Content uploaded by Tanmoy Mukhopadhyay
Author content
All content in this area was uploaded by Tanmoy Mukhopadhyay on Jul 23, 2020
Content may be subject to copyright.
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Vol.:(0123456789)
1 3
Archives of Computational Methods in Engineering
https://doi.org/10.1007/s11831-020-09438-w
ORIGINAL PAPER
Stochastic Oblique Impact onComposite Laminates: AConcise Review
andCharacterization oftheEssence ofHybrid Machine Learning
Algorithms
T.Mukhopadhyay1 · S.Naskar2 · S.Chakraborty3· P.K.Karsh4· R.Choudhury5· S.Dey5
Received: 21 September 2019 / Accepted: 29 April 2020
© CIMNE, Barcelona, Spain 2020
Abstract
Due to the absence of adequate control at different stages of complex manufacturing process, material and geometric proper-
ties of composite structures are often uncertain. For a secure and safe design, tracking the impact of these uncertainties on
the structural responses is of utmost significance. Composite materials, commonly adopted in various modern aerospace,
marine, automobile and civil structures, are often susceptible to low-velocity impact caused by various external agents. Here,
along with a critical review, we present machine learning based probabilistic and non-probabilistic (fuzzy) low–velocity
impact analyses of composite laminates including a detailed deterministic characterization to systematically investigate the
consequences of source- uncertainty. While probabilistic analysis can be performed only when complete statistical descrip-
tion about the input variables are available, the non-probabilistic analysis can be executed even in the presence of incom-
plete statistical input descriptions with sparse data. In this study, the stochastic effects of stacking sequence, twist angle,
oblique impact, plate thickness, velocity of impactor and density of impactor are investigated on the crucial impact response
parameters such as contact force, plate displacement, and impactor displacement. For efficient and accurate computation, a
hybrid polynomial chaos based Kriging (PC-Kriging) approach is coupled with in-house finite element codes for uncertainty
propagation in both the probabilistic and non- probabilistic analyses. The essence of this paper is a critical review on the
hybrid machine learning algorithms followed by detailed numerical investigation in the probabilistic and non-probabilistic
regimes to access the performance of such hybrid algorithms in comparison to individual algorithms from the viewpoint of
accuracy and computational efficiency.
1 Introduction
Due to the high specific strength, stiffness, rigidity, fatigue,
corrosion resistance and other outstanding mechanical
characteristics (with tunable characteristics) compared to
standard metallic structural materials, laminated composite
plates have a broad application in the spacecraft, marine,
automotive, mechanical and civil sectors. Composite struc-
tures are often susceptible to low-velocity impact caused
by various external agents, leading to a significant influ-
ence on the intended performance of the system. Therefore,
investigating the behaviour of composite structures sub-
jected to impact load is of utmost importance. On the other
hand, uncertainties in a composite material may arise due
to presence of voids in between the laminate, incomplete
knowledge about the fibre parameters, porosity, alternation
in ply thickness and various other inevitable issues involved
in the complex manufacturing process. Quite naturally, the
low-velocity impact responses are affected by the presence
T. Mukhopadhyay, S. Naskar, S. Chakraborty, P. K. Karsh, R.
Choudhury and S. Dey have contributed equally to this work.
* T. Mukhopadhyay
tanmoy@iitk.ac.in
S. Naskar
susmitanaskar@iitb.ac.in
1 Department ofAerospace Engineering, Indian Institute
ofTechnology Kanpur, Kanpur, India
2 Department ofAerospace Engineering, Indian Institute
ofTechnology Bombay, Mumbai, India
3 Department ofApplied Mechanics, Indian Institute
ofTechnology Delhi, NewDelhi, India
4 Department ofMechanical Engineering, Parul Institute
ofEngineering andTechnology, Parul University, Vadodara,
India
5 Mechanical Engineering Department, National Institute
ofTechnology Silchar, Silchar, India
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
A1
A2
A3
A4
A5
A6
A7
A8
A9
A10
A 11
A12
A13
A14
A15
A16
A17
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
of these uncertainties. A general overview of the sources
of uncertainty in the computational framework of a struc-
tural system is presented in Fig.1 [1]. One ad-hoc way to
deal with these uncertainties is to introduce the so- called
partial safety factors at the design stage. However, a more
rigorous method will demand quantification of the effect of
the material and the geometric uncertainties on the output
responses. To this end, we would pursue both probabilistic
and non-probabilistic low-velocity impact assessment of
composite laminates to cover two possible instances of get-
ting an adequate statistical report on the input parameters,
or unavailability of the same owing to restrictions on per-
forming experiments involving a large number of samples.
Researchers, over the years, have studied the behaviour of
composite structures under the action of impact load. While
Xu and Chen [2] conducted low-velocity impact analysis
of carbon- epoxy laminates for damage detection, Liu etal.
[3] studied the influence of shape of impactor (such as
conical, hemispherical and flat) on the low-velocity impact
responses of sandwich plate. In both cases, experimental
as well as numerical analyses were performed. Jagtap etal.
[4] carried out finite element (FE) simulation for damage
identification of laminated plates due to impact loading.
The effect of boundary condition and velocity of impactor
were determined. Similarly, Balasubramani etal. [5] per-
formed numerical investigation to determine the effect of
boundary conditions, the thickness of laminate, impactor’s
mass and velocity on transverse and longitudinal stress of
the composite laminate due to low-velocity impact loading.
Tan and Sun [6] and Sun and Chen [7] also used the finite
element method with Newmark time integration scheme to
investigate low-velocity impact on composite structures. A
comprehensive review on low-velocity impact loading on
composite structures can be found in [8]. Ahmed and Wei
[9] also reviewed numerical and experimental methods for
computing dynamic and static responses of composite plates
subjected low-velocity impact and quasi-static loads.
Several works dealing with failure mechanism of compos-
ite plates subjected to low-velocity impact load can be found
in the literature. While Yuan etal. [10] used an analytical
model based on the theory of first order shear deformation
for the analysis of damage and deformation of laminated
glass under low-velocity impact, Zhang and Zhang [11]
applied FE model for damage detection in composite struc-
tures due to low-velocity impact. Feng and Aymerich [12],
Maio etal. [13] and Kim etal. [14] developed and applied
Fig. 1 General overview of the sources of uncertainty in the computational framework of a structural system
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
progressive damage models to investigate the failure mecha-
nism of laminated composite due to the low-velocity impact.
Lipeng etal. [15] investigated delamination failure due to
impact load by using a self-adapting delamination element
method. Johnson etal. [16] presented different models for
failure analysis of composite plates by considering internal
damage and delamination due to impact loading. Coutellier
etal. [17] developed a model for delamination detection in
thin composite structures. Jih and Sun [18], on the other
hand, investigated experimentally the delamination in lami-
nated composite plates due to low-velocity impact.
Despite the vast literature on low-velocity impact analysis
of composite structures, none of these studies consider the
presence of uncertainties in the system. Due to the complexity
of manufacturing, accurate design specifications of composite
structures cannot be achieved in real life. As a consequence,
uncertainties in a composite structure are unavoidable. In
composite material, the main sources of uncertainties are due
to variation in material properties and inaccurate geometrical
properties. Such uncertainties are introduced in the elementary
input level (elemental mass and stiffness matrix), and propa-
gate to the global level (global mass and stiffness matrix) of
composite structures and hence, leads to a significant devia-
tion from the deterministic value of impact responses. In the
present paper, the effects such source-uncertainties on the low-
velocity oblique impact (refer to Fig.2a) response of compos-
ite plates are aimed to be addressed. The analysis is divided
into three sections namely deterministic, probabilistic and non-
probabilistic, the later two sections being dedicated to stochas-
tic analysis and uncertainty quantification (UQ). Only when
the probabilistic distributions of uncertain input parameters
are accessible can the probabilistic analyses be performed. In
many instances though, it is not possible to obtain the complete
probabilistic distributions of the input variables. In such cases,
non-probabilistic fuzzy analysis can be employed to portray
the effects of uncertainty. It is to be noted that both conven-
tional probabilistic and non-probabilistic analysis techniques
involve significant computational efforts due to the require-
ment of performing thousands of expensive finite element
simulations. One way to circumvent this issue is to develop a
machine learning model on the basis of representative origi-
nal finite element simulations. It is worthy to note here that
machine learning is a broad domain. A schematic diagram
showcasing the various aspects of machine learning techniques
and its relationship with data science is shown in Fig.3. In this
work, we are only interested in supervised learning techniques.
Popular supervised learning techniques include Gaussian pro-
cess or Kriging [19–22], Polynomial chaos expansion (PCE)
[23–25], analysis-of-variance decomposition [26–29], Polyno-
mial chaos based Kriging (PC- Kriging) [30–33] etc. In this
work, we review three machine learning techniques in the con-
text of stochastic low-velocity impact analysis. The machine
learning techniques reviewed here are polynomial chaos
expansion, Kriging and polynomial chaos based Kriging.
This paper is composed of six sections in the order of
chronological inter-dependence including the current intro-
duction section. Section2 describes governing equations for
the analysis of the transient low-velocity oblique impact of
composite plates that includes the descriptions of dynamic
equations, contact law and Newmark’s integration scheme. In
Sect.3, detailed description of the surrogate model based on
PC-Kriging is provided. Section4 provides both probabilis-
tic and non-probabilistic stochastic approaches for the impact
analysis of low-velocity. The numerical results are presented
in Sect.5 (deterministic, probabilistic and fuzzy based non-
probabilistic results including the comparative performance
of three different surrogate models i.e. PCE, Kriging and PC-
Kriging). Finally, in Sect.6, major observations and conclu-
sion are provided along with an overview of the current level
of development in relevant research fields.
2 Review oftheGoverning Equations
forLow-Velocity Impact onLaminated
Composites
A laminated composite plate is considered with length L,
width b, and thickness t subjected to normal and oblique
impact loading (as shown in Fig.2). The dynamic equation
[34] of such system can be expressed as
where
M(̃
𝜍)
,
(̃
𝜍)
,
𝛿
and
̈
𝛿
are the randomized mass matrix,
randomized stiffness matrix, displacement vector and accel-
eration vector, respectively, while {F} is externally applied
force vector. Here,
(̃
𝜍)
indicates the degree of randomization.
The force vector including the contact force
(
F
C)
in case of
impact can be expressed as
The equation of motion for the rigid impactor is given by
where
mimp(̃
𝜍)
is the mass of impactor while
̈
𝛿imp
is the accel-
eration of impactor.
2.1 Contact Law
Modified Hertzian contact law can be utilized to calculate
the contact force between impactor and the composite plate
[35]. The impactor is assumed as a spherical elastic solid
body.
The contact force can be obtained during loading as
(1)
[M(̃
𝜍)]{
̈
𝛿}+[K(̃
𝜍)]{𝛿}={F(̃
𝜍)}
(2)
F
(̃
𝜍)=
{
000F
C
(̃
𝜍)000
}
(3)
mimp
(̃
𝜍)
̈
𝛿
imp
+F
c
(̃
𝜍)=
0
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
Fig. 2 a Laminated composite plate subjected to normal and oblique impact load by a spherical mass. b A typical example of twisted plate. c
Geometric details of twist in the plate
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
where
𝛾
denotes the local indentation and k is the modified
contact stiffness [36] which can be expressed by contact
theory as
where
Ei
is the elastic modulus of the impactor,
Eyy
is the
elastic modulus of laminated composite plate of the upper-
most ply in the transverse direction, while
Rimp
and
𝜐
are the
radius and Poisson’s ratio of impactor, respectively. At the
time of loading and unloading the contact force
(
F
C)
can be
estimated as
where
Fm
and
𝛾m
are the maximum contact force and maxi-
mum indentation, respectively. The permanent indentation
(𝛾0)
in loading and unloading cylce is given by
where
𝛽c
is the constant, while
𝛾Cr
is the critical identation.
