11 and E22, as well as the
MCS 100,000 0.009131 — maximum strengths YT/ZT and YC/ZC, length L, width l, and
AK-MCS [53] 12 + 159 = 171 0.009596 5.09 force P exhibit almost the same level of impact. Among
Proposed method 12 + 90 = 102 0.009339 2.27 these, E22, L, l, and P significantly contribute to the failure
probability, while E11, YT/ZT, and YC/ZC have a reducing

Figure 6: Evolution of probability for example 1.

16  Najib Zemed et al.

Figure 7: Evolution of weight contribution of metamodels for example 1.

effect on the failure probability. As for the remaining vari- becomes feasible to significantly diminish the probability
ables, their impact on the failure probability is negligible. of failure. This proactive approach will bolster the overall
To enhance reliability and reduce the probability of resilience of the structure and promote a more dependable
failure of this composite plate, it is imperative to take performance of the composite plate.
appropriate measures. Based on the results of our sensi-
tivity analysis, it is evident that enhancing the robustness
of critical variables is essential. Specifically, by prioritizing 7.4 Laminates with hole under in-plane
the optimization of the thickness t of the laminate layers loading
and selecting high-quality composite materials character-
ized by high Young’s modulus E11 and optimal maximum The second example is a laminated composite plate com-
strengths (YT/ZT and YC/ZC), while ensuring that dimensions prising 4 layers, arranged in a symmetric cross-ply config-
(length and width) align with the required specifications, it uration [0°, 90°, 90°, 0°], made of graphite/epoxy T300/5208,

Figure 8: Sensitivity of failure probability to the mean of variables for example 1.

 Reliability and sensitivity assessment of laminated composite plates  17

Figure 9: Sensitivity of failure probability to the standard deviation of variables for example 1.

and featuring a circular hole at the center of the plate. This

Nx
laminate is clamped on one side and simply supported on
the other, while being subjected to a linear tensile load on
x that side, as illustrated in Figure 10.
In the context of this example, given the absence of an
t analytical solution to validate the FE model, the structure was
L modeled using two types of predefined FEs in the ANSYS
software, namely, shell181 and shell281. A comparison of
Figure 10: Composite laminates with a hole under in-plane loading. the displacement results for the node located at the center

Table 5: Random variable distribution of the symmetric cross-ply [0°, 90°, 90°, 0°]

Variable Description Unit Mean value Standard deviation Distribution

E11 Elastic modulus in the 11 direction MPa 13,2500 13,250 Normal

E22 Elastic modulus in the 22 direction MPa 10,800 1,080 Normal
v12 Poisson’s ratio in the 12 direction — 0.24 0.024 Normal
G12 Elastic modulus in the 12 direction MPa 5,700 570 Normal
G13 Elastic modulus in the 13 direction MPa 5,700 570 Normal
G23 Elastic modulus in the 23 direction MPa 3,400 340 Normal
L Length of the laminate m 0.2286 0.02286 Normal
l Width of the laminate m 0.127 0.002 Normal
d Subtracted hole diameter m 0.0254 0.00254 Normal
t1 Thickness of laminate 1 m 0.000127 0.0000127 Normal
t2 Thickness of laminate 2 m 0.000127 0.0000127 Normal
t3 Thickness of laminate 3 m 0.000127 0.0000127 Normal
t4 Thickness of laminate 4 m 0.000127 0.0000127 Normal
θ1 Fiber orientation in laminate 1 ° 0 5 Normal
θ2 Fiber orientation in laminate 2 ° 90 5 Normal
θ3 Fiber orientation in laminate 3 ° 90 5 Normal
θ4 Fiber orientation in laminate 4 ° 0 5 Normal
Nx Linear force applied MN/m 0.045 0.0045 Gumbel
18  Najib Zemed et al.

Table 6: Results of reliability analysis for the symmetric cross-ply [0°, of the edge under a uniformly distributed load revealed con-
90°, 90°, 0°] sistency between the two approaches.
Regarding meshing, a specific strategy was employed,
Method Number of calls Pf Error (%) with an increase in density near the central hole and a
MCS 50,000 0.0339 — decrease in other areas. To assess mesh convergence, var-
AK-MCS [53] 12 + 241 = 253 0.0396 16.81 ious mesh sizes were utilized, gradually decreasing. The
Proposed method 12 + 142 = 154 0.0308 9.14 optimal mesh size was chosen by observing that the lateral
displacement results of the specified node showed no

Figure 11: Evolution of probability for example 2.

