Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
a r t i c l e i n f o a b s t r a c t
Article history: In this study a reliable shape optimization for laminated plate structures has been attempted. For a xed
Received 2 October 2014 higher order plate model, a simple a-posteriori strain recovery algorithm, following ZZ type patch
Accepted 14 February 2015 recovery technique, has been developed. The recovery is seen to be accurate. The effect of higher
approximation order and mesh renement on the quality of the obtained solution quantities like stress
components and displacements, is studied in detail. The shape of the cutout is optimized with weight
Keywords: minimization as the objective function and the rst-ply failure criterion as the constraint. It is observed
Higher order shear deformation model
that control of the discretization error (via adaptive mesh renements) leads to vastly different nal
Strain recovery
a-posteriori error estimation
designs, as compared to those obtained using reasonably rened meshes, but without adaptivity. It is
Discretization error seen that without adaptivity, the design obtained is unsafe, as either more material removal is predicted
Shape optimization or failure is predicted at higher loads, as compared to that obtained using adaptivity.
First-ply failure 2015 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compstruc.2015.02.020
0045-7949/ 2015 Elsevier Ltd. All rights reserved.
20 P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935
discretization error on the nal optimal design was not studied. and design, some specic applications of laminated composites
The early work which involves the study of effect of discretization require unsymmetric laminates. For example, the coupling
error, on the nal optimum shape, was seen due to Kikuchi et al. between bending and extension exhibited by this type of laminates
[17]. In the shape optimization procedure, due to change in shape, is an essential feature of jet turbine fan blades with pre-twist. It
the mesh gets distorted and the nal design is sensitive to the can be noted that the theories for unsymmetric laminates are
approximation error associated with the discretization. Thus, applicable to symmetric laminates as a special case. Unlike sym-
improvement in the quality of the approximation is needed [17]. metric laminates, unsymmetric laminates exhibit membrane-ex-
In this work global error estimates were developed for adaptive ure coupling phenomenon, which necessitates the use of a
renement strategies. A similar work was carried out by Hinton displacement eld containing both, membrane as well as exure
et al. [18]. Weck et al. [19] have worked on saving the computa- deformation terms which contribute to the overall response of a
tional cost during optimization of composite structures with laminate. The analysis of laminated plates is based on the choice
ply-orientations and thickness as optimization variables using of a plate theory. Several plate theories have been developed with
adaptive meshing technique. Benneet and Botkin [20] aimed to assumed variation of the displacement eld in the transverse
provide more accurate estimate of the true optimal solution with direction. For example, see plate theories in [3,4,12,3438]. These
the effect of adaptive meshing on stresses used in constraint theories attempt to give a higher order representation of strains
evaluation. Schleupen et al. [21] developed both global (based on in the laminate thickness direction. In the following, we present
error of the strain energy of overall structure) and local error esti- the details of plate theory due to Reddy [27] implemented in the
mates (based on error in a particular quantity of interest like dis- present study. The displacement eld
placement or stress component) for global and local adaptive
renements separately. The potential of these two techniques were ux; y; z ux; y; z v x; y; z wx; y; zT 1
then compared through two dimensional shape optimization prob- is derived from the expanded Taylors series in terms of thickness
lems. Morin et al. [22] developed an algorithm based on adaptive coordinate z. Here, ux; y; z; v x; y; z and wx; y; z are the
nite element method to equidistribute the errors due to shape displacement components along x; y and z axes, respectively.
optimization and discretization to optimize the computational These components, following the work of Reddy [27], are given as
cost. An application to X-FEM based structural optimization can
be seen in [23]. An evolutionary technique was used along with ux; y; z u0 x; y zhx x; y z2 /x x; y z3 wx x; y
sensitivity analysis, for a low cost adaptive remeshing, in shape
optimization problems by Bugeda et al. [24].
v x; y; z v 0 x; y zhy x; y z2 /y x; y z3 wy x; y 2
The application of adaptive meshing using goal oriented error wx; y; z w0 x; y
control for topology optimization was done by Bruggi and Verani
[25]. Another application of adaptive renement approach to In the expansion in Eq. (2), it is assumed that transverse normal
topology optimization can be seen in Wang et al. [26]. strain zz is zero. The linear straindisplacement relationships
In the present work a design of laminated composite plate, with using small deformation theory can be obtained from this
a centrally located cutout for minimum weight, subjected to a con- equation.
straint that the plate should not fail under rst-ply failure load cri- The condition that the transverse shear stresses vanish on the
terion, has been studied. Here, an attempt is made to demonstrate plates top and bottom faces (see Fig. 1) is equivalent to the
the effect of reliability of constraints on the nal optimal solution. requirement that the corresponding strains be zero on these
Initially, the nal optimal solution is obtained without considering surfaces, i.e.
reliability of the computed data used in the evaluation of rst-ply
d d
failure load constraint. The process is then repeated with a control cyz x; y; cxz x; y; 0 3
2 2
on the reliability of the computed data, i.e. effect of discretization
error control on nal optimal shape. In the present work a higher On introduction of these conditions in the expressions for trans-
order shear deformable plate theory proposed by Reddy [27] has verse shear strains, the following relations are obtained.
been adopted and implemented in a nite element code. Further, 4 4
Zienkiewicz-Zhu (ZZ) [2831] type a-posteriori patch recovery /x /y 0; wy 2
hy w0 ;y and wx 2 hx w0 ;x
3d 3d
based error estimator is developed for strain eld corresponding
4
to the plate model considered. Although, the use of Genetic
Algorithms (GA) (for example, [12]) and evolutionary algorithms The displacement eld of Eq. (2) is modied by setting /x and /y
(for example [24]) is very popular in optimization studies, in the to be zero according to conditions of Eq. (4). The resulting displace-
present study we have used a conventional optimization algorithm ment eld is now written as
- Complex Search [32] to obtain an optimal design. Finally, the
effect of evaluation of rst-ply failure load constraint, with and ux; y; z u0 x; y z hx x; y z3 wx x; y
without control in discretization error, is studied. Here, the Tsai-
v x; y; z v 0 x; y z hy x; y z3 wy x; y 5
Wu rst-ply failure criterion [33] has been used as a constraint.