For the oblique impact, the local indentaion is given by
where
𝛾imp
and
𝛾plt
are impactor’s displacement and targeted
plate displacements, respectively, while
𝛽and 𝜓
are the
oblique impact angle and twist angle, respectively, along the
(4)
Fc
(̃
𝜍)=k(̃
𝜍)𝛾(̃
𝜍)
1.5
0
≤
𝛾
≤
𝛾
m
(5)
k
(̃
𝜍)=
4
3
√
Rimp
1
[1−𝜐2
i(̃
𝜍)]
E
i
(̃
𝜍)+1
E
yy
(̃
𝜍)
(6)
F
c(̃
𝜍)=Fm
[
𝛾(̃
𝜍)−𝛾0
𝛾
m
−𝛾
0]5∕2
and Fc(̃
𝜍)=Fm
[
𝛾(̃
𝜍)−𝛾0
𝛾
m
−𝛾
0]3∕2
(7)
𝛾
0
=0when 𝛾
m
<𝛾
Cr
𝛾0
=𝛽
c(
𝛾
m
−𝛾
Cr)
when 𝛾
m≥
𝛾
Cr
(8)
𝛾(t)(
̃
𝜍)=𝛾imp(t)(
̃
𝜍)cos 𝛽+𝛾plt(xc,yc,t)(
̃
𝜍)cos 𝜓
global z-direction, respectively. The contact force elements
at the global direction of contact point can be described as
2.2 Newmark’s Time Integration Scheme
The contact force involved in the equilibrium Eqs.(1) and
(3) is generally transient in nature for the dynamic response
of a laminated composite plate under the impact by a spheri-
cal impactor. The time integration scheme of Newmark [37]
is used to solve the equations that depend on time. Use of
above scheme with time interval
t
gives the subsequent
relations at the time
t+Δt
where
[̄
K]
and
[̄
k]
are the effective stiffness matrix of the
plate and impactor, respectively, and given by
Effective contact forces at time
t+Δt
can be derived as
(9)
Fix =0, Fiy =Fc(̃
𝜍)sin 𝜓,Fiz =Fc(̃
𝜍)cos 𝜓.
(10)
[̄
K]𝛿(t+t)={
̄
F}(t+t)
(11)
k
imp𝛿
(t+Δt)
imp
={FC}(t+Δt
)
(12)
[̄
K]=K+a0M
(13)
[̄
k]=a
0
m
imp
(14)
{
F}
(t+Δt)
={F}
(t+𝛿t)
+[M]
(
a
0
𝛿
(t)
+a
1̇
𝛿
(t)
+a
2⃛
𝛿
(t))
(15)
̄
F
(t+Δt)
C
e
=F(t+Δt)
C+mimp
(
a0𝛿(t)
imp +a1̇
𝛿(t)
imp +a2̈
𝛿(t)
imp
)
Fig. 3 Different facets of
machine learning techniques
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
2 11
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
The acceleration and velocity can be derived from displace-
ment at time
t+Δt
as
The initial boundary condition considered as
where
V0
is the initial velocity of the impactor. The time
integration constants can be expressed as
For the present study, the value of
′′
and
𝛽′′
are considered
as 0.5 and 0.25, respectively.
3 Hybrid Machine Learning Based
onKriging andPCE
Let, x=
{
x
1
,…,x
N}
∈ℝN to be the input variables and
y
ℝ
O
to be the output responses. We also assume
M(⋅)
to be the computational model (FE model in present case)
such that
For impact analysis, the model
M(
⋅
)
is computationally
expensive to evaluate and hence, the task of quantifying the
uncertainties in the output response
y
becomes difficult. One
way to deal with this issue is to replace the computationally
expensive finite element model
M(
⋅
)
with a surrogate
⌢
M(
⋅
)
.
It can be noted here that we have used the words surrogate
modelling and machine learning in identical sense keeping
in mind its purpose in the context of this article. We would
review three methods of machine learning in this section that
can be used as a surrogate of the original simulation model.
3.1 Polynomial Chaos Expansion
Polynomial chaos expansion (PCE) is one of the most pop-
ular methods available in literature. This was first imple-
mented by Wiener [38] and hence, is also known as ‘Wiener
Chaos expansion’. Xiu and Karniadakis [23] subsequently
generalized the technique and proved its effectiveness for
different continuous and discrete systems from the so called
(16)
{̈
𝛿}
(t+Δt)
=a0({𝛿}
(t+Δt)
−{𝛿}
(t)
)−a1{
̇
𝛿}
(t)
−a2{
̈
𝛿}
(t)
̈
𝛿
(t+Δt)
imp =a0(𝛿(t+Δt)
imp −𝛿(t)
imp)−a1̇
𝛿(t)
imp −a2̈
𝛿(t)
imp
{
̇
𝛿}(t+Δt)={̇
𝛿}(t)+a3{̈
𝛿}(t)+a4{̈
𝛿}(t+Δt)
̇
𝛿
(t+Δt)
imp
=̇
𝛿(t)
imp
+a3̈
𝛿(t)
imp
+a4̈
𝛿(t+Δt)
imp
(17)
𝛿
=
̇
𝛿=
̈
𝛿=0, 𝛿
imp
=
̈
𝛿
imp
=0and
̇
𝛿
imp
=V
0
(18)
a
0=
1
𝛽��𝛿t2,a1=
1
𝛽��Δt
,a2=
1
2𝛽�� −
1,
a3
=
(
1−��
)
Δtand a
4
=��Δt
(19)
y=M(x)
Askey-scheme,
L2
convergence in the corresponding Hilbert
space.
Assuming
𝐢
=
(
i
1
,i
2
,…,i
N)
∈ℕ
N
0
to be a multi-index
with
|i|=i1+i2+
⋯
+iN,
and let n ≥ 0 be an integer. The
nth order PCE of g(X) is given as:
where {ai} are unknown coefficients that must be deter-
mined. Φi(X) are N-dimensional orthogonal polynomials
with maximum order of and satisfies
Here, δij denotes the multivariate kronecker delta function.
It is to be noted that the orthogonal polynomials are depend-
ent on the PDF ϖ(x) of input variables. Table1 presents the
orthogonal polynomial type and the random variable type
correspondence [23].
Over last two decades, researchers have developed and
utilized different variants of PCE. Xiu and Karniadakis
[23] proposed the Wiener–Askey PCE where the unknown
coefficients associated with the coefficients were deter-
mined by using either collocation method or the Galerkin
projection. With this method, it is possible to solve sto-
chastic partial differential equations in an efficient way.
However, Wiener–Askey PCE is intrusive in nature and
hence, knowledge about the governing partial differential
equation of the system is required. As a consequence, this
method is not applicable to cases where the user only have
some data and no knowledge about the process from which
the data is generated.
To tackle the above-mentioned problem, researchers
focused on developing nonintrusive (data-driven) PCE.
The easiest and most popular way to train a data-driven
PCE is by minimizing the least square error of the system
(20)
̂
g
(X)=
n
∑
|i|=0
ai𝛷i(X
)
(21)
E
(𝛷i(X)𝛷j(X)) =
�
𝛺
𝛷i(X)𝛷j(X)𝜛(x)=𝛿ij,0
≤|
i
|
,
|
j|
≤N
Table 1 The Correspondence of the type of orthogonal polynomial
with distribution pattern
Type Random variables Type of orthogo-
nal polynomial
Support
Continuous Gaussian Hermite (− ∞,∞)
Gamma Laguerre [0,∞)
Beta Jacobi [a, b]
Uniform Legendre [a, b]
Discrete Poisson Charlier
{0, 1, …}
Binomial Krawtchouk
{0, 1, …,N}
Negative binomial Meixner
{0, 1, …}
Hypergeometric Hahn
{0, 1, …,N}
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
3 11
312
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
[39, 40]. However, this method is susceptible to overfitting
and as a result often performs poorly. Methods for train-
ing a PCE model by using the quadrature rule can also be
found in the literature [41, 42]. However, both these train-
ing algorithms suffer from the curse of dimensionality and
hence, are only applicable to small-scale problems with
limited number of input variables.
To address the curse of dimensionality associated with
least-square and quadrature based training algorithms,
Blatman and Sudret [24, 43] two adaptive sparse PCEs
which can used for solving problems having hundreds of
input variables. Both the methods proposed follow similar
flow where an iterative algorithm is used to determine the
importance of the terms involved in PCE and the lesser
important terms are removed. In the first method, the
important terms in PCE are determined by tracking the
change in coefficient of determination 2 (due to addition/
removal of a term). In the second approach, a more rigor-
ous framework, referred to as the least-angle regression
is used to determine the important terms of PCE. With
both these approaches, there is a significant reduction in
the number of unknown coefficients associated with PCE
and thereby, issues with hundreds of input variables can
be solved.
Jacquelin etal. [44] identified that for lightly damped sys-
tems, the convergence of PCE is very poor. It was proposed
that integrating Aitken’s transformation into the framework
of PCE can improve its convergence significantly. Pascual
and Adhikari [45] hybridized the basic formulation of PCE
by coupling it with perturbation method. Four variants of the
hybrid perturbation-PCE was proposed and reduced spectral
method was used to identify unknown coefficients associ-
ated with the bases. The proposed approaches were utilized
to solve the stochastic eigenvalue problem. It was observed
that the approaches proposed lead to a better approximation
of larger eigenvalues.
3.2 Kriging
In today’s time, one of the most popular machine learning
technique is perhaps the Gaussian process, a.k.a. Krig-
ing is a Bayesian machine learning technique where we
assume that the response y, conditioned on input x is a
sample from a Gaussian process.
where
𝜇(
⋅
;𝐁)
is the mean function and
R(
⋅
,
⋅
;𝜽)
is the cor-
relation kernel.
𝐁,𝜎
and
𝜽
are the hyperparameters of the
Gaussian process respectively, denotes the unknown coef-
ficients related to the mean function, the process variance
and the length-scale parameter associated with the correla-
tion kernel. In order to use Gaussian process as a machine
(22)
y|
x;𝐁,𝜎,𝜽∼GP
(
𝜇(x;𝐁),𝜎
2
R
(
x
1
,x
1
;𝜽
))
learning technique, the hyperparameters needs to be esti-
mated based on some training data. This can either be
achieved by maximizing the likelihood [21] or by using the
Bayes rule [46–49].
The most popular form of Gaussian process is the zero
mean Gaussian process or the simple Kriging. In this vari-
ant, we assume
𝜇(
⋅
;𝐁)=0
. As a consequence, only
𝜎
and
𝜽
are the only hyperparameters associated with the system.
An improvement to the simple Kriging is the ordinary
Kriging where we assume the mean function is assumed
to be constant,
𝜇(
⋅
;𝐁)=a0
where
a0
is a constant. Unfor-
tunately, the fact that the mean function is modelled as a
constant often results in erroneous models.
To enhance the Kriging model’s precision, universal
Kriging was developed [50, 51]. In universal Kriging, the
mean function represented as a linear regression model by
using multivariate polynomials
where
bi(x)
represents the ith basis function and ai denotes
the coefficient associated with the ith basis function. With
this setup, the mean function captures the largest variance
in the data and the correlation function interpolates the
residual. Considering,
x
=
{
x
1
,x
2
,…,x
n}
to be input sam-
ples and
g={g1,g2,…,gn}
to be the responses, the design
matrix and the correlation matrix can be represented. The
regression portion can be written as a n × p model matrix F,
whereas the stochastic process is defined using a n × n cor-
relation matrix Ψ
where ψ(∙,∙) is a correlation function, parameterised by a
set of hyperparameters θ. As already stated, the hyperpa-
rameters are identified either by using maximum likelihood
estimation (MLE) or by using the Bayes rule.
