FGM DL

Composites Part B 159 (2019) 300–326
Contents lists available at ScienceDirect
Composites Part B
journal homepage: www.elsevier.com/locate/compositesb
Material optimization of functionally graded plates using deep neural T

network and modified symbiotic organisms search for eigenvalue problems
Dieu.T.T. Do, Dongkyu Lee, Jaehong Lee∗
Department of Architectural Engineering, Sejong University, 209 Neungdong-ro, Gwangjin-gu, Seoul, 05006, Republic of Korea
A R T I C LE I N FO A B S T R A C T
Keywords: The paper is aimed at improving computational cost enhanced by a new combination of deep neural network
Deep neural network (DNN) and modified symbiotic organisms search (mSOS) algorithm for optimal material distribution of func-
Modified symbiotic organisms search tionally graded (FG) plates. The material distribution is described by control points, in which coordinates of
Functionally graded plates these points are located along the plate thickness using B-spline basis functions. In addition, DNN is used as an
Buckling
analysis tool to supersede finite element analysis (FEA). By using DNN, solutions can directly be predicted by an
Free vibration
optimal mapping which is defined by learning relationship between input and output data of a dataset in training
process. Each of dataset is randomly created from analysis through iterations by using isogeometric analysis
(IGA). The mSOS being a robust metaheuristic algorithm is employed to solve two optimization problems:
buckling and free vibration with various volume constraints. Moreover, the power of mSOS is verified by
comparing to other algorithms in the open literature. Finally, optimal results in all examples generated by the
proposed method are compared to those of a combination of IGA and mSOS to demonstrate its effectiveness and
robustness.
1. Introduction extensively applied to various engineering fields. For example, Zhang

et al. [23] used IGA to formulate finite elements for arbitrary spatially
Functionally graded materials (FGMs) are a kind of smart materials curved beams. As shown in the paper, the order of interpolations used
and their properties change continuously along certain directions. in the analysis is increased to tackle issues in curved beam elements
FGMs are usually made from two distinct component materials such as such as shear and membrane locking. IGA is also used for vibration
metal and ceramic, in which ceramic phase is perfectly efficient in analysis of pipes conveying fluid by Zare et al. [24]. Due to continuous
withstanding high temperatures while metal phase is of excellent properties of NURBS, gained results with a smaller number of DOFs
fracture toughness. Overcoming limitations of laminated composites, were much better than those of standard finite element method. Luu
the FGMs perfectly discard undesired stress discontinuity appearing et al. [25] presented bending and buckling analyses of shear deformable
between two layers in laminated composites. For these reasons, FGMs laminated composite curved beams with arbitrary lay-up and curvature.
are widely used in many various fields such as aircraft engineering, In which, behaviors of curved beams were analyzed based on first order
nuclear plants, electrical engineering, etc. A large number of research shear deformation theory and NURBS was used to model displacement
publications could be easily found in the literature as Refs. [1–18]. The field, geometry and curvature. The effectiveness and robustness of the
effective material properties of the FGMs can be defined by different high-order NURBS elements were demonstrated via convergence study.
rules such as the three-phase model by Frohlich et al. [19], Mori-Ta- Tran et al. [26] studied behaviors of FG plates based on IGA and HSDT
naka technique [20], or mixture [21], etc. to enhance the accuracy and depict precisely shear stress distribution.
Recently, IGA was proposed by Hughes et al. [22] to integrate Menasria et al. [27] proposed a novel HSDT to investigate the thermal
computer aided design (CAD) and FEA. In IGA, the same non-uniform buckling responses of FG sandwich plates. In which, the theory con-
rational B-spline (NURBS), which is smoother and higher, is used for siders a trigonometric variation of transverse shear stress and verifies
symbolizing exact CAD geometry and approximating to FEA solution the traction free boundary conditions without employing the shear
fields. Moreover, the exact geometry is still conserved at the coarsest correction factors. The efficiency of that method was proven in com-
discretization level and this method is effective in reducing degrees of parison with other HSDTs. Nguyen-Xuan et al. [28] proposed a com-
freedoms (DOFs) for high-order elements. Therefore, IGA has been bination between fifth-order shear deformation theory and IGA for
∗
Corresponding author.
E-mail address: jhlee@sejong.ac.kr (J. Lee).
https://doi.org/10.1016/j.compositesb.2018.09.087
Received 25 June 2018; Received in revised form 3 September 2018; Accepted 24 September 2018
Available online 05 October 2018
1359-8368/ © 2018 Published by Elsevier Ltd.
D.T.T. Do et al. Composites Part B 159 (2019) 300–326
composite sandwich plates. Besides, Carrera et al. [29] proposed re- output data is linear while practical problems are almost non-linear. For
fined theories based on Carrera's Unified Formulation (CUF) for the this reason, DNN being an ANN with multiple hidden layers between
linear static analysis of beam structures. In which, higher order models input and output layers is developed to solve more complex nonlinear
that account for shear deformations and in- and out-of-plane warping relationships. Accordingly, DNN is extensively applied in various fields
can be formulated straightforwardly. The accuracy of solution and ef- such as computer vision, speech recognition, robots, self-driving cars,
ficiency of the proposed methods were demonstrated by comparing to translation, etc. In particular, Gopalakrishnan et al. [44] used DNN
the three-dimensional finite element models. Consequently, several models trained on database including millions of images and transfer
remarkable applications of that method to other problems could be their learning capacity to automatic pavement crack detection from
found in Refs. [30–35]. Therefore, HSDT will be used in this study. digitized pavement surface images attained from the FHWA/LTPP da-
In most of the above-mentioned studies for FGMs, ceramic volume tabase. Vesperini et al. [45] suggested DNN for multi-room speaker
fraction of FG plates is defined by a specified mathematical function localization. In which, multi-layer perceptron and convolutional neural
with gradient indexes such as the exponential law or power law [36]. networks architectures was examined. Spille et al. [46] used DNN for
Generally, indexes are easily adjusted to attain desired material prop- predicting speech intelligibility over a wide range. In that study, the
erties; however, such material distributions are idealized by a pre- proposed method did not demand separated noise signal and clean
defined shape function. Therefore, it can be restricted in applying to speech input. From obtained results, it was concluded that DNNs can
practical problems. To circumvent this limitation, researchers have listen in the dips in modulated maskers. Lee et al. [47] applied recent
suggested several advanced techniques for describing volume fraction. DNN methods for ten-bar truss structure. The efficiency and accuracy of
For instance, Cho et al. [37,38] proposed piecewise bilinear Lagrangian DNN was proven in comparison with conventional neural network.
shape functions for symbolizing volume fraction distribution. From that They indicated that over the last decade, applications of ANN for
papers, the approximated volume fraction with bilinear elements for structural analysis problems have been reduced dramatically because of
finite difference method performed efficiently and satisfied conceptual restrictions of traditional methods [47]. Similar to ANN, DNN also has
C 0 -continuity of the volume fraction. In complex profiles, a large two main issues: overfitting and computational time. DNN is prone to
number of volume fraction points were demanded to achieve an even overfitting due to added layers of abstraction; therefore, in this study,
better approximation. As a result, the number of design variables in- dropout technique [48] is used to tackle this problem. Furthermore,
creases in optimization problem and optimal process will consume a many training parameters such as number of layers, units, learning rate,
significant amount of computational cost. Otherwise, if higher-order etc, are required in DNN. Batching trick also takes time in computation
Lagrange functions are used, then physical boundary constraints on in training process. Therefore, mini-batch [49] has been utilized to
design variable in the interval [0,1] may be violated. Because they are supersede batching. It helps DNN to increase convergence rate and
pointwise negative over some parts of a given domain. Goupee et al. escape local solutions. Moreover, computational time is well enhanced
[39–41] used piecewise Hermite polynomial functions to define volume by GPU tool in DNN training process. From the above studies, it can be
fraction of FG structures. In this work, material distribution is C1-con- seen that DNN applications to computational mechanics and structural
tinuity for the entire domain. Similar to piecewise Lagrangian func- engineering fields have been limited, specially for FG plates. This paper
tions, cubic Hermite splines are negative over some parts of a given hence fills this gap by proposing DNN for predicting behaviors of FG
domain. Therefore, necessary and sufficient conditions are demanded to plates. In which, input-output data pairs prepared for training are
tackle this issue. Recently, Lieu et al. [42,43] successfully applied B- created by the combination of IGA and HSDT for analysis. The combi-
spline basis functions to describe material distribution of FG plates. In nation of IGA and HSDT is applied to ensure the accuracy of dataset.
which, distribution of ceramic volume fraction is controlled by chan- Recently, optimal design of FG plates has attracted substantial at-
ging coordinates of control points along the thickness plate. It is worth tention from researchers as in studies [50–53]. The optimization algo-
noting that these functions are always pointwise positive in the whole rithms are usually divided into two categories: gradient- and non-gra-
domain; therefore, physical boundary constraints are not violated dient-based methods. Gradient-based method includes several typical
without any conditions. Moreover, a good approximation can be algorithms such as optimality criterion (OC) [54], force method [55],
achieved with a low number of DOFs. Consequently, this approach is sequential linear programming (SLP) [56], sequential quadratic pro-
utilized to describe material distribution of FG plates in this work. gramming (SQP) [57], etc. The convergence speed of the first group is
From the above-mentioned discussions, it is easily seen that IGA is very fast; however, for searching optimal solution, sensitivity analyses
suitable for free vibration and buckling analysis of FG plates. However, of the objective function and constraints are compulsory steps. More-
in line with the “no free lunch” theorem, such methods entail more over, performance of such analyses is fairly complex and expensive. In
computational cost for numerical analysis. Specially, in optimization addition, these algorithms generally depend on a suitable starting point
problems, a large number of structural analyses are required to find a which is given by users. In the highly nonlinear problems with many
globally optimal solution; therefore, numerical analysis of the plates design variables, choosing that kind of point is not easy. From above-
based on IGA must be executed many times. To overcome this short- mentioned reasons, these algorithms are often stuck in local solutions.
coming, DNN is proposed to quickly predict behavior of the FG plates. To circumvent theses disadvantages, algorithms in the second group
DNN is used to supersede FEA in optimal process to significantly save such as genetic algorithm (GA) [58], differential evolution (DE) [59],
time for calculation, but still ensuring the quality of solution. particle swarm optimization (PSO) [60], symbiotic organism search
Artificial neural network (ANN) is a computational model inspired (SOS) [61], flower pollination algorithm (FPA) [62], several improved
by architecture of biological neural networks and which is intended to versions of non-gradient-based algorithms [63–68] have been devel-
simulate the way that humans learn from given data. In most cases, oped. In this group, potential candidates are randomly generated in the
ANN includes a hidden, input and output layers in which each layer searching process without using gradient informations; therefore, these
consists of units. Each connection between units will transmit in- algorithms easily reach global solution. However, they have a drawback
formation or signal from one to another. To start training process in that usually get highly computational cost, particular for optimization
ANN, a dataset including a lot of input-output data pairs is required. problems with many design variables. SOS algorithm, which is inspired
From these data pairs, an optimal mapping approximating relationship by symbiotic interactions between organisms in the nature, is first in-
between input and output data is defined. Therefore, with a given troduced by Cheng et al. [61] for benchmark problems. SOS is not only
input, ANN is able directly predict corresponding output data by using simple by mathematical operations in three phases but also does not
the defined mapping. With limitation of the number of hidden layer, require specific algorithm parameters. From that, this algorithm be-
traditional ANN methods only successfully apply model problems that comes general and robust. Therefore, a large number of studies have
final value is a simple function or the relationship between input and applied SOS to solve optimization problems as Refs. [69–76]. However,
301
similar to other non-gradient-based methods, a major limitation of SOS