Figure 12: Evolution of weight contribution of metamodels for example 2.

 Reliability and sensitivity assessment of laminated composite plates  19

Figure 13: Sensitivity of failure probability to the mean of variables for example 1.

significant improvement for smaller meshes, while main- 7.4.1 Reliability analysis of the composite plate
taining stability for the larger mesh.
In this example, the maximum strengths in different The reliability results for the laminate are presented in
directions are considered constant and are presented in Table 6. The exact probability was computed using the
Table 1. The variables in this problem are the length, the MCS method with 50,000 simulations. It is observed in
width, the thickness of each layer, the radius of the hole, the table that the proposed method yielded an acceptable
the fiber orientation for each layer, the elastic modulus in result with an error of 9.14%, and it achieved this with only
different directions, Poisson’s ratio in direction 12, and the 154 simulations. In contrast, the AK-MCS method required
applied linear force. The probability distributions of these a larger number of simulations to produce a result with a
variables are presented in Table 5. greater error than the proposed method.

Figure 14: Sensitivity of failure probability to the standard deviation of variables for example 2.
20  Najib Zemed et al.

The evolution of failure probability as a function of the 8 Conclusion

number of simulations is presented in Figure 11, illus-
trating the trajectory of the estimated failure probability The new reliability methodology for laminated composites
using both our proposed method and the AK-MCS method. was presented, cleverly combining MCS with a metamodel
It is remarkable that our method converged to the exact resulting from the fusion of three approaches: ANN, SVM,
solution by the 154th iteration, while the AK-MCS method and Kriging. This methodology demonstrated remarkable
required 253 simulations to reach the same stopping cri- accuracy while minimizing the number of simulations
terion. Therefore, it is encouraging to note that our method required to estimate the probability of failure. Its efficiency
required a relatively modest number of simulations, espe- is undeniable. However, it is important to note that the
cially considering the large number of random variables accuracy of the estimation may vary depending on the quality
involved, totaling 18 variables. of the metamodels used, as shown in our second example
Figure 12 shows the evolution of the weighting coeffi- where the precision of the metamodels slightly decreased.
cients of the three meta-models used in the proposed Thus, ensuring the robustness of the metamodels is essential
method. In this example, the Kriging meta-model does to guarantee reliable results. It is important to emphasize that
not dominate as much as in the previous example. In the failure criteria used in our analysis may not always accu-
fact, it can be observed that the three metamodels almost rately reflect the actual behavior of composite materials in
equally contributed throughout the process. This shows the complex situations. The accuracy of these criteria should be
advantage of this approach, which combines the strengths considered in the interpretation of the results.
of each meta-model by assigning them different weights In addition, local sensitivity was evaluated, highlighting
based on their accuracy. the critical variables that most influence the probability of
failure. This rigorous analysis has allowed us to recommend
essential optimization measures to strengthen reliability
7.4.2 Sensitivity analysis of the composite plate and reduce the probability of failure of composite plates.
In particular, it is imperative to prioritize the optimization
Figures 13 and 14, respectively, illustrate the sensitivity of of the thickness of the laminated layers, to choose high-
failure probability concerning the mean and standard quality composite materials, and to ensure that the dimen-
deviation of variables. It is evident that the mechanical sions strictly meet the required specifications. These actions
properties of the laminates have a negligible impact on aim to guarantee optimal performance of laminated com-
the probability of failure. However, regarding the thick- posite materials, even in uncertain situations, and to
ness of the laminates, only the thicknesses t2 and t3, corre- strengthen their structural integrity.
sponding to laminates with fibers oriented at 90°, have In summary, this study opens valuable perspectives for
moderate and negative effects on the probability of failure. improving design and decision-making in the field of laminated
Concerning the dimensions of the laminate, the length L of composite materials, while ensuring reliable performance in
the laminate has no effect, while the width and hole dia- the face of uncertainty. It encourages further research to
meter have a significant impact on the probability of further refine the reliability and sensitivity methodologies
failure, with negative and positive effects, respectively. applicable to these crucial materials in modern engineering.
As for the fiber orientation angles, they have no effect
on the probability of failure when their mean changes. Funding information: The authors state no funding involved.
However, for variations in their standard deviations, angles
θ2 and θ3 have an impact on the probability of failure. Author contributions: All authors have accepted responsi-
Finally, just like in the first example, the applied force has bility for the entire content of this manuscript and approved
a substantial impact as a contributor to the probability of its submission.
failure.
According to the results of our sensitivity analysis, Conflict of interest: The authors state no conflict of interest.
to enhance the reliability and reduce the probability of
failure of this composite plate, it is essential to decrease
the hole diameter and increase the width of the composite,
as these adjustments have a significant impact on the prob- References
ability of failure. In addition, a slight increase in the thick-
ness of the 90° oriented fiber layers can also contribute to [1] Timoshenko S, Woinowsky-Krieger S. Theory of plates and shells.
strengthening the structure. Vol. 2. New York: McGraw-hill; 1959.
 Reliability and sensitivity assessment of laminated composite plates  21