Similar to PCE discussed in previous section, univer-
sal Kriging also suffers from the curse of dimensional-
ity. To address this issue, blind Kriging was proposed in
[51–54]. In blind Kriging, the polynomial order used to
represent the mean function of the Gaussian process is
selected in an adaptive manner. Bayes rule is used to train-
ing the blind Kriging model. It is worthwhile to mention
that blind Kriging satisfies both the hierarchy criterion
and the heredity criterion. As per the hierarchy criterion,
(23)
𝜇
(⋅,𝐁)=
P
∑
i=1
aibi(x
)
(24)
F
=
b1
x
1
…bp
x
1
⋮⋱⋮
b
1
(xn)⋯b
p
(xn)
(25)
𝜓
=
𝜓
x1,x1
…𝜓
x1,xn
⋮⋱⋮
𝜓
xn,x1
⋯𝜓(xn,xn)
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
lower order effects in the mean function are selected before
the higher order effects. Whereas, as per the heredity cri-
terion, an effect can only be important if its parent effects
are already important. Other variants of Kriging includes
Co-Kriging [55] and stochastic Kriging [56–58]. A com-
parative assessment from the viewpoint of accuracy and
computational efficiency can be found in ref [19], where
both high and low dimensional input parameter space was
considered for a comprehensive analysis.
The choice of suitable correlation function is a crucial
element for all the Kriging variants [59–61]. Correlation
function that are commonly used with Gaussian process
are mostly stationary and hence,
With such as correlation function, it is possible to represent
multivariate functions as product of one-dimensional corre-
lations. Popular stationary correlation functions includes: (a)
exponential correlation function (b) generalised exponential
correlation function (c) Gaussian correlation function (d)
linear correlation function (e) spherical correlation function
(f) cubic correlation function and (g) spline correlation func-
tion. The mathematical forms of all the correlation functions
are provided below:
1. Exponential correlation function:
2. Generalised exponential correlation function:
3. Gaussian correlation function:
4. Linear correlation function:
5. Spherical correlation function:
6. Cubic correlation function:
7. Spline correlation function:
(26)
𝜓
(x,x
�
)=
∏
j
𝜓j(𝜃,xi−x
�
i
)
(27)
𝜓j(𝜃;dj)=exp(−𝜃j|dj|)
(28)
𝜓j
(𝜃;d
j
)=exp(−𝜃
j|
d
j|𝜃
n+1),0<𝜃
n+1≤2
(29)
𝜓
j(𝜃;dj)=exp(−𝜃jd
2
j)
(30)
𝜓j(𝜃;dj)=max{0, 1 −𝜃j|dj|}
(31)
𝜓
j(𝜃;dj)=1−1.5𝜉j+0.5𝜉
2
j
,𝜉j=min{0, 𝜃j
|
dj
|}
(32)
𝜓
j(𝜃;dj)=1−3𝜉
2
j
+2𝜉
3
j
,𝜉j=min{1, 𝜃j
|
dj
|}
where
𝜉j=𝜃j|dj|
For all the correlation functions described above,
dj
=x
i
−x
�
i
. The hyperparameters associated with the
covariance functions are determined either by using the
maximum likelihood estimate (MLE) or by using the
Bayes rule. A detailed account of MLE in the context of
Kriging is given in [21].
3.3 Polynomial Chaos Based Kriging (PC-Kriging)
Finally, we discuss about a hybrid machine learning tech-
nique, referred to as the polynomial chaos based Kriging
(PC-Kriging) [30–32]. PC-Kriging is a novel surrogate
model that combine two well-known surrogates, namely,
polynomial chaos expansion (PCE) [23, 25] and Kriging [19,
20]. PC-Kriging can be viewed as a Kriging model where
the mean/trend function is modelled by using PCE. With
this setup, it is possible to achieve a higher order accuracy
as compared to PCE and Kriging.
PC-Kriging is a special kind of Kriging where the mean
function of the Gaussian process is modelled by using poly-
nomial chaos expansion. More specifically,
𝝁(⋅)
in Eq.(23)
is represented by using Eq.19. Under limiting condition,
PC-Kriging converges either to PCE or to Kriging. Similar
to Kriging, the hyperparameters in PC-Kriging are learned
by maximizing the likelihood. For further details, interested
readers may refer [19, 62].
Despite PC-Kriging’s benefit over its individual PCE
and Kriging, the hybrid metamodel suffers from the curse
of dimensionality due to the factorial growth of unknown
coefficients with a rise in the number of input parameters
N. This limitation originates from the PCE component
of PC-Kriging. To address this problem, a variant of PC-
Kriging, referred to as Optimal PC- Kriging (OPC-Kriging)
[31] was proposed. In OPC-Kriging, least angle regression
(LAR) is used to only retain the important components of
PCE. The OPC-Kriging follows an iterative algorithm where
each polynomial can be added to the trend part one-by-one.
Figure4 presents a flowchart depicting the algorithm of
OPC-Kriging.
(33)
𝜓
j(𝜃;dj)=
1−5𝜉2
j+30𝜉3
j,0
≤
𝜉j
≤0.2
1.251−𝜉3
j, 0.2 ≤𝜉j≤
1
0, 𝜉j>1
409
410
4 11
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
4 Machine Learning Based Stochastic
Impact Analysis
A major objective of the present study is to determine sto-
chastic response of low-velocity impact loading on com-
posite plates following probabilistic and non-probabilistic
frameworks. Both geometric and material uncertainties are
considered in this work. To be specific, uncertainties in the
composite plate stem from variation in the material prop-
erties, fibre orientation angle, twist angle, oblique impact
angle, initial velocity of impactor, mass density of impac-
tor, and thickness of target plates, inclusion of which in
the analysis (following probabilistic and non-probabilistic
approaches) is discussed here.
4.1 Probabilistic Impact Analysis
For probabilistic impact analysis, statistical descriptions of
the stochastic inputs are necessary. To that end, the machine
learning techniques discussed in previous section have been
coupled with our in-house FE code for low-velocity impact
analysis. For quantifying the uncertainty in the output
responses, first input training samples are obtained using
an appropriate design of experiment (sampling) scheme.
Due to its simplicity and already proven superior perfor-
mance, Sobol sequence [63, 64] has been used in this study.
In the next step, the training outputs are obtained by using
the actual FE solver. In the third step, the machine learning
models are trained and the hyperparameters associated with
the models are computed. Finally, Monte Carlo simulation
is carried out based on the trained ML model to compute
the probability density function of the output responses. A
flowchart depicting the ML based probabilistic uncertainty
quantification algorithm is presented in Fig.5. For the cur-
rent study, the following cases of uncertainties are consid-
ered at each lamina level (layer-wise uncertainty modelling)
1. Variation of fibre-orientation angle:
2. Variation of twist angle:
3. Variation of oblique impact angle:
4. Variation of initial velocity of impactor:
5. Variation of mass density of impactor:
6. Variation of thickness of the plate:
7. Variation in location of loading point:
𝜓1{
𝜃
,
E
,
G
,
𝜐
,
𝜌
}=
𝛩
[{
𝜃
(̃
𝜍
)},{
E
(̃
𝜍
)},{
G
(̃
𝜍
)},{
𝜐
(̃
𝜍
)},{
𝜌
(̃
𝜍
)}]
𝜓2{𝜓,𝜃,E,G,𝜐,𝜌}=𝛩[𝜓,{𝜃(̃
𝜍)},{E(̃
𝜍)},{G(̃
𝜍)},{𝜐(̃
𝜍)},{𝜌(̃
𝜍)}]
𝜓3{,𝜃,E,G,𝜐,𝜌}=𝛩[,{𝜃(̃
𝜍)},{E(̃
𝜍)},{G(̃
𝜍)},{𝜐(̃
𝜍)},{𝜌(̃
𝜍)}]
𝜓4{V,𝜃,E,G,𝜐,𝜌}=𝛩[V,{𝜃(̃
𝜍)},{E(̃
𝜍)},{G(̃
𝜍)},{𝜐(̃
𝜍)},{𝜌(̃
𝜍)}]
𝜓5{𝜌imp,𝜃,E,G,𝜐,𝜌}=𝛩[𝜌imp ,{𝜃(̃
𝜍)},{E(̃
𝜍)},{G(̃
𝜍)},{𝜐(̃
𝜍)},{𝜌(̃
𝜍)}]
𝜓
6
{tplt,𝜃,E,G,𝜐,𝜌}=𝛩[tplt ,{𝜃(̃
𝜍)},{E(̃
𝜍)},{G(̃
𝜍)},{𝜐(̃
𝜍)},{𝜌(̃
𝜍)}]
Fig. 4 Flowchart for OPC-
Kriging
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
5 11
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
Here
̃
𝜁
is used to denote the stochastic representation
of the system parameters. The parameters
E,G,𝜐,𝜌
are
the set of Young’s moduli, shear moduli, mass den-
sity and Poisson’s ratio in different directions, where
the entire set of stochastic material properties is
{
E
1,
E
2,
E
3,
G
12,
G
13,
G
23,
𝜐
12,
𝜐
13,
𝜐
23,
𝜐
32,
𝜐
21,
𝜐
31,
𝜌
}
. Unless
otherwise mentioned, the degree of stochasticity from the
respective deterministic values is taken as
±10%
(as per
standard design practice) for each of the components in
the set of material properties.
𝜓7{Lp,𝜃,E,G,𝜐,𝜌}=𝛩[Lp,{𝜃(̃
𝜍)},{E(̃
𝜍)},{G(̃
𝜍)},{𝜐(̃
𝜍)},{𝜌(̃
𝜍)}]
4.2 Fuzzy Impact Analysis
Although probabilistic analysis is more rigorous as it pro-
vides the probability distribution of the output responses,
it is limited by the fact that we require probability distribu-
tion of the input variables for carrying out such analysis.
In the real-life scenario, we may not have knowledge about
the probability distribution of the input variables due to
the requirement of extensive experimental characterization
of the materials involving thousands of physical samples.
Under such circumstances of sparse data availability, we
have to opt for non- probabilistic analysis. Out of differ-
ent non- probabilistic analysis methods available in litera-
ture, fuzzy based non-probabilistic analysis is employed
Fig. 5 Flowchart for probabilistic impact analysis based on hybrid machine learning models coupled with FE simulations
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
for uncertainty quantification and propagation in low-
velocity impact analysis of the laminated composite plate.
The fuzzy theory is employed in the intermediate stage
between non-members and members known as member-
ship function
[
𝜇
pi]
that signifies the degree to which each
component in the territory leads to the fuzzy set [65]. The
triangular membership function is employed for the fuzzy
number
[Pi(̃
𝜍𝛼)]
and expressed as
where
PM
i
,
PU
i
,
PL
i
denotes the mean value, upper bound and
lower bound, respectively. Here
̃
𝜍𝛼
indicates the fuzzified
variations corresponding to each α-cut, where α is known
as the degree of fuzziness or membership grade ranging
from 0 to 1. As an example, the Gaussian distribution can
be approximated by using the triangle as shown in Fig.6a,
where the area under the Gaussian distribution is equal to
the area under the triangular function [66]. The triangular
fuzzy membership function is written as
where
𝜆=(2
𝜋𝜎
x)1∕2
, Xi and
𝜎x
represents the mean and
standard deviation (S.D.) of the Gaussian distribution. In
the present study, triangular membership function [µP(i)] is
employed as
By applying the α-cut method, the fuzzy input number Pi
can be grouped into the set
̄
Pi
of (n + 1) intervals Pi
(j)
where n is the number of α-cut levels. The interval of j-th
level of i-th fuzzy number can be expressed
where Pi
(j,U) and Pi
(j,L) represent the upper and lower bound
of the interval at the j-th level, respectively. At j = n,
P(n,U)
i
=P
(n,L)
i
=P
(n,M)
i
. The superscript U represents the upper
bound, while L denotes lower bound. The fuzzy input num-
bers are considered as the uncertain model parameters for
the uncertainty analysis and an interval analysis is carried
out at different α-levels [67].