is computational cost. Consequently, the authors proposed mSOS [77]
with five modifications in three phases to reduce computational cost
but still guarantee the accuracy of solution. This algorithm successfully
applied to pin-jointed optimization problems with both discrete and
continuous design variables. Moreover, its effectiveness and robustness
have completely proven in terms of convergence rate and quality of
solution. Thus, mSOS will be extended for material optimization of FG
plates in this work.
Based on the above investigations, it can be seen that both DNN and
mSOS have not been applied for optimizing the material distribution of
the FG plates yet. For this reason, a combination of DNN and mSOS is
first proposed for these optimization problems to improve the compu-
tational cost and enhance the adaptability. In the procedure, DNN is
implemented to quickly predict the behavior of FG plates without using Fig. 1. The functionally graded plate model.
any FEA tools. In which, solutions are estimated by the optimal map-
ping obtained by learning relationship between input and output data n
pairs of a given dataset. Each of dataset is randomly generated from FG Vc (ξ ) = ∑ Bip (ξ ) Vc,i (z ), ξ ∈ [0,1],
i=1 (4)
plate analysis through iterations by using IGA based on HSDT to ensure
the accuracy of solution. Moreover, the superiorities of B-spline basis
function are utilized to describe the distribution of ceramic in the plate. Vm (ξ ) + Vc (ξ ) = 1, (5)
In order to reach the global solution, mSOS as an optimizer is proposed
where Vc (ξ ) and Vm (ξ ) are ceramic and metal volume fractions, re-
to reduce computational time. In the present study, objective function is
spectively; Vc, i (z ) is ceramic volume fraction at the ith control point; n is
maximized the buckling load and natural frequency for eigenvalue
the number of control points; Bip (ξ ) is the ith B-spline basic function.
problems. The locations of control points are considered as continuous
These functions are not only good at approximating properties with a
design variables in both problems. The robustness and reliability of the
low number of DOFs but also pointwise non-negative over the entire
proposed methods are proven through several examples in terms of the
domain, and thus physical boundary constraints on interpolated volume
accuracy and computational cost.
fraction always lie in the range [0,1] (Ref. [42]).
In this work, the Mori-Tanaka technique [20] is used to estimate the
2. Isogeometric analysis for FG plates effective properties of the FGM. The effective bulk and shear modulus
are determined as follows:
2.1. B-splines and NURBS
K e − Km Vc
= K −K
,
p K c − Km 1 + Vm K +c 4 / 3mμ (6)
The B-spline basis functions Bi (ξ ) (p = 0) are determined by the m m
Cox-de Boor recursion formula [78] as follows
μe − μm Vc
1, if ξi ≤ ξ < ξi + 1, = μ −μ ,
Bi0 (ξ ) = ⎧ μc − μm 1 + Vm μc + fm
⎨ m 1 (7)
⎩ 0, otherwise. (1)
μ (9K + 8μ )
where f1 = m6(K m+ 2μ m) .
In general, the B-spline basis functions (p ≥ 1) can be defined as m m
From Eqs. (6) and (7), the effective Young's modulus and Poisson's
ξ − ξi ξi + p + 1 − ξ ratio are defined as follows
Bip (ξ ) = Bip − 1 (ξ ) + Bip+−1 1 (ξ ),
ξi + p − ξi ξi + p + 1 − ξi + 1 (2) 9K e μe 3K e − 2μe
E= ,ν= .
3K e + μe 2(3K e + μe ) (8)
where ξi ∈  is the ith knot with i = 1,2, …, n + p + 1; p is the poly-
nomial order and n is the number of basis functions or control points
which are used to describe the B-spline curve.
The NURBS basis functions for a NURBS surface are described as 2.3. Governing equation
Bi B jq ωi, j
p
According to third-order shear deformation theory [80], the dis-
Ni,pj, q (ξ , η) = n m ,
∑iˆ = 1 ∑ jˆ = 1 Biˆp B jqˆ ωiˆ, jˆ (3) placement field at any point of the plate is described as follow
4 3
where Bip (ξ ) and B jq (η) are the B-spline basis functions of order p and q, u (x , y, z ) = u 0 + zβx −
3h2
z (βx + w , x ),
corresponding to the knot vectors Ξ = {ξ1, ξ2, …, ξn + p + 1} and
H = {η1, η2, …, ηj , …, ηm + q + 1} , respectively; ωi denotes the weight referring
v (x , y, z ) = v0 + zβy −
4 3
3h2
z (βy + w , y ), ( −h
2
≤z≤
h
2 ),
to the ith control point. w (x , y ) = w0, (9)
T
where u 0 = {u 0 v0 }T are the membrane displacements; β = { βx βy }
2.2. Functionally graded materials are the rotations in the x − y , y − z planes, respectively; and w0 is the
deflection on the mid-plane of the plate.
In this study, functionally graded plate made of ceramic and metal The strain-displacement relations are of the forms
phases is considered as shown in Fig. 1. The material distribution of the
FG plate is symbolized by the ceramic volume fractions at control points ε = [ εxx εyy γxy ]T = ε0 + zκ1 + z 3κ2,
in which their locations are determined through the plate thickness by γ = [ γxz γyz ]T = εs + z 2κs, (10)
Greville abscissae [79]. The ceramic and metal volume fractions of the
FGM are determined based on B-spline basic functions as follows: (
where c =
−4
3h2 )
302
T
u 0, x ⎡ βx , x ⎤ ⎡ βx , x + w0, xx ⎤ where qI = {u 0I v0I w0I βxI βyI } is the DOFs vector at the Ith control
⎡ ⎤ ⎢ β ⎥, κ = c ⎢ ⎥, point.
ε0 = ⎢ v0, y ⎥, κ1 = y, y 2 βy, y + w0, yy
⎢ ⎥ ⎢ ⎥ Substituting Eq. (20) into Eq. (11), the element strains can be re-
⎢u 0, y + v0, x ⎥ ⎢ βx , y + βy, x ⎥ ⎢ βx , y + βy, x + 2w0, xy ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ written as follows:
⎡ βx + w0, x ⎤ ⎡ βx + w0, x ⎤ m×n
εs = ⎢ , κ = 3c ⎢ .
β + w0, y ⎥ s β + w0, y ⎥ { ε0 κ1 κ2 εs κs }T = ∑ { Bm b1
I BI BbI 2 BsI1 BsI 2 } qI ,
⎣ y ⎦ ⎣ y ⎦ I=1 (21)
(11)
in which
A weak form of the dynamic model for the plates can be expressed
as 0 0 0 0⎤
⎡ NI , x ⎡ 0 0 0 NI , x 0 ⎤
BmI =
⎢ 0 NI , y 0 0 0 ⎥, BbI 1 = ⎢ 0 0 0 0 NI , y ⎥,
∫Ω ¯ εdΩ
δεT D + ∫Ω ¯ s γdΩ
δγT D = ∫Ω ˜ T mu
δu ˜ dΩ, (12) ⎢
NI , x 0 0 0 ⎦
⎥ ⎢ ⎥
⎣ NI , y ⎣ 0 0 0 NI , y NI , x ⎦
where
⎡0 0 NI , xx NI , x 0 ⎤
A B E⎤ BbI 2 = ⎢ 0 0 NI , yy 0 NI , y ⎥,
¯ =⎡
S S ⎢ ⎥
¯ s = ⎡ A B ⎤,
D ⎢ B D F ⎥, D
⎢ BS D S ⎥ ⎣0 0 2NI , xy NI , y NI , x ⎦
⎣ E F H⎦ ⎣ ⎦ (13)
0 0 NI , x NI 0 ⎤ 0 0 NI , x NI 0 ⎤
BsI1 = ⎡ , BsI 2 = 3c ⎡
in which ⎢ 0 0 NI , y 0 NI ⎥ ⎢ 0 0 NI , y 0 NI ⎥
⎣ ⎦ ⎣ ⎦ (22)
h /2
(A, B, D, E, F, H) = ∫ −h /2
Q (1, z , z 2, z 3, z 4 , z 6) dz , Substituting Eq. (21) into Eqs. (12) and (19), the weak forms of
(AS, BS , DS ) =∫
h /2
G (1, z 2, z 4 ) dz , eigenvalue problems can be expressed as
−h /2 (14)
(K − ω2 M) d = 0, (23)
and the material matrices are described as
(K − λ cr K g) d = 0, (24)
⎡ 1 ν (z ) 0 ⎤
E (z )
Q= ⎢ ν (z ) 1 0 ⎥,
2
1 − ν (z ) ⎢ ⎥ where the global stiffness matrix K is given by
⎣ 0 0 (1 − ν (z ))/2 ⎦
m T m
⎧ B ⎫ ⎡ A B E ⎤ ⎧ B ⎫ ⎧ Bs1 ⎫ ⎡ As Bs ⎤ ⎧ Bs1 ⎫
T
G=
E (z ) ⎡1 0⎤
, K= ∫Ω B b1 B D F⎥ B 1 +
b dΩ,
2(1 + ν (z )) ⎢ 0 1⎥ ⎢
⎨ b2 ⎬ E F B ⎨ b2 ⎬ ⎨ s
B ⎬2 ⎢ Bs D s ⎥ ⎨ Bs2 ⎬
⎣ ⎦ (15) B
⎩ ⎭ ⎣ ⎦ B
⎩ ⎭ ⎩ ⎭ ⎣ ⎦ ⎩ ⎭ (25)
and the mass matrix is expressed as and the global mass matrix M is given as
⎡ I0 0 0 ⎤ M= ∫Ω N˜ T mN˜ dΩ,
m = ⎢ 0 I0 0 ⎥ , (26)
⎢ 0 0 I0 ⎥ (16)
⎣ ⎦ in which
where
⎧ N1 ⎫ ⎡ NI 0 0 0 0⎤
˜ = N2 ,
N N1 = ⎢ 0 NI 0 0 0 ⎥,
⎡ I1 I2 cI4 ⎤ ⎨ ⎬ ⎢ 0 0 NI
I0 = ⎢ I2 I3 cI5 ⎥, (I1, I2, I3, I4 , I5, I7) ⎩ N3 ⎭ ⎣ 0 0⎥⎦
⎢ cI cI c 2I ⎥
⎣ 4 5 7⎦ 0 0 0 NI 0 ⎤ ⎡ 0 0 NI , x NI 0 ⎤
⎡
h /2 N2 = ⎢ 0 0 0 0 NI ⎥, N3 = ⎢ 0 0 NI , y 0 NI ⎥,
= ∫ −h/2 ρ (z )(1, z, z2, z 3, z 4, z7) dz (17) ⎢
⎣0 0 0 0 0 ⎥
⎢ ⎥
⎦ ⎣0 0 0 0 0⎦ (27)
and
and geometric stiffness matrix is expressed as
u0 v0
u ⎧ ⎫ w
⎧ 1⎫
˜ = u2 ,
u in which u1 =
⎧
βx
⎫
, u2 = βy ⎧ 0⎫
, u3 = 0 .
Kg = ∫Ω (Bg)T N0BgdΩ, (28)
⎨ u3 ⎬ ⎨ ⎬ ⎨ ⎬ ⎨0⎬
⎩ ⎭ ⎩ βx + w0, x ⎭ β + w0, y
⎩ y ⎭
⎩ ⎭ where
(18)
0 0 NI , x 0 0 ⎤
For the buckling analysis, a weak form of the plate under in-plane B gI = ⎡ ,
⎢ 0 0 NI , y 0 0 ⎥ (29)
force is given as ⎣ ⎦
and ω, λ cr ∈  are natural frequency and critical buckling load, re-
∫ δεTDε
¯ d Ω+ ∫ δγT D
¯ sγd Ω+ ∫ ∇T δw0 N0 ∇w0 d Ω= 0,
spectively.
Ω Ω Ω (19)
0 0
⎡ N x N xy ⎤ 3. Deep neural network
here ∇T = [ ∂/ ∂x ∂/ ∂y ]T is the gradient operation and N0 = ⎢ 0 ⎥
N N y0
⎣ xy ⎦
is a matrix regarded to the pre-buckling loads. A feed-forward neural network as an ANN is shown in Fig. 2
wherein connections between the units do not form a cycle. This net-
2.4. Isogeometric analysis work consists of N layers, the information transfer in only one direction
from input layer, through (N − 1) hidden layers and to Nth output
The generalized displacement u of the eth NURBS element is ap- layer. In the neighboring layers, every two-unit has a connection, to
proximated as which a corresponding connection weight is added. Each unit in the
m×n present layer receives a sum of output values of the units in previous
u (ξ , η) = ∑ NI (ξ , η) qI , layer multiplied with corresponding connection weights and provides
I=1 (20) output value of the activation function for the sum as follows
303
Fig. 2. Feed forward neural network structure.
Table 1 stochastic optimization. This method only needs first-order gradients