[2] Reddy JN. Theory and analysis of elastic plates and shells. Boca [24] Bucher CG, Bourgund U. A fast and efficient response surface
Raton: CRC Press; 2006. approach for structural reliability problems. Struct Saf.
[3] Reddy JN. A refined nonlinear theory of plates with transverse 1990;7(1):57–66.
shear deformation. Int J Solids Struct. 1984;20(9–10):881–96. [25] Gavin HP, Yau SC. High-order limit state functions in the response
[4] Hosseini-Hashemi S, Fadaee M, Atashipour SR. Study on the free surface method for structural reliability analysis. Struct Saf.
vibration of thick functionally graded rectangular plates according to a 2008;30(2):162–79.
new exact closed-form procedure. Compos Struct. 2011;93(2):722–35. [26] Huh J. Reliability analysis of nonlinear structural systems using
[5] Reddy JN. A refined shear deformation theory for the analysis of response surface method. KSCE J Civ Eng. 2000;4(3):135–43.
laminated plates. National Aeronautics and Space [27] Rajashekhar MR, Ellingwood BR. A new look at the response sur-
Administration; 1986. face approach for reliability analysis. Struct Saf. 1993;12(3):205–20.
[6] Tornabene F, Viscoti M, Dimitri R. Generalized higher order layerwise [28] Deng J, Gu D, Li X, Yue ZQ. Structural reliability analysis for implicit
theory for the dynamic study of anisotropic doubly -curved shells performance functions using artificial neural network. Struct Saf.
with a mapped geometry. Eng Anal Bound Elem. 2022;134:147–83. 2005;27(1):25–48.
[7] Tornabene F, Viscoti M, Dimitri R. Static analysis of doubly-curved [29] Papadrakakis M, Papadopoulos V, Lagaros ND. Structural reliability
shell structures of smart materials and arbitrary shape subjected to analyis of elastic-plastic structures using neural networks and
general loads employing higher order theories and generalized Monte Carlo simulation. Comput Methods Appl Mech Eng.
differential quadrature method. Comput Model Eng Sci. 1996;136(1):145–63.
2022;133(3):719–98. [30] Shao S, Murotsu Y. Structural reliability analysis using a neural
[8] Brischetto S. Hygrothermal loading effects in bending analysis of network. JSME Int J Ser A. 1997;40(3):242–6.
multilayered composite plates. Comput Model Eng Sci (CMES). 2012 [31] Dai HZ, Zhao W, Wang W, Cao ZG. An improved radial basis func-
Nov;88(5):367–417. tion network for structural reliability analysis. J Mech Sci Technol.
[9] Civalek Ö. Free vibration and buckling analyses of composite plates 2011;25(9):2151–9.
with straight-sided quadrilateral domain based on DSC approach. [32] Deng J. Structural reliability analysis for implicit performance
Finite Elem Anal Des. 2007;43(13):1013–22. function using radial basis function network. Int J Solids Struct.
[10] Abdo T, Rackwitz R. A new beta-point algorithm for large time- 2006;43(11):3255–91.
invariant and time-variant reliability problems. In: Der [33] Jing Z, Chen J, Li X. RBF-GA: An adaptive radial basis function
Kiureghian A, Thoft-Christensen P, editors. Reliability and metamodeling with genetic algorithm for structural reliability
Optimization of Structural Systems. Vol. 90. Berlin, Heidelberg: analysis. Reliab Eng Syst Saf. 2019;189:42–57.
Springer; 1991. p. 1–12 (Lecture Notes in Engineering). [34] Bichon BJ, Eldred MS, Swiler LP, Mahadevan S, McFarland JM.
[11] Rackwitz R, Flessler B. Structural reliability under combined Efficient global reliability analysis for nonlinear implicit perfor-
