0% found this document useful (0 votes)
25 views13 pages

Newton Method

This document presents a methodology for unconstrained optimization of laminated thin plates using a decomposition method combined with a regularized Newton method. The approach focuses on optimizing fiber orientation to minimize maximum displacement in specific plate regions, employing the pb-2 Rayleigh-Ritz method for solution approximation. The results are discussed in relation to existing literature, highlighting the effectiveness of the proposed optimization framework.

Uploaded by

StructureClass
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
25 views13 pages

Newton Method

This document presents a methodology for unconstrained optimization of laminated thin plates using a decomposition method combined with a regularized Newton method. The approach focuses on optimizing fiber orientation to minimize maximum displacement in specific plate regions, employing the pb-2 Rayleigh-Ritz method for solution approximation. The results are discussed in relation to existing literature, highlighting the effectiveness of the proposed optimization framework.

Uploaded by

StructureClass
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Decomposition Method and Regularized Newton’s

Method Applied to Unconstrained Optimization:


Laminated Thin Plates Fiber Orientation
Tales de Vargas Lisbôa, Ph.D. 1; Filipe Paixão Geiger 2; and Rogério José Marczak 3
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

Abstract: This work aims at the unconstrained optimization of laminate plate problems by introducing a methodology that uses a decom-
position method along with regularized Newton method. The framework is applied to fiber orientation optimization in laminated thin
plates regarding the minimization of the maximum displacement of specific plate regions. First, the problem’s domain is divided into
isotropic and remainder parts, with the latter grouping laminar anisotropic behavior. This leads to an evaluation of the gradient and Hessian
of the objective function in such fashion that inverses are defined only at the first step of the procedure and they are related to the isotropic
part only, although the Hessian is updated each step. The fiber orientation sensibility on a plate’s solution can be evaluated separately,
leading to an analytic and straightforward way of deriving the first and second derivatives of the design variables. The pb-2 Rayleigh-Ritz
Method is used to approximate the solution space and to determine the problem’s semianalytical response. The results obtained are dis-
cussed and compared to those found in the literature. DOI: 10.1061/(ASCE)EM.1943-7889.0001615. © 2019 American Society of Civil
Engineers.

Introduction including shell and plate elements, is difficult, and the use of brick
elements can elevate the computational cost and, therefore, time of
Composite plates are structures with two or more materials with analysis to impractical levels (Carrera 2002). Thus, a numerical
certain desirable characteristics that improve the mechanical prop- method that can aid in this matter is important for engineers and
erties of each individual material (Alternbach 1998; Ferreira et al. designers.
2003). One matrix is used to gather fibers together, which is the one The Adomian decomposition method (ADM) (Adomian 1986,
responsible for carrying loads. Stacking layers makes it possible to 1994) is a powerful tool in such cases and can be implemented to
achieve better mechanical characteristics (Carrera 2002), for exam- solve linear problems in three steps: (1) differential/matrix operator
ple, the high strength-to-weight relation and stiffness (Reddy decomposition into two terms, linear and remainder; (2) a series
2004). It has been extensively used especially in specific demand- expansion of the differential equation solution; and (3) the deter-
ing engineering areas where performance is considered the main mination of each term of the expansion in a recursive manner.
factor, such as the aerospace and automotive industries (Thai Gabet (1993) has discussed the implications of applying the
et al. 2013). Therefore, optimization of structure tailoring is impor- ADM to partial differential equations (PDEs), while Gárcia-
tant to obtain the best configuration using the least amount of Olivares (2003) has employed it to obtain analytic solutions of
material or any other desired design variable. nonlinear PDEs. Biazar et al. (2004) have applied the method to
Layered structures are not simple to work with owing to their solve systems of ordinary differential equations and obtained
complexity, especially at the different scale levels that arise from excellent results. Moreover, several researchers have presented
their composition (Alternbach 1998). There are few analytic solu- proofs for convergence (Cherruault and Adomian 1993; Abbaoui
tions that allow for composite plate evaluation (Demasi 2009a; Nik and Cherruault 1995), proposed modifications (Wazwaz 2000;
and Tahani 2009) since they are restricted to certain cases, so most Abdelwahid 2003), and shown the theoretical basis of the method
solutions are obtained via finite elements and are widely used by (Gabet 1994). Other considerations on the method may be found in
industry and researchers (Rao and Meyer-Piening 1990; Raghu Duan et al. (2012) and Almazmumy et al. (2012).
et al. 2018). However, the implementation of such techniques, Herein, a different approach (Lisbôa and Marczak 2017; Lisbôa
1
et al. 2017) regarding the main application of the Adomian decom-
Dept. of Mechanical Engineering, Federal Univ. of Rio Grande do Sul, position method is considered; it concerns the differential/matrix
Rua Sarmento Leite, 425, Porto Alegre, RS CEP 90050-170, Brazil
(corresponding author). Email: tales.lisboa@ufrgs.br; taleslisboa@daad
operator decomposition. The idea is to guide this procedure using
-alumni.de a constitutive hierarchy concept (e.g., Cowin and Mehrabadi 1995;
2
Dept. of Mechanical Engineering, Federal Univ. of Rio Grande do Chadwick et al. 2001; Ting 2003) and a constitutive decomposition
Sul, Rua Sarmento Leite, 425, Porto Alegre, RS CEP 90050-170, Brazil. (e.g., Browaeys and Chevrot 2004; Tu 1968). Consequently, the
Email: filipe.geiger@ufrgs.br linear part of the decomposition, normally defined (only) by the
3
Professor, Dept. of Mechanical Engineering, Federal Univ. of Rio differential operator of the highest degree, is composed of other
Grande do Sul, Rua Sarmento Leite, 425, Porto Alegre, RS CEP differential operator terms. As proposed by Lisbôa and Marczak
90050-170, Brazil. Email: rato@mecanica.ufrgs.br
(2017) and Lisbôa et al. (2017), one may decompose the constit-
Note. This manuscript was submitted on March 6, 2018; approved on
November 19, 2018; published online on April 5, 2019. Discussion period utive matrix into isotropic and anisotropic terms, resulting in
open until September 5, 2019; separate discussions must be submitted isotropic and anisotropic differential operators. Therefore, the re-
for individual papers. This paper is part of the Journal of Engineering cursive procedure is, essentially, the problem’s isotropic solution
Mechanics, © ASCE, ISSN 0733-9399. being recursively enhanced by anisotropic contributions.