Material properties of the FG plate. with little memory requirement. Moreover, its advantages are com-
Al SUS304 Si3N4 Al2O3 bined ones of two currently well-known methods: AdaGrad [83] and
RMSProp [84]. Specifically, Adam method consists of advantages [82]
E(GPa) 70 201.04 348.43 380 as follow:
ν 0.3 0.3 0.3 0.3
• The magnitudes of parameter updates are invariant to rescaling of

ρ(kg/m3) 2707 8166 2370 3800
the gradient.
Table 2 • Its stepsizes are approximately bounded by the stepsize hyper-
Non-dimensional buckling load of SSSS Al/Al2O3 square plate (a/ h = 10) . parameter.
• It does not require a stationary objective.
n RPT-1 [28] RPT-2 [92] RPT-3 [93] Present
• It works with spare gradients.
1 8.0851 8.0946 8.0807 8.0806 • It performs a form of step size annealing naturally.
2 6.8126 6.8208 6.8095 6.8108
5 5.8461 5.8554 5.8423 5.8431 Firstly, biased first moment estimate and second raw moment esti-
mate are required to update by using equations as below
np − 1 mt = β1 mt − 1 + (1 − β1) gt , (33)
Ojp = f (Ujp) = f ⎛⎜ ∑ wjip − 1 × Oip − 1 + θjp⎞
⎟,
vt = β2 vt − 1 + (1 − β2) gt2 , (34)
⎝ i=1 ⎠ (30)
where Ujp and Ojp