random load sequences. Comput Struct. 1978;9(5):489–94. mance functions. AIAA J. 2008;46(10):2459–68.
[12] Hartini E, Adrial H, Pujiarta S. Reliability analysis of primary and [35] Kaymaz I. Application of kriging method to structural reliability
purification pumps in RSG-GAS using Monte Carlo simulation problems. Struct Saf. 2005;27(2):133–51.
approach. J Teknol Reakt Nukl Tri Dasa Mega. 2019;21(1):15–22. [36] Vapnik VN. An overview of statistical learning theory. IEEE Trans
[13] Au SK, Beck JL. A new adaptive importance sampling scheme for Neural Netw. 1999;10(5):988–99.
reliability calculations. Struct Saf. 1999;21(2):135–58. [37] Vapnik V. The nature of statistical learning theory. Springer Science
[14] Engelund S, Rackwitz R. A benchmark study on importance sam- & Business Media; 1999. p. 340.
pling techniques in structural reliability. Struct Saf. [38] Onkar AK, Upadhyay CS, Yadav D. Probabilistic failure of laminated
1993;12(4):255–76. composite plates using the stochastic finite element method.
[15] Ibrahim Y. Observations on applications of importance sampling in Compos Struct. 2007;77(1):79–91.
structural reliability analysis. Struct Saf. 1991;9(4):269–81. [39] Lopes PAM, Gomes HM, Awruch AM. Reliability analysis of lami-
[16] Melchers RE. Importance sampling in structural systems. Struct nated composite structures using finite elements and neural net-
Saf. 1989;6(1):3–10. works. Compos Struct. 2010;92(7):1603–13.
[17] Tokdar ST, Kass RE. Importance sampling: a review. WIREs Comput [40] Dey S, Mukhopadhyay T, Sahu SK, Li G, Rabitz H, Adhikari S.
Stat. 2010;2(1):54–60. Thermal uncertainty quantification in frequency responses of
[18] Au SK, Ching J, Beck JL. Application of subset simulation methods to laminated composite plates. Compos Part B Eng. 2015;80:186–97.
reliability benchmark problems. Struct Saf. 2007;29(3):183–93. [41] Chen W, Jia P. Interlaminar stresses analysis and the limit state
[19] Au SK, Beck JL. Estimation of small failure probabilities in high function approximating methods for composite structure reliability
dimensions by subset simulation. Probab Eng Mech. assessment: A selected review and some perspectives. J Compos
2001;16(4):263–77. Mater. 2013;47(12):1535–47.
[20] Bjerager P. Probability integration by directional simulation. J Eng [42] Momeni Badeleh M, Fallah NA. Probabilistic reliability analysis for
Mech. 1988;114(8):1285–302. active vibration control of piezoelectric laminated composite plates
[21] Nie J, Ellingwood BR. Directional methods for structural reliability using mesh-free finite volume method. J Intell Mater Syst Struct.
analysis. Struct Saf. 2000;22(3):233–49. 2023;34(7):836–60.
[22] Zio E, Pedroni N. Functional failure analysis of a thermal–hydraulic [43] Martinez JR, Bishay PL, Tawfik ME, Sadek EA. Reliability analysis of
passive system by means of line sampling. Reliab Eng Syst Saf. smart laminated composite plates under static loads using artificial
2009;94(11):1764–81. neural networks. Heliyon. 2022;8(12):e11889.
[23] Pradlwarter HJ, Schuëller GI, Koutsourelakis PS, Charmpis DC. [44] Haeri A, Fadaee MJ. Efficient reliability analysis of laminated com-
Application of line sampling simulation method to reliability posites using advanced Kriging surrogate model. Compos Struct.
benchmark problems. Struct Saf. 2007;29(3):208–21. 2016;149:26–32.
22  Najib Zemed et al.