wx; y; z w0 x; y
Symmetric laminates nd many applications in the aircraft The corresponding straindisplacement relations, using
industry. Although symmetric laminates are simple to analyze innitesimal strains, are:
P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935 21
y
Initial element of interest
Top surface y
d T,2 L,
1
z
Y
Bottom surface
l
4
l x
3 x
l
2
l
1
X
(a) (b)
Fig. 1. Laminate geometry, coordinate systems, stacking sequence and layer numbering.
xx u0 ;x z hx ;x z3 wx ;x uh . Note that in this study Dirichlet means the part of lateral bound-
yy v 0 ;y z hy; z3 wy ;y ary where geometric constraints are imposed, while Neumann
y
stands for the parts of the lateral boundary where in-plane traction
cxy u0 ;y v 0 ;x z hx ;y z hy ;x z3 wx ;y z3 wy ;x 7 is applied. Further, M is a generic representation of displacement
2
cyz hy 3 z wy w0 ;y constraints on the Dirichlet boundary edge. For example, the
cxz hx 3 z2 wx w0 ;x boundary conditions can be clamped ui 0; i 1; 2; 3; soft
simple-support un ; u3 0; hard simple-support ut ; u3 0; etc.
Here, comma ; denotes the partial derivative. Here, un and ut denote in-plane displacement components normal
and tangential to an edge, respectively.
2.2. Finite element formulation Triangular elements are used in the nite element
approximation employed in this study, along with hierarchic shape
The total potential, Pp , for the structure is given by functions of order p (p 6 4). The mesh generation is done using
Z Z advancing front method based automatic mesh generator. A typical
1
Pp u ru eu dV T 3 u3 ds mesh generated over the plate domain is shown in Fig. 2.
2 V R [R
Z A detailed study on various plate models for laminate applica-
T 1 u1 T 2 u2 ds 8 tions and their nite element implementation can be seen in
CN Pandya and Kant [3638].
where V is the volume enclosed by the plate domain; ru and eu
are the engineering stress and strain vectors, respectively. R and R 3. A-posteriori recovery of pointwise strains and error
denote the top and bottom faces of the laminated plate and T 3 x; y estimation
is the applied transverse load on these faces; C are the lateral faces
with C CN [ CD . Here, CN denotes the Neumann boundary and CD In a typical engineering analysis a mathematical model for
denotes Dirichlet boundary; T 1 ; T 2 are the in-plane tractions speci- physical problem is rst selected, such that it incorporates the
ed on the lateral faces along 1 and 2 directions, respectively. Here, essential features of the actual physical problem. The nite
u1 ; u2 and u3 denote the three components of the displacement eld element method determines an approximation to the exact
u in 1, 2 and 3 directions, respectively. Using the model described solution of the mathematical model. The computed solution should
by (5), the total potential Pu can be dened by substituting u in be compared with exact solution of the mathematical model which
Eq. (8). is being solved. Hence, the computed results can be used to make
The approximate solution to the problem, uh , is the minimizer engineering decisions only when one can guarantee that nite
of the total potential Pp uh and is obtained from the solution of
the following weak problem:
Find uh 2 H0 V such that
B uh ; v h F v h 8 vh H 0 V 9
where
Z
h T h
B uh ; v h ru e v dV; 10
V
Z Z
y
F vh T 3 v h3 ds T 1 v h1 T 2 v h2 ds 11
R [R CN
and H0 V v h j Pp v h < 1 and M v h 0 on CD ; v h is the test
function and has the same form as uh given by (2). We will further x
p
dene Bv ; v kv kE as the energy norm. Note that Fig. 2. Rectangular domain with a circular cutout, meshed with an advancing front
h h
B u ; u 2U uh where U uh is the strain energy for the solution method based automatic mesh generator.
22 P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935
^x x xc
tionally expensive. Hence, the more economical recovery (or pro- 16
^
y y yc
jection) based error estimators have been employed in this study.
It was found in [42,43] that the error estimator based on stress The monomials qi are given as,
recovery (dened in [29]) was reliable locally for patches at the
^; q4 ^x2 ; q5 ^xy
q1 1; q2 ^x; q3 y ^2 ; . . .
^ ; q6 y 17
boundary of the domain, as well as the interior of the domain.
