Académique Documents
Professionnel Documents
Culture Documents
Composite Structures
journal homepage: www.elsevier.com/locate/compstruct
a r t i c l e i n f o a b s t r a c t
Article history: A technique for the design optimization of composite laminated structures is presented in this work. The
Available online 17 May 2008 optimization process is performed using a genetic algorithm (GA), associated with the finite element
method (FEM) for the structural analysis. The GA is adapted with special operators and variables codifi-
Keywords: cation for the specific case of composite laminated structures optimization. Some numerical examples are
Multiobjective optimization presented to show the flexibility of this tool to solve different kinds of problems. Two cases of multiob-
Genetic algorithms jective optimization of plates under transverse or in-plane load are studied. In these examples the min-
Composites laminated structures
imization of two objectives, such as weight and deflection or weight and cost, are simultaneously
Finite element analysis
performed and a pareto-optimal set is obtained by shifting the optimization emphasis using a weighting
factor. The stiffness maximization of a composite shell under pressure load is presented in the last exam-
ple, where the geometrically nonlinear behavior of the structure is considered. Some aspects of the opti-
mization performance, such as the apparent reliability and the computational cost, are investigated in
each application.
Ó 2008 Elsevier Ltd. All rights reserved.
0263-8223/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruct.2008.05.004
444 F.S. Almeida, A.M. Awruch / Composite Structures 88 (2009) 443–454
The organism fitness is obtained directly from an objective func- very common constraint adopted in most of optimization prob-
tion using simple structure information and gradient evaluations lems, and it is used in this work too. The Tsai–Wu failure criterion
are not required. [7] is used for the failure prediction in a ply. A safety factor against
In real designs cases, when the structural geometry is usually failure kf can be obtained in linear problems with the Tsai–Wu fail-
complex and the prediction of the structural behavior must be ure function using the material strength parameter for traction,
accurate, it is necessary to use numerical tools, such as the FEM, compression and shearing at each of the principal material axes.
for the structural analysis. These methods are computationally In geometrically nonlinear analysis the structural failure is verified
expensive and may turn the optimization processes with GA at each load step of the incremental solution method, which is
impracticable when a large amount of analyses is required. Many stopped if material failure is detected.
researchers have proposed modifications to the classical GA struc-
ture to take advantage of composite laminates characteristics and
minimize the computational cost. Some of these new strategies 3. Genetic algorithms for composite laminate structure
are applied in this work, consisting essentially in a GA restructur- optimization
ing of the variable codification and the genetic operators.
When more than one objective is handled in the design process, This work uses a GA provided with many modifications with re-
a multiobjective optimization problem may be solved considering spect to the classical GA structure stated by Goldberg [4]. Although
all the objectives simultaneously, providing a set of optimum de- the main concepts and the sequence of operations remain similar
signs (pareto-optimal set), depending of the emphasis given to to the original formulation, a new scheme for the variable codifica-
each one of the objectives. The pareto-optimal set may be very use- tion and special genetic operators are introduced, increasing the
ful when the critical objective is not known a priori. Various performance of the method in the case of composite structures
researchers have studied the problem of multiobjective optimiza- optimization.
tion of laminated structures, but the use of this approach together An initial population containing P organism is first created in a
with GA and the finite element method (FEM) have not been random process. In order to create successive generations, parents
widely explored. are chosen from the current population based on their fitness.