Even though in the present study we have considered
triangular membership functions for the input parameters,
(34)
Pi
(̃
𝜍
𝛼
)=[P
U
i
,P
M
i
,P
L
i]
(35)
𝜇
P(i)=max
⎡
⎢
⎢
⎣
0, 1 −
�
�
�
X(j)
i−Xi
�
�
�
𝜆
⎤
⎥
⎥
⎦
(36)
𝜇
P(i)=1−
(
P
M
i−Pi
)/(
P
M
i−P
L
i
)
,for P
L
i
≤
Pi
≤
P
M
i
𝜇
P(i)=1−(Pi−PM
i)/(PU
i−PM
i),for PM
i≤Pi≤P
U
i
𝜇P(i)
=0, otherwise
(37)
̄
Pi
(̃
𝜍
𝛼
)=[P
(0)
i
,P
(1)
i
,P
(2)
i
,P
(3)
i
,…,P
(j)
i
,…,P
(n)
i]
(38)
P
(j)
i=
[
P(j,L)
i,P(j,U)
i
]
the input membership functions can be augmented further
depending on the availability of limited number of input
dataset. In this work, we start by evaluating the deterministic
solution at
=1
level first and continue towards the lower
−
cut levels using an interval analysis. As a special case,
if the input–output relation of the problem in hand is mono-
tonic in nature, computing the bounds of the fuzzy outputs
becomes trivial. Unfortunately, for most real-life problems,
the input–output relation is not monotonic in nature. Under
such circumstances, a maximization and minimization algo-
rithm involving multiple simulations is necessary. In this
work, we proceed by first formulating the machine learn-
ing models as a surrogate to the actual FE code. Then we
perform MCS on the trained machine learning models to
compute the maximum and minimum values of the response
quantities of interest for a particular α-cut level. It is to be
noted that only a single machine learning model is required
in this case corresponding to
=0
as the same model can
be reused for other α-cut levels. The number of actual FE
simulations required in this study is therefore equal to the
number of training samples needed to train the models of
machine learning. The procedure of the present fuzzy impact
approach is summarized in Figs.6b and 7.
5 Numerical Investigation andDiscussion
In this work a glass–epoxy laminated composite plate
having dimensions
L=1m, b=1m
and
t=0.002 m
is considered. Unless otherwise mentioned, the plate is
considered to be subjected to normal and oblique impact
loadings at the centre of the plate. The deterministic
material properties of glass–epoxy are
E1
=38.6 ×10
9
Pa,
E2=8.27 ×109
Pa,
G12 =G13 =4.144 ×109
Pa,
G23
=1.657 ×10
9
Pa,
𝜌=2600 kg/m3
,
𝜐=0.26
[68]. The
diameter of spherical steel ball (impactor) is considered
as
0.0127 m
. It is assumed that the fibre orientation angle
may have a variation of 5% and the material properties
may have a variation of
10%
with respect to the determin-
istic values. Such variations are considered as per stand-
ard industrial practices; however, the current analysis can
be extended to other percentages of variation, if required.
Contact force (CF), impactor displacement (ID) and
plate displacement (PD) are considered to be the output
response variables. The in-house deterministic finite ele-
ment code for impact analysis is validated with results of
Sun and Chen [7] (refer to Fig.8), wherein it is observed
that the current results are extremely close to the results
of literature.
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
6 11
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
5.1 Deterministic Impact Analysis
Deterministic numerical results of the low-velocity impact
are discussed in this subsection (Tables2, 3, 4, 5, 6, 7, 8)
to study the basic and fundamental influence of different
system parameters such as fibre-orientation angle, oblique
impact angle, twist angle, initial velocity of impactor, mass
density of impactor, thickness of plate and location of
impact loading. Here we study four different crucial stacking
sequences of the composite laminate: bending stiff laminate
([0°/0°/30°/− 30°]s), cross ply laminate ([90°/0°/90°/0°]s),
torsion stiff laminate). The effects of stacking sequence on
Fig. 6 a Triangular membership function approximated from Gaussian distribution. b Fuzzy analysis for different value of α-cuts
659
660
661
662
663
664
665
666
667
668
669
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
low-velocity impact responses are furnished in Table2. It is
observed that the peak CF is highest for the torsion stiff lam-
inates. On the other hand, peak ID and peak PD are found
to be minimum for torsion stiff laminates and maximum for
bending stiff laminates. Table3 shows the variation of peak
impact responses with the change in twist angle. The peak
CF is found to increase with increase in twist angle. On the
contrary, peak ID and peak PD decrease with the increase
in twist angle. The influence of oblique impact angle on the
responses is shown in Table4. While peak CF and peak PD
Fig. 7 Flowchart for non-probabilistic impact analysis based on fuzzy approach (Machine learning models are used instead of direct FE model,
as indicated using a blue colour box)
670
671
672
673
674
675
676
677
678
679
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
decrease with increase in the impact angle, peak ID is found
to follow a reverse trend. All the peak responses are found
to increase with increase in the initial velocity, as shown
in Table5. Effect of mass density on the impact response
is shown in Table6. In this case, increase in mass density
raises the peak responses. The effect of plate thickness, as
presented in Table7, reveals that peak CF increases with the
increase in plate thickness, while peak PD and peak ID show
an opposite trend. The effect of impact point on the critical
impact responses is shown in Table8, where it is found that
peak CF, PD and ID are maximum at point 2, point 3 and
point 3, respectively.
Fig. 8 Time histories of a contact force and b deflection of glass
epoxy composite plates considering a centrally impacted bending
stiff laminated composite plate
(±0
◦
∕±30
◦
)
with dimension L = 1m,
b = 1 m, and t = 0.002 m, ѱ = 0°, β = 0°, initial velocity of impac-
tor = 5m/s, diameter of spherical steel ball = 0.0127 m, mass density
of impactor
(𝜌)=0.0085
N−s
cm
4 [7]
Table 2 Effect of stacking sequence (quasi-isotropic stiff, tor-
sion stiff, cross ply and bending stiff laminates on low-velocity
impact responses considering t = 0.002 m, ѱ = 0°, β = 0°, V = 5 m/s,
ρ = 0.0085
N−s
cm
4
Stacking sequence Impact responses (maximum value)
CF (N) ID (m) PD (m)
Bending stiff 744.7855 0.000225 0.090134
Quasi-isotropic stiff 770.0546 0.000221 0.0854
Cross ply 770.45 0.000219 0.08794
Torsion stiff 773.31 0.000219 0.08548
Table 3 Effect of twist angle
(𝜓)
on low-velocity impact responses
with considering t = 0.002m, β = 0°, V = 5 m/s, ρ = 0.0085
N−s
cm
4 , bend-
ing stiff laminate (02°/± 30°)s
Twist angle Impact responses (maximum value)
CF (N) ID (m) PD (m)
𝜓=0◦
744.7855 0.000227 0.090134
𝜓=15◦
776.3958 0.000221 0.0882
𝜓=30◦
874.1484 0.000206 0.0833
𝜓=45◦
1053.9 0.000188 0.07413
Table 4 Effect of oblique impact angle
(𝛽)
on low-velocity impact
responses with considering t = 0.002 m, ѱ = 0°, V = 5m/s, ρ = 0.0085
N−s
cm4
, bending stiff laminate (02°/± 30°) s
Oblique impact
angle
Impact responses (maximum value)
CF (N) ID (m) PD (m)
𝛽=0◦
744.7855 0.000225 0.090134
𝛽=15◦
724.1631 0.000232 0.08965
𝛽=30◦
661.4398 0.000251 0.087791
𝛽=45◦
553.4121 0.00029 0.084967
Table 5 Effect of initial velocity of impactor on low-velocity impact
responses with considering t = 0.002m, ѱ = 0°, β = 0°, ρ = 0.0085
N−s
cm
4 ,
bending stiff laminate (02°/± 30°)s
Initial velocity of
impactor (m/s)
Impact responses (maximum value)
CF (N) ID (m) PD (m)
V = 5 744.7855 0.000227 0.090134
V = 10 1549.402 0.00042 0.177863
V = 15 2365.073 0.000606 0.263738
V = 20 3193.182 0.000789 0.349809
Table 6 Effect of mass density of impactor (ρ in
N−s
cm
4 ) on low-veloc-
ity impact responses with considering t = 0.002 m, ѱ = 0°, β = 0°,
V = 5m/s, bending stiff laminate (02°/± 30°)s
Mass density of
impactor
Impact responses (maximum value)
CF (N) ID (m) PD (m)
𝜌
= 75 × 10−4 719.9314 0.00021 0.08149
𝜌
= 80 × 10−4 733.6016 0.000219 0.085852
𝜌
= 85 × 10−4 744.7855 0.000227 0.090134
𝜌
= 90 × 10−4 755.3778 0.000235 0.094109
𝜌
= 95 × 10−4 766.9816 0.000242 0.098161
680
681
682
683
684
685
686
687
688
689
690
691
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
5.2 Stochastic Impact Analysis
In this section, results corresponding to the probabilistic
and non-probabilistic impact analysis are presented. The
formation of surrogate models based on PCE-Kriging is
discussed first including comparative assessment of other
related surrogates. After validating the accuracy of the sur-
rogate models, detailed stochastic analyses are carried out
in the subsequent subsections.
5.2.1 Surrogate Modelling andValidation
In this section, first we discuss about training the machine
learning models. To be specific, convergence studies to
determine the optimal number of training samples are pre-
sented. Second, we perform a comparative assessment of
PCE, Kriging and PC-Kriging.
5.2.1.1 Design ofExperiments One important task in sur-
rogate modelling is to generate suitable training samples
for training the surrogate model. As already stated in the
preceding section, Sobol sequence is adopted in this study
to generate samples for training ML model. However, the
optimal number of training samples required still needs to
be determined. To that end, a study by varying the number
of training samples has been carried out. Figure9 shows the
PDF of responses (for direct MCS and PCE-Kriging based
MCS) with respect to training sample size of 32, 64, 128,
256, 512 and 1024. For all the three output responses, the
results obtained using 512 training samples are almost iden-
tical to those obtained using 1024 samples. Based on this
observation, we conclude that 512 is the optimal number
of training samples. Note that all the subsequent results are
obtained by training the surrogate with 512 training sam-
ples.
5.2.1.2 PCE Versus Kriging Versus PC-Kriging: A Compara-
tive Study The surrogate PC-Kriging is developed by
combine PCE and Kriging. In this section, we examine the
performance of the three surrogate models (PCE, Kriging
and PC-Kriging) in the context of probabilistic low-velocity
impact analysis. To that end, coefficient of determination
(
R2
)
and root mean square error (RMSE) have been com-
puted corresponding to training sample size of 32, 64, 128,
256, 512 and 1024. Figure10 shows the
R2
and RMSE cor-
responding to the different training sample size and the three
surrogate models.
It is observed that PC-Kriging consistently outperforms
PCE and Kriging; although the results obtained using PCE
are found to be extremely close to the PC-Kriging results.
Moreover, similar to the observations in previous section,
the results obtained corresponding to sample size of 512
and 1024 are almost identical (with
R2
close to 1), indi-
cating that the surrogate models converge at 512 training
samples. Figure11 shows the probability density functions
obtained using PCE, Kriging and PC-Kriging, wherein the
results are compared with benchmark Monte Carlo simula-
tion results. For all the three cases, PC-Kriging is found to
yield best results followed by PCE, establishing the superior-
ity of PC-Kriging over PCE and Kriging. All the subsequent
results in this paper are obtained using PC-Kriging trained
with 512 training samples. It can be noted in this context
that stochastic analysis of composite structures leading to
the uncertainty quantification of different global responses
have recently received significant attention from the scien-
tific community [69–79]. However, most of these studies
consider a single machine learning algorithm to map the
stochastic input–output domain. The current investigation
is the first attempt to investigate the performance of hybrid
machine learning algorithms for any structural response of
composite structures.