Several denitions of such projections are possible (see [42,44]). Now, to recover a smoothened strain eld we should nd
In this study a simple procedure for the recovery of strains, from smoothened strain components. To get these coefcients, as afore-
the nite element solution, using patchwise data is proposed (as an said, the strain recovery procedure uses the principle of minimiza-
extension of the method in [45]). More details of this estimator can tion of energy norm of the error, i.e. the energy due to errors in
be seen in earlier works of authors [46,47]. These recovered strains strain and stress components, over the patch considered. A typical
will be then used to design a simple error estimator. patch, with a layer of one element neighborhood, over element s is
shown in Fig. 3. In this, the strain components of the nite element
3.1. Procedure for recovery of strains solution are known. The material properties and other relevant
information about patch is also available. From this, the strain
Following the representation of the solution by Eq. (7), we energy of the error can be computed as,
re-write the components of strain in the following form as Z Z d
8 0 9 1 2
8 9 8 1 9 8 9 8 2 9 J h Q h dz dA 18
>
> exx >
> >
>
>
> e >
exx >
> xx> >
>>
> 0 > > >
> exx > > 2 Apatch z2d
>
> >
> >
> >
>
1 >
> >
0 >
>> >
> >
> >
2 >
>
> >
< eyy >
=
>
> e >
>
< yy =
<
e >
> >
yy >
>
<>
= 0 >
>
=
>
< eyy >
> >
=
1 where and h are the recovered and nite element strain vectors
cyz cyz0 z 0 z2 cyz z3 0 12
>
> > > > > > > > > > and Q is material stiffness matrix, respectively. Note that J is the
>
>
> cxz >
>
>
>
>
>
> 0 >
> cxz >
>
>
>
>
>
> 0 >
>
>
>
>
>
>
>
>
cxz1 >
>
>
>
>
>
> 0 >
>
>
>
>
: ; >
> > >
> >
: >
; >
: >
; >
: >
2 ;
strain energy of the error in the strain, h . In Eq. (18), although,
c : 0 ; c 1
0 c
xy cxy xy xy
the material stiffness matrix Q varies over the laminate thickness, it
The recovered strain is also assumed to have the same form is constant over a given lamina thickness. Hence, the integration
(in terms of z) as the exact one, Eq. (7). Thus, the recovered strain is over laminate thickness can be written as sum of integration over
also represented as the individual lamina thicknesses. Further, the integration over
8 9 8 ;0 9
8 ;1 9 8 9 8 ;2 9 patch area can be written as sum of integration over the element
>
> exx >
>
>
>
>
exx >
>
> exx >>
> > >
> 0 > > >
> exx > > areas in the patch. Thus, the above equation can be written as
>
> e >> >
> > > > > > > >
>
>
< yy >
>
=
>
>
> ;1 >
e;0 >
>
< eyy >
>
< yy >
>
=
>
=
>
>
< 0 >
>
>
>
=
>
> ;2 >
< eyy >
> >
= N P 1 Z
"
LAY Z zl n
#
;0 ; 1 1X X
N
o
cyz cyz
z 0 z cyz
2
z 3
0 J h l h
Q dz dA 19
>
> > > > > > > > > > 2 i0 sP l1 zl1
>
>
> cxz >
>
>
>
>
>
>
>
;0 >
cxz >
>
>
>
>
>
> 0 >
>
>
>
>
>
> ;1 >
> cxz >
>
>
>
>
>
> 0 >
>
>
>
i
>
: ; >> > >
> >
: >
; >
: >
; >
: >
;
cxy : ;0 ; c; 1
0 c ; 2
cxy xy xy
13
Given the representation of as in Eq. (13), it is now desired to Material 1 Material 2
obtain the recovered strain eld as a polynomial, element by p3
4p 2p p
element, such that the recovered strain components are poly-
1
nomials that are one order higher than the corresponding nite p p
5p 1p 3 2
element strain components. Thus, if the order of approximation = p0 11
p p
=
for elements p is employed then all the recovered in-plane strain 6p p
4 p 0
components will be polynomials of degree p and the out of plane 10
p 5
7p 8p
p
strain components will be polynomials of degree p 1. The 9p 6
representation of the recovered strain components in terms of
0
polynomials are given below. For example, xx is given as (a) (b)
p1X
p2=2
0 Fig. 3. One nelemento neighborhood patch P s over an element n o s consisting of
xx0 exx;i qi 14 p p
elements (a) sj ; j 0; 1; 2; . . . ; 11 in one material and (b) sj ; j 0; 1; 2; . . . ; 6
i1 in two different materials.
P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935 23
where NP is the number of elements in the patch over an element s, for testing the quality of a-posteriori error estimators for general
NLAY is the number of laminae in the laminate, Q l is the material second order elliptic problems. A detailed study was carried out
stiffness matrix and zl and zl1 are the top and bottom coordinates by the authors in [46] to ascertain the quality of the estimator.
of lth lamina, respectively. The procedure was based on the basic idea given in [44]. The pro-
The minimization of the function J with respect to each cedure is presented in the following for the sake of completeness.
unknown coefcient of recovered strain term gives as many For more details, see the earlier work of authors in [46].
linearly independent equations as the number of coefcients in Let x be a small subregion of interest, lying inside the domain
Eq. (18). The coefcients are solved for each element patch and X. Then asymptotically, for x sufciently small, the nite element
the values are retained for the element s. solution is essentially the best approximation of the local p 1th
The above step is followed by reconstruction. In this, using order Taylor series expansion of the exact solution u, over a region
these coefcients, smoothened strain components can be slightly bigger than x . It was assumed for the asymptotic error
constructed over the patch. These strains, in turn, can be used to analysis that all global contributions to error in the local region
construct the stress components. The in-plane stresses are x (i.e. pollution error) were negligible. Further, it was assumed
constructed using material properties, i.e. from the constitutive that the dominant part of the local error was due to the p 1th
equation. Although, the recovered shear strains are improvements degree terms of the local Taylor series expansion of the exact
over the nite element strains, equilibrium equations are used to solution.
get the transverse stress eld as the model used itself will not give However, for laminated composite plates, limited detailed
good transverse stress components. This is because the strain interior analysis of local error exists (for example see the earlier
continuity through the laminate thickness has been assumed and works of authors in [4650]). In [46], the work of [51,52] was fol-
it will lead to discontinuity of stresses at the interfaces. lowed to get the global component of the error (for a rectangular
The strain energy from recovered and nite element strains are plate) in a local region x due to only boundary layer effect. The
calculated over the whole domain. effect of the thickness of the plate, d, on the convergence rate is
Z seen through a slowing down in the setting of asymptotic behavior,
1XNEL
i.e. a more rened mesh may be required to get asymptotic behav-
U r dV patch
2 j1 j ior. This phenomenon is also known as locking effect. It is known
Z 20 that the h-version of the nite element method can be used to con-
1XNEL
trol the boundary layer effect by using sufcient mesh renements
Uh rh h dV patch
2 j1 j near the boundaries. Assuming that the thickness d is xed (away
from zero), for the error e u uh , we can write
where U and U h are the strain energies over the whole domain
l
from recovered strains and nite element strains, respectively. r jjejjEX 6 C d h 23
and rh are the recovered and nite element stresses, and h are
the recovered and nite element strains, respectively and NEL is where l minp; r and r depends on the regularity of the solution
the number of elements in the mesh constructed over the domain. u of the plate model; k kEx is the energy norm given by
p
Remark: When there is a material discontinuity, the averaging kukEx 2 U x u. Here, U x is the strain energy of u over region
is done over the elements with same material. For example, given x.