are the input and output values of activation function in which mt and vt designate exponential moving averages of gradient
of the jth unit in the pth layer, respectively; wjip − 1 is connection weight and squared gradient, respectively; hyper-parameters β1, β2 ∈ [0,1]
between the ith unit in the (p − 1) th layer and the jth unit in the pth control exponential decay rates of these moving averages; gt is the
layer; θjp is the bias of the jth unit in the pth layer; f is the activation vector of partial derivatives of gradient of loss function (∇Et ) with re-
function chosen for each problem type. gard to weight function set evaluated at step t.
In this study, an identity mapping is chosen for activation function Then, the bias-corrected first and second raw moment estimates are
in regression problem. Herein, the simplest non-linear and widely-used calculated as follows:
one is the Rectified Linear Unit (ReLU) utilized for training in neural mt
ˆt =
m ,
network. The ReLU function can be written as follows 1 − β1t (35)
f (U ) = max(U , 0). (31) vt
vˆt = .
A training pattern is a pair of the input data and the corresponding 1 − β2t (36)
output data. From an input data in the training pattern, a corresponding
The weight functions can be updated by the Adam method as:
target data will be predicted. Such target data are compared with the
corresponding output ones to calculate the value of some pre- η
Wt + 1 = Wt − ˆ t,
m
determined error functions. In those error functions, mean square error vˆt + ε (37)
(MSE) is one of the most generally utilized process for evaluating re-
where W is the weight matrix; η is a learning rate indicating a step size
gression problems [81] and is defined as follows
and ε is taken to be constant 10−8 .
n
1 Batch gradient descent (BGD) computes error for each example in
EMSE =
n
∑ (Oj − Tj)2, the training datdaset. However, it only updates the model after all
j=1 (32)
training examples have been evaluated i.e it updates at end of each
where n is the result of number of training elements multiplied by training epoch. Therefore, the additional complexity of accumulating
number of output neurons; Oj is the output data in the training pattern; prediction errors across all training examples is required. And some
Tj is the target data predicted from input data in the training pattern. restrictions still exist such as memory, training speed and stability.
In the training phase, a loss function as in Eq. (32) is minimized by Therefore, in order to overcome above-mentioned shortcomings, mini-
an optimizer of gradient descent method to find optimal weight para- batch gradient descent (mBGD) have been proposed [49]. This method
meter. In multi-layer networks, back-propagation, one of the most aims to split the training datdaset into small batches to calculate model
popular methods, is used to compute the effect of each weight with error and update model coefficients [85]. mBGD is created from bal-
regard to the loss function. Recently, adaptive moment estimation ance between efficiency of BGD and robustness of the stochastic gra-
(Adam) firstly proposed by Kingma et al. [82] is a method for powerful dient descent. Some advantages of mBGD are found. Particularly, the
304
Fig. 3. Data sets of SSSS Al/Al2O3 square plate under uni-axial compression.
Table 3
Mean square error for training and test sets with various activation functions and optimizers.
Optimizer Adam RMSprop Adagrad SGD Adadelta
Error (10−6) Training Test Training Test Training Test Training Test Training Test
Linear 226.8837 230.0037 305.2928 294.614 227.3904 226.7866 221.4227 223.4005 228.1582 231.2314
Sigmoid 2.9146 2.8558 7.8125 7.5044 112.6907 111.0888 275.3472 284.2783 58.1913 54.3531
Softmax 0.8121 0.8937 2.5682 2.492 56.2627 53.1702 14500.79 14795.82 14394.06 14693.87
Softplus 1.0777 1.0229 19.1361 19.3227 117.246 116.0133 211.8015 216.693 13.8106 12.8662
Tanh 21.1124 20.9497 19.5649 18.8004 23.6053 2.8373 226.2581 228.507 25.3934 23.3518
ReLU 0.4677 0.5116 13.6763 13.7491 1.9634 2.2834 65.3821 63.796 3.6131 3.5606
effectiveness of this approach is shown in the memory without all be found. With the limited training data, these complicated relations
training data because the frequency updated by this model is higher are caused by noise samples existing in the training set but not in real
than BGD. mBGD may converge faster and escape local solutions. It test data; therefore, overfitting may occur in the training process. To
provides a more effective computational process than the stochastic prevent this issue, dropout technique being an abstract of dropping out
gradient descent. For that reason, mDGB is utilized in this study. units in a neural network is proposed by Srivastava et al. [48]. In this
Deep neural networks include multiple nonlinear hidden layers; technique, some randomly selected units with all their incoming and
therefore, the complex relations between their inputs and outputs can outcoming connections are temporarily removed from the network to
305
Fig. 4. The convergence history of loss function of FG plate under uni-axial compression in the sym-5cpts case.
Table 4
Optimal results of SSSS Al/Al2O3 square plate under uni-axial compression obtained by different optimization algorithms for various volume constraints.
Design variables f1 = 35% f1 = 50% f1 = 65%
DE SOS mSOS DE SOS mSOS DE SOS mSOS
1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 1 1
3 0.4003 0.4003 0.4003 1 1 1 1 1 1
4 0 0 0 0.0001 0.0001 0.0001 0.4502 0.4502 0.4502
5 0 0 0 0 0 0 0 0 0
Best value 11.2171 11.2171 11.2171 14.6997 14.6997 14.6997 16.2238 16.2238 16.2238
Worst value 11.2171 11.2171 11.2171 14.6997 14.6997 14.6997 16.2238 16.2238 16.2238
Mean value 11.2171 11.2171 11.2171 14.6997 14.6997 14.6997 16.2238 16.2238 16.2238
Standard deviation 0 0 0 0 0 0 0 0 0
Average numbers of analysis 10650 13000 2175 7870 12300 990 9010 13400 2070
reduce overfitting. In this study, dropout technique is also used to im- number of control points.
prove the performance of neural nets.
4.1.2. Maximization of frequency of functionally graded plate
4. Optimization procedure This subsection presents the optimal material distribution for a FG
plate subjected to frequency constraints. In which, objective function is
4.1. Problem statement to maximize the first frequency. The mathematical statements of the
optimization problem are of the form
In this study, two eigenvalue problems are investigated. The natural
Maximize ω¯ (Vc, i )
frequency and buckling load are maximized in the controlled range of
∫ −hh/2/2 Vc dz
total volume fraction of ceramic. Subject to g2 = − f2 ≤ 0
h
0 ≤ Vc, i ≤ 1, i = 1,2, …, n (39)
4.1.1. Maximization of buckling loads of functionally graded plate
The problem aims to maximize the critical load of a FG plate that all where ω̄ and f2 stand for the non-dimensional natural frequency and
constraints on design variables are satisfied. The design variables are total ceramic ratio, respectively. In which, f2 is an user-defined aspect.
locations of control points that used to describe material distribution To tackle constraints, non-death penalty method, one of the most
through-the-thickness. This problem can mathematically be expressed popular constraint-handling approach [86], is used. As a result, the
as problems can now be rewritten as
nin 2
Maximize P¯cr (Vc, i ) Maximize fpenalty (Vc, i ) = f (Vc, i ) + λ ∑ [max(0, gj )] ,
j=1
∫ −hh/2/2 Vc dz neq 2
Subject to g1 = h
− f1 ≤ 0 + λ ∑k = 1 [max(0, |hk | − ε )] (40)
0 ≤ Vc, i ≤ 1, i = 1,2, …, n (38)
in which f is the critical buckling load or the non-dimensional natural
where P̄cr is the critical buckling load; Vc, i is ceramic volume fraction at frequency; gj and hk are the jth inequality and kth equality constraints,
the ith control point; h is thickness of plate; f1 is a desired ceramic; n is respectively; nin and neq are the number of inequality and equality
306
Fig. 5. Convergence history of the FG plate obtained by the DE, SOS and mSOS.
Table 5
Optimal results of SSSS Al/Al2O3 square plate under uni-axial compression in the sym-5cpts case obtained by different methods.
DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS
1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 0.392 0.4003 0.988 1 1 1
4 0 0 0 0.0001 0.442 0.450
5 0 0 0 0 0 0
P̄cr Prediction 11.162 11.217 14.590 14.700 16.057 16.224
IGA 11.185 14.603 16.193
Error 0.206% 0.490% 0.089% 0.748% 0.840% 1.029%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.348 0.497 0.647
Error 0.571% 0.000% 0.600% 0.000% 0.462% 0.000%
Time (seconds) 9 4941 7 4172 10 4414
307
Fig. 6. Optimal material distribution of SSSS Al/Al2O3 square plate under uni-axial compression in the sym-5cpts case obtained by different methods for various
volume constraints.
308
Table 6
1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 1 1 1 1 1 1
4 0.248 0.250 1 1 1 1
5 0 0 0.006 0.001 0.756 0.750
6 0 0 0 0 0 0
7 0 0 0 0 0 0
P̄cr Prediction 12.057 12.241 15.437 15.757 16.915 17.510
IGA 12.236 15.769 17.531
Error 1.463% 1.503% 2.105% 2.031% 3.514% 3.398%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.350 0.501 0.651
Error 0.000% 0.000% 0.200% 0.000% 0.154% 0.000%
Time (seconds) 6 6378 16 4618 14 4748
constraints, respectively; λ is penalty parameter and chosen as 105; ε is a 4.2.1. Mutualism phase
small positive tolerance which is chosen to be 10−4 in this study. In the mutualism phase, two interacting organisms receive benefit
from the relationship. Xi and Xj symbolize the ith and jth organisms of
the ecosystem, respectively. In the ecosystem, Xi interacts with Xj being
4.2. The symbiosis organisms search algorithm a randomly selected organism to create new candidate organisms as
follows
The symbiotic organisms search algorithm, first proposed by Cheng
Xinew = Xi + rand (0,1) × (X best − Mutual_Vector × BF 1), (41)
et al. [61], is a simple and effective metaheuristic algorithm. The SOS
has effectively and robustly been demonstrated for both benchmark and
Xjnew = Xj + rand (0,1) × (X best − Mutual_Vector × BF 2), (42)
real-word problems [61,75,87–89]. The SOS algorithm is inspired by
the symbiotic interactions between organisms in an ecosystem where X i + Xj
each organism represents a potential solution in a search domain. Mutual_Vector = ,
2 (43)
Symbiotic relation is defined as reliance-based relationship among or-
ganism in order to satisfy their sustenance or even survive in an eco- where rand (0,1) is a function that returns an uniformly distributed
system. The most common symbiotic relationships such as mutualism, random number in the range [0,1]; X best is the best organism in the
commensalism and parasitism are found in the nature. The SOS algo- ecosystem symbolizing the highest degree of adaption; the benefit
rithm initiates with a randomly created population comprising N or- factors BF1 and BF2 randomly created either 1 or 2 symbolize the level
ganisms in the ecosystem. Then, population is updated through mutu- of benefit to each organism, and Mutual_Vector expresses the re-
alism, commensalism and parasitism phases after each generation. lationship characteristic between two organisms Xi and Xj .
After that, the fitnesses of Xinew and Xjnew are compared with those
of Xi and Xj organisms to select the most suitable organism in each pair
309
volume constraints.
310
Table 7
1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 1 1 1 1 1 1
4 0.954 0.951 1 1 1 1
5 0 0 1 1 1 1
6 0 0 0 0 1 1
7 0 0 0.002 0 0.05 0.049
8 0 0 0 0 0 0
9 0 0 0 0 0 0.002
P̄cr Prediction 13.174 13.125 15.835 16.132 17.585 18.146
IGA 13.114 16.138 18.152
Error 0.455% 0.372% 1.878% 1.841% 3.124% 3.092%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.351 0.500 0.650
Error 0.285% 0.000% 0.000% 0.000% 0.000% 0.000%
Time (seconds) 22 7837 20 7066 31 7283
which is going to go the next phase, respectively. duplication Parasite_Vector which symbolizes a parasite organism.
Then, a new candidate organism is generated by modifying some
4.2.2. Commensalism phase components of the original Parasite_Vector . A randomly selected or-
The commensalism takes place when an organism receives benefits ganism Xj from the ecosystem representing a host of the parasite is
from the relationship while the other is not affected. The interaction compared to Parasite_Vector via objective function value to gain the
between two distinct organisms Xi and Xj is performed to generate a best one for the next generation.
new organism based on the best organism. In this phase, only an or-
ganism gets benefits from the interaction; therefore, only Xinew is newly 4.3. The modified symbiotic organisms search algorithm
generated as the following relation
Xinew = Xi + rand (−1,1) × (X best − Xj), (44) The balance between local and global search abilities plays a sig-
nificant role in the success of the population-based metaheuristic al-
where rand (−1,1) is an uniformly distributed random number in the gorithms as designated by Das et al. [90]. In which, the local search
interval [−1,1]. helps to improve convergence rate or computational cost but it is often
Finally, a better organism in the pair of Xi and Xinew is chosen based trapped at local optimum solutions. Conversely, the global search is
on their fitnesses to go to the next phase. good at exploring global optimum solution but it requires a large
amount of computational cost in the searching process. In the original
4.2.3. Parasitism phase SOS, both the mutualism and commensalism phases enhance the local
The parasitism takes place when only one organism gets benefits search ability of the algorithm while the parasitism phase improve the
while the other is harmed. Firstly, the ith organism Xi creates a global search ability. However, the parasitism phase requires a large
311
volume constraints.
312
Fig. 11. The convergence history of loss function of the asymmetric FG plate under uni-axial compression.
Table 8
Optimal results of SSSS Al/Al2O3 square plate under uni-axial compression in the asym-13cpts case obtained by different methods.
1 0 0 0 0 0 0
2 1 1 1 1 1 1
3 1 1 1 1 1 1
4 0.7518 0.7512 1 1 1 1
5 0 0 0.2479 0.2505 1 1
6 0 0 0 0 0 0
7 0 0 0 0 0 0
8 0 0 0 0 0 0
9 0 0 0 0 0.7487 0.7507
10 0 0 1 1 1 1
11 1 1 1 1 1 1
12 1 1 1 1 1 1
13 1 1 1 1 1 1
P̄cr Prediction 11.3142 11.4284 14.1386 14.4604 15.8096 16.4255
IGA 11.4313 14.4577 16.4231
Error 1.024% 0.999% 2.207% 2.225% 3.736% 3.750%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.3502 0.5 0.6499
Error 0.057% 0.000% 0.000% 0.000% 0.000% 0.000%
Time (seconds) 38 8445 45 10855 43 9875
amount of computational cost. If the last phase is removed to save time, Xinew = Xi + rand (0,1) × (X k − Mutual_Vector × BF 1), (45)
the SOS is easily stuck into local solutions. For this reason, the others
should be adjusted to guarantee the effectively balance between local and
and global abilities. To tackle these shortcoming, Do et al. [77] pro-
posed a newly modified version of SOS called mSOS for pin-jointed Xjnew = Xj + rand (0,1) × (X k − Mutual_Vector × BF 2). (46)
structures. The mSOS is recommended five modifications on three
phases of SOS. Moreover, unlike most of metaheuristic algorithms, the If an organism gets full benefit from the interaction, e.g the benefit
SOS and mSOS not only simple but also do not require any specification value is 2, the terms of Mutual_Vector × BF 1 and
of algorithmic parameters. Mutual_Vector × BF 2 also (X k − Mutual_Vector × BF 1) and
(X k − Mutual_Vector × BF 2) in Eqs. (45) and (46) will increase di-
4.3.1. Modification of the mutualism phase versity of the organisms. For this reason, new candidate organisms may
The modified mutualism phase is developed to reinforce the global be out of feasible regions. This leads to deviation of the fitness between
ability in the searching optimal process. Accordingly, the best organism the best organism and the entire domain also decrease the stability of
X best in Eqs. (41) and (42) is replaced by a randomly chosen one X k to the algorithm. Moreover, a larger computational cost is required to
extend search space and increase diversity of the ecosystem. Candidate reach the global solution. As a result, each organism will be get a part of
organisms can be written as benefit from the interaction, i.e the benefit value is 1.
313
Fig. 12. Optimal material distribution of SSSS Al/Al2O3 square plate under uni-axial compression in asym-13cpts case obtained by different methods for various
volume constraints.
314
Table 9 5. Numerical examples