© ASCE 04019036-1 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


The pb-2 Rayleigh-Ritz method (RRM) (Bhat 1985; Liew and ADM with RNM
Wang 1993a, b; Liew et al. 1993) is adopted due to its good per-
formance in solving plate problems and consists in the modification Initially it is necessary to solve the problem of the anisotropic
of the usual RRM to simplify the implementation of the boundary laminated plate. To obtain the displacements in a certain point of
conditions (Bhat 1985). the component, the ADM is applied considering the unconstrained
Classical laminated plate theory (CLPT) is assembled by the optimization problem
utilization of classical plate theory (CPT) and the equivalent
min uðρÞ
single layer (ESL) concept (Reddy 2004, 2007; Alternbach 1998). ρ∈Rn
A laminated plate is a thin structure that is assembled by stacking
thinner laminae, normally constructed of orthotropic material (a where the objective function u∶Rn → R, u is at least C2 in Ω, and ρ
matrix with aligned fibers). By arranging the fiber orientation and are the design variables. One may expand u in Taylor series, with
the corresponding stacking in a specific fashion, one may obtain respect to ρ, as
 
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

different properties with respect to, for example, displacement or


∂uðρÞ T 1 ∂2 uðρÞ
strength. The definition of such an assembly naturally leads to uðρ þ ΔρÞ ¼ uðρÞ þ Δρ þ ΔρT Δρ þ OðΔρ3 Þ
an optimization problem, for example, the fiber orientation for each ∂ρ 2 ∂ρ∂ρT
lamina, in order to extremize a particular (mechanical) property.
To obtain the best fiber configuration to improve the plate re- where OðΔρ3 Þ → 0 with Δρ → 0 in a region around ρ. One aims
sponse, the Newton’s method (NM) (Grippo et al. 1989; Lewis and at a null gradient of uðρÞ, resulting in
Overton 2013; Ueda and Yamashita 2014) is implemented, since it  2 
∂ uðρÞ −1 ∂uðρÞ
is a well-known procedure that has been utilized to solve several Δρ ¼ − ð1Þ
optimization problems. The method, however, has drawbacks when ∂ρ∂ρT ∂ρ
applied to nonconvex objective functions (Grippo et al. 1989). To
name two, the initial guess should be close enough to the minimum/ ρnew ¼ ρold þ Δρ ð2Þ
maximum of the objective function (in the convex region of the
objective function), and the method seeks minima or maxima Eq. (1) describes the Newton equation while Eq. (2) describes
equally. Some procedures are considered to extend the method its step size (Dembo and Steihaug 1983) in the method’s original
to nonconvex problems—regularization—and to seek an objective version form. Now consider that uðρÞ is obtained by solving the
function’s minima—line search or trust region (Grippo et al. 1989; boundary value problem as
Lewis and Overton 2013). Researchers have used the NM along
Lð∂; ρÞuðρÞ ¼ fðρÞ; ρ∈Ω ð3Þ
with ADM (Chun 2005; Abbasbandy 2003; Kang et al. 2015), aim-
ing at the improvement of root-search algorithms for nonlinear
homogeneous equations. Procedures for determining the best stack- Bð∂ÞuðρÞ ¼ gðρÞ; ρ ∈ ∂Ω ð4Þ
ing sequence in composite materials are discussed in Ghiasi et al.
in which Lð∂; ρÞ and Bð∂Þ are the governing equations and boun-
(2009, 2010).
dary differential operator, respectively, while fðρÞ and gðρÞ are
In this work, the equations of the CLPT are briefly revisited.
their inhomogeneous parts. To obtain the gradient and the Hessian,
The pb-2 RRM is then considered and the governing linear system
one can differentiate Eq. (3) with respect to ρ twice as follows:
derived. The objective function, gradient, and Hessian are then
obtained recursively by ADM. The procedure starts with the de- ∂uðρÞ ∂f ∂Lð∂; ρÞ
composition of the reduced constitutive tensor at the lamina level. Lð∂; ρÞ ¼ − uðρÞ
∂ρ ∂ρ ∂ρ
This implies the decomposition of an isotropic and a remainder
stiffness (Lisbôa et al. 2017). The latter is associated to the aniso-
∂2 uðρÞ ∂2 f ∂2 Lð∂; ρÞ
tropic influence and, as a result, is the only part used to determine Lð∂; ρÞ T ¼ T− uðρÞ
both gradient and Hessian. Furthermore, only the nonhomogene- ∂ρ∂ρ ∂ρ∂ρ ∂ρ∂ρT
 T
ous term in the resultant linear system changes through the recur- ∂Lð∂; ρÞ ∂uðρÞ ∂Lð∂; ρÞ ∂uðρÞ
− −
sive routine since the isotropic stiffness is the same in the entire ∂ρT ∂ρ ∂ρ ∂ρT
optimization process. The plate transverse displacement is mini-
mized for different cases of boundary conditions, loading, and Using the ADM, the differential operator is decomposed into
stacking sequences. The novelty of the present work lies in the two terms, linear and remainder, and the solution is expanded into
application of ADM to an optimization procedure aimed at a re- series, resulting in
duction of the computational cost of the Hessian update at each
optimization step. Both matrix and index notation are adopted X

Lð∂; ρÞ ¼ LF ð∂Þ þ LV ð∂; ρÞ and uðρÞ ¼ uðiÞ ðρÞ ð5Þ
throughout the text. Greek and small-cap Arabic indices vary from
i¼0
1 to 2 and from 1 to 3, respectively. All others vary by use.
This article is divided into the following sections. The section where the superscripts F and V denote fixed and variable, respec-
“ADM with RNM” presents ADM along with a regularized NM tively. By construction, as detailed further in subsequent sections,
(RNM). The section “RRM Applied to CLPT” summarizes the pb-2 the fixed part is made independent of ρ. Thus, one may write
RRM applied to CLPT. The section “Optimization Procedure”
applies the method to the optimization of laminated plates. The ∂LF ð∂Þ ∂Lð∂; ρÞ ∂LV ð∂; ρÞ
¼0⇒ ¼
section “Numerical Applications” shows the numerical results ∂ρ ∂ρ ∂ρ
and comparisons with results found in the literature. The section
“Conclusions” presents the conclusions drawn in the study. The solution and first and second derivatives with respect to ρ,
Appendix I presents the analytical derivatives of the transforma- already considering the operator’s decomposition and the solution’s
tion matrix. expansion [Eq. (5)], are then defined as

© ASCE 04019036-2 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


X
∞  X
∞ 
uðρÞ − uðρ þ β l ΔρÞ ≥ −αβ l gT Δρ ð13Þ
fobj ¼ uðiÞ ¼ ðLF Þ−1 f − LV uðiÞ ð6Þ
i¼0 i¼0 where 0 < α, β < 1. As a “perturbation”, uðρ þ β l ΔρÞ may also be
evaluated via ADM [Eq. (9)]. Eq. (13) ensures that the step is not
X  

∂uðiÞ ∂f ∂LV X ∞ X∞
∂uðiÞ excessively large. The new set of design variables [modifying
g¼ ¼ ðLF Þ−1 − uðiÞ − LV ð7Þ Eq. (2)] is then defined as
i¼0
∂ρ ∂ρ ∂ρ i¼0 i¼0
∂ρ
ρnew ¼ ρold þ β l Δρ ð14Þ
X 
∞ 2 ðiÞ
∂ u ∂2 f ∂2 LV X

H¼ ¼ ðLF Þ−1 − uðiÞ þ The methodology (ADM along with RNM) just presented is
i¼0
∂ρ∂ρT ∂ρ∂ρT ∂ρ∂ρT i¼0
 T  general. It is applied to laminated thin plates to obtain the best fiber
∂LV X