the element sP0 , the patch consists of elements fsPi gi 0; 1; . . . ; 6 Further, assume that for a subregion x 2 X, sufciently away
(see Fig. 3(b)). from the boundary
Remark: No attempt is made here to obtain a smoothened
p
stress or strain eld (as prescribed in [30,31]). It is expected, fol- jjejjEx 6 C d h 24
lowing the work of [42,45], that the recovered strain eld will be
more accurate than that obtained by the nite element solution. in the absence of boundary layer effects, and for a xed d. The
readers are referred to the work of [53] for a detailed proof on
convergence of local error for isotropic plates. Thus, if the nite
4. Denition of a-posteriori error estimator based on strain
element solution is obtained over the same mesh using p 1
recovery p1
order elements, the error ep1 u uFE in the nite element
p1
The recovered strain can be used to dene an a-posteriori solution uFE satises
estimate of the error. The element error indicator gs , for an p1
element s is given as: jjep1 jjEx 6 C d h 25
Z "NLAY
XZ n
#
zi o Hence, we can obtain,
g2s h Q i h dz dA 21
p1 p1 p1
s i1 zi1 jjejjEx jju uFE uFE uh jjEx 6 jju uFE jjEx
p1 p1
26
The element error indicators can be used to dene the global jjuFE h
u jjEx jjuFE h
u jjEx
error estimator nX as:
v thus, Eq. (26) means that the error is essentially the difference
u NEL
uX between the p 1th order solution and the pth order solution,
nX t g2s 22 p1
s1 when h ! 0. Here it is to be noted that uFE denotes the nite ele-
ment solution uh obtained with an approximation of order p 1.
The error estimator based on the recovered strain developed Letting nx be the error estimator for subregion x
, we dene
above has to be tested for robustness and accuracy. Following
the work of [42,45], it is imperative to subject an estimator to rig- nx nx
jx 27
orous bench-marking tests in order to ascertain the quality of the jjejjEx jjuFEp1 uh jjEx
estimator for the class of materials, domains, loading and boundary
conditions of interest. In [44,45], a rigorous mathematical proof where jx is the effectivity index for the subregion x
. Ideally,
was given, which leads to a simple computer-based procedure jx 1 is desired.
24 P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935
The details of this work for laminated plates with various stack- 5.1. Optimization algorithm
ing sequences, boundary conditions, plate length to thickness
ratios, materials, mesh patterns, etc. can be seen in earlier work Complex Search Method algorithm is used for solving this
of authors [46]. The quality of the proposed error estimator was optimization problem [32]. The algorithm begins with a number
also studied in [46]. of feasible points created at random. If a point is found to be infea-
sible, a new point is created using previously generated points.
Usually, the infeasible point is pushed towards the centroid of the
4.1. Adaptivity
previously found feasible points. Once a set of feasible points is
found, the worst point is reected about the centroid of rest of the
The energy norm of the error obtained using the recovered
points to nd a new point. Depending on the feasibility and function
strain eld is used to rene the mesh. The procedure developed
value of the new point, the point is further modied or accepted. If
is very simple. It involves the computation of the energy norm of
the new point falls outside the variable boundaries, the point is
the error in each element, followed by a ranking of the elements
modied to fall on the violated boundary. If new point is infeasible,
in the order of highest contributions to the total error. The
the point is retracted towards the feasible points. The worst point in
elements contributing eighty percent or more of the maximum
the simplex is replaced by this new feasible point and algorithm
error are rened. This procedure is repeated till convergence to
continues for next iteration. Here, the reection parameter and con-
within the specied tolerances is obtained. It should be noted that
vergence parameter are chosen as 1.3 and 0.01, respectively.
adaptive analysis requires repeated solution of the boundary value
The algorithm given above is presented in the following.
problem, with the modied meshes. Thus, the cost of computation
increases (see [39,45] for details).
Step 1: Assume a bound in x xL ; xU , a reection parameter a
and termination parameters and d.
5. Optimization problem formulation Strep 2: Generate an initial set of P feasible points
For p 1; . . . ; P 1.
The plate with dimensions of X; Y, thickness d and a centrally p
(a) Randomly generate xi ; i 1; . . . ; N
located cutout of elliptical shape with initial size 2a; 2b and p
(b) If xi is infeasible then reset
oriented at h degrees with respect to x axis, is shown in Fig. 4.
Here, 2a is major axis, 2b is minor axis.
1
xp xp x xp 29
The objective of the optimization problem is to minimize the 2
weight of the plate. Since all the laminae of same material are where x is the centroid of previously generated feasible points.
taken, the objective becomes the minimization of the plate mate- Repeat this process until xp becomes feasible.
rial area. Hence, the design parameters are, semi-major axis - a, (c) Else if xp is feasible then continue with (a) until P feasi-
semi-minor axis b and orientation of the elliptical cutout. ble points are generated.
Objective function : Minimize weight of the material; W
(d) For all feasible points evaluate the function value f xp
Or
Set the iteration counter k = 1.
Step 3: Reection step
Minimize area of the material; A X Y p ab
Subject to : State of stress is such that the Tsai Wu failure index 6 0:8 and
(a) Select the point xR such that
Y f xR Max f xp Fmax 30
6 a;b 6 0:4Y
8
28 x of all feasible points except xR
(b) Calculate the centroid
and the new point
The second constraint given above is purely from geometric
constraints point of view. xm x a x xR 31
(c) If xm is feasible and f xm P Fmax , retract half the distance
to the centroid x. Continue until f xm < Fmax ;
y
Else if x is feasible and f xm < Fmax , go to Step 5.
m
MeshI MeshII
MeshIII MeshIV
Fig. 5. Mesh types or patterns used for plate validation study.