Comparison of the non-dimensional buckling load of SSSS Al/Al2O3 square
plate based on optimal design and power-law distribution for the same volume In this study, two eigenvalue problems are investigated to demon-
constraints. strate the effectiveness and robustness of the proposed method. In
Buckling load Model No. control f1 which, an Al/Al2O3 square plate under uni-axial compression and a
points SUS304/Si3N4 square plate under free vibration are considered.
35% 50% 65% Material properties of these FG plates are shown in Table 1. The
properties of FGMs are estimated by Mori-Tanaka scheme. First of all,
P̄crpl − − 6.9384 8.0806 9.7719
behavior of the FG plates is analyzed by using IGA with HSDT. Then,
P̄crop Symmetric FG 5cpts 11.162 14.590 16.057
plate datasets containing input-output data pairs are randomly created from
7cpts 12.057 15.437 16.915 the above behavior analyses. By using these data, DNN starts training to
9cpts 13.174 15.835 17.585 explore the mapping rules over them. Therefore, with a given input
Asymmetric FG 13cpts 11.314 14.138 15.810 data, corresponding output data are directly predicted by the obtained
plate
mapping. In the training phase, the influence of activation functions
pl based on the power law volume fraction distribution. and optimizers on the loss value and computational time are executed
op
based on the optimal volume fraction distribution. for the first example. From acquired results, an appropriate activation
function and optimizer will be utilized in all subsequent examples. In all
Table 10 examples, 10,000 input-output data pairs in a dataset are learned after
Non-dimensional natural frequency of SSSS SUS304/Si3N4 square plate 1000 epochs to gain the optimal mapping. Note that, the dataset is
(a/ h = 10) . divided into two groups: training and test sets corresponding 8000 and
2000 data pairs. In which, the training dataset is used for learning and
Mode n RPT-1 [28] RPT-2 [92] RPT-3 [93] Quasi-3D [93] Present
the test dataset is used only to assess the performance, i.e. general-
1 1 0.0542 0.0543 0.0542 0.0538 0.0542 ization of a process. If a model fits to the training dataset also fits well
2 0.0485 0.0485 0.0485 0.0480 0.0485 the test dataset, minimal overfitting has taken place. Otherwise, a better
5 0.0438 0.0438 0.0438 0.0433 0.0440
fitting of the training dataset as opposed to the test dataset usually
10 0.0416 0.0416 0.0416 0.0411 0.0419
2 1 0.1294 0.1296 0.1293 0.1259 0.1293 points to overfitting.
2 0.1155 0.1157 0.1154 0.1122 0.1156 Next, optimal material distribution of the plates will be determined
5 0.1043 0.1045 0.1042 0.1010 0.1046 by using optimization algorithms based on the found mapping in the
10 0.0991 0.0993 0.0990 0.0959 0.0997 previous step. In this phase, the original SOS and DE are examined as a
reference to verify the effectiveness of mSOS via statistical informations
such as average value, worst value, best value, standard deviation and
4.3.2. Modification of the commensalism phase
average numbers of structural analyses. The searching process of all
In this phase, two modifications improving the local exploitation
algorithms will be stopped either the relative error between the best
ability are suggested. In the first modification, rand (−1,1) coefficient is
and the average objective function values is less than or equal to the
superseded by rand (0.4, 0.9) one. From Eq. (44), it can be seen that
value of tolerance or the maximum number of structural analyses
rand (−1,1) coefficient considerably affects on the convergence speed.
reaches (MaxEval). In all examples, the values of ecosystem, tolerance
With wide range from −1 to 1, the search space will be extended
and MaxEval are set to be 50, 10−6 and 300,000, respectively. In this
further and this leads to a slow convergence speed. Therefore, a suffi-
study, optimization algorithms are stochastic methods; therefore, each
cient range in [0.4, 0.9] is proposed to reduce computational time but
algorithm will perform ten independent runs for each case. All nu-
still ensure the accuracy. So, Eq. (44) can be rewritten as
merical examples are executed on a desktop computer core i5-4460
Xinew = Xi + rand (0.4, 0.9) × (X best − Xj). CPU of 3.2 GHz clock speed and 8 GB of RAM.
(47)
In the original commensalism phase, the worst organism in each 5.1. Buckling problem
pair Xi and Xinew maybe better than the other ones in the whole present
ecosystem. This selection can cause a slow convergence speed. In the first problem, let us consider a simply supported Al/Al2O3
Therefore, an elitist selection technique [91] is utilized to circumvent square plate with a/ h = 10 . An uniform mesh of 7 × 7 cubic NURBS
these shortcomings. This process is described as follows. Firstly, N new elements is used to describe geometry of the domain. In order to verify
organisms Xinew and N pre-interaction organisms Xi are merged to the present results, non-dimensional critical buckling load P¯cr = Pcr a/ Dm
generate a new ecosystem consisting of 2N organisms. Then, N best of the Al/Al2O3 plate with different volume fraction indexes n obtained
organisms in the new ecosystem are chosen for the next step. Thus, new by IGA are compared with those of other methods in the literature
organisms are inherited good characteristics of the previous genera- [28,92,93] as shown in Table 2. The material distributions are de-
tions. As a result, this contributes to the local search ability. scribed by the power-law function as following form
1 z n
Vc (z ) = ⎛ + ⎞ .
⎝2 h⎠ (48)
4.3.3. Modification of the parasitism phase
The original parasitism phase plays an important role in the global As Table 2 shows, obtained results by the present method are close
exploration ability; however, it requires a high computational cost, to the reference solutions. Therefore, it can be concluded that the
because a new candidate organism Parasite_Vector is created by re- current method is accurate to create datasets for training in DNNs. In
placing some random components of the duplication of the organism Xi each of dataset, input data are the locations of control points through-
with other ones. Then, Parasite_Vector is compared to randomly se- the-thickness plate and output data consist of non-dimensional critical
lected organism in the ecosystem. This leads to a larger search space buckling load and ceramic volume fraction. The training and optimal
and a slow convergence rate. Besides, the effective balance between process will be examined in two models: symmetric and asymmetric
local and global abilities has been guaranteed by the modified mutu- material distributions.
alism and commensalism phases. Accordingly, the parasitism phase is
removed in mSOS to save computational cost but still acquires the 5.1.1. The FG plate with symmetric material distribution
global solution. From a given boundary condition, material distributions of the plate
315
Fig. 13. Data sets of SSSS SUS304/Si3N4 square plate under free vibration.
are symmetric at mid-plane. Therefore, only half of the number of tackled. Hence, the combination of Adam optimizer and ReLU function
control points along thickness plate are considered as design variables. is suggested in all subsequent examples in this study.
In this example, material distribution is described by five, seven or nine Next, the effectiveness and robustness of mSOS are demonstrated by
control points in a half of the plate. comparing to DE and SOS in terms of statistical informations. Obtained
First of all, a symmetric material distribution of the plate defined by results are presented in Table 4 in case of various volume constraints.
five control points (sym-5cpts) is carried out. Each data pair in dataset As seen that the optimal results attained by mSOS agree well with those
consisting of five input data and two output data are randomly gener- gained by DE and SOS. In the first case ( f1 = 35%), mSOS only requires
ated as shown in Fig. 3(a). This figure addresses the correlation be- 2175 analyses while DE and SOS need 10,650 and 13,000 analyses,
tween non-dimensional buckling load and ceramic volume fraction of respectively. In the next case ( f1 = 50%), mSOS uses the number of
the plate. The influence of various activation functions and optimizers analyses less than DE and SOS (990 for mSOS; 7870 for DE and 12,300
on training is investigated. For this investigation, six activation func- analyses for SOS). And in the final case with 65% ceramic volume
tions: Linear, Sigmoid, Softmax, Softplus, Tanh, ReLU and five opti- fraction, mSOS requires 2070 analyses which are less than DE (9010
mizers: Adam, RMSprop, Adagrad, SGD, Adadelta are considered. Ob- analyses) and SOS (13,400 analyses). All algorithms remain stable in all
tained MSE for training and test sets in 5-100-100-2 architecture with cases with a zero standard deviation. mSOS converges much faster than
dropout 0.001 after 1000 epoch are shown in Table 3. This table shows DE and SOS as depicted in Fig. 5. Therefore, it reveals that mSOS sig-
that the lowest MSE for training and test sets (0.4677 × 10−6 and nificantly reduces the computational time but still guarantee the global
0.5116 × 10−6) are achieved when Adam optimizer and ReLU function is optimal solution. Consequently, mSOS will be used in all subsequent
simultaneously applied. Convergence history of loss function of training examples while DE and SOS are not.
and test sets obtained by the optimal combination is depicted in Fig. 4. Optimal material distribution of the Al/Al2O3 plate gained by mSOS
From displayed results, it can be seen that MSE of both training and test based on DNN (DNN-mSOS) are compared to those of combination of
sets gradually converge to zeros. Moreover, overfitting is also properly mSOS and IGA (IGA-mSOS) in all cases. Obtained results are shown in
316
Fig. 14. The convergence history of loss function of FG plate under free vibration in the sym-5cpts case.
Table 11
Optimal results of SSSS SUS304/Si3N4 square plate under free vibration in the sym-5cpts case obtained by different methods.
1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 0.4 0.399 1 1 1 1
4 0 0 0.001 0 0.45 0.45
5 0 0 0 0.001 0 0
ω̄ Prediction 0.0526 0.0526 0.0594 0.0594 0.0662 0.0662
IGA 0.0526 0.0595 0.0662
Error 0.000% 0.000% 0.168% 0.000% 0.000% 0.000%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.350 0.5004 0.650
Error 0.000% 0.000% 0.080% 0.000% 0.000% 0.000%
Time (seconds) 27 8606 19 6989 25 9801
Table 5. From this table, it is easily seen that DNN predicts quite good in all cases, and less than 3.4% for DNN-mSOS and IGA-mSOS. DNN
when the relative errors between DNN and IGA are smaller than 1.0%. takes 440 s for training and DNN-mSOS only spends 6 s for case
Moreover, optimal solutions from DNN-mSOS and IGA-mSOS are very f1 = 35%, 16 s for 50% and 14 s for 65% to find optimal solution while
close to each other (relative errors are less than 1.1%). Furthermore, IGA-mSOS requires 6378, 4618 and 4748 s for cases 35%, 50% and
DNN-mSOS only takes 9 s, 7 s and 10 s in three cases to attain optimal 65%, respectively. Material distributions of the plate are depicted in
solutions and 446 s for training while IGA-mSOS takes more than 4000 s Fig. 8. The figure reveals that optimal distributions obtained by DNN-
to reach the goal. Fig. 6 reveals optimal material distribution of the mSOS are in good agreement with those from IGA-mSOS.
plate obtained by two methods. It is clear that the optimal material In the last case as plotted in Fig. 3(c), optimal mapping is obtained