In the present work, two examples of multiobjective optimiza- Next, the genetic operators are applied to create children in order
tion of composite laminate plates using GA and FEM are studied. A to form a children population. An elitist selection scheme is used to
third application deals with the single optimization of a semi- obtain the new generation taking organisms from the current pop-
cylindrical shell considering the geometrically nonlinear behavior ulation and from the children population just created. This process
of the structure. is repeated until the convergence criterion is met. A description of
the variable codification, the genetic operators and the selection
scheme for the construction of new generations are given in the
2. Composite structure analysis following subsections.
Real problems of composite structure design depend of reliable 3.1. Composite laminate codification
structural analysis. In the case of composite laminates, the deter-
mination of the mechanical behavior is difficult even for simple A pair of chromosomes is adopted for the representation of each
geometric configurations. It happens because of some complex laminate in this work, similar to the scheme implemented by Sore-
mechanisms inherent to the material, like coupling between mekun [8]. As only symmetric laminates are studied in this work,
stretching, bending and twisting deformations, depending on the just half of the layers need to be coded in the gene strings, and
stacking sequence. These coupling effects and the usually complex so the total number of genes in a chromosome is proportional to
structural geometric configurations inhibit the use of closed-form half of the maximum admissible number of layers in a specific de-
solutions in the structural analysis. In consequence, numerical sign problem. Each layer is represented by a pair of genes (with one
method must be adopted for the accurate prediction of the struc- gene in each chromosome), being the first pair referred to the out-
tural behavior. ermost layer and the succeeding pairs of genes referred to the in-
In this work, a triangular flat plate and shell element with 18 ner layers. In the first chromosome, named ‘‘orientation
degrees of freedom called DKT (discrete Kirchhoff triangle) is used. chromosome”, is stored information about fiber orientation and
This element was developed by Bathe and Ho [5] for the nonlinear about the number of plies contained by each layer of the laminate.
analysis of isotropic plates and shells. The original formulation was The second chromosome, called ‘‘material chromosome”, is used to
modified by the introduction of specific constitutive matrices of store the material properties of the layer, containing information
the membrane and bending parts in order to allow the analysis like ply thickness, elastic and strength constants.
of symmetric composite laminated structures. Any other element Each design variable must assume one of the admissible dis-
formulation could be adopted, since in GA based optimization no crete values defined in the optimization process. These values are
gradient evaluation is needed, and so no additional modification represented by positive integer numbers. They are used to code
to the analysis tool is necessary. In the solution of geometrically de variable value in its correspondent gene. Since there are two
nonlinear problems an incremental iterative scheme referred as kinds of variables, the orientation and the material variable, two
generalized displacement control method (GDCM) is used (see code alphabets are used in the codification. An empty stack code,
[6]). This method is able to solve nonlinear problems with multiple represented by the number zero, is used in problems where the
critical points and snap-back points allowing the complete deter- variation of the number of layers is considered. This code is intro-
mination of the post buckling behavior of the structure. The only duced in a pair of genes, representing the layer to be deleted, by a
parameter to be set in the GDCM is the basic load increment ki, specific genetic operator. Another genetic operator acts adding lay-
which corresponds to the ratio between the first load increment ers to the laminate by changing the value zero of a pair of genes by
and the complete load. This parameter determines the sensibility any other admissible value. The maximum number of layers in the
of the method to the nonlinear characteristic of the problem. laminate is limited by the number of genes in the chromosome.
Additionally to the structure displacements, the analysis tool In Fig. 1 is presented an example of the decodification of a pair
must be able to determinate accurately the stress components at of chromosomes in a laminate stacking sequence, based in a given
the composite layers in order to predict material failure. This is a orientation and material gene codification alphabets. These chro-
F.S. Almeida, A.M. Awruch / Composite Structures 88 (2009) 443–454 445
3.2.2. Mutation
Mutation is the class of genetic operators responsible to main-
tain the genetic diversity of the population by introducing new
information in the chromosomal strings of each child after it is cre-
ated by the crossover operation. These operations provide a ran-
dom search capability to GA, which may be useful to find
promising areas in the design space, and prevent crossover to lose
its effect due to a standardization of the population. In the classical
implementation of this operator, to each gene is given a small
probability to switch to any other permissible value, excepting
its current value [3].
In spite of the randomness of the mutation process, it is possible
to incorporate to this operator some knowledge about the response
of composite laminates with respect to the alteration of one or Fig. 3. (a) Orientation and material alteration; (b) ply addition; (c) ply deletion.