Table 6
Comparison of stresses for rectangular cross-ply thick plate under sinusoidal loading with mesh renement for xed p 3.
Table 7
Comparison of stresses for rectangular cross-ply thick plate under sinusoidal loading with p renement for xed Mesh-III.
Table 12 Table 13
Comparison of non-dimensionalized transverse displacement for thin cross-ply Comparison of non-dimensionalized transverse displacement for thin cross-ply
laminate with mesh renement for p 3 xed. laminate with p renement for xed Mesh-III.
Mesh pattern 0
w Pp p 0
w Pp
Mesh-I 0.4336 0.0605 1 0.0388 0.0530
Mesh-II 0.4352 0.0612 2 0.4255 0.0600
Mesh-III 0.4353 0.0614 3 0.4353 0.0614
Mesh-IV 0.4353 0.0614 4 0.4353 0.0614
Kant et al. [38] 0.4344 Kant et al. [38] 0.4344
P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935 29
Table 18
Comparison study for rst-ply failure load for 45=45= 45=45 laminate with h renement for p 2 xed.
Table 19
Comparison study for rst-ply failure load for 45=45= 45=45 laminate with p renement for xed Mesh-II.
Table 20
Comparison study for rst-ply failure load for 0=90S laminate with h renement for p 2 xed.
Table 21
Comparison study for rst-ply failure load for 0=90S laminate with p renement for xed Mesh-II.
y 2
1
*
1
229 mm
127 mm
*
2
1 2
Fig. 6. Laminate showing the failure locations predicted from present study and reported in [57].
P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935 31
Remark: In above example cases the transverse normal stress, transversely with uniformly distributed load, is analyzed for shape
that is r33 , is not considered while evaluating the failure index in optimization. The intensity of the transverse uniform load is
Eq. (34). This is because this component cannot be computed q0 2:0 N/mm2. The lamination scheme considered is 0=90S with
either from recovered strain data or nite element data, directly. S 10. Here, Y X2 is taken.
Further, no attempt has been made to evaluate this component All laminae are made of T300/5208 Graphite/Epoxy (pre-preg)
using equilibrium equations. and have a thickness of 0.127 mm. The material mechanical prop-
erties are given in Table 14 and strength parameters are given in
Table 17. Here, p 2 is used in the study.
6.3. Shape optimization
The optimum cutout shape results are tabulated in Table 22 and
the corresponding cutout shapes during optimization process are
Here, with few typical examples, we will demonstrate the effect
shown in Fig. 7. From these results we observe that:
of control of discretization error on the nal optimal shape of the
cutout and value of failure index attained. In this study, cross-ply
1. The discretization errors are below 5% for all the feasible
and angle-ply symmetric laminates with cutouts have been
shapes. Hence, adaptive renement is not needed.
studied.
2. The initial conguration with the given loading had a maximum
Remark: In each of the problems, we start with an initial prole
failure index of 0.4095. Thus, there was scope for reduction in
of the elliptical cutout to be circular with a b Y=8, and h 0
.
weight by increasing the size of cutout, which in turn may
Since the p 2 approximation, with sufciently rened mesh,
increase the failure index value. This fact is seen through the
gives reasonable failure load values, this is taken as the order of
intermediate conguration during the optimization procedure.
approximation for all future computations.
The size of the cutout is increased, thereby increasing the maxi-
mum value of failure index.
6.3.1. A rectangular 0=90S laminate under transverse uniform loading 3. The maximum failure index for optimal solution is very close to
A rectangular symmetric laminated plate, with a cutout, the allowable value set (FLD 6 0:8).
clamped along one of the smaller edge (that is, x 0) and loaded 4. The reduction in weight for this study is 3.68%.
Table 22
Optimal shape for cutout in 0=90S laminate without adaptive renement and p 2.
y y
x x
Mesh I a=b=0.3175, max. index=0.4095, error=3.658% Mesh II a=0.5799, b=0.4109, max. index=0.7709, error=4.867%
y y
x x
Mesh III a=0.6057, b=0.41, max. index=0.7959, error=4.8293% Mesh IV a=0.6015, b=0.4119, max. index=0.7899, error=4.385%
Fig. 7. Cutout shapes during optimization in 0=90S laminate under uniform transverse load.
Table 23
Optimal shape for cutout in 45= 45S laminate without adaptive renement and p 2.
y y
x x
Mesh I Mesh II
y y
x x
Mesh III Mesh IV
y y
x x
Mesh V Mesh VI
Fig. 8. Adaptive renements in optimum shape for 45= 45S laminate under uniform transverse load.
Table 24
Effect of adaptivity on nal optimal shape in 45= 45S laminate under transverse
uniform load. Table 26
Effect of adaptivity on optimal shape in 45= 45S laminate subjected to combined
Renement level error (%) Max. index loading.
Initial 15.6503 0.7712 Adaptivity a b Weight Max. index error (%)
First 15.0825
Second 13.5012 No 1.0155 0.8753 10.1103 0.7997 8.3792
Third 12.4125 Yes 1.0157 0.8258 10.2681 0.7999 6.4297
Fourth 11.0823
Fifth 10.4591 1.0089
5. The failure occurs at the edges of cutout boundary. When the 1. The discretization errors are above 10% in most of the feasible
contribution to the failure index from each of the stress compo- shapes.
nent is studied in detail, it was seen that the stress component 2. The reduction in weight obtained is about 22%.
ryy contributes more as compared to other components. Thus, 3. The nal optimum shape has reached the upper bound on the
the mode of failure can be said to be a matrix failure. This result design variables, that is, the minor and major axes of the ellip-
is in accordance with the results obtained by Ericson et al. [58]. tical cutout. Although, the value of failure index is very close to
6. The optimal shape of the cutout is an ellipse with an orientation the allowable value set for this optimum shape, there is scope of
angle of 0
with respect to x axis. additional material removal from this design.