distribution of two approaches are almost identical. From this figure, it by training for 9-100-100-2 architecture after 1000 epochs. Loss values
can be observed that a major metal tends towards mid-plane while of both sets also converge to zeros as indicated in Fig. 9. And optimal
ceramic has the trend to move elsewhere. weights are found only in 475 s. As expected, both DNN and mSOS save
From the above discussion, it indicates that DNN not only helps to a lot of time as pointed out in Table 7. Moreover, its accuracy is de-
dramatically save computational cost but also still guarantees solution monstrated by comparing to IGA approach as shown in Table 7 and
accuracy. Fig. 10.
Similar to sym-5cpts, a model with 7 (sym-7cpts) or 9 control points
(sym-9cpts) are executed. In the sym-7cpts case, dataset shown in
Fig. 3(b) is trained for 7-100-100-2 architecture with the same previous 5.1.2. The FG plate with asymmetric material distribution (asym-13cpts)
parameters. Fig. 7 indicates that loss values of training and test sets A FG plate where metal is reinforced at the bottom surface and
converge to zeros (5.9843 × 10−7 for training set and 7.3578 × 10−7 for ceramic on the top surface of the plate is considered. Material dis-
test set). Optimal results of the proposed method are compared with tribution of the plate will be described by 13 control points along the
those of IGA-mSOS and summarized in Table 6. As seen that the relative plate thickness. The number of design variables is thus 13. A dataset
errors of optimal solutions between DNN and IGA are smaller than 3.6% consists of 10,000 data pairs created for learning in DNN is shown in
Fig. 3(d). Optimal mapping is defined after 1000 epochs and model loss
317
Fig. 15. Optimal material distribution of SSSS SUS304/Si3N4 square plate under free vibration in the sym-5cpts case obtained by different methods for various
volume constraints.
318
Table 12
1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 1 1 1 1 1 1
4 0.248 0.246 1 1 1 1
5 0 0.003 0.007 0 0.753 0.750
6 0 0.001 0 0 0 0
7 0 0.001 0 0.002 0 0
ω̄ Prediction 0.0531 0.0532 0.0598 0.0599 0.0668 0.0669
IGA 0.0531 0.060 0.0669
Error 0.000% 0.188% 0.333% 0.170% 0.149% 0.149%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.350 0.501 0.651
Error 0.000% 0.000% 0.200% 0.000% 0.154% 0.000%
Time (seconds) 19 15130 19 12030 38 15049
is plotted as in Fig. 11. From this figure, it can be seen that MSE of symmetric model. Besides, a symmetric plate has a buckling capacity
training and test sets gradually converge to zeros after 500 epochs. better than an asymmetric one with the same number of control points.
A comparison of the gained results by DNN-mSOS and IGA-mSOS is
tabulated in Table 8. It can be seen that DNN ensures the accuracy of 5.2. Free vibration problem
solution with a very small error. In addition, it enables DNN to sig-
nificantly reduce computational cost. For instance, DNN requires 461 s A SUS304/Si3N4 square plate with simply supported boundary
to find the optimal mapping. For optimization, DNN-mSOS takes 38, 45 condition is taken into consideration in this section. The geometry of
and 43 s for cases 35%, 50% and 65%, respectively. While IGA-mSOS the plate is given as a/ h = 10 . Table 10 shows the non-dimensional
needs 8,445, 10,855 and 9875 s for the first, second and last cases, natural frequency ω¯ = ωh ρc / Gc of the plate with different volume
respectively. Fig. 12 shows optimal material distributions of the plate in fraction indexes n. The obtained results are compared with those of
case of three different constraints. It is observed that the results of two published papers [28,92,93]. As shown in the table, the present results
methods are almost identical. Moreover, on the bottom surface of the agree well with those of other methods. Therefore, datasets for free
plate, ceramic volume fraction is zero, i.e. metal is reinforced. However, vibration problems are created by this approach. In which, locations of
ceramic has a major contribution on the top surface of the plate. control points are considered as input data; non-dimensional natural
Therefore, optimal designs satisfy all constraints. frequency ω̄ and ceramic volume fraction are considered as output data.
Table 9 shows a comparison of buckling loads of the plate based on Similar to buckling problems, the architecture with two hidden layers is
optimal design and power-law distribution (Eq. (48)) for the same vo- used to find optimal mapping. Datasets for these examples are depicted
lume constraints. From this table, it can be seen that all optimal results in Fig. 13.
are greater than those of power law. Therefore, B-splines are effective
for describing material distribution of the plate. Better optimal designs 5.2.1. The FG plate with symmetric material distribution
are found when increasing the number of control points in the Dataset for sym-5cpts case shown in Fig. 13(a) are used for training
319
volume constraints.
320
Table 13
1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 1 0.999 1 1 1 1
4 0.942 0.949 1 1 1 1
5 0 0 1 0.997 1 1
6 0 0 0.010 0.003 1 1
7 0 0 0 0 0.055 0.047
8 0 0 0 0.001 0 0.001
9 0 0 0 0 0 0.008
ω̄ Prediction 0.0536 0.0536 0.060 0.060 0.0669 0.0671
IGA 0.0535 0.0601 0.0672
Error 0.187% 0.000% 0.166% 0.000% 0.446% 0.298%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.349 0.502 0.651
Error 0.286% 0.000% 0.398% 0.000% 0.154% 0.000%
Time (seconds) 16 17231 29 11349 33 12155
in DNN to get optimal weights. Convergence history of MSE are in- DNN predicts very well because relative errors between gained solu-
dicated in Fig. 14. As expected, DNN converges after 1000 epochs and tions by DNN and IGA are smaller than 0.4% in all cases. Furthermore,
MSE of training and test sets are 4.0162 × 10−7 and 3.9631 × 10−7 , re- material distributions gained by the proposed method agree well with
spectively. To determine optimal mapping, DNN only requires 444 s. those from IGA-mSOS as shown in Fig. 17. Moreover, DNN takes 458 s
Regarding to the accuracy of the DNN's solutions, the obtained re- to find the optimal mapping and DNN-mSOS only requires 19, 19 and
sults are tabulated as shown in Table 11 and plotted in Fig. 15. It is 38 s. Thus, the proposed method helps to dramatically reduce compu-
evident from the table that DNN is able to reach to good results and tational time in this problem faster approximately (31 times for the first
completely identical with those of IGA. Thus, the optimal solutions (ω̄ case, 25 times for the second case and 30 times for the last case).
and vf ) obtained by DNN-mSOS and IGA-mSOS are similar to each In the sym-9cpts case, the effectiveness and reliability of both DNN
other. However, DNN-mSOS saves a lot of time. For example, it only and mSOS are demonstrated via obtained results. For instance, con-
spends 27 s for the first case, 19 and 25 s for the last two cases while vergence of DNN is depicted in Fig. 18 and it needs 470 s to calculate.
IGA-mSOS takes very long (8,606, 6989 and 9801 s for the corre- The accuracy and computational cost of the current method are shown
sponding three cases). Generally, the proposed method offers a very in Table 13 and Fig. 19.
good tool to significantly reduce computational cost.
In the next case with 7 control points, loss model describing con-
5.2.2. The FG plate with asymmetric material distribution (asym-13cpts)
vergence of MSE of training and test sets are depicted as in Fig. 16. It
Similar to buckling problem, a metal-ceramic FG plate is considered.
can be seen that loss values of both training and test sets converge to
Fig. 13(d) depicts a given dataset and its each data pair consists of 13
zeros and stable after 500th epoch. Table 12 summarizes optimal so-
input and two output. Fig. 20 describes convergence history of DNN in
lutions obtained by two approaches. From the table, it can be seen that
which MSE of training and test sets are 3.6396 × 10−7 and 3.7068 × 10−7 ,
321
volume constraints.
322
Fig. 20. The convergence history of loss function of the asymmetric FG plate under free vibration.
Table 14
Optimal results of SSSS SUS304/Si3N4 square plate under free vibration in the asym-13cpts case obtained by different methods.
1 0 0 0 0 0 0
2 1 1 1 1 1 1
3 1 1 1 1 1 1
4 0.7243 0.7485 1 1 1 1
5 0 0 0.2340 0.2494 1 1
6 0 0 0 0 0 0
7 0 0 0 0 0 0
8 0 0 0 0 0 0
9 0 0 0 0 0.7461 0.7491
10 0 0.0013 1 1 1 1
11 1 1 1 1 1 1
12 1 1 1 1 1 1
13 1 1 1 1 1 1
ω̄ Prediction 0.05259 0.05258 0.05887 0.05906 0.06583 0.066
IGA 0.05241 0.05899 0.06599
Error 0.342% 0.019% 0.203% 0.322% 0.242% 0.258%
vf Prediction 0.350 0.350 0.500 0.500 0.650 0.650
IGA 0.3457 0.4984 0.6496
Error 1.229% 0.000% 0.320% 0.000% 0.062% 0.000%
Time (seconds) 55 35389 43 34374 49 42934
respectively. The training process requires 451 s to gained the optimal From Table 15, it can be seen that natural frequencies of the plate
mapping. gained from the optimal results are larger than those directly solved by
The results obtained by the present method are tabulated in the power law with same constraints. Moreover, a FG plate with sym-
Table 14 in comparison with those achieved by IGA-mSOS. Material metric material distribution is stiffer than an asymmetric one.
distributions of the plate with different volume constraints are shown in
Fig. 21. As can be seen from the figure, optimal distributions obtained 6. Conclusions
by DNN-mSOS are in an excellent agreement with those from IGA-
mSOS. The precision of the present method is also ascertained when the A new combination of DNN and mSOS is successfully developed as
relative errors between solutions obtained by DNN and IGA are almost well as applied for optimization of the material distribution of the FG
smaller than 1.3% in all cases. Furthermore, the robustness of the plates. Objective function is to maximize buckling load and natural
present method is proven via computational cost in three cases. For frequency of eigenvalue problems. The locations of control points are
example, DNN-mSOS requires 55, 43 and 49 s for cases 35%, 50% and considered as continuous design variables. DNN is employed to replace
65%, respectively; while IGA-mSOS spends 35,389, 34,374 and FEA for buckling and free vibration analyses of the FG plates. A recently
42,934 s for that cases. From these displayed results, it can be con- developed mSOS is proposed to reach better global solutions for several
cluded that DNN-mSOS method remarkably saves the computational volume constraints. Based on the investigated numerical examples,
cost but still guarantees the accuracy of solution. several conclusions are made as follows:
323
Fig. 21. Optimal material distribution of SSSS SUS304/Si3N4 square plate under free vibration in the asym-13cpts case obtained by different methods for various
volume constraints.
(i) The B-spline basis functions are more suitable and effective than solving eigenvalue problems.
other functions for describing material distribution of the FG (iii) The mSOS outperforms the DE and SOS in terms of computational
plates. cost and the accuracy.
(ii) Based on the optimal mapping which is found through learning the (iv) The combination of DNN and mSOS helps to significantly save
relationship between input and output data, behaviors of the plate computational time for optimization problems in comparison with
can quickly be predicted with high accuracy without directly the combination of IGA and mSOS.
324
Table 15 graded carbon nanotube-reinforced composite shells. Compos Struct