PN i
i¼1 X g P
An ¼ ; ð1Þ
N
where X ig is the total number of generations analyzed in the ith opti-
mization procedure. In GA optimization procedures it is very com-
mon to occur repeated analyses because one specific design may
Fig. 4. Gene swap operation.
appear in many generations during the process. When a memory
containing information about the performed analyses is used, asso-
ciated to the GA, these repeated analyses can be avoided, resulting
3.2.3. Gene swap in an important reduction of the computational cost. In these cases,
The main characteristic of the gene swap operator is the ability another measurement of the computational cost can be stated tack-
to modify laminate stack sequence without changes of the total ing the average number of the effectively performed analyses (Ar),
number of plies with fibers oriented on each permissible direction. obtained dividing the number of analyses that were effectively car-
This allows GA to change the bending behavior of the laminate ried out by the number of GA executions (N).
without modifying its in-plane mechanical response. The imple- The criterion to stop the optimization process, which was used
mentation of this operator follows Refs. [8,9], where two pairs of in all examples presented here, is based in two parameters: the
genes are randomly chosen and have their position shifted in the upper limit of the number of generations (NLG) and the maximum
chromosome, resulting on a new stacking sequence. Such opera- number of generations with no improvement of the best design
tion occurs at a given probability pgs, usually with a larger value (NSD). Once one of these limits is reached, the optimization process
than those corresponding to mutation operators probabilities. A is stopped and the best laminate of the last generation is taken as
description of the gene swap operation is given in Fig. 4. the optimization result. NLG and NSD are defined in each optimiza-
tion procedure, depending on the complexity of a specific problem.
3.3. Selection scheme
4.1. Weight and deflection minimization of a composite laminated
There are many ways to obtain the population of successive plate under transverse load
generations in a GA. In classical algorithms new generations are
formed only by children created by genetic operators applied to This example deals with the design of a composite laminated
the current population. This process has many drawbacks since square plate, subjected to a uniform pressure load on its surface.
there is no warranty of improvement or maintenance of achieved Minimization of the structural weight and deflection are the design
evolution when all the old organisms are replaced. To solve this objectives. The two objectives must be considered in conjunction
problem new selection schemes were created, being one of them with the constraints imposed by material failure and maximum
the elitism scheme, which consists in transfer some good organ- values of contiguous plies thickness with the same fiber orienta-
isms from the old population to a new generation, preserving tion. These are two opposite objectives, since improvements in
desirable genetic information. This papers deals with a multiple one of them leads to depreciation of the other, and they must be
elitist scheme proposed by Soremekun [8]. In the implementation, considered at the same time in the optimization, requiring a mul-
both, parent and child populations of size P are independently tiobjective approach. In its formulation, the objective function
ranked from best to worst fitness. These two populations are then must contain both objectives, which are weighted by a factor
combined and ranked together, resulting in a combined population which controls the emphasis given to each one of the objectives
with 2P organisms. Then, the best Ne individuals of the combined in the optimization. As a result of the variation of the weighting
population are transferred to the new generation. The best individ- factor used in the objective function, this problem has a set of opti-
uals of child population that have not already been used are taken mal solutions (pareto-optimal set) instead of a single solution.
to fill the remainder of the new generation. The number of top ele- The structure geometry, boundary conditions and the mechan-
ments (Ne) to be transferred to the new generation is a GA param- ical properties of the composite material are presented in Fig. 5.