∂uðiÞ ∂LV X∞
∂uðiÞ X∞ 2 ðiÞ
∂ u orientation, given a certain number of laminae. Consequently, the
− − − LV ð8Þ objective function is the norm of the displacement, and the dem-
∂ρ i¼0 ∂ρ
T ∂ρ i¼0 ∂ρ T
∂ρ∂ρ T
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

i¼0
onstrated approach is slightly modified.
where f obj , g, and H correspond to the objective function, gradient,
and Hessian. In Eqs. (6)–(8) the dependence of the variables on
ρ and ∂ is omitted. RRM Applied to CLPT
The ADM recursive system is then defined as
The application of composite structures with more than one layer
uðnÞ ¼ −ðLF Þ−1 LV uðn−1Þ ð9Þ has been increasing in recent decades, as has interest in their
mechanical analysis (Nik and Tahani 2009). Here the classical
 V  theory was applied due to it scapacity to produce good results,
∂uðnÞ −1 ∂L ðn−1Þ ∂uðn−1Þ and the first-order derivative simplifies the final system to be solved
¼ −ðL Þ
F u þLV ð10Þ
∂ρ ∂ρ ∂ρ (Alternbach 1998). The displacement field of CLPT is written
(Reddy 2004)
 2 V  V ðn−1Þ T
∂2 uðnÞ F −1 ∂ L ∂L ∂u
¼ −ðL Þ u ðn−1Þ
þ U 1 ðx1 ; x2 ; x3 Þ ¼ u1 ðx1 ; x2 Þ − x3 u3;1 ðx1 ; x2 Þ
∂ρ∂ρT ∂ρ∂ρT ∂ρT ∂ρ

∂LV ∂uðn−1Þ 2 ðn−1Þ
V∂ u
U 2 ðx1 ; x2 ; x3 Þ ¼ u2 ðx1 ; x2 Þ − x3 u3;2 ðx1 ; x2 Þ
þ þL ð11Þ
∂ρ ∂ρT ∂ρ∂ρT
U 3 ðx1 ; x2 Þ ¼ u3 ðx1 ; x2 Þ
where
where U i = plate displacement in êi direction, where xi denotes the
uð0Þ ¼ ðLF Þ−1 f Cartesian coordinates; uα and u3 = tangential and transverse
displacements, respectively.
∂uð0Þ ∂f By the linear kinematic relation and generalized Hooke’s law,
¼ ðLF Þ−1
∂ρ ∂ρ the plate strains, ϵ, and plate stresses, σ, are described respectively
∂2 uð0Þ 2
F −1 ∂ f
as
T ¼ ðL Þ ð12Þ
∂ρ∂ρ ∂ρ∂ρT ϵ ¼ ε − x3 κ; σ ðkÞ ¼ QðkÞ ϵ
The procedure outlined by Eqs. (9)–(11) stops when a required/ in which
desired convergence criterion is met and demonstrates that the up-
dating of the Hessian matrix depends only on the inverse of a term ϵT ¼ f ϵ1 ϵ2 ϵ6 gT ; εT ¼ f u1;1 u2;2 u1;2 þ u2;1 gT ;
that is independent of the design variable—“fixed”—and is carried κT ¼ f u3;11 u3;22 2u3;12 gT
out at the first optimization step.
The NM converges to the minima in the objective function con- where ε and κ respectively denote the plate’s membrane strain and
vex region. For a global convergence property, Eq. (1) is modified curvature and
as follows (Ueda and Yamashita 2010):
−1
QðkÞ ¼ RðkÞ ĈRðkÞ
Δρ ¼ −ðH þ μIÞ−1 g
QðkÞ is the stiffness of lamina k in global coordinates, while Ĉ
in which corresponds to the reduced constitutive tensor for the plane stress
state, defined in the lamina’s local coordinates. R transforms Ĉ to
μ ¼ c1 Λ þ c2 kgkδ ; Λ ¼ maxð0; −σmin ðHÞÞ global coordinates, and R and its derivatives are presented in
Appendix I.
where c1 , c2 , and δ are given constants such that c1 > 1, c2 > 0,
In general, the lamina constitutive properties describe the behav-
and δ ≥ 0 (Ueda and Yamashita 2010). σmin corresponds to the
ior of an orthotropic material. Thus, one writes the reduced constit-
smallest eigenvalue of H, and k · k defines the L2 -norm. The regu-
utive as (Reddy 2004, 2007)
larization gives to the method global convergence, even for non-
convex problems, due to the fact that the modified Hessian is 2 3
Ĉ11 Ĉ12 0
always positive (semi)-definite (Ueda and Yamashita 2010). More- 6 7
over, Armijo’s step size rule (Ueda and Yamashita 2010)—an Ĉ ¼ 4 Ĉ22 0 5 ð15Þ
inexact line search method (Ahookhosh and Ghaderi 2017)—is symm: Ĉ66
used herein to weight the step size obtained by Newton’s method.
One searches for the smallest integer l ≥ 0 such that where

© ASCE 04019036-3 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


E11 E22 The external potential, V ext , is determined as usual:
Ĉ11 ¼ ; Ĉ22 ¼
1 − ν 12 ν 21 1 − ν 12 ν 21 Z
Ĉ12 ¼ ν 12 E22 ; Ĉ66 ¼ E12 ; ν 21 E11 ¼ ν 12 E22 ð16Þ V ext ¼ − qT udΛ; qT ¼ f q 1 q2 q 3 gT
Λ
and Eαβ denotes the elastic moduli (longitudinal and transversal)
and ν 12 and ν 21 correspond to the Poisson coefficients. Eq. (15) has where qi describes the applied loadings to the reference surface
to do with an orthorhombic/orthotropic symmetry. with respect to the degrees of freedom. For the sake of simplicity,
Since the pb-2 RRM is applied to determine the stationary concentrated forces and moments are not explicitly considered.
point through the minimization of the total potential energy func- The CLPT has two important characteristics:
tional, the strain energy V int and the external potential V ext are • Since it is based on the kinematic model of classical plate theory
defined. V int can be written through the extensional stiffness, A, (Kirchhoff plate theory), all its hypotheses are transferred
extensional-bending coupling stiffness, B, and bending stiffness, to CLPT.
• The homogenization of material properties transforms the lami-
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

D, as (Lisbôa et al. 2017; Reddy 2004)


nate in a single anisotropic plate (ABD matrix), and conse-
Z
1 quently, the results are obtained with sufficient accuracy for
V int ¼ εT Aε − 2εT Bκ þ κT DκdΩ ð17Þ
2 Ω global properties, such as plate displacements and natural
frequencies (Alternbach 1998).
where the aforementioned stiffnessrd are There exist other plate theories as well as frameworks to
consider a plate’s nonhomogeneity. Carrera (2003) and Demasi
X
L
1X L (2009a, b, c, d) have developed Carrera’s and the Generalized Uni-
QðkÞ ; x2 jzkþ1 QðkÞ ;
z z
A¼ x3 jzkþ1 B¼
k¼1
k
2 k¼1 3 k fied Formulation, respectively. Through them, one can construct
virtually all types of plate theories. Moreover, one may construct
1X L
in a straightforward manner layerwise and zigzag theories (Demasi
x3 jzkþ1 QðkÞ
z