Comparison of the non-dimensional natural frequency of SSSS SUS304/Si3N4 2018;203:349–60.
square plate based on optimal design and power-law distribution for the same [17] Nguyen TN, Thai CH, Nguyen-Xuan H. A novel computational approach for func-
tionally graded isotropic and sandwich plate structures based on a rotation-free
volume constraints. meshfree method. Thin-Walled Struct 2016;107:473–88.
Frequency Model No. control [18] Nguyen TN, Thai CH, Nguyen-Xuan H, Lee J. Geometrically nonlinear analysis of
f2
functionally graded material plates using an improved moving Kriging meshfree
points
method based on a refined plate theory. Compos Struct 2018;193:268–80.
35% 50% 65% [19] Fröhlich A, Sack R. Theory of the rheological properties of dispersions. Proc R Soc
Lond A 1946;185(1003):415–30.
ω̄ pl − − 0.0490 0.0542 0.0609 [20] Mori T, Tanaka K. Average stress in matrix and average elastic energy of materials
ω̄op Symmetric FG plate 5cpts 0.0526 0.0594 0.0662 with misfitting inclusions. Acta Metall 1973;21(5):571–4.
7cpts 0.0531 0.0598 0.0668 [21] RJ N. Analysis of functionally graded plates. Int J Numer Methods Eng
9cpts 0.0536 0.0600 0.0669 2000;47(1–3):663–84.
Asymmetric FG plate 13cpts 0.0526 0.0589 0.0658 [22] Hughes TJR, Cottrell JA, Bazilevs Y. Isogeometric analysis: CAD, finite elements,
NURBS, exact geometry and mesh refinement. Comput Methods Appl Mech Eng
2005;194(39):4135–95.
pl based on the power law volume fraction distribution.
op [23] Zhang G, Alberdi R, Khandelwal K. Analysis of three-dimensional curved beams
based on the optimal volume fraction distribution. using isogeometric approach. Eng Struct 2016;117:560–74.
[24] Zare A, Eghtesad M, Daneshmand F. Numerical investigation and dynamic behavior
Both DNN and mSOS have been shown the effectiveness and ro- of pipes conveying fluid based on isogeometric analysis. Ocean Eng
2017;140:388–400.
bustness, it can be extended to more complex engineering problems [25] Luu AT, Kim NI, Lee J. Bending and buckling of general laminated curved beams
such as for bi-directional FGMs or shells. Moreover, topology optimi- using NURBS-based isogeometric analysis. Eur J Mech A Solids 2015;54:218–31.
zation also awaits further attention. [26] Tran LV, Ferreira AJM, Nguyen-Xuan H. Isogeometric analysis of functionally
graded plates using higher-order shear deformation theory. Compos B Eng
2013;51:368–83.
Acknowledgments [27] Menasria A, Bouhadra A, Tounsi A, Bousahla AA, Mahmoud SR. A new and simple
HSDT for thermal stability analysis of FG sandwich plates. Steel Compos Struct
2017;25(2):157–75.
This research was supported by Grant (NRF-2017R1A4A1015660) [28] Nguyen-Xuan H, Thai CH, Nguyen-Thoi T. Isogeometric finite element analysis of
from NRF (National Research Foundation of Korea) funded by MEST composite sandwich plates using a higher order shear deformation theory.
(Ministry of Education and Science Technology) of Korean government. 2013;55:558–74.
[29] Carrera E, Giunta G. Refined beam theories based on a unified formulation. Int J
Appl Mech 2010;02(01):117–43.
References [30] Ibrahim SM, Carrera E, Petrolo M, Zappino E. Buckling of thin-walled beams by a
refined theory. J Zhejiang Univ - Sci 2012;13(10):747–59.
[1] Attia A, Bousahla AA, Tounsi A, Mahmoud S, Alwabli AS. A refined four variable [31] Pagani A, Petrolo M, Colonna G, Carrera E. Dynamic response of aerospace struc-
plate theory for thermoelastic analysis of FGM plates resting on variable elastic tures by means of refined beam theories. Aero Sci Technol 2015;46:360–73.
foundations. Struct Eng Mech 2018;65(4):453–64. [32] Carrera E, Filippi M, Mahato PK, Pagani A. Free-vibration tailoring of single- and
[2] El-Haina F, Bakora A, Bousahla AA, Tounsi A, Mahmoud SR. A simple analytical multi-bay laminated box structures by refined beam theories. Thin-Walled Struct
approach for thermal buckling of thick functionally graded sandwich plates. Struct 2016;109:40–9.
Eng Mech 2017;63(5):585–95. [33] Carrera E, Miguel AGd, Pagani A. Extension of MITC to higher-order beam models
[3] Khetir H, Bouiadjra MB, Houari MSA, Tounsi A, Mahmoud M. A new nonlocal tri- and shear locking analysis for compact, thin-walled, and composite structures. Int J
gonometric shear deformation theory for thermal buckling analysis of embedded Numer Methods Eng 2017;112(13):1889–908.
nanosize FG plates. Struct Eng Mech 2017;64(4). [34] Carrera E, Kaleel I, Petrolo M. Elastoplastic analysis of compact and thin-walled
[4] Abdelaziz HH, Meziane MAA, Bousahla AA, Tounsi A, Mahmoud S, Alwabli AS. An structures using classical and refined beam finite element models. Mech Adv Mater
efficient hyperbolic shear deformation theory for bending, buckling and free vi- Struct 2017;0(0):1–13.
bration of FGM sand wich plates with various boundary conditions. Steel Compos [35] Pagani A, Carrera E. Unified formulation of geometrically nonlinear refined beam
Struct 2017;25(6). theories. Mech Adv Mater Struct 2018;25(1):15–31.
[5] Mouffoki A, Bedia EAA, Houari MSA, Tounsi A, Mahmoud SR. Vibration analysis of [36] Suresh S, Mortensen A. Fundamentals of functionally graded materials: processing
nonlocal advanced nanobeams in hygro-thermal environment using a new two- and thermomechanical behaviour of graded metals and metal-ceramic composites.
unknown trigonometric shear deformation beam theory. Smart Struct Syst IOM Communications Limited; 1998.
2017;20(3):369–83. [37] Cho JR, Ha DY. Volume fraction optimization for minimizing thermal stress in
[6] Beldjelili Y, Tounsi A, Mahmoud SR. Hygro-thermo-mechanical bending of S-FGM Ni–Al2o3 functionally graded materials. Mater Sci Eng A Struct Mater
plates resting on variable elastic foundations using a four-variable trigonometric 2002;334(1):147–55.
plate theory. Smart Struct Syst 2016;18(4):755–86. [38] Cho JR, Ha DY. Optimal tailoring of 2d volume-fraction distributions for heat-re-
[7] Bouderba B, Houari MSA, Tounsi A, Mahmoud SR. Thermal stability of functionally sisting functionally graded materials using FDM. Comput Methods Appl Mech Eng
graded sandwich plates using a simple shear deformation theory. Struct Eng Mech 2002;191(29):3195–211.
2016;58(3):397–422. [39] Goupee AJ, Vel SS. Two-dimensional optimization of material composition of
[8] Bousahla AA, Benyoucef S, Tounsi A, Mahmoud SR. On thermal stability of plates functionally graded materials using meshless analyses and a genetic algorithm.
with functionally graded coefficient of thermal expansion. Struct Eng Mech Comput Methods Appl Mech Eng 2006;195(44):5926–48.
2016;60(2):313–35. [40] Goupee AJ, Vel SS. Optimization of natural frequencies of bidirectional functionally
[9] Bourada M, Kaci A, Houari MSA, Tounsi A. A new simple shear and normal de- graded beams. Struct Multidiscipl Optim 2006;32(6):473–84.
formations theory for functionally graded beams. Steel Compos Struct [41] Goupee AJ, Vel SS. Multi-objective optimization of functionally graded materials
2015;18(2):409–23. with temperature-dependent material properties. Mater Des 2007;28(6):1861–79.
[10] Hamidi A, Houari MSA, Mahmoud SR, Tounsi A. A sinusoidal plate theory with 5- [42] Lieu QX, Lee J. Modeling and optimization of functionally graded plates under
unknowns and stretching effect for thermomechanical bending of functionally thermo-mechanical load using isogeometric analysis and adaptive hybrid evolu-
graded sandwich plates. Steel Compos Struct 2015;18(1):235–53. tionary firefly algorithm. Compos Struct 2017;179:89–106.
[11] Hebali H, Tounsi A, Houari MSA, Bessaim A, Bedia EAA. New quasi-3d hyperbolic [43] Lieu QX, Lee D, Kang J, Lee J. NURBS-based modeling and analysis for free vi-
shear deformation theory for the static and free vibration analysis of functionally bration and buckling problems of in-plane bi-directional functionally graded plates.
graded plates. J Eng Mech 2014;140(2):374–83. Mech Adv Mater Struc 2018;0(0):1–17.
[12] Meziane MAA, Abdelaziz HH, Tounsi A. An efficient and simple refined theory for [44] Gopalakrishnan K, Khaitan SK, Choudhary A, Agrawal A. Deep Convolutional
buckling and free vibration of exponentially graded sandwich plates under various Neural Networks with transfer learning for computer vision-based data-driven pa-
boundary conditions. J Sandw Struct Mater 2014;16(3):293–318. vement distress detection. Constr Build Mater 2017;157:322–30.
[13] Zidi M, Tounsi A, Houari MSA, Adda Bedia EA, Anwar Bég O. Bending analysis of [45] Vesperini F, Vecchiotti P, Principi E, Squartini S, Piazza F. Localizing speakers in
FGM plates under hygro-thermo-mechanical loading using a four variable refined multiple rooms by using Deep Neural Networks. Comput Speech Lang
plate theory. Aero Sci Technol 2014;34:24–34. 2018;49:83–106.
[14] Tounsi A, Houari MSA, Benyoucef S, Adda Bedia EA. A refined trigonometric shear [46] Spille C, Ewert SD, Kollmeier B, Meyer BT. Predicting speech intelligibility with
deformation theory for thermoelastic bending of functionally graded sandwich deep neural networks. Comput Speech Lang 2018;48:51–66.