P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935 33
4. Failure occurs at the edges of cutout boundary. This is due to are taken. Here, N xx and N xy denote uniform axial stress along x
high stress concentration because of constraint and free-edge direction and shear stress in x y plane per unit length, respec-
effects. Further, it was observed that the location of rst-ply tively, applied on the edges.
failure changes from one iteration to another in the cutout In this study rst the optimal cutout shape is obtained without
vicinity. controlling the discretization error and then the optimal shape is
5. When the contribution to failure index from each of the stress obtained with control of discretization error. Again, the mesh is
component at the point of failure was analyzed, it was seen that rened adaptively till either the tolerance in discretization error
the stress components ryy and sxy contribute most. Thus, it can is reduced below 7% or for ve levels of renement, whichever is
be said that the mode of failure in this case is matrix failure. attained earlier. The results are reported in Table 26. Here, we
This result is in accordance with experimental study reported see that the nal optimal shapes obtained with and without adap-
by Herakovich [54]. tivity are closer to each other in all respects. This is because the
discretization error in both the cases is close. Further, it is seen that
In the above study the discretization error was not controlled. failure occurs near the cutout boundary, but the location changes
The discretization error for the optimal shape is over 15%. Now, during the iterations. Thus, it can be concluded that if the dis-
without changing the optimal shape obtained, an attempt is made cretization error is within specied tolerance, then there is no
to control the discretization error below 7% or ve renement effect of adaptivity on nal design. Note that using higher p
levels, whicever is attained earlier. The sequence of adaptive mesh (p P 3) will not result in signicant change in the conclusions, as
renements is shown in Fig. 8. Mesh-I is the initial mesh used. The the discretization error is controlled.
discretization errors for these rened meshes are reported in Remark: The optimization algorithm used here works with a set
Table 24. of feasible points. In the present study the initial set of feasible
The error estimator predicts that the elements of the clamped shapes are chosen such that the failure index is closer to 0.8.
edge and the cutout boundaries have high discretization error. Further, it is seen that during each iteration the feasible shape
Hence, a relatively rened mesh is seen at these locations. This attained is such that failure index reaches close to the maximum
result is in agreement with the earlier study of the authors [46]. value of 0.8. Therefore, the evolution of failure index with iterations
The higher discretization error near the constraint and free edges will be almost a horizontal line. Further, it is interesting to see the
of the cutout are expected due to constraint and free-edge effects. evolution of optimal shape with iterations. Here, this has been
Thus, due to adaptive renement the errors are reduced from depicted for Examples 6.3.1 and 6.3.2 for some of the iterations.
15.6503% to 10.4591%. However, the failure index has increased Remark: The proposed approach is very efcient as it simultane-
from 0.7712 to 1.0089. This shows that the optimal solution ously estimates and controls the discretization error in each itera-
obtained without control over discretization error is unsafe. tion. The cost of computation involved with this approach, with
Hence, control of discretization error is essential in shape lower value of approximation, can be signicantly lower than that
optimization study. without the control of discretization error but a highly rened mesh
Now we will see the effect of discretization error control, and/or higher value of approximation. It was shown in [42,45] that
through adaptive mesh renements, on the outcome of the recovered strain eld is accurate than that obtained by the nite
optimization process. The mesh is rened adaptively till either element solution. This leads to the accurate computations of stres-
the tolerance in discretization error is reduced below 7% or for ve ses used in the optimization constraints evaluation. Further, it was
levels of renement, whichever is attained earlier. The comparison shown in [42,4547] that the ZZ type approach of estimating and
of results with and without adaptivity is shown in Table 25. controlling the discretization error is computationally economical.
From these results it is seen that: Therefore, the overall approach used for shape optimization is
computationally economical and accurate as well.
1. The nal optimal shape obtained with and without adaptivity Remark: Here, all computations have been carried out approx-
are signicantly different. The optimal shape of the cutout with- imation order p 2. This has been done intentionally because in all
out discretization error control is circular in nature while the practical purpose computations a lower order of approximation is
one obtained with discretization error controlled throughout used. When a lower approximation order is used and discretization
the optimization procedure is elliptic in shape with orientation error is not controlled then nal solution may not be accurate.
of 0
with x-axis. Here, we wanted to demonstrate this fact through the optimization
2. The discretization error for the nal optimal shape with adap- study. If the approximation order is increased then the recovered
tive procedure is well below the tolerance mentioned. strains will be more accurate. Therefore, the discretization error
3. Failure occurs at the cutout boundary, but the location changes will be lesser as compared to the error when obtained with
from one iteration to another. p 2. However, the nal optimal shape obtained with higher
4. The value of the maximum failure index attained is very close to approximation order may not be different from that one obtained
the allowable value. with p 2.
5. The weight reduction in the optimal design obtained without
control of discretization error (about 22%) is more compared
to one obtained with control of discretization error (about 7. Conclusions
2%). Thus, from this example it is seen that if the discretization
error is not controlled, more material removal will be predicted In this study the shape optimization of laminated composites
erronouysly, hence compromising safety of design. plates with cutout has been studied with a special emphasis on
control of the nite element discretization error. A third order
shear deformation plate theory has been used for the laminate
6.3.3. A rectangular 45= 45S laminate under combined loading analysis and is implemented in a nite element code. A strain eld
A rectangular symmetrically laminated composite plate, having same representation as the exact solution of the displace-
45= 45S clamped along a smaller edge, x 0 is taken. The plate ment eld of the plate model considered is recovered using energy
is loaded under uniformly distributed load, in-plane tensile loading projections over the element patches similar to ZZ type patch
and shear load. The laminate properties are as in Section 6.3.1. For recovery. Then these recovered strains are used to dene a-posteri-
this case, q0 0:15 N/mm2, N xx 20 N/mm and N xy 0:35 N/mm ori error estimator, which drives a simple mesh renement
34 P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935
algorithm. The plate model used is tested for its accuracy in pre- [3] Abrate S. Optimal design of laminated plates and shells. Compos Struct
1994;29:26986.
dicting transverse displacement component, stress components
[4] Kam TY, Lai FM, Liao SC. Minimum weight design of laminated composite
and rst-ply failure load obtained using Tsai-Wu failure theory plates subject to strength constraint. AIAA J 1996;34(8):1699708.
under various boundary; transverse loading conditions, plate [5] Shin YS, Haftka RT, Watson LT, Plaut RH. Design of laminated plates for
thickness ratios and lamination schemes. An optimization algo- maximum buckling load. J Compos Mater 1989;23:34869.