3 k¼1 3 k 2009b, d), along with a predefined interpolation of displacement.
Such theories produce better results with respect to local compo-
nents, such as transverse shear stress, although it increases the
and L defines the number of layers and zk the distance of the bot-
required computational costs to solve problems. As mentioned pre-
tom of the kth lamina to the reference surface. The integrand of
viously, CPLT is used herein due to its simplicity and its perfor-
Eq. (17) can be further simplified as
mance in global variables.
Z
1 With the strain energy and the external potential defined, one
V int ¼ χ T Tχ dΩ interpolates the plate displacements as
2 Ω

in which u ¼ Φλ

χ T ¼ f u1;1 u2;2 u1;2 þ u2;1 u3;11 u3;22 2u3;12 gT ⇒ χ ¼ ∂u where


2 3 8 ð1Þ 9
where gð1Þ ϕT 0 0 <c >
> =
2 3 6 7
∂ ∂ Φ¼4 0 gð2Þ ϕT 0 5 ; λ ¼ cð2Þ ð20Þ
>
: ð3Þ >
;
6 ∂x1 0 0 0 0 7 ð3Þ
6 ∂x2 7 ( ) 0 0 g ϕ T
3×3n c 3n
6 ∂ ∂ 7 u1
∂T ¼ 6
6
0 0 0 0 7;
7 u¼ u2
6 ∂x2 ∂x1 7 in which gðiÞ = Bhat’s beam functions (Lisbôa et al. 2017; Bhat
4 ∂2
∂ 2
∂ 2 5 u3 1985; Liew and Wang 1993b) assigned to the boundary conditions
0 0 0 2 (controlled by pb-2 modification); ϕT = approximation basis; and
∂x21 ∂x22 ∂x1 ∂x2
cðiÞ = weighting constants vector. The product gðiÞ ϕT generates kin-
ematically admissible functions, a requirement of RRM. n is the
Matrix T is well known in CLPT (Reddy 2004; Lisbôa et al. span of each interpolation basis, and λ groups the constants vectors
2017) as the ABD matrix and can be constructed in an alternative for all degrees of freedom.
fashion, more suitable to the present study, as follows: Both strain energy and external potential are then rewritten as
22 3 3 8
1 2 zkþ1 R
XL 66 x3 x
2 37 7 X L XL
χ ¼ Nλ
 < V int ¼ 1 λT NT TNdΛλ
>
T¼ 66 7 ⊗ QðkÞ 7 ¼ ½ZðkÞ ⊗ QðkÞ  ¼ TðkÞ ⇒ 2 Λ
44 1 5 5
k¼1 2 1 3 k¼1 k¼1 N ¼ ∂Φ >
: R T
x x V ext ¼ Λ q ΦdΛλ
2 3 3 3 zk
ð18Þ
The stationary point is determined by the minimization of the
total energy potential functional as follows:
¼ Tð1Þ þ Tð2Þ þ Tð3Þ þ · · · þ TðkÞ ð19Þ
∂Π
which is essentially the contribution of each lamina stiffness to the Π ¼ V int þ V ext ; δΠ ¼ 0 ∴ δλ ¼ 0
∂λ
laminated stiffness as if it were an anisotropic single layer. This
alternative method makes the differentiation of T with respect to resulting in the linear system
the fiber orientation easier. In Eq. (18), the symbol ⊗ denotes the
Kronecker product. Bλ ¼ F

© ASCE 04019036-4 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


where ∂TV ∂QðiÞ
Z Z ¼ ZðiÞ ⊗
∂θi ∂θi
B¼ NT TNdΛ; F¼ qT ΦdΛ
Λ Λ 8
> 2 ðiÞ
2
∂ T V < ZðiÞ ⊗ ∂ Q ; i¼j
¼ ∂θi ∂θj
∂θi ∂θj :>
Optimization Procedure 0; i≠j
Let the vector that groups the fiber orientation of all laminae be in which repeated indices do not sum. The second derivative is null
defined as θ (design variables), whose size is equal to the number for different θi [one simply observes Eq. (18)].
of laminae L. Let the constitutive decomposition be defined as In assembling the weighting vector and its first and second
follows (Lisbôa and Marczak 2017; Lisbôa et al. 2017): derivatives with respect to the design variables [Eqs. (21)–(23)],
one constructs the objective function, the gradient, and the Hessian
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

Ĉ ¼ Ĉiso þ Ĉani by multiplying them by the interpolation functions. Consequently,


the presented procedure obtains these entities globally.
which implies As mentioned previously, the procedure is used to minimize
plate displacements with respect to prechosen regions or points.
TðθÞ ¼ TF þ TV ðθÞ ⇒ BðθÞ ¼ BF þ BV ðθÞ Thus, the objective function, the gradient, and the Hessian are
slightly modified (see the section “ADM with RNM”) so as to
where TF and TV are constructed using Ĉiso and Ĉani , respectively. be defined through a displacement norm
Consequently, as stated in Eq. (5), only entities with superscript V
vary with the design variables since isotropic materials are invariant min juðρÞj ⇒ minn uðρÞ2
ρ∈Rn ρ∈R
to spatial transformations (Lisbôa et al. 2017).
Now, consider the expansion of the solution, as presented which results in
in Eq. (5). As observed in Lisbôa et al. (2017) and Lisbôa and
Marczak (2017), this implies the same expansion of the RRM X
m X
m
objective function: ū3 ¼ λðiÞ Θðx̄1 ; x̄2 Þ λðiÞ
T

constants vector [Eq. (20)] as


i¼0 i¼0
X  X
λ¼ λð0Þ þ λð1Þ þ λð2Þ þ ··· þ λðmÞ þ :::
m
∂λðiÞ m
gradient: ḡ ¼ 2 Θðx̄1 ; x̄2 Þ λðiÞ
i¼0
∂θ i¼0
As a result of these definitions, Eqs. (9)–(11) are rewritten as X
m
∂2 λðiÞ X
m
Hessian: H̄ ¼ 2 Θðx̄1 ; x̄2 Þ λðiÞ
∂θ∂θ T
λðnÞ ¼ −ðBF Þ−1 BV λðn−1Þ ð21Þ i¼0 i¼0
X  X 
m
∂λðiÞ m
∂λðiÞ
 V  þ2 Θðx̄1 ; x̄2 Þ
∂λðnÞ ∂B ðnÞ ∂λðn−1Þ i¼0
∂θ i¼0
∂θ
¼ −ðBF Þ−1 λ þ BV ð22Þ
∂θ ∂θ ∂θ ð26Þ