eter to be adjusted at each application. The elastic constants are the Young’s modulus in the fiber direction
(E1) and transverse to the fiber direction (E2), the shear modulus
4. Numerical examples and discussion (G12) and the Poisson’s ratio (m12), respectively. Strength parame-
ters for traction and compression for longitudinal and transversal
Three examples of composite laminated structures design using directions are given by F1t, F1c, F2t, and F2c, respectively. The
GA optimization and FEM analysis are presented in the following remainder parameters are the shear strength (F6) and the specific
sections. To prove the success of the optimization procedure and weight (q). The structure must support a design pressure load of
to characterize the design space of the problems, all the possible 0.1 MPa with no material failure (the Tsai–Wu failure function
laminate configurations are previously analyzed. Additionally, N must be lower than 1.0 for the whole plate) and thickness of con-
optimizations are carried out for each example to obtain the algo- tiguous plies with the same fiber orientation must not be greater
rithm reliability and computational cost. This is possible to accom- than 2 mm or less than 0.75 mm.
plish because results of previous analyses of the design space are In the plate design, the laminate is restricted to be symmetric
stored. This information is used by the GA to evaluate the objective with 8 layers, being represented in the GA by a pair of chromo-
function. somes, with 4 genes each one. The fiber orientation angle and
The apparent reliability (R) is determined by taking the number the thickness of each layer are the optimization variables. They
of optimizations where the GA finds at least one global optimum must assume one of the discrete values given in Table 1, where
(No), divided by the total number of applications of the GA (N). It the codification adopted in the GA is also given. The number of
defines the chances of obtaining the global optimum in a single genes combined with the number of possible values of each vari-
optimization process. As the structural analysis employing the able leads the size of the design space (SDS) to be equal to 65536.
FEM is usually the most time consuming task in the optimization In this example the fitness evaluation FIT must consider both,
procedure, the GA cost (An), which is determined by the average weight and deflection reduction, at the same time. It is done by
number of analyses required in a single optimization process, is Eq. (2a), where the fitness value is taken as the inverse of the objec-
given by the following expression: tive function, and the weighting factor a is introduced to allow the
F.S. Almeida, A.M. Awruch / Composite Structures 88 (2009) 443–454 447
where
W W min D Dmin
W ¼ þ 1; D ¼ þ 1: ð2bÞ
W max W min Dmax Dmin
The parameters FF and Tv are introduced in Eq. (3a) to penalize
unfeasible designs. The first one represents the maximum value of
the failure function evaluated in the structure, while the parameter Fig. 6. Weight and central displacement of feasible designs.
448 F.S. Almeida, A.M. Awruch / Composite Structures 88 (2009) 443–454
figure are the designs that form the pareto-optimal set, which must of repeated structures is avoided using a memory, since Ar is found
be obtained by the GA, according to the emphasis given to each of to be about 40–57% of An, depending on the value of a. Further-
the objectives. Details of the pareto-optimal set are presented in more, the number of analyses effectively performed Ar is about
Table 2. 4.82–7.09% of the SDS, which means that only a small part of the
Good reliability levels were obtained for the optimization with design space is explored, but the total number of analyses is still
most of the a values, as can be observed in Table 3, where the col- huge. A reduction of the average value Ar can be obtained when
umn r represents the standard deviation of the apparent reliability the parameters of GA are modified. However, in all tests accom-
R. The loss of reliability in the optimizations using a equal to 0.55 plished here this reduction was obtained together with a reduction
occurs because the GA finds many times the design identified by L on the reliability levels, mainly for values of a presenting critical
instead of the design identified by J, which is the correct solution in reliability levels.