[6] Lee MS, Kikuchi N, Scott RA. Shape optimization laminated composite plates.
rithm using complex search method has been employed. Finally, Comput Methods Appl Mech Eng 1989;72:2955.
the effect of discretization error control over the nal optimal [7] Adali S, Richter A, Verijenko VE, Summers EB. Optimal design of hybrid
design is demonstrated through simple, yet effective examples. laminates with discrete ply angles for maximum buckling load and minimum
cost. Compos Struct 1995;32:40915.
The key conclusions from this study can be given as follows: [8] Adali S, Walker M, Verijenko VE. Multiobjective optimization of laminated
plates for maximum prebuckling, buckling and postbuckling strength using
1. For thick laminates, the plate model used needs moderately continuous and discrete ply angles. Compos Struct 1996;35:11730.
[9] Botkin ME. Shape optimization with buckling and stress constraints. AIAA J
rened mesh and the cubic approximation order used is Technical Notes 1996;34(2):4235.
sufcient. [10] Nagendra S, Haftka RT, Grdal Z. Design of a blade-stiffened composite panel
2. For moderately thick laminates, proper mesh renement with a hole. Compos Struct 1991;18:195219.
[11] Naik NK, Ganesh VK. Optimum design studies on FRP beams with holes.
and approximation orders should be chosen for a better Compos Struct 1993;24:5966.
convergence. [12] Sivakumar K, Iyengar NGR, Deb K. Free vibration of laminated composite
3. For thin laminates, boundary layer effect is severe. Mesh plates with cutout. J Sound Vib 1999;221(3):44370.
[13] Haftka RT, Grandhi RV. Structural shape optimization a survey. Comput
renements are needed in order to eliminate the effect of
Methods Appl Mech Eng 1986;57:91106.
locking. Higher approximation orders p P 3 is also needed [14] Ding Y. Shape optimization of structures: a literature survey. Comput Struct
in order to reduce the locking effect. Thus, for a thin plate 1986;24(6):9851005.
greater care should be taken (with respect to the approx- [15] Sobieszczanski-Sobieski J, Haftka RT. Multidisciplinary aerospace design
survey of recent developments. Struct Optim 1997;14:123.
imation) in order to ensure reliability of the solution [16] Kaufmann M, Zenkert D, Wennhage P. Integrated cost/weight optimization of
quantities. aircraft structures. Struct Multidisc Optim 2010;41:32534.
4. The transverse displacement component obtained for thick, [17] Kikuchi N, Chung KY, Torigaki T, Taylor JE. Adaptive nite element methods for
shape optimization of linearly elastic structures. Comput Methods Appl Mech
moderately thick and thin cross-ply and angle ply laminates Eng 1986;57:6789.
for various boundary conditions under uniform as well as [18] Hinton E, zakca M, Rao NVR. An integrated approach to structural shape
sinusoidal transverse loads are accurate. Further, the stres- optimization of linearly elastic structures. Part II: Shape denition and
adaptivity. Comp Sys Eng 1991;2:4156.
ses obtained using recovered strains are reasonably [19] Weck M, Nottebaum T. Adaptive meshing saving computational costs during
accurate. the optimization of composite structures. Struct Optim 1993;6:10815.
5. For angle ply laminates the failure starts at the top or bottom [20] Bennett JA, Botkin ME. Structural shape optimization with geometric
description and adaptive mesh renement. AIAA J 1985;23(3):45864.
edge of the laminate. The failure is dominated by ryy and sxy [21] Schleupen A, Maute K, Ramm E. Adaptive FE-procedures in shape
stress components. This results in matrix type failure in the optimization. Struct Multidisc Optim 2000;19:282302.
laminate. [22] Morin P, Nochetto RH, Pauletti MS, Verani M. Adaptive nite element method
for shape optimization. ESAIM: Control Optim Calculus Variations
6. For cross-ply laminates, the failure starts at the top or bot-
2012;18:112249.
tom edge of the laminate. The failure mainly occurs due to [23] Tian X, Wang MY. h-Adaptive extended nite element method for structural
stress component ryy . Again, this gives rise to matrix type optimization. 10th World Congress on Structural and Multidisciplinary
Optimization. USA: Orlando, Florida; 2013.
failure in the laminate.
[24] Bugeda G, Rdenas JJ, Oate E. An integration of a low cost adaptive remeshing
7. In case of laminates with cutout, the failure starts at the cut- strategy in the solution of structural shape optimization problems using
out boundary. This is because of high stresses due to free- evolutionary methods. Comput Struct 2008;86:156378.
[25] Bruggi M, Veranio M. A fully adaptive topology optimization algorithm with
edge effect at the cutout boundaries.
goal-oriented error control. Comput Struct 2011;89:148193.
8. With the use of stresses obtained using recovered strains, [26] Wang Y, Kang Z, He Q. An adaptive renement approach for topology
failure is predicted earlier than that obtained using the nite optimization based on separated density eld description. Comput Struct
element stresses. These failure loads are 518% lesser than 2013;117:10122.