 2 V  V ðnÞ T where Θðx̄1 ; x̄2 Þ is constructed by the regions or points in which a


∂2 λðnÞ F −1 ∂ B ðnÞ ∂B ∂λ
¼ −ðB Þ λ þ displacement is minimized as
∂θ∂θT ∂θ∂θT ∂θT ∂θ
 X
np nr Z
X
∂BV ∂λðnÞ ∂2 λðn−1Þ Θðx̄1 ; x̄2 Þ ¼ Φnp ðx̄1 ; x̄2 Þ þ ΦAp ðx1 ; x2 ÞdAp
þ þB V ð23Þ
∂θ ∂θT ∂θ∂θT i¼1 i¼1 Ap

with the initial step [Eq. (12)] as Φnp and ΦAp denote the approximation functions evaluated at a spe-
cific point and over a specific area Ap . m is the number of iterations
λð0Þ ¼ ðBF Þ−1 f to achieve ADM convergence.
The regularization enforces the search direction to be defined as
∂λð0Þ ∂f
¼ ðBF Þ−1
∂θ ∂θ Δθ ¼ −ðH̄ þ μIÞ−1 ḡ; μ ¼ c1 Λ þ c2 kḡkδ ;
∂2 λð0Þ F −1 ∂ f
2
Λ ¼ maxð0; −σmin ðH̄ÞÞ ð27Þ
T ¼ ðB Þ ð24Þ
∂θ∂θ ∂θ∂θT
and the Armijo’s step size rule (Ueda and Yamashita 2010)
To avoid the integration of BV at each optimization step, the described as the smallest l ≥ 0 subject to
follow equation is considered:
R ū3 − ū#3 ≥ −αβ l ḡT Δθ ð28Þ
Z  F
B ¼ TF ∶ Ω N ∘ NdΩ
B ¼ T∶ N ∘ NdΩ ⇒ R ð25Þ in which the perturbed solution ū#3 is obtained using the recursive
Ω BV ¼ TV ∶ Ω N ∘ NdΩ procedure of Eq. (21) and BV defined for θ þ β l Δθ. The design
variables of the next iteration are then defined as
where the fourth-order tensor (N ∘ N) has the two minor sym-
metries. The symbol ∘ denotes a dyadic product. Eq. (25) eases θnew ¼ θold þ β l Δθ
the differentiation process as well, that is, the first and second deriv-
atives of BV are then determined through the derivatives of TV and the optimization procedure restarts until some convergence
[Eq. (18)] as criteria are met.

© ASCE 04019036-5 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


cases of stacking and various positions of assessing the transverse
top
displacement to be minimized are considered.

Case 1: Square SSSS and CCCC Plates under Uniform


Loading
right
left The first case refers to SSSS and CCCC plates with one lamina
under uniform loading. The material properties considered are γ 1 ¼
20 and γ 2 ¼ 0.6 [Eq. (29)]. The objective function of both cases is
assessed at the center of the plate. For SSSS plates, it is well known
bottom that there exists a local minimum at θ ¼ ½ð2n þ 1Þπ=4 [rad],
Fig. 1. Stacking and definition of boundary edges. n ∈ N. The global minimum is located at θ ¼ ½ðn − 1Þπ=2 [rad],
n ∈ N. For CCCC plates, there exists only the global minimum,
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

located at θ ¼ ½ðn − 1Þπ=2 [rad], n ∈ N.


In SSSS plates, three different initial guess are evaluated:
Numerical Applications θaini ¼ 15°, θbini ¼ 35°, and θcini ¼ 70°. As one may observe in
Figs. 2(a and b), the procedure with initial guess θbini ¼ 35° reaches
Examples are chosen to assess the characteristics of the methodol- the local minimum. For the other two guesses, the method con-
ogy introduced herein. In all cases, the plate is considered square verged to the global minimum. All three cases required 18 itera-
(1½m × 1½m). The boundary condition types are the usual ones tions (nit ) to reach the tolerance value.
(Lisbôa et al. 2017; Bhat 1985; Liew and Wang 1993b): In CCCC plates, two initial guesses are considered: θaini ¼ 40°
• Free: no imposition on the edge is made; and θbini ¼ 70°. Both displacement and fiber orientation are shown
• Simply supported: un ¼ u3 ¼ 0;
in Figs. 2(c and d). As expected, the initial guess θaini tends to the
• Clamped edge: un ¼ ut ¼ u3 ¼ u3;n ¼ 0.
minimum at 0°, while θbini tends to 90°. The first and second initial
Four letters express a plate’s boundary conditions: the first one
defines the south edge (Fig. 1) while the other rotates counterclock- guesses, θaini ¼ 40° and θbini ¼ 70°, required four and five iterations,
wise (east, north, and west). In Fig. 1, one observes the stacking respectively, to achieve the minimum.
configuration ðθ1 ; θ2 ; : : : Þ of each lamina, from top to bottom of It is important to note that even for these simple cases, the ob-
the plate. jective function is nonconvex with respect to the design variable
The lamina elastic properties [Eq. (16)] are as follows: space. As result, the RNM can be used along with ADM, and con-
sequently, both gradient and Hessian can be obtained by the expan-
E11 ¼ γ 1 E22 ; E12 ¼ γ 2 E22 ; ν 12 ¼ 0.25 ð29Þ sion of Eqs. (22)–(24).

where γ i are given constants. For the RNM and Amrijo’s step size
Case 2: Square FFFC Plate under Torsion
[Eqs. (27) and (28)], the parameter values used in all presented
analyses were The second case evaluates the best stacking configuration concern-
ing a FFFC plate under a unitary torsion at the free edge opposite to
α ¼ 0.5 β ¼ 0.5 c1 ¼ 1.25 c2 ¼ 0.25 δ¼2 the clamped one (the loading is applied to the right edge while the
left one is clamped). Three different conditions are considered:
The tolerances used for the ADM and the optimization pro- minimization of the transverse displacement at (1) a vertex located
cedure were at ðx1 ; x2 Þ ¼ ð1; 0Þ, (2) two vertices located at ðx1 ; x2 Þ ¼ ð1; 0Þ and
ðx1 ; x2 Þ ¼ ð1; 1Þ, and (3) the free edge where the torque is applied.
toladm ¼ 10−7 ; tolopt ¼ 10−6 The procedure changes little as just Θðx̄1 ; x̄2 Þ in Eqs. (26) is modi-
fied when considering all approaches. The material properties con-
where the error in the ADM is evaluated as the difference between sider γ 1 ¼ 25 and γ 2 ¼ 0.5. Moreover, the values of θi are depicted
two subsequent steps, ðm þ 1Þ and ðmÞ, in the entire domains as from −90° to 90° due to the modular symmetry with respect to these
" #1=2 bounds. One notices that for this case, there is no information con-
T R
ðmþ1Þ λðmþ1Þ Λ ΦT ΦdΛλðmþ1Þ cerning the global minimum. Thus, a multistart procedure is
eADM ¼ Pm ðmÞT R P ðmÞ Þ considered.
ð n¼0 λ Þ Λ ΦT ΦdΛð m n¼0 λ
Case 2A
and the error in the optimization procedure as In this case—minimization of the transverse displacement of the
vertex (0, 1)—two stacking configurations are considered: one
ðmþ1Þ
eOPT ¼ kθðmþ1Þ − θðmÞ k and two laminae. The initial guesses are
• for one lamina: θaini ¼ 50°, θbini ¼ −60°, and θcini ¼ 10° and
For all solutions presented herein, n ¼ 16 [Eq. (20)]. The inter- • for two laminae: θaini ¼ f −70° −10° g, θbini ¼ f 50° −60° g,
polation basis is constructed by polynomial functions in an iden- and θcini ¼ f −10° 10° g.
tical fashion as in Lisbôa and Marczak (2017), Lisbôa et al. (2017), One notices that just one point of the domain is considered in the
and Liew et al. (1993). minimization; thus, the fibers will be aligned so as to nullify
Two cases are selected to show the methodology results. The the transverse displacement at the assessed position. Therefore,
first is related to square SSSS and CCCC plates with one lamina the global minimum is achieved when ū ¼ 0. Several configura-
under uniform transverse loading. This case is selected because it tions achieve such a minimum. As a result, this problem can be
presents known local/global minima (obtained analytically). The considered as ill-posed. Nevertheless, the problem’s solutions
second one concerns FFFC plates subjected to unitary torque at are presented here so as to examine the methodology under such
the free edge located at the opposite side of the clamped edge. Four circumstances.