[45] Mathew TV, Prajith P, Ruiz RO, Atroshchenko E, Natarajan S. Adaptive probability. In: Proceedings of the 3rd International Conference on
importance sampling based neural network framework for reliability Uncertainty Quantification in Computational Sciences and
and sensitivity prediction for variable stiffness composite laminates Engineering (UNCECOMP 2019); 2019 June 24–26; Crete, Greece:
with hybrid uncertainties. Compos Struct. 2020;245:112344. Institute of Structural Analysis and Antiseismic Research School of
[46] Zhou C, Li C, Zhang H, Zhao H, Zhou C. Reliability and sensitivity Civil Engineering National Technical University of Athens (NTUA)
analysis of composite structures by an adaptive Kriging based Greece; 2019. p. 343–56.
approach. Compos Struct. 2021;278:114682. [57] Cheng K, Lu Z. Structural reliability analysis based on ensemble
[47] Morio J. Global and local sensitivity analysis methods for a physical learning of surrogate models. Struct Saf. 2020;83:101905.
system. Eur J Phys. 2011;32(6):1577. [58] Goel T, Haftka RT, Shyy W, Queipo NV. Ensemble of surrogates.
[48] Wu YT. Computational methods for efficient structural reliability Struct Multidiscip Optim. 2007;33(3):199–216.
and reliability sensitivity analysis. AIAA J. 1994;32(8):1717–23. [59] Hoeting JA, Madigan D, Raftery AE, Volinsky CT. Bayesian model
[49] Wei P, Tang C, Yang Y. Structural reliability and reliability sensitivity averaging: a tutorial with comments by M. Clyde, David Draper and
analysis of extremely rare failure events by combining sampling E. I. George, and a rejoinder by the authors. Stat Sci.
and surrogate model methods. Proc Inst Mech Eng Part O J Risk 1999;14(4):382–417.
Reliab. 2019;233(6):943–57. [60] Zerpa LE, Queipo NV, Pintos S, Salager JL. An optimization metho-
[50] Kaw AK. Mechanics of composite materials. Boca Raton: CRC Press; dology of alkaline–surfactant–polymer flooding processes using
2005. p. 491. field scale numerical simulation and multiple surrogates. J Pet Sci
[51] Beltrami E. Sulle condizioni di resistenza dei corpi elastici. Il Nuovo Eng. 2005;47(3):197–208.
Cimento. 1885;18(1):145–55. [61] Rafiq MY, Bugmann G, Easterbrook DJ. Neural network design for
[52] Wang Z, Shafieezadeh A. REAK: Reliability analysis through engineering applications. Comput Struct. 2001;79(17):1541–52.
Error rate-based Adaptive Kriging. Reliab Eng Syst Saf. [62] Karush W. Minima of functions of several variables with inequal-
2019;182:33–45. ities as side conditions. Department of Mathematics, University of
[53] Echard B, Gayton N, Lemaire M. AK-MCS: An active learning relia- Chicago; 1939.
bility method combining Kriging and Monte Carlo simulation. [63] Matheron G. The intrinsic random functions and their applications.
Struct Saf. 2011;33(2):145–54. Adv Appl Probab. 1973;5(3):439–68.
[54] Wu YT, Mohanty S. Variable screening and ranking using sampling- [64] Sacks J, Schiller SB, Welch WJ. Designs for computer experiments.
based sensitivity measures. Reliab Eng Syst Saf. 2006;91(6):634–47. Technometrics. 1989;31(1):41–7.
[55] Martinez JR, Bishay PL. On the stochastic first-ply failure analysis of [65] Reddy J, Pandey A. A first-ply failure analysis of composite lami-
laminated composite plates under in-plane tensile loading. nates. Comput Struct. 1987;25(3):371–93.
Compos Part C Open Access. 2021;4:100102. [66] Reddy Y, Reddy J. Linear and non-linear failure analysis of com-
[56] Amrane C, Mattrand C, Beaurepaire P, Bourinet JM, Gayton N. On posite laminates with transverse shear. Compos Sci Technol.
the use of ensembles of metamodels for estimation of the failure 1992;44(3):227–55.