Table 7 Effect of thickness of plate (t) on low-velocity impact
responses with considering ѱ = 0°, β = 0°, V = 5 m/s, ρ = 0.0085
N−s
cm
4 ,
bending stiff laminate (02°/± 30°) s
Thickness of plate
(m)
Impact responses (maximum value)
CF (N) ID (m) PD (m)
t = 0.002 322.9597 0.000548 0.266644
t = 0.004 744.7855 0.000225 0.090134
t = 0.006 1054.777 0.000188 0.050085
t = 0.008 1248.632 0.000176 0.033335
Table 8 Effect of location of impactor contacting point on low-
velocity impact responses with dimension t = 0.002m, ѱ = 0°, β = 0°,
V = 5 m/s, ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°)s (loca-
tion of impact points on the laminated composite plate is indicated in
the inset of Fig.19a)
Location of impactor Impact responses (maximum value)
CF (N) ID (m) PD (m)
Location 1 731.8873 0.000228 0.075777
Location 2 744.7855 0.000225 0.090134
Location 3 735.177 0.000229 0.124411
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
7 11
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
5.2.2 Probabilistic Impact Analysis
Having established the superiority of PC-Kriging over
PCE and Kriging, we present results for probabilistic
impact analysis in this subsection based on the PC-Kriging
assisted approach. The results presented here correspond to
the impact location at the centre of the plate, unless other-
wise mentioned. Figures12 shows the variation of contact
force, displacements of impactor and plate, and velocity of
impactor with respect to time history for different stacking
sequences. The figure also shows the corresponding stochas-
tic response bounds arising due to the source- uncertainties.
It is found that contact force initially increases at a signifi-
cant rate with time and then decreases up to zero gradually.
Impactor and plate displacements are noticed to gradually
increase to a peak value and then reduce with the elapse of
time. The velocity of the impactor reduces gradually over
time and becomes constant after a certain duration.
The influence of fibre orientation angle in composite
laminates is shown in Fig.13. It is observed that the peak
CF occurs for the torsion stiff laminates. The effects of
twist angle on the critical impact responses are furnished in
Fig.14. In this case, the CF increases with the increase in
twist angle, while peak ID and peak PD have a reverse trend.
In case of impact loading, impact angle has a significant
effect on the critical impact responses as shown in Fig.15.
The peak CF and peak PD decreases with the increase in
impact angle from 0° to 45° while peak PD is found to have
a reverse trend. All the impact responses increase with the
increase in the initial velocity of the impactor as shown in
Fig. 9 Convergence study for PC-Kriging with respect to the number of training samples. For all the three responses, PC-Kriging converges at
512 training samples
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
Fig.16 due to the increase in kinetic energy. The standard
deviation of the response parameters is also found to follow
a similar trend for initial velocity of impactor. The increase
in impactor mass density also leads to an increase of all
impact responses for the same reason as shown in Fig.17.
The effect of plate thickness on the impact responses are
shown in Fig.18, wherein contact force is found to increase
with the increase in plate thickness. On the other hand, the
displacement of the impactor and plate displacement reduce
as the plate thickness increases. The standard deviation of
the response parameters is also found to follow a simi-
lar trend for thickness. The effect of location of impactor
Fig. 10 PCE vs Kriging versus
PC-Kriging (PC-Kriging is
found to yield the best results)
787
788
789
790
791
792
793
794
795
796
797
798
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
contacting point on the impact responses is shown in Fig.19.
It is observed that contact force is maximum at the loca-
tion 2 i.e. centre of the plate, while plate displacement and
impactor displacement are maximum at location 3. The rela-
tive coefficient of variation is shown for various influencing
system parameters in Fig.20 to understand about their rela-
tive degree of influence on the impact response parameters.
The coefficient of variation (COV) is obtained by taking the
ratio of standard deviation to mean of the responses. Here
the relative coefficient of variation (RCOV) is computed by
Fig. 11 Comparison of PCE, Kriging and PC-Kriging results. All the three models are trained with 512 training samples
799
800
801
802
803
804
805
806
807
808
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
Fig. 12 Stochastic variation of the time history of low-velocity impact responses for different stacking sequences of the composite plate a–d
for torsion stiff laminate (45°, − 45°, 45°, − 45°)s, e–h for bending stiff laminate (0°, 0°, 30°, − 30°)s considering t = 0.002m, ѱ = 0°, β = 0°,
V = 5m/s, ρ = 0.0085
N−s
cm
4 , and Δt = 1 micro-second. Stochastic variation of the time history of low-velocity impact responses for different stack-
ing sequences a–d for cross ply laminate (90°, 0°, 90°, 0°)s, e–h for quasi-Isotropic stiff laminate (0°, 45°, − 45°, 90°)s considering t = 0.002m,
ѱ = 0°, β = 0°, V = 5m/s, ρ = 0.0085
N−s
cm4
, and Δt = 1 micro-second
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
normalizing the COVs with respect to the sum of all COVs.
The relative sensitivity [67] of critical impact responses for
the six cases indicated in Sect.4.1 (considering impact at
the centre of the plate) can be clearly understood from this
analysis.
5.2.3 Fuzzy Based Non-probabilistic Impact Analysis
In this sub-section, we present numerical results correspond-
ing to the non-probabilistic assessment based on fuzzy
analysis, which is beneficial if the complete description of
the probability distribution of the input variables is not avail-
able. In this paper, the fuzzy approach is used to find out the
Fig. 13 Effect of variation of stacking sequence (quasi-isotropic stiff
laminate (0°, 45°, − 45°, 90°)s, torsion stiff laminate (45°, − 45°, 45°,
− 45°)s, cross ply laminate (90°, 0°, 90°, 0°)s and bending stiff lami-
nate (0°, 0°, 30°, − 30°)s on low-velocity impact responses consider-
ing t = 0.002m, ѱ = 0°, β = 0°, V = 5m/s, ρ = 0.0085
N−s
cm
4
Fig. 14 Effect of variation of twist angle (
𝜓
) on PDF plots of low-
velocity impact responses considering t = 0.002m, β = 0°, V = 5 m/s,
ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°)s
809
810
8 11
812
813
814
815
816
817
818
819
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
non-probabilistic responses by means of a predefined inter-
val of input parameters. The membership grade is considered
0–1 at a level of 0.25. Similar to probabilistic analysis, PC-
Kriging models trained with 512 training samples are used.
Fuzzy triangular membership function of the stochastic input
parameters is formed to address the variation of contact force,
plate displacement, and impactor displacement corresponding
to each level of α- cut. It is found that the resulting output
membership functions show a deviation from the triangular
distribution of input membership functions.
Similar to the probabilistic analysis, Figs.21, 22, 23, 24,
25, 26 and 27 show the influence of different input vari-
ables on the low-velocity impact responses following the
fuzzy based approach. In Fig.21, influence of ply-angle on
low velocity impact responses are shown. For torsion stiff
Fig. 15 Effect of variation of impact angle (β) on PDF plots of low-
velocity impact responses considering t = 0.002m,
𝜓=0◦
, V = 5m/s,
ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°)s
Fig. 16 Effect of variation of initial velocity of impactor (V) on
PDF plots of low-velocity impact responses considering t = 0.002m,
ѱ = 0°, β = 0°, ρ = 0.0085
N−s
cm4
, bending stiff laminate (02°/± 30°)s
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
Fig. 17 Effect of variation of mass density of impactor (ρ in
N−s
cm
4 ) on
PDF plots of low-velocity impact responses considering t = 0.002m,
ѱ = 0°, β = 0°, V = 5m/s, bending stiff laminate (02°/± 30°)s
Fig. 18 Effect of variation of thickness of plate (m) on PDF of low-
velocity impact responses considering ѱ = 0°, β = 0°, V = 5 m/s,
ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°)s
and bending stiff laminate configurations, the maximum and
minimum values of contact forces are identified respectively.
On the other hand, maximum plate displacement and impac-
tor displacement are observed for bending stiff laminate.
Figure22 shows the effect of the twist angle on fuzzy low
velocity impact response behaviour of laminated composite
plates. The contact force peak value is noticed to increase
as the angle of twist increases. On the other hand, as the
angle of twist increases the plate displacement is found to
reduce. The influence of oblique impact angle is shown in
Fig.23. The contact force and plate displacement decrease
with the increase in the oblique impact angle; impact dis-
placement is found to have a reverse trend. Figures24 and
25 show the effect of the mass and initial velocity of the
835
836
837
838
839
840
841
842
843
844
845
846
847
848
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
impactor on the transient impact responses, respectively. All
the critical responses increase with increase in the impac-
tor mass and initial impactor velocity. The influence of the
thickness of plate is shown in Fig.26, where the contact
force increases with the increase in thickness of plate. The
plate displacement and impactor displacement are found
to decrease with the increase in the thickness of laminate.
Finally, Fig.27 shows the effect of location of impactor on
the fuzzy responses considering three different points on
the plate surface. The maximum value of contact force is
observed when the impact occurs at the centre of the plate.
On the other hand, both plate-displacement and impactor
displacement are observed to have maximum value when
the impact is on location 3.
6 Remarks andPerspective onHybrid
Machine Learning Models
In this paper, we reviewed the possibility of using a hybrid
machine learning technique (PC-Kriging) for stochastic
computational mechanics considering a critical impact
problem. Note that the concept of hybrid machine learn-
ing approaches is not new; in fact, there exist a plethora of
hybrid machine learning approaches in the literature. The
primary idea of these methods is to combine more than one
machine learning models so as to exploit the advantages of
both (or, all of them). The first use of hybrid machine learn-
ing model is perhaps the ‘ensemble method’ proposed in
[80, 81]. The primary premise of this work was to represent
the response as a weighted combination of more than one
machine learning techniques. The ‘ensemble of surrogate’
method has gained significant attention and its applica-
tion can be found in different domain [82–83]. Analysis of
variance (ANOVA) decomposition [84], also known as the
high-dimensional model representation (HDMR) [85], is a
popular choice among researchers for hybridization. Over
the years, researchers have come up with different variants
of HDMR/ANOVA by combining it with other machine
learning techniques. For example, Shan and Wang [86,
87] combined radial basis function (RBF) with cut-HDMR
(aka anchored ANOVA) to formulate RBF-HDMR. Within
this framework, the basis functions in cut HDMR are rep-
resented by using RBF. In an independent study, Chowd-
hury etal. [88–90] formulated moving least square based
cut-HDMR (MLS-HDMR) for solving structural reliability
analysis problems. The formulation for MLS-HDMR and
RBF-HDMR are similar; the only difference resides in the
Fig. 19 Effect of variation of impactor contacting point on PDF plots
of low-velocity impact responses considering t = 0.002 m, ѱ = 0°,
β = 0°, V = 5 m/s, ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°) s
(Location of impact points on the laminated composite plate is indi-
cated in the inset of Fig.19a)
Fig. 20 Relative coefficient of variation (RCOV) for peak con-
tact force, plate displacement, and impactor displacement of cen-
trally impacted glass–epoxy laminated composite plates consider-
ing bending stiff laminate (02°/± 30°)s, ѱ = 45°, β = 30°, V = 10 m/s,
ρ = 0.0090
N−s
cm
4 , t = 0.004m
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
Fig. 21 Effect of variation of stacking sequence (quasi-isotropic stiff
laminate (0°, 45°, − 45°, 90°)s, torsion stiff laminate (45°, − 45°, 45°,
− 45°)s, cross ply laminate (90°, 0°, 90°, 0°)s and bending stiff lami-
nate (0°, 0°, 30°, − 30°)s on low-velocity impact responses consider-
ing t = 0.002m, ѱ = 0°, β = 0°, V = 5m/s, ρ = 0.0085
N−s
cm
4
Fig. 22 Effect of variation of twist angle (ѱ) on low-velocity impact
responses considering t = 0.002 m, β = 0°, V = 5 m/s, ρ = 0.0085
N−s
cm
4 ,
bending stiff laminate (02°/± 30°)s
fact that the basis functions for MLS-HDMR are represented
by using MLS based regression. As an improvement over
MLS-HDMR and RBF-HDMR, Huang etal. [91] proposed
support vector regression HDMR (SVR-HDMR) in 2015. It
was illustrated that the performance of SVR-HDMR outper-
forms RBF-HDMR. Note that all the HDMR based hybrid
machine learning algorithms discussed above are based on
cut-HDMR.