© ASCE 04019036-6 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

(c) (d)

Fig. 2. Case 1: (a) ū of SSSS plates; (b) θ during minimization procedure of SSSS plates; (c) ū of CCCC plates; and (d) θ during minimization
procedure of CCCC plates.

(a) (b)

Fig. 3. Case 2A for one lamina: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

© ASCE 04019036-7 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

Fig. 4. Case 2A for two laminae: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

(a) (b)

Fig. 5. Case 2B for one lamina: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

The objective function and the fiber orientation for the cases • one lamina: θaini ¼ 10°, θbini ¼ 50°, and θcini ¼ 40°;
concerning one lamina are depicted in Fig. 3. As one observes, all • two laminae: θaini ¼ f 20° 80° g, θbini ¼ f 30° −60° g, and
cases have achieved the global minimum. However, θaini ¼ 50° and θcini ¼ f −70° 70° g;
θbini ¼ −60° stabilize at θ ¼ −54.276°, while θcini ¼ 10° does so • three laminae: θaini ¼ f 80° 50° 20°g, θbini ¼ f−70° 30° 20° g,
at θ ¼ −39.361°. and θcini ¼ f 20° 80° 80° g;
Analogously to the one-lamina case, the objective function and • four laminae: θaini ¼ f 50° 80° −30° 30° g, θbini ¼
the fiber orientation through the optimization process are shown in f −30° 30° 50° −50° g, and θcini ¼ f40° −50° 80° −30°g
Fig. 4. As noted, the set θbini ¼ f 50° −60° g has achieved a local The results for one lamina are depicted in Fig. 5. As in the pre-
minimum while the other sets have reached the global one (ū ¼ 0), vious case, one of the initial guesses achieved a local minimum at
although the fiber orientations of these sets are different. In this θ ¼ 57.6°. The other initial guesses reached a minimum, both with
case and the next ones, in which more than one lamina is consid- θ ≈ 0°. In this case, the ū ≠ 0 at the minimum as the effects of
ered, the sets of the design variable are related to the symbol (circle, pointing the stiffness (or the material principal directions) in one
square, or triangle), while the lines (solid or dashed types) corre- direction, in order to minimize the transverse displacement of a
spond to a particular variable. vertex, has the opposite effect at the other one.
The results for two laminae are presented in Fig. 6. All initial
Case 2B guesses reached the same minimum, with very close sets of stack-
For this case, four stacking configurations are tested: one to four ing configuration θ ≈ f42.5°∓42.5°g. For the worst case, 10 iter-
laminae. In each evaluation, three initial guesses are considered as ations were required to reach convergence, although this case
follows: started far from the minimum and in a nonconvex region.

© ASCE 04019036-8 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

Fig. 6. Case 2B for two laminae: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

(a) (b)

Fig. 7. Case 2B for three laminae: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

The results obtained for three laminae are similar to the previous • three laminae: θaini ¼ f−85° 50° −60°g, θbini ¼ f−40° 5° 40°g,
case and are shown in Fig. 7. All initial guesses have reached and θcini ¼ f85° 5° −85°g.
the same minimum, with similar stacking configurations. The The results of this approach are similar to Case 2B. In the case
initial guess θaini clearly starts in a nonconvex region, as is evident of one lamina (Fig. 9), two initial guesses reach a minimum,
in the initial iterations in Fig. 7(a). The obtained sets are with θ ¼ 0°. A local minimum was found at θ ¼ 66.42°. For
θ ≈ f 35° 0 ∓35° g, and the worst case of initial guess re- two laminae, as shown in Fig. 10, all initial guesses reached the
quired 14 iterations to reach convergence. global minimum, with θ ¼ f 43.5° ∓43.5° g. This result is sim-
Similarly to the previous examples, the case with four laminae ilar for the previous case (for two vertices). The last case (Fig. 11)
(Fig. 8) reached a minimum in two “opposite” sets. They are also showed an initial guess that reached a local minimum at
θ ≈ f 53° ∓38° ∓38° 53° g. The worst case required θ ¼ f −60.7° 38.8° −61.6° g. The minimum was achieved by
eight iterations for the convergence. θ ¼ f 32.9° 0° ∓32.7° g, which again is similar to the results
Case 2C of the previous study.
Analogously to the previous case, three stacking configurations are
tested: one to three laminae. In each evaluation, three initial guesses
Rate of Convergence
are considered as follows:
• one lamina: θaini ¼ −15°, θbini ¼ −70°, and θcini ¼ 20°; As mentioned, RNM along with Amrijo’s step size develop global
• two laminae: θaini ¼ f −15° 15° g, θbini ¼ f 20° 60° g, and convergence, even with the Hessian with a negative eigenvalue.
θcini ¼ f −85° 65° g; The convergence was demonstrated in Case 1 (Fig. 2). In the

© ASCE 04019036-9 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

Fig. 8. Case 2B for four laminae: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

(a) (b)

Fig. 9. Case 2C for one lamina: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

(a) (b)

Fig. 10. Case 2C for two laminae: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

© ASCE 04019036-10 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

Fig. 11. Case 2C for three laminae: (a) objective function ū; and (b) fiber orientation θ for three different initial guesses.

subsequent convergence examples (Figs. 3–11), some oscillations


are a consequence of the Hessian.
The rate of convergence shows a different behavior with respect
to the number of design variables and the function to be minimized.
Three cases are considered so as to show such behavior. Thus, let
the log of the difference between subsequent steps be defined by

ek ¼ log10 ðūkþ1 − ūk Þ

where k is an integer. The rate of convergence can be evaluated in a


straightforward fashion with the ek × k curve tendency.
The rate of convergence of the simplest case—FFFC with one
point of evaluation and one lamina (Example 2A for one lamina)—
is shown in Fig. 12. As one observes and expects, the convergence
is superlinear. The oscillations are related to the Hessian negative
eigenvalues (in θaini and θcini at the first iterations).
The second example regards a FFFC with three laminae and
evaluation on two points (Case 2). Despite the fact that the rate
Fig. 13. Rate of convergence for FFFC plates with three laminae and
varies considerably through the iterations, for θbini and θcini it has
evaluation on two points.
a similar multilinear behavior, as depicted in Fig. 13. As shown

Fig. 12. Rate of convergence for a FFFC plate with one lamina and one Fig. 14. Rate of convergence for FFFC plates with two laminae and
point of evaluation. evaluation on free edge.