[27] Reddy JN. A simple higher-order theory for laminated plates. J Appl Mech
the reported results for angle-ply laminates. Further, these 1984;51:74552.
values for failure load are about 12% lesser than the values [28] Zienkiewicz OC, Zhu JZ. A simple error estimator and the adaptive procedure
reported in literature for cross-ply laminates. for practical engineering analysis. Int J Numer Methods Eng 1987;24:
33757.
9. Adaptive renement procedure based on a-posteriori error [29] Zienkiewicz OC, Zhu JZ. The superconvergent patch recovery and a-posteriori
estimation affects the nal optimal solution. The nal design error estimates. Part 1: The recovery technique. Int J Numer Methods Eng
obtained with a cruder approximation is dangerous, because 1992;33:133164.
[30] Rank E, Zienkiewicz OC. A simple error estimator in the nite element method.
more material removal is predicted if the discretisation error Comm Appl Numer Methods 1987;3:2439.
is high (20% more material removal predicted). This is [31] Zhu JZ, Zienkiewicz OC. Superconvergence recovery technique and a-posteriori
because the plate behavior is stiffer with a crude approx- error estimators. Int J Numer Methods Eng 1990;30:132139.
[32] Deb K. Optimization for Engineering Design, Algorithms and Examples. New
imation, as compared to that obtained with proper rene-
Delhi: Prentice Hall of India Private Limited; 1998.
ments and/or approximation order. [33] Tsai SW, Wu EM. A general theory of strength for anisotropic materials. J
10. This study brings out the important point that it is impera- Compos Mater 1971;5:5880.
[34] Lo KH, Christensen RM, Wu EM. A higher-order theory of plate deformation
tive to have some indication of the discretisation error, in
Part-2: Laminated plates. J Appl Mech 1977;44(4):61669.
order to accept or reject the nal design obtained from the [35] Actis RL, Szab BA, Schwab C. Hierarchic models for laminated plates and
shape optimization. shells. Comput Methods Appl Mech Eng 1999;172(14):79107.
[36] Pandya BN, Kant T. A rened higher-order generally orthotropic C0 plate
bending element. Comput Struct 1988;28:11933.
[37] Pandya BN, Kant T. Finite element stress analysis of laminated composite
References plates using a higher-order displacement model. Comp Sci Technol
1988;32:13755.
[38] Pandya BN. Higher-order theories and nite element evaluations for
[1] Park WJ. An optimal design of simple symmetric laminates under the rst ply
multilayer composite plates. Ph.D. dissertation, Indian Institute of
failure criterion. J Compos Mater 1982;16:34155.
Technology Bombay, India; 1988.
[2] Fukunaga H, Sekine H. Optimum design of composite structures for shape,
[39] Babuka I, Strouboulis T. The nite element method and its
layer angle and layer thickness distributions. J Compos Mater
reliability. Oxford: Oxford Science Publications; 2001.
1993;27(15):147992.
P.M. Mohite, C.S. Upadhyay / Computers and Structures 153 (2015) 1935 35
[40] Verfrth R. A review of a-posteriori error estimation and adaptive mesh- [49] Mohite PM, Upadhyay CS. Accurate computation of critical local quantities in
renement. New York: Wiley; 1996. composite laminated plates under transverse loading. Comput Struct
[41] Verfrth R. A review of a-posteriori error estimation techniques for elasticity 2006;84(1011):65775.
problems. Comput Methods Appl Mech Eng 1999;176:41940. [50] Mohite PM, Upadhyay CS. A generalized adaptive nite element analysis of
[42] Babuka I, Strouboulis T, Upadhyay CS, Gangaraj SK, Copps K. Validation of a- laminated plates. Comput Struct 2012;112113:21734.
posteriori error estimators by numerical approach. Int J Numer Methods Eng [51] Suri M, Babuka I, Schwab C. Locking effects in the nite element
1994;37:1073123. approximation of plate models. Math Comput 1995;64(210):46182.
[43] Babuka I, Strouboulis T, Mathur A, Upadhyay CS. Pollution error in the h [52] Yunus SM, Pawlak TP, Wheeler MJ. Application of the Zienkiewicz-Zhu error
version of the nite element method and the local quality of a-posteriori error estimator for plate and shell analysis. Int J Numer Methods Eng
estimators. Finite Elem Anal Des 1994;17:273321. 1990;29:128198.
[44] Babuka I, Strouboulis T, Upadhyay CS. A model study of the quality of a- [53] Gastaldi L. Uniform interior error estimates for the ReissnerMindlin plate
posteriori error estimators for linear elliptic problems. Error estimation in the model. Math Comp 1993;61(204):53967.
interior of patchwise uniform grids of triangles. Comput Methods Appl Mech [54] Herakovich CT. Mechanics of brous composites. Delhi: John Wiley and Sons,
Eng 1994;114:30778. Inc; 1998.
[45] Upadhyay CS. Computer-based analysis of error estimation and [55] Mohite PM. Composite materials and structures, NPTEL course. http://nptel.ac.
superconvergence in nite element computations. Ph.D. dissertation at Texas in/courses/101104010/.
A&M University, Texas; May 1997. [56] Reddy JN, Miravete A. Practical analysis of composite laminates. New
[46] Mohite PM, Upadhyay CS. A simple strain-recovery based a-posteriori error York: CRC Press; 1995.
estimator for laminated composite plates. J Aerospace Sci Technol [57] Reddy YSN, Reddy JN. Linear and non-linear failure analysis of composite
2007;59(3):21222. laminates with transverse shear. Comp Sci Technol 1992;44:22755.
[47] Mohite PM, Upadhyay CS. Local quality of smoothening based a-posteriori [58] Ericson K, Persson M, Carlsson L, Gustavsson A. On the prediction of the
error estimators for laminated plates under transverse loading. Comput Struct initiation of delamination in a 0=90s laminate with a circular hole. J Compos
2002;80(1819):147788. Mate 1984;18:495506.
[48] Mohite PM, Upadhyay CS. Focussed adaptivity for laminated plates. Comput
Struct 2003;81(5):28798.