Hybrid machine learning approaches based on random
sampling HDMR, also known as the ANOVA decomposi-
tion, can also be found in the literature. Chakraborty and
Chowdhury [92] developed a sequential experimental design
based generalized ANOVA by coupling polynomial chaos
expansion [23, 25, 43] with RS-HDMR [93, 94]. An adap-
tive version of this algorithm was also proposed [95]. Later
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
generalized ANOVA was further hybridized by coupling
Gaussian process [96–98] with it. This was referred to as the
hybrid polynomial correlated function expansion (H-PCFE).
In essence, H-PCFE is a fusion of three machine learning
algorithms, namely PCE, RS-HDMR and Gaussian process
[22, 46, 47, 99]. The primary idea of H-PCFE is to represent
the mean function of Gaussian process by using general-
ized ANOVA. Adaptive variants of H-PCFE was proposed
in [100, 101]. It was illustrated that with hybridization (or
fusion), the accuracy of the machine learning algorithm
improves.
Apart from HDMR, hybrid machine learning algorithms
based on Gaussian process, also known as the Kriging [20,
21, 62] is also popular in the literature. In [102], a new
hybrid machine learning algorithm was developed by com-
bining fuzzy logic, artificial neural network and Kriging.
In another work, Pang etal. [103] combined Gaussian pro-
cess with neural network. The two methods differ in how
Fig. 23 Effect of variation of impact angle (β) on low-velocity impact
responses considering t = 0.002 m, ѱ = 0°, V = 5 m/s, ρ = 0.0085
N−s
cm
4 ,
bending stiff laminate (02°/± 30°)s
Fig. 24 Effect of variation of mass density of impactor (ρ in
N−s
cm
4 ) on
low-velocity impact responses considering t = 0.002m, ѱ = 0°, β = 0°,
V = 5m/s, bending stiff laminate (02°/± 30°)s
909
910
9 11
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
the neural network and Gaussian processes are combined.
While the former uses neural network and Gaussian process
sequentially, separated by fuzzy logic, the latter uses neural
network to represent the covariance function of the Gauss-
ian process. The method used in the current paper is also a
hybrid machine learning technique, referred to as the poly-
nomial chaos based Kriging (PC-Kriging). This method was
Fig. 25 Effect of variation of initial velocity of impactor (V in m/s) on
low-velocity impact responses considering t = 0.002m, ѱ = 0°, β = 0°,
ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°)s
Fig. 26 Effect of variation of plate thickness (t) on low-velocity
impact responses considering ѱ = 0°, β = 0°, V = 5 m/s, ρ = 0.0085
N−s
cm4
, bending stiff laminate (02°/± 30°)s
first proposed in [30] and was then further improved in [31].
In this method, the mean function of Kriging is represented
by using polynomial chaos expansion. It is argued that poly-
nomial chaos expansion performs a global approximation by
using basis function and Kriging performs local approxima-
tion by using the covariance kernel.
927
928
929
930
931
932
933
934
935
936
937
938
939
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
Based on the literature and the results presented in this
paper, it is safe to conclude that hybrid machine learning
approaches are generally more accurate as compared to a
single machine learning approach (note that such single
machine learning approaches have been shown to predict
accurately in various engineering problems [104–133]).
However, there is no free lunch and this enhancement in
the accuracy normally comes at a cost of the efficiency. For
instance, H-PCFE discussed above is more accurate but less
efficient as compared to the generalized ANOVA. Similarly,
polynomial chaos based Kriging used in this paper is more
accurate than polynomial chaos and Kriging; however, the
computational time necessary for training a polynomial
chaos based Kriging is more. To address this issue, research-
ers over the last few years have developed different adaptive
algorithms. Having said that, there is still a significant scope
for further developments when it comes to hybrid machine
learning algorithms.
7 Conclusions
This paper deals with the effects of input-uncertainty on
low-velocity impact responses of composite laminates,
which is investigated based on an efficient machine learning
algorithm. The Newmark’s time integration scheme is imple-
mented to solve time–histories of transient responses, while
the modified Hertzian contact law is employed to obtain
the contact force and other parameters. First, a determinis-
tic analysis is carried to investigate the effects of different
system parameters (such as stacking sequence, twist angle,
impact angle, initial velocity of impactor, mass density of
impactor and thickness of plate). Subsequently a proba-
bilistic analysis is presented to characterize the complete
probabilistic descriptions of low-velocity impact responses.
Finally, to address the scenario where complete statistical
descriptions of the input data are not available (sparse input
data), a fuzzy based non-probabilistic approach is presented
for low-velocity impact analysis of composites. Since con-
ventional methods for probabilistic and non-probabilistic
analyses are exorbitantly computationally expensive, we
integrated a hybrid polynomial chaos based kriging (PC-
Kriging) approach with the conventional framework to
obtain a high level of computational efficiency. By hybridiz-
ing the two powerful metamodelling techniques, polynomial
chaos and kriging (to capture the global and local behaviour
of a system, respectively), it is possible to exploit the com-
plementary advantages of these two models in a single com-
putational framework. In essence, here we have presented
a numerical demonstration of the superiority of hybrid
machine learning algorithms over individual models in a
systematic way including a critical review of the algorithms.
The novelty of this paper lies in characterizing the effect
of source-uncertainty on low- velocity impact of compos-
ite plates as well as development of the hybrid simulation
approach based on PC-Kriging coupled with the finite
element model of composite laminates to achieve compu-
tational efficiency. We have presented a comprehensive
study following both the probabilistic and non-probabilistic
approaches that covers every possible scenario of the avail-
ability or unavailability of the statistical distributions of the
input parameters. The stochastic results (both probabilistic
and non-probabilistic) in this paper show that the inevita-
ble effect of uncertainty has significant effect on the critical
impact responses of composite laminates. Thus it is impor-
tant to adopt an inclusive design approach by quantifying
the stochastic variation of the global responses to ensure
adequate safety and serviceability of the structure under low-
velocity impact. Besides that, the hybrid PC-Kriging based
approach adopted in this study to achieve computational effi-
ciency in the expensive and time-consuming process of mod-
elling impact in complex structural forms like composite
Fig. 27 Effect of variation of impactor contacting point on low-
velocity impact responses considering t = 0.002 m, ѱ = 0°, β = 0°,
V = 5m/s, ρ = 0.0085
N−s
cm
4 , bending stiff laminate (02°/± 30°) s (Loca-
tion of impact points on the laminated composite plate is indicated in
the inset of Fig.19a)
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
structures can be useful for other computationally intensive
problems of structural analyses and mechanics.
Acknowledgements TM and SN acknowledge the initiation grants
received from IIT Kanpur and IIT Bombay, respectively. PKK and RC
are grateful for the financial support from MHRD, India during the
research work. SC acknowledges the support of XSEDE and Center for
Research Computing, University of Notre Dame for providing compu-
tational resources required for carrying out this work.
Data availability All data used to generate these results is available in
the main paper. Further details could be obtained from the correspond-
ing author(s) upon request.
Compliance with Ethical Standards
Conflict of interest The authors declare that they have no conflict of
interest
References
1. Naskar S (2018) Spatial variability characterisation of laminated
composites, University of Aberdeen
2. Xu S, Chen PH (2013) Prediction of low velocity impact damage
in carbon/epoxy laminates. Procedia Eng 67:489–496. https ://
doi.org/10.1016/j.proen g.2013.12.049
3. Liu J, He W, Xie D, Tao B (2017) The effect of impactor shape
on the low-velocity impact behavior of hybrid corrugated core
sandwich structures. Compos Part B Eng 111:315–331. https ://
doi.org/10.1016/j.compo sites b.2016.11.060
4. Jagtap KR, Ghorpade SY, Lal A, Singh BN (2017) Finite element
simulation of low velocity impact damage in composite lami-
nates. Mater Today Proc 4:2464–2469. https ://doi.org/10.1016/j.
matpr .2017.02.098
5. Balasubramani V, Boopathy SR, Vasudevan R (2013) Numeri-
cal analysis of low velocity impact on laminated composite
plates. Procedia Eng 64:1089–1098. https ://doi.org/10.1016/j.
proen g.2013.09.187
6. Tan TM, Sun CT (1985) Use of statical indentation laws in the
impact analysis of laminated composite plates. J Appl Mech
52:6. https ://doi.org/10.1115/1.31690 29
7. Sun CT, Chen JK (1985) On the impact of initially stressed
composite laminates. J Compos Mater 19:490–504. https ://doi.
org/10.1177/00219 98385 01900 601
8. Richardson MOW, Wisheart MJ (1996) Review of low-veloc-
ity impact properties of composite materials. Compos Part A
Appl Sci Manuf 27:1123–1131. https ://doi.org/10.1016/1359-
835X(96)00074 -7
9. Ahmed A, Wei L (2015) The low velocity impact dam-
age resistance of the composite structures. Rev Adv Mater
40:127–145
10. Yuan Y, Xu C, Xu T, Sun Y, Liu B, Li Y (2017) An analyti-
cal model for deformation and damage of rectangular laminated
glass under low-velocity impact. Compos Struct 176:833–843.
https ://doi.org/10.1016/j.comps truct .2017.06.029
11. Zhang J, Zhang X (2015) An efficient approach for predicting
low-velocity impact force and damage in composite laminates.
Compos Struct 130:85–94. https ://doi.org/10.1016/j.comps truct
.2015.04.023
12. Feng D, Aymerich F (2014) Finite element modelling of dam-
age induced by low-velocity impact on composite laminates.
Compos Struct 108:161–171. https ://doi.org/10.1016/j.comps
truct .2013.09.004
13. Maio L, Monaco E, Ricci F, Lecce L (2013) Simulation of low
velocity impact on composite laminates with progressive failure
analysis. Compos Struct 103:75–85. https ://doi.org/10.1016/j.
comps truct .2013.02.027
14. Kim E-H, Rim M-S, Lee I, Hwang T-K (2013) Composite dam-
age model based on continuum damage mechanics and low
velocity impact analysis of composite plates. Compos Struct
95:123–134. https ://doi.org/10.1016/j.comps truct .2012.07.002
15. Lipeng W, Ying Y, Dafang W, Hao W (2008) Low-velocity
impact damage analysis of composite laminates using self-adapt-
ing delamination element method. Chin J Aeronaut 21:313–319.
https ://doi.org/10.1016/S1000 -9361(08)60041 -2
16. Johnson A, Pickett A, Rozycki P (2001) Computational meth-
ods for predicting impact damage in composite structures.
Compos Sci Technol 61:2183–2192. https ://doi.org/10.1016/
S0266 -3538(01)00111 -7
17. Coutellier D, Walrick JC, Geoffroy P (2006) Presentation of
a methodology for delamination detection within laminated
structures. Compos Sci Technol 66:837–845. https ://doi.
org/10.1016/j.comps citec h.2004.12.037
18. Jih CJ, Sun CT (1993) Prediction of delamination in composite
laminates subjected to low velocity impact. J Compos Mater
27:684–701. https ://doi.org/10.1177/00219 98393 02700 703
19. Mukhopadhyay T, Chakraborty S, Dey S, Adhikari S, Chowd-
hury R (2017) A critical assessment of kriging model variants
for high-fidelity uncertainty quantification in dynamics of com-
posite shells. Arch Comput Methods Eng 24:495–518. https ://
doi.org/10.1007/s1183 1-016-9178-z
20. Biswas S, Chakraborty S, Chandra S, Ghosh I (2017) Kriging-
based approach for estimation of vehicular speed and passen-
ger car units on an urban arterial. J Transp Eng Part A Syst
143:04016013
21. Kaymaz I (2005) Application of Kriging method to structural
reliability problems. Struct Saf 27:133–151
22. Nayek R, Chakraborty S, Narasimhan S (2019) A Gaussian
process latent force model for joint input-state estimation in
linear structural systems. Mech Syst Signal Process 128:497–
530. https ://doi.org/10.1016/j.ymssp .2019.03.048
23. Xiu D, Karniadakis GE (2002) The Wiener–Askey polynomial
chaos for stochastic differential equations. SIAM J Sci Comput
24:619–644
24. Blatman G, Sudret B (2010) An adaptive algorithm to build
up sparse polynomial chaos expansions for stochastic finite
element analysis. Probab Eng Mech 25:183–197
25. Sudret B (2008) Global sensitivity analysis using polynomial
chaos expansions. Reliab Eng Syst Saf 93:964–979
26. Chakraborty S, Chowdhury R (2017) Hybrid framework for the
estimation of rare failure event probability. J Eng Mech. https
://doi.org/10.1061/(asce)em.1943-7889.00012 23
27. Chakraborty S, Goswami S, Rabczuk T (2019) A surrogate
assisted adaptive framework for robust topology optimization.