© ASCE 04019036-11 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


in Ahookhosh and Ghaderi (2017), Newton’s method along with an Acknowledgments
inexact step length method (Amrijo’s step size) may lose efficiency
when trapped close to the bottom of a narrow curve of the objective The authors would like to acknowledge CNPq and CAPES for their
function. Until the seventh iteration, the behavior is similar to a financial support (Projects 1303477 and 310649).
superlinear one. Subsequently, it becomes linear. For θaini, the rate’s
general behavior is superlinear.
The third example—FFFC with two laminae and evaluation on References
the free edge (Case 3)—presents different tendencies, as depicted in
Fig. 14. For θaini and θbini , the rate’s general characteristic is linear, Abbaoui, K., and Y. Cherruault. 1995. “New ideas for proving convergence
whereas for θcini it is superlinear. of decomposition methods.” Comput. Math. Appl. 29 (7): 103–108.
https://doi.org/10.1016/0898-1221(95)00022-Q.
In considering these three examples, one notices that the rate of
Abbasbandy, S. 2003. “Improving Newton-Raphson method for non-linear
convergence of the ADM + RNM depends on the number of design equations by modified Adomian decomposition method.” Appl. Math.
variables and the objective function—initial guesses and, conse-
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

Comput. 145 (2–3): 887–893. https://doi.org/10.1016/S0096-3003(03)


quently, the Hessian. However, the rate of convergence is, at least, 00282-0.
linear. Abdelwahid, F. 2003. “A mathematical model of Adomian polynomials.”
Appl. Math. Comput. 141 (2–3): 447–453. https://doi.org/10.1016
/S0096-3003(02)00266-7.
Conclusions Adomian, G. 1986. Nonlinear stochastic operator equations. Orlando, FL:
Academic Press.
The ADM is introduced along with RNM to obtain the minimum of Adomian, G. 1994. Solving frontier problems of physics: The decomposi-
a particular objective function. The methodology was applied to tion method. Dordrecht, Netherlands: Kluwer.
optimize the fiber orientation of thin laminated plates. Two cases Ahookhosh, M., and S. Ghaderi. 2017. “On efficiency of nonmonotone
were considered so as to demonstrate its efficiency. The method- Armijo-type line searches.” Appl. Math. Modell. 43: 170–190. https://
doi.org/10.1016/j.apm.2016.10.055.
ology performed as designed: there are no derivatives of inverted
Almazmumy, M., F. Hendi, H. Bakodah, and H. Almuzi. 2012. “Recent
matrices, making the procedure much simpler to implement and modifications of Adomian decomposition method for initial value prob-
even create existent codes/routines. As a result, the determination lem in ordinary differential equations.” Am. J. Comput. Math. 2 (3):
of both gradient and Hessian was facilitated, suggesting that the 228–234. https://doi.org/10.4236/ajcm.2012.23030.
Hessian could be calculated in each iteration, increasing the con- Alternbach, H. 1998. “Theories for laminated and sandwich plates.” Mech.
vergence rate of the optimization process. Compos. Mater. 34 (3): 243–252. https://doi.org/10.1007/BF02256043.
The ADM is designed to solve nonlinear differential equations, Bhat, R. B. 1985. “Plate deflections using orthogonal polynomials.” J. Eng.
which could be useful for fiber optimization in nonlinear bending. Mech. 111 (11): 1301–1309. https://doi.org/10.1061/(ASCE)0733
Furthermore, constrained optimization could be considered along -9399(1985)111:11(1301).
with a penalty method as well (such as Lagrange multipliers). It is Biazar, J., E. Babolian, and R. Islam. 2004. “Solution of the system of ordi-
nary differential equations by Adomian decomposition method.” Appl.
also important to note that the methodology can be expanded to
Math. Comput. 147 (3): 713–719. https://doi.org/10.1016/S0096-3003
other plate theories. For example, one may consider the Carrera (02)00806-8.
or Generalized Unified Formulation along with ADM and RNM Browaeys, J. T., and S. Chevrot. 2004. “Decomposition of the elastic tensor
to evaluate a plate’s mechanical response and optimize the fiber and geophysical applications.” Geophys. J. Int. 159 (2): 667–678.
orientation of different theories. https://doi.org/10.1111/j.1365-246X.2004.02415.x.
Carrera, E. 2002. “Theories and finite elements for multilayered, aniso-
tropic, composite plates and shells.” Arch. Comput. Methods Eng.
Appendix. Analytical Derivatives of Transformation 9 (2): 87–140. https://doi.org/10.1007/BF02736649.
Matrix Carrera, E. 2003. “Theories and finite elements for multilayered plates and
shells: A unified compact formulation with numerical assessment and
This matrix is the result of fourth-order reduced constitutive tensor benchmarking.” Arch. Comput. Methods Eng. 10 (3): 215–296. https://
(2 × 2 × 2 × 2) transformation and the contraction R4 → R2 and doi.org/10.1007/BF02736224.
is written Chadwick, P., M. Vianello, and S. C. Cowin. 2001. “A new proof that the
number of linear elastic symmetries is eight.” J. Mech. Phys. Solids
2 3
cos2 ðθk Þ sin2 ðθk Þ sinðθk Þcosðθk Þ 49 (11): 2471–2492. https://doi.org/10.1016/S0022-5096(01)00064-3.
6 7 Cherruault, Y., and G. Adomian. 1993. “Decomposition methods: A new
RðkÞ ¼ 4 sin2 ðθk Þ cos2 ðθk Þ −sinðθk Þ cosðθk Þ 5 proof of convergence.” Math. Comput. Modell. 18 (12): 103–106.
2sinðθk Þ cosðθk Þ −2sinðθk Þ cosðθk Þ sin2 ðθk Þ − cos2 ðθk Þ https://doi.org/10.1016/0895-7177(93)90233-O.
Chun, C. 2005. “Iterative methods improving newton’s method by the de-
where k varies from 1 to L (lamina number). Moreover, RðkÞ ¼ composition method.” Comput. Math. Appl. 50 (10–12): 1559–1568.
T

−1 https://doi.org/10.1016/j.camwa.2005.08.022.
RðkÞ .
Cowin, S. C., and M. M. Mehrabadi. 1995. “Anisotropic symmetries of
The first and second derivatives of R are
linear elasticity.” Appl. Mech. Rev. 48 (5): 247–285. https://doi.org/10
2 3 .1115/1.3005102.
− sin 2θt sin 2θt 2 cos 2θt
∂R ðtÞ
6 7 Demasi, L. 2009a. “∞6 Mixed plate theories based on the Generalized Uni-
¼ δ tj 4 sin 2θt − sin 2θt −2 cos 2θt 5 fied Formulation. Part I: Governing equations.” Compos. Struct. 87 (1):
∂θj 1–11. https://doi.org/10.1016/j.compstruct.2008.07.013.
− cos 2θt cos 2θt −2 sin 2θt
2 3 Demasi, L. 2009b. “∞6 Mixed plate theories based on the Generalized Uni-
−2 cos 2θt 2 cos 2θt −4 sin 2θt fied Formulation. Part II: Layerwise theories.” Compos. Struct. 87 (1):
2 ðtÞ
∂ R 6 7 12–22. https://doi.org/10.1016/j.compstruct.2008.07.012.
¼ δ tj δ tk 4 2 cos 2θt −2 cos 2θt 4 sin 2θt 5
∂θj ∂θk Demasi, L. 2009c. “∞6 Mixed plate theories based on the Generalized Uni-
2 sin 2θt −2 sin 2θt −4 cos 2θt fied Formulation. Part III: Advanced mixed high order shear deforma-
tion theories.” Compos. Struct. 87 (3): 183–194. https://doi.org/10.1016
where the t do not sum. /j.compstruct.2008.07.011.