plates. Aero Sci Technol 2013;24(1):209–20. [47] Lee S, Ha J, Zokhirova M, Moon H, Lee J. Background information of deep learning
[15] Bouderba B, Houari MSA, Tounsi A. Thermomechanical bending response of FGM for structural engineering. Arch Comput Mech Eng 2018;25(1):121–9.
thick plates resting on Winkler-Pasternak elastic foundations. Steel Compos Struct [48] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a
2013;14(1):85–104. simple way to prevent neural networks from overfitting. J Mach Learn Res
[16] Nguyen TN, Thai CH, Nguyen-Xuan H, Lee J. NURBS-based analyses of functionally 2014;15:1929–58.
[49] Hinton GE. A practical guide to training restricted Boltzmann machines vol 7700.
325
Berlin, Heidelberg: Springer Berlin Heidelberg; 2012. hydrothermal scheduling. Ain Shams Eng J 2016http://dx.doi.org/10.1016/j.asej.
[50] Moita JS, Araújo AL, Correia VF, Mota Soares CM, Herskovits J. Material dis- 2016.04.002.
tribution and sizing optimization of functionally graded plate-shell structures. [71] Ezugwu AES, Adewumi AO. Discrete symbiotic organisms search algorithm for
Compos B Eng 2018;142:263–72. travelling salesman problem. Expert Syst Appl 2017;87:70–8.
[51] Franco Correia VM, Aguilar Madeira JF, Araújo AL, Mota Soares CM. Multiobjective [72] Ezugwu AE, Adewumi AO. Soft sets based symbiotic organisms search algorithm for
optimization of ceramic-metal functionally graded plates using a higher order resource discovery in cloud computing environment. Future Gener Comput Syst
model. Compos Struct 2018;183:146–60. 2017;76:33–50.
[52] Lieu QX, Lee S, Kang J, Lee J. Bending and free vibration analyses of in-plane bi- [73] Nanda SJ, Jonwal N. Robust nonlinear channel equalization using WNN trained by
directional functionally graded plates with variable thickness using isogeometric symbiotic organism search algorithm. Appl Soft Comput 2017;57:197–209.
analysis. Compos Struct 2018;192:434–51. [74] Sadek U, Sarjaš A, Chowdhury A, Svečko R. Improved adaptive fuzzy backstepping
[53] Lieu QX, Lee J, Lee D, Lee S, Kim D, Lee J. Shape and size optimization of func- control of a magnetic levitation system based on Symbiotic Organism Search. Appl
tionally graded sandwich plates using isogeometric analysis and adaptive hybrid Soft Comput 2017;56:19–33.
evolutionary firefly algorithm. Thin Wall Struct 2018;124:588–604. [75] Yu VF, Redi AANP, Yang CL, Ruskartina E, Santosa B. Symbiotic organisms search
[54] Khot LBNS, Venkayya VB. Comparison of optimality criteria algorithms for and two solution representations for solving the capacitated vehicle routing pro-
minimum weight design of structures. AIAA J 1979;17(2):182–90. blem. Appl Soft Comput 2017;52:657–72.
[55] Kaveh A, Kalatjari V. Topology optimization of trusses using genetic algorithm, [76] Tran DH, Luong-Duc L, Duong MT, Le TN, Pham AD. Opposition multiple objective
force method and graph theory. Int J Numer Methods Eng 2003;58:771–91. symbiotic organisms search (OMOSOS) for time, cost, quality and work continuity
[56] Lamberti L, Pappalettere C. Comparison of the numerical efficiency of different tradeoff in repetitive projects. J Comput Des Eng 2018;5(2):160–72.
sequential linear programming based algorithms for structural optimisation pro- [77] Do DTT, Lee J. A modified symbiotic organisms search (mSOS) algorithm for op-
blems. Comput Struct 2000;76(6):713–28. timization of pin-jointed structures. Appl Soft Comput 2017;61:683–99.
[57] Sedaghati R. Benchmark case studies in structural design optimization using the [78] de Boor C. On calculating with B-splines. J Approx Theor 1972;6(1):50–62.
force method. Int J Solids Struct 2005;42(21):5848–71. [79] Piegl L, Tiller W. B-spline curves and surfaces. Monographs in visual communica-
[58] Ootao Y, Tanigawa Y, Ishimaru O. Optimization of material composition of functions. Berlin, Heidelberg: Springer; 1995.
tionally graded plate for thermal stress relaxation using a genetic algorithm. J [80] Reddy JN. A simple higher-order theory for laminated composite plates. J Appl
Therm Stresses 2000;23(3):257–71. Mech 1984;51(4):745–52.
[59] Mayer DG, Kinghorn BP, Archer AA. Differential evolution – an easy and efficient [81] Makridakis S, Andersen A, Carbone R, Fildes R, Hibon M, Lewandowski R, et al. The
evolutionary algorithm for model optimisation. Agric Syst 2005;83(3):315–28. accuracy of extrapolation (time series) methods: results of a forecasting competi-
[60] Kou XY, Parks GT, Tan ST. Optimal design of functionally graded materials using a tion. J Forecast 1982;1(2):111–53.
procedural model and particle swarm optimization. Comput Aided Des [82] Kingma DP, Ba J. Adam: a method for stochastic optimization. International
2012;44(4):300–10. Conference for Learning Representations; 2014.
[61] Cheng MY, Prayogo D. Symbiotic Organisms Search: a new metaheuristic optimi- [83] Duchi J, Hazan E, Singer Y. Adaptive subgradient methods for online learning and
zation algorithm. Comput Struct 2014;139:98–112. stochastic optimization. J Mach Learn Res 2011;12:2121–59.
[62] Yang XS. Flower pollination algorithm for global optimization. Unconv Comput Nat [84] Tieleman T, Hinton G. Lecture 6.5-RMSProp, COURSERA: neural networks for
Comput 2012;7445:240–9. machine learning University of Toronto; 2012. Technical Report.
[63] Do DTT, Lee S, Lee J. A modified differential evolution algorithm for tensegrity [85] B rownlee J. A gentle introduction to mini-batch gradient descent and how to
structures. Compos Struct 2016;158:11–9. configure batch size. 2017.
[64] Ho-Huu V, Do-Thi TD, Dang-Trung H, Vo-Duy T, Nguyen-Thoi T. Optimization of [86] Simon D. Evolutionary optimization algorithms. John Wiley & Sons; 2013.
laminated composite plates for maximizing buckling load using improved differ- [87] Panda A, Pani S. A Symbiotic Organisms Search algorithm with adaptive penalty
ential evolution and smoothed finite element method. Compos Struct function to solve multi-objective constrained optimization problems. Appl Soft
2016;146:132–47. Comput 2016;46:344–60.
[65] Ho-Huu V, Nguyen-Thoi T, Truong-Khac T, Le-Anh L, Vo-Duy T. An improved [88] Das B, Mukherjee V, Das D. DG placement in radial distribution network by sym-
differential evolution based on roulette wheel selection for shape and size optimi- biotic organisms search algorithm for real power loss minimization. Appl Soft
zation of truss structures with frequency constraints. Neural Comput Appl Comput 2016;49:920–36.
2018;29(1):167–85. [89] Verma S, Saha S, Mukherjee V. A novel symbiotic organisms search algorithm for
[66] Ho-Huu V, Vo-Duy T, Luu-Van T, Le-Anh L, Nguyen-Thoi T. Optimal design of truss congestion management in deregulated environment. J Exp Theor Artif Intell
structures with frequency constraints using improved differential evolution algo- 2017;29(1):59–79.
rithm based on an adaptive mutation scheme. Auto Constr 2016;68:81–94. [90] Das S, Abraham A, Chakraborty UK, Konar A. Differential evolution using a
[67] Lieu QX, Do DTT, Lee J. An adaptive hybrid evolutionary firefly algorithm for shape neighborhood-based mutation operator. IEEE Trans Evol Comput
and size optimization of truss structures with frequency constraints. Comput Struct 2009;13(3):526–53.
2018;195:99–112. [91] Padhye N, Bhardawaj P, Deb K. Improving differential evolution through a unified
[68] Secui DC. A modified Symbiotic Organisms Search algorithm for large scale eco- approach. J Global Optim 2013;55(4):771–99.
nomic dispatch problem with valve-point effects. Energy 2016;113:366–84. [92] Nguyen-Xuan H, Tran LV, Thai CH, Kulasegaram S, Bordas SPA. Isogeometric
[69] Abdullahi M, Ngadi MA, Abdulhamid SM. Symbiotic Organism Search optimization analysis of functionally graded plates using a refined plate theory. 2014;64:222–34.
based task scheduling in cloud computing environment. Future Gener Comput Syst [93] Nguyen NT, Hui D, Lee J, Nguyen-Xuan H. An efficient computational approach for
2016;56:640–50. size-dependent analysis of functionally graded nanoplates. 2015;297:191–218.
[70] Das S, Bhattacharya A. Symbiotic organisms search algorithm for short-term
326

FGM DL

Uploaded by

Copyright:

Available Formats

FGM DL

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FGM DL

Uploaded by

Copyright:

Available Formats

Composites Part B 159 (2019) 300–326

Contents lists available at ScienceDirect

Material optimization of functionally graded plates using deep neural T

1. Introduction extensively applied to various engineering ﬁelds. For example, Zhang

similar to other non-gradient-based methods, a major limitation of SOS

Fig. 2. Feed forward neural network structure.

Table 1 stochastic optimization. This method only needs ﬁrst-order gradients

• The magnitudes of parameter updates are invariant to rescaling of

where Ujp and Ojp

DE SOS mSOS DE SOS mSOS DE SOS mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

Table 9 5. Numerical examples

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS DNN-mSOS IGA-mSOS

Table 15 graded carbon nanotube-reinforced composite shells. Compos Struct

You might also like