Comput Methods Appl Mech Eng 346:63–84. https ://doi.
org/10.1016/j.cma.2018.11.030
28. Chakraborty S, Chatterjee T, Chowdhury R, Adhikari S (2017)
A surrogate based multi- fidelity approach for robust design opti-
mization. Appl Math Model 47:726–744
29. Chakraborty S, Chowdhury R (2016) Polynomial correlated
function expansion. https ://doi.org/10.4018/978-1-5225-0588-4.
ch012
30. Schobi R, Sudret B, Wiart J (2015) Polynomial chaos based Krig-
ing. Int J Uncertain Quantif 5:171–193. https ://doi.org/10.1615/
Int.J.Uncer taint yQuan tific ation .20150 12467
31. Kersaudy P, Sudret B, Varsier N, Picon O, Wiart J (2015) A
new surrogate modeling technique combining Kriging and
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
10 74
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
110 0
1101
110 2
110 3
110 4
110 5
110 6
110 7
110 8
110 9
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
112 0
112 1
112 2
112 3
112 4
112 5
112 6
112 7
112 8
112 9
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
polynomial chaos expansions: application to uncertainty analysis
in computational dosimetry. J Comput Phys 286:103–117. https
://doi.org/10.1016/j.jcp.2015.01.034
32. Goswami S, Chakraborty S, Chowdhury R, Rabczuk T (2019)
Threshold shift method for reliability-based design optimization.
http://arxiv .org/abs/1904.11424
33. Naskar S, Sriramula S (2017) Random field based approach for
quantifying the spatial variability in composite laminates. In:
20th International conference on composite structures (ICCS20)
34. Dey S, Mukhopadhyay T, Spickenheuer A, Adhikari S, Heinrich
G (2016) Bottom up surrogate based approach for stochastic fre-
quency response analysis of laminated composite plates. Compos
Struct 140:712–727
35. Dey S, Karmakar A (2014) Effect of oblique angle on low veloc-
ity impact response of delaminated composite conical shells.
Proc Inst Mech Eng Part C J Mech Eng Sci 228:2663–2677.
https ://doi.org/10.1177/09544 06214 52179 9
36. Yang S, Sun C (1982) Indentation law for composite laminates.
In: Composite materials: testing and design (6th conference), p
425. https ://doi.org/10.1520/stp28 494s
37. Bathe KJ (1996) Finite element procedures. Prentice Hall, New
Jersey
38. Wiener N (1938) The homogeneous chaos. Am J Math
60:897–936
39. Hampton J, Doostan A (2015) Coherence motivated sampling
and convergence analysis of least squares polynomial Chaos
regression. Comput Methods Appl Mech Eng 290:73–97
40. Coelho RF, Lebon J, Bouillard P (2011) Hierarchical stochastic
metamodels based on moving least squares and polynomial chaos
expansion. Struct Multidiscip Optim 43:707–729
41. Madankan R, Singla P, Patra A, Bursik M, Dehn J, Jones M,
Pavolonis M, Pitman B, Singh T, Webley P (2012) Polynomial
chaos quadrature-based minimum variance approach for source
parameters estimation. Procedia Comput Sci 9:1129–1138
42. Zhang Z, El-Moselhy TA, Elfadel IM, Daniel L (2014) Calcula-
tion of generalized polynomial-chaos basis functions and Gauss
quadrature rules in hierarchical uncertainty quantification. IEEE
Trans Comput Des Integr Circuits Syst 33:728–740
43. Blatman G, Sudret B (2011) Adaptive sparse polynomial chaos
expansion based on least angle regression. J Comput Phys
230:2345–2367
44. Jacquelin E, Adhikari S, Sinou JJ, Friswell MI (2015) Polynomial
chaos expansion in structural dynamics: accelerating the conver-
gence of the first two statistical moment sequences. J Sound Vib
356:144–154
45. Pascual B, Adhikari S (2012) Hybrid perturbation-polynomial
chaos approaches to the random algebraic eigenvalue problem.
Comput Methods Appl Mech Eng 217–220:153–167
46. Bilionis I, Zabaras N (2012) Multi-output local Gaussian process
regression: applications to uncertainty quantification. J Comput
Phys 231:5718–5746
47. Bilionis I, Zabaras N, Konomi BA, Lin G (2013) Multi-output
separable Gaussian process: towards an efficient, fully Bayes-
ian paradigm for uncertainty quantification. J Comput Phys
241:212–239
48. Krige DG (1951) A statistical approach to some basic mine valu-
ation problems on the witwatersrand. J Chem Metall Min Soc S
Afr 52:119–139
49. Krige DG (1951) A statisitcal approach to some mine valua-
tions and allied problems at the Witwatersrand, University of
Witwatersrand
50. Olea RA (2011) Optimal contour mapping using Kriging. J Geo-
phys Res 79:695–702
51. Warnes JJ (1986) A sensitivity analysis for universal kriging.
Math Geol 18:653–676
52. Joseph VR, Hung Y, Sudjianto A (2008) Blind Kriging: a new
method for developing metamodels. J Mech Des 130:031102
53. Hung Y (2011) Penalized blind kriging in computer experiments.
Stat Sin 21:1171–1190
54. Couckuyt I, Forrester A, Gorissen D, De Turck F, Dhaene T
(2012) Blind Kriging: implementation and performance analysis.
Adv Eng Softw 49:1–13
55. Kennedy M, O’Hagan A (2000) Predicting the output from a
complex computer code when fast approximations are available.
Biometrika 87:1–13
56. Kamiński B (2015) A method for the updating of stochastic krig-
ing metamodels. Eur J Oper Res 247:859–866
57. Qu H, Fu MC (2014) Gradient extrapolated stochastic kriging.
ACM Trans Model Comput Simul 24:1–25
58. Wang B, Bai J, Gea HC (2013, Stochastic Kriging for random
simulation metamodeling with finite sampling. In: 39th Design
automation conference, vol 3B, ASME, p V03BT03A056. https
://doi.org/10.1115/detc2 013-13361
59. Rivest M, Marcotte D (2012) Kriging groundwater solute con-
centrations using flow coordinates and nonstationary covariance
functions. J Hydrol 472–473:238–253
60. Putter H, Young GA (2001) On the effect of covariance func-
tion estimation on the accuracy of Kriging predictors. Bernoulli
7:421–438
61. BiscayLirio R, Camejo DG, Loubes JM, MuñizAlvarez L (2013)
Estimation of covariance functions by a fully data-driven model
selection procedure and its application to Kriging spatial inter-
polation of real rainfall data. Stat Methods Appl 23:149–174
62. Saha A, Chakraborty S, Chandra S, Ghosh I (2018) Kriging
based saturation flow models for traffic conditions in Indian
cities. Transp Res Part A Policy Pract 118:38–51. https ://doi.
org/10.1016/j.tra.2018.08.037
63. Sobol IM (1976) Uniformly distributed sequences with an
additional uniform property. USSR Comput Math Math Phys
16:236–242
64. Bratley P, Fox BL (1988) Implementing Sobol’s quasirandom
sequence generator. ACM Trans Math Softw 14:88–100
65. Witteveen JAS, Bijl H (2006) Modeling arbitrary uncertainties
using gram-schmidt polynomial chaos. In: 44th AIAA aero-
space sciences meeting and exhibition, American Institute of
Aeronautics and Astronautics, Reston, Virigina. https ://doi.
org/10.2514/6.2006-896
66. Hanss M, Willner K (2000) A fuzzy arithmetical approach to
the solution of finite element problems with uncertain param-
eters. Mech Res Commun 27:257–272. https ://doi.org/10.1016/
S0093 -6413(00)00091 -4
67. Moens D, Hanss M (2011) Non-probabilistic finite element
analysis for parametric uncertainty treatment in applied
mechanics: recent advances. Finite Elem Anal Des 47:4–16.
https ://doi.org/10.1016/j.finel .2010.07.010
68. Kollár LP, Springer GS (2003) Mechanics of composite struc-
tures. Cambridge University Press, Cambridge. https ://doi.
org/10.1017/cbo97 80511 54714 0
69. Kalita K, Mukhopadhyay T, Dey P, Haldar S (2020) Genetic
programming assisted multi- scale optimization for multi-
objective dynamic performance of laminated composites: the
advantage of more elementary-level analyses. Neural Comput
Appl 32:7969–7993
70. Kumar RR, Mukhopadhyay T, Pandey KM, Dey S (2019) Sto-
chastic buckling analysis of sandwich plates: the importance
of higher order modes. Int J Mech Sci 152:630–643
71. Naskar S, Mukhopadhyay T, Sriramula S (2019) Spatially vary-
ing fuzzy multi-scale uncertainty propagation in unidirectional
fibre reinforced composites. Compos Struct 209:940–967
72. Dey S, Mukhopadhyay T, Naskar S, Dey TK, Chalak HD,
Adhikari S (2019) Probabilistic characterization for dynamics
113 0
113 1
113 2
113 3
113 4
113 5
113 6
113 7
113 8
113 9
114 0
114 1
114 2
114 3
114 4
114 5
114 6
114 7
114 8
114 9
115 0
115 1
115 2
115 3
115 4
115 5
115 6
115 7
115 8
115 9
116 0
116 1
116 2
116 3
116 4
116 5
116 6
116 7
116 8
116 9
117 0
117 1
117 2
117 3
1174
117 5
1176
117 7
117 8
117 9
118 0
118 1
118 2
118 3
118 4
118 5
118 6
118 7
118 8
118 9
119 0
119 1
119 2
119 3
119 4
119 5
119 6
119 7
119 8
119 9
1200
12 01
1202
1203
1204
1205
1206
1207
1208
1209
1210
12 11
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
Stochastic Oblique Impact onComposite Laminates: AConcise Review andCharacterization of…
1 3
and stability of laminated soft core sandwich plates. J Sand-
wich Struct Mater 21(1):366–397
73. Mukhopadhyay T, Naskar S, Karsh PK, Dey S, You Z (2018)
Effect of delamination on the stochastic natural frequencies of
composite laminates. Compos B Eng 154:242–256
74. Naskar S, Mukhopadhyay T, Sriramula S (2018) Probabilistic
micromechanical spatial variability quantification in laminated
composites. Compos B Eng 151:291–325
75. Karsh PK, Mukhopadhyay T, Dey S (2019) Stochastic low-
velocity impact on functionally graded plates: probabilistic and
non-probabilistic uncertainty quantification. Compos B Eng
159:461–480
76. Karsh PK, Mukhopadhyay T, Chakraborty S, Naskar S, Dey
S (2019) A hybrid stochastic sensitivity analysis for low-
frequency vibration and low-velocity impact of functionally
graded plates. Compos B Eng 176:107221
77. Kumar RR, Mukhopadhyay T, Naskar S, Pandey KM, Dey S
(2019) Stochastic low-velocity impact analysis of sandwich
plates including the effects of obliqueness and twist. Thin
Walled Struct 145:106411
78. Naskar S, Mukhopadhyay T, Sriramula S (2017) Non-prob-
abilistic analysis of laminated composites based on fuzzy
uncertainty quantification. In: 20th International conference
on composite structures (ICCS20)
79. Naskar S, Sriramula S (2017) Vibration analysis of hollow
circular laminated composite beams: a stochastic approach.