this case. It happens because both designs J and L have practically As can be seen in the Tables 2 and 3, the GA is successful in find-
the same fitness value for a equal to 0.55. The same occurs for a ing most of the pareto-optimal designs, but the designs identified
equal to 0.70, when the GA finds the design identified by O in many by the points B and D are not obtained. These points are possible
optimization processes instead of the design identified by N, which solutions of the optimization, but they are located out of a convex
is the correct solution. In optimizations where a is taken as being curve defined by the other optimal points. Due to this fact the GA
equal to 0.80, 0.85 and 0.95, the low level of reliability occurs be- does not find the points B and D, since the fitness is evaluated as a
cause the designs obtained in these cases have values of the fitness convex combination of the objectives. Figs. 7a and b show the dif-
function very close to the optimal designs identified by O and P. ference of the fitness values of the points B and D with respect to
The average number of analyses required (An), the average their neighbor points in a range of a where the optimal solution
number of analyses effectively performed (Ar) and the ratio be- changes from point A to C and from point C to E, respectively
tween Ar and the size of the design space (SDS), which is equal (see Table 3). The figures show that the fitness of the points B
to 65,536, are also shown in Table 3 for each value of a. An expres- and D are never greater than those of their neighbor points at
sive reduction of the computational cost is observed if the analysis the same time and so they cannot be obtained by the GA, no matter
the value of the weighting parameter.
All optimal designs obtained in the pareto optimal set present
Table 2 outer layers with fibers oriented at 90° and the maximum admis-
Pareto-optimal designs
sible thickness (2 mm). These layers are the most important for
Optimal Laminate Weight Deflection Weighting the plate bending behavior, and the previous characteristics give
design (N) (mm) factor a the best stiffness properties. The inner layers of the staking se-
A [902,0, +452,0, 902,0, 451,0]S 219.7 7.9 0.0–0.20 quence of the different designs vary to obtain the different results
B [902,0, 452,0, 902,0, +450,75]S 211.9 8.8 – for the structural weight and deflection, according to the emphasis
C [902,0, 451,75, 902,0, +451,0]S 204.0 9.6 0.25
given for each of the optimization procedures. The lightest design
D [902,0, 451,75, 902,0, +450,75]S 196.2 10.8 –
E [902,0, 451,0, 902,0, +451,0]S 188.4 11.8 0.30 obtained here has less than the half of the weight of the heaviest;
F [902,0, 450,75, 902,0, +451,0]S 180.5 13.2 0.35 however, the last one presented a displacement more than seven
G [902,0, 450,75, 902,0, +450,75]S 172.7 15.0 0.40 times lower than the first one. The chose of the design to be used
H [902,0, 450,75, 901,75, +451,0]S 164.8 17.3 0.45 in a specific application depends on how critical is the weight for
I [902,0, 450,75, 901,75, +450,75]S 157.0 19.9 0.50
J [902,0, 450,75, 901,0, +451,0]S 149.1 23.3 0.55
such application and on the magnitude of the allowable
L [902,0, 450,75, 901,0, +450,75]S 141.3 27.2 0.60 displacement.
M [902,0, 450,75, 900,75, 450,75]S 133.4 32.1 0.65
N ½901;0
2 ; 45
1;0
901;0 S 125.6 38.5 0.70 4.2. Cost and weight minimization of an in-plane loaded composite
O ½901;0
2 ; þ45
0;75
451;0 S 117.7 46.6 0.75–0.85
laminate plate
P ½901;0
2 ; 0
0;75
þ 450;75 S 109.9 59.5 0.90–1.0
Fig. 7. Differences of the fitness values of points B and D with respect to their neighbor points. (a) When FIT(B) > FIT(A), then FIT(C) > FIT(B); (b) when FIT(D) > FIT(C), then
FIT(E) > FIT(D).
The laminated is supposed to be symmetric and formed by 6–12 the unfeasible designs, driving the GA to feasible areas of the de-
layers. Each layer has 2 plies that may assume the following orien- sign space. A small ‘‘bonus”, proportional to the safety factor, is
tations: 0°2, ±45° and 90°2. They are represented in the orientation incorporated to the objective function in the case of feasible
gene alphabet by the codes 1, 2 and 3, respectively. The material
gene alphabet is formed by the codes 1 and 2, corresponding to
Kevlar-epoxy and Graphite-epoxy, respectively. Empty stacks code
can be used to reduce the number of layers. Due to the symmetry,
only 6 genes are necessary in each of the two chromosomes used to
represent the composite material in the GA. Considering the num-
ber of variables and the number of possible values of these