© ASCE 04019036-12 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036


Demasi, L. 2009d. “∞6 Mixed plate theories based on the Generalized Uni- Liew, K. M., Y. Xiang, and S. Kitipornchai. 1993. “Transverse vibration of
fied Formulation. Part IV: Zig-zag theories.” Compos. Struct. 87 (3): thick rectangular plates. Part I: Comprehensive sets of boundary con-
195–205. https://doi.org/10.1016/j.compstruct.2008.07.010. ditions.” Comput. Struct. 49 (1): 1–29. https://doi.org/10.1016/0045
Dembo, R. S., and T. Steihaug. 1983. “Truncated-Newton algorithms -7949(93)90122-T.
for large-scale unconstrained optimization.” Math. Programm. 26 (2): Lisbôa, T. V., F. P. Geiger, and R. J. Marczak. 2017. “A recursive meth-
190–212. https://doi.org/10.1007/BF02592055. odology to determine the mechanical response of thin laminated plates
Duan, J. S., R. Rach, D. Băleanu, and A. M. Wazwaz. 2012. “A review of in bending.” J. Aerosp. Technol. Manage. 9 (4): 409–422. https://doi
the Adomian decomposition method and its applications to fractional .org/10.5028/jatm.v9i4.761.
differential equations.” Commun. Fract. Calculus 3 (2): 73–99. Lisbôa, T. V., and R. J. Marczak. 2017. “A recursive methodology for
Ferreira, A., C. Roque, and P. Martins. 2003. “Analysis of composite plates semi-analytical rectangular anisotropic thin plates in linear bending.”
using higher-order shear deformation theory and a finite point formu-
Appl. Math. Modell. 48: 711–730. https://doi.org/10.1016/j.apm.2017
lation based on the multiquadric radial basis function method.” Com-
.04.020.
posites Part B 34 (7): 627–636. https://doi.org/10.1016/S1359-8368
Nik, A. M. N., and M. Tahani. 2009. “Analytical solutions for bending
(03)00083-0.
Downloaded from ascelibrary.org by University of Central Florida on 12/17/20. Copyright ASCE. For personal use only; all rights reserved.

analysis of rectangular laminated plates with arbitrary lamination and


Gabet, L. 1993. “The decomposition method and linear partial differential
equations.” Math. Comput. Modell. 17 (6): 11–22. https://doi.org/10 boundary conditions.” J. Mech. Sci. Technol. 23 (8): 2253–2267.
.1016/0895-7177(93)90191-Z. https://doi.org/10.1007/s12206-009-0511-4.
Gabet, L. 1994. “The theoretical foundation of the Adomian method.” Raghu, P., A. Rajagopal, and J. Reddy. 2018. “Nonlocal nonlinear finite
Comput. Math. Appl. 27 (12): 41–52. https://doi.org/10.1016/0898 element analysis of composite plates using TSDT.” Compos. Struct.
-1221(94)90084-1. 185: 38–50. https://doi.org/10.1016/j.compstruct.2017.10.075.
Gárcia-Olivares, A. 2003. “Analytical solution of nonlinear partial differ- Rao, K., and H. Meyer-Piening. 1990. “Analysis of thick laminated aniso-
ential equations of physics.” Kybernetes 32 (4): 548–560. https://doi.org tropic composite plates by the finite element method.” Compos. Struct.
/10.1108/03684920310463939. 15 (3): 185–213. https://doi.org/10.1016/0263-8223(90)90031-9.
Ghiasi, H., K. Fayazbakhsh, D. Pasini, and L. Lessard. 2010. “Optimum Reddy, J. N. 2004. Mechanics of laminated composite plates and shells.
stacking sequence design of composite materials. Part II: Variable stiff- Boca Raton, FL: CRC Press.
ness design.” Compos. Struct. 93 (1): 1–13. https://doi.org/10.1016/j Reddy, J. N. 2007. Theory and analysis of elastic plates and shells.
.compstruct.2010.06.001. Boca Raton, FL: CRC Press.
Ghiasi, H., D. Pasini, and L. Lessard. 2009. “Optimum stacking sequence Thai, C. H., A. Ferreira, E. Carrera, and H. Nguyen-Xuan. 2013.
design of composite materials. Part I: Constant stiffness design.” Com- “Isogeometric analysis of laminated composite and sandwich plates us-
pos. Struct. 90 (1): 1–11. https://doi.org/10.1016/j.compstruct.2009.01 ing a layerwise deformation theory.” Compos. Struct. 104: 196–214.
.006. https://doi.org/10.1016/j.compstruct.2013.04.002.
Grippo, L., F. Lampariello, and S. Lucidi. 1989. “A truncated newton Ting, T. C. T. 2003. “Generalized Cowin-Mehrabadi theorems and a direct
method with nonmonotone line search for unconstrained optimization.” proof that the number of linear elastic symmetries is eight.” Int. J. Solids
J. Optim. Theor. Appl. 60 (3): 401–419. https://doi.org/10.1007 Struct. 40 (25): 7129–7142. https://doi.org/10.1016/S0020-7683(03)
/BF00940345.
00358-5.
Kang, S. M., W. Nazeer, M. Tanveer, Q. Mehmood, and K. Rehman. 2015.
Tu, Y. O. 1968. “The decomposition of an anisotropic elastic tensor.”
“Improvements in newton-raphson method fir nonlinear equations us-
Acta Crystallogr. Sect. A 24 (2): 273–282. https://doi.org/10.1107
ing modified adomian decomposition method.” Int. J. Math. Anal.
9 (39): 1919–1928. https://doi.org/10.12988/ijma.2015.54124. /S0567739468000458.
Lewis, A. S., and M. L. Overton. 2013. “Nonsmooth optimization via Ueda, K., and N. Yamashita. 2010. “Convergence properties of regular-
quasi-newton methods.” Math. Programm. 141 (1–2): 135–163. ized newton method for the unconstrained nonconvex optimization.”
https://doi.org/10.1007/s10107-012-0514-2. Appl. Math. Optim. 62 (1): 27–46. https://doi.org/10.1007/s00245
Liew, K. M., and C. M. Wang. 1993a. “Pb-2 Rayleigh-Ritz method for gen- -009-9094-9.
eral plate analysis.” Eng. Struct. 15 (1): 55–60. https://doi.org/10.1016 Ueda, K., and N. Yamashita. 2014. “A regularized Newton method without
/0141-0296(93)90017-X. line search for unconstrained optimization.” Comput. Optim. Appl.
Liew, K. M., and C. M. Wang. 1993b. “Vibration studies on skew plates: 59 (1–2): 321–351. https://doi.org/10.1007/s10589-014-9656-x.
Treatment of internal line supports.” Comput. Struct. 49 (6): 941–951. Wazwaz, A.-M. 2000. “A new algorithm for calculating adomian polyno-
https://doi.org/10.1016/0045-7949(93)90005-X. mials for nonlinear operators.” Appl. Math. Comput. 111 (1): 33–51.

© ASCE 04019036-13 J. Eng. Mech.

J. Eng. Mech., 2019, 145(6): 04019036

You might also like