In: 12th International conference on structural safety and
reliability
80. Goel T, Haftka RT, Shyy W, Queipo NV (2007) Ensemble of
surrogates. Struct Multidiscip Optim 33:199–216
81. Müller J, Shoemaker CA (2014) Influence of ensemble sur-
rogate models and sampling strategy on the solution quality
of algorithms for computationally expensive black-box global
optimization problems. J Glob Optim 60:123–144. https ://doi.
org/10.1007/s1089 8-014-0184-0
82. Müller J, Piché R (2011) Mixture surrogate models based
on Dempster–Shafer theory for global optimization prob-
lems. J Glob Optim 51:79–104. https ://doi.org/10.1007/s1089
8-010-9620-y
83. Viana FAC, Haftka RT, Watson LT (2013) Efficient global opti-
mization algorithm assisted by multiple surrogate techniques.
J Glob Optim 56:669–689. https ://doi.org/10.1007/s1089
8-012-9892-5
84. Yang X, Choi M, Lin G, Karniadakis GE (2012) Adaptive
ANOVA decomposition of stochastic incompressible and com-
pressible flows. J Comput Phys 231:1587–1614
85. Rabitz H, Aliş ÖF (1999) General foundations of high dimen-
sional model representations. J Math Chem 25:197–233
86. Shan S, Wang GG (2010) Survey of modeling and optimization
strategies to solve high-dimensional design problems with com-
putationally-expensive black-box functions. Struct Multidiscip
Optim 41:219–241. https ://doi.org/10.1007/s0015 8-009-0420-2
87. Shan S, Wang GG (2011) Turning black-box functions into white
functions. J Mech Des. https ://doi.org/10.1115/1.40029 78
88. Chowdhury R, Rao BN (2009) Assessment of high dimensional
model representation techniques for reliability analysis. Probab
Eng Mech 24:100–115
89. Chowdhury R, Rao BN, Prasad AM (2007) High dimensional
model representation for piece-wise continuous function approxi-
mation. Commun Numer Methods Eng 24:1587–1609
90. Chowdhury R, Rao BN, Prasad AM (2009) High-dimensional
model representation for structural reliability analysis. Commun
Numer Methods Eng 25:301–337
91. Huang Z, Qiu H, Zhao M, Cai X, Gao L (2015) An adap-
tive SVR-HDMR model for approximating high dimensional
problems. Eng Comput 32:643–667. https ://doi.org/10.1108/
EC-08-2013-0208
92. Chakraborty S, Chowdhury R (2016) Sequential experimental
design based generalised ANOVA. J Comput Phys 317:15–32
93. Chakraborty S, Chowdhury R (2017) Polynomial correlated
function expansion. In: Modeling and simulation techniques in
structural engineering, IGI Global, pp 348–373
94. Chakraborty S, Chowdhury R (2015) Polynomial correlated
function expansion for nonlinear stochastic dynamic analysis.
J Eng Mech 141:04014132. https ://doi.org/10.1061/(ASCE)
EM.1943-7889.00008 55
95. Chakraborty S, Chowdhury R (2017) Towards ‘h-p adap-
tive’ generalized ANOVA. Comput Methods Appl Mech Eng
320:558–581
96. Chakraborty S, Chowdhury R (2016) Moment independent sen-
sitivity analysis: H-PCFE–based approach. J Comput CivEng
31:06016001-1–06016001-11. https ://doi.org/10.1061/(asce)
cp.1943-5487.00006 08
97. Majumder D, Chakraborty S, Chowdhury R (2017) Probabilis-
tic analysis of tunnels: a hybrid polynomial correlated function
expansion based approach. Tunn Undergr Space Technol. https
://doi.org/10.1016/j.tust.2017.07.009
98. Chatterjee T, Chakraborty S, Chowdhury R (2016) A bi-level
approximation tool for the computation of FRFs in stochastic
dynamic systems. Mech Syst Signal Process 70–71:484–505
99. Chakraborty S, Chowdhury R (2019) Graph-theoretic-approach-
assisted gaussian process for nonlinear stochastic dynamic analy-
sis under generalized loading. J Eng Mech 145:04019105. https
://doi.org/10.1061/(ASCE)EM.1943-7889.00016 85
100. Chakraborty S, Chowdhury R (2017) An efficient algorithm for
building locally refined hp—adaptive H-PCFE: application to
uncertainty quantification. J Comput Phys 351:59–79
101. Chakraborty S, Chowdhury R (2017) Hybrid framework for
the estimation of rare failure event probability. J Eng Mech
143:04017010. https ://doi.org/10.1061/(ASCE)EM.1943-
7889.00012 23
102. Tapoglou E, Karatzas GP, Trichakis IC, Varouchakis EA (2014)
A spatio-temporal hybrid neural network-Kriging model for
groundwater level simulation. J Hydrol 519:3193–3203. https ://
doi.org/10.1016/j.jhydr ol.2014.10.040
103. Pang G, Yang L, Karniadakis GE (2019) Neural-net-induced
Gaussian process regression for function approximation and PDE
solution. J Comput Phys 384:270–288. https ://doi.org/10.1016/j.
jcp.2019.01.045
104. Dey S, Mukhopadhyay T, Khodaparast HH, Adhikari S (2016) A
response surface modelling approach for resonance driven reli-
ability based optimization of composite shells. Periodica Poly-
technica Civ Eng 60(1):103–111
105. Naskar S, Mukhopadhyay T, Sriramula S (2018) A comparative
assessment of ANN and PNN model for low-frequency stochastic
free vibration analysis of composite plates Handbook of proba-
bilistic models for engineers and scientists, Elsevier Publication,
pp 527–547
106. Mukhopadhyay T, Dey TK, Dey S, Chakrabarti A (2015) Optimi-
zation of fiber reinforced polymer web core bridge deck: a hybrid
approach. Struct Eng Int 25(2):173–183
107. Dey S, Mukhopadhyay T, Sahu SK, Adhikari S (2018) Stochas-
tic dynamic stability analysis of composite curved panels sub-
jected to non-uniform partial edge loading. Eur J Mech A Solids
67:108–122
108. Dey S, Mukhopadhyay T, Adhikari S (2017) Metamodel based
high-fidelity stochastic analysis of composite laminates: a con-
cise review with critical comparative assessment. Compos Struct
171:227–250
109. Naskar S, Mukhopadhyay T, Sriramula S, Adhikari S (2017)
Stochastic natural frequency analysis of damaged thin-walled
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
12 74
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
13 01
1302
1303
1304
1305
1306
1307
1308
1309
1310
13 11
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
13 74
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
REVISED PROOF
Journal : Large 11831 Article No : 9438 Pages : 32 MS Code : 9438 Dispatch : 23-7-2020
T.Mukhopadhyay et al.
1 3
laminated composite beams with uncertainty in micromechanical
properties. Compos Struct 160:312–334
110. Dey S, Mukhopadhyay T, Sahu SK, Adhikari S (2016) Effect
of cutout on stochastic natural frequency of composite curved
panels. Compos B Eng 105:188–202
111. Dey S, Mukhopadhyay T, Spickenheuer A, Gohs U, Adhikari S
(2016) Uncertainty quantification in natural frequency of com-
posite plates: an artificial neural network based approach. Adv
Compos Lett 25(2):43–48
112. Dey TK, Mukhopadhyay T, Chakrabarti A, Sharma UK (2015)
Efficient lightweight design of FRP bridge deck. Proc Inst Civ
Eng Struct Build 168(10):697–707
113. Dey S, Mukhopadhyay T, Khodaparast HH, Adhikari S (2016)
Fuzzy uncertainty propagation in composites using Gram-
Schmidt polynomial chaos expansion. Appl Math Model
40(7–8):4412–4428
114. Mukhopadhyay T, Naskar S, Dey S, Adhikari S (2016) On quan-
tifying the effect of noise in surrogate based stochastic free vibra-
tion analysis of laminated composite shallow shells. Compos
Struct 140:798–805
115. Dey S, Naskar S, Mukhopadhyay T, Gohs U, Sriramula S, Adhi-
kari S, Heinrich G (2016) Uncertain natural frequency analysis
of composite plates including effect of noise: a polynomial neural
network approach. Compos Struct 143:130–142
116. Naskar S, Sriramula S (2018) On quantifying the effect of noise
in radial basis based stochastic free vibration analysis of lami-
nated composite beam. In: 8th European conference on compos-
ite materials
117. Dey S, Mukhopadhyay T, Khodaparast HH, Kerfriden P, Adhi-
kari S (2015) Rotational and ply-level uncertainty in response of
composite shallow conical shells. Compos Struct 131:594–605
118. Mukhopadhyay T, Dey TK, Chowdhury R, Chakrabarti A,
Adhikari S (2015) Optimum design of FRP bridge deck: an
efficient RS-HDMR based approach. Struct Multidiscip Optim
52(3):459–477
119. Dey S, Mukhopadhyay T, Adhikari S (2018) Uncertainty quanti-
fication in laminated composites: a meta-model based approach.
CRC Press, Boca Raton
120. Vaishali Mukhopadhyay T, Karsh PK, Basu B, Dey S (2020)
Machine learning based stochastic dynamic analysis of function-
ally graded shells. Compos Struct 237:111870
121. Mukhopadhyay T (2018) A multivariate adaptive regression
splines based damage identification methodology for web core
composite bridges including the effect of noise. J Sandwich
Struct Mater 20(7):885–903
122. Karsh PK, Mukhopadhyay T, Dey S (2018) Stochastic dynamic
analysis of twisted functionally graded plates. Compos B Eng
147:259–278
123. Maharshi K, Mukhopadhyay T, Roy B, Roy L, Dey S (2018)
Stochastic dynamic behaviour of hydrodynamic journal bear-
ings including the effect of surface roughness. Int J Mech Sci
142–143:370–383
124. Metya S, Mukhopadhyay T, Adhikari S, Bhattacharya G (2017)
System reliability analysis of soil slopes with general slip sur-
faces using multivariate adaptive regression splines. Comput
Geotech 87:212–228
125. Mukhopadhyay T, Mahata A, Dey S, Adhikari S (2016) Proba-
bilistic analysis and design of HCP nanowires: an efficient sur-
rogate based molecular dynamics simulation approach. J Mater
Sci Technol 32(12):1345–1351
126. Mukhopadhyay T, Chowdhury R, Chakrabarti A (2016) Struc-
tural damage identification: a random sampling-high dimensional
model representation approach. Adv Struct Eng 19(6):908–927
127. Mahata A, Mukhopadhyay T, Adhikari S (2016) A polynomial
chaos expansion based molecular dynamics study for probabilis-
tic strength analysis of nano-twinned copper. Mater Res Express
3:036501
128. Dey S, Mukhopadhyay T, Sahu SK, Li G, Rabitz H, Adhikari S
(2015) Thermal uncertainty quantification in frequency responses
of laminated composite plates. Compos B Eng 80:186–197
129. Dey S, Mukhopadhyay T, Khodaparast HH, Adhikari S (2015)
Stochastic natural frequency of composite conical shells. Acta
Mech 226(8):2537–2553
130. Mukhopadhyay T, Dey TK, Chowdhury R, Chakrabarti A (2015)
Structural damage identification using response surface based
multi-objective optimization: a comparative study. Arab J Sci
Eng 40(4):1027–1044
131. Naskar S, Sriramula S (2017) Effective elastic property of ran-
domly damaged composite laminates, Engineering postgraduate
research symposium, Aberdeen, United Kingdom
132. Dey S, Mukhopadhyay T, Adhikari S (2015) Stochastic free
vibration analyses of composite doubly curved shells: a Kriging
model approach. Compos B Eng 70:99–112
133. Dey S, Mukhopadhyay T, Adhikari S (2015) Stochastic free
vibration analysis of angle-ply composite plates: a RS-HDMR
approach. Compos Struct 122:526–536
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
1392
1393
1394
1395
1396
1397
1398
1399
1400
14 01
1402
1403
1404
1405
1406
1407
1408
1409
1410
14 11
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
14 74
1475
1476
1477
1478
1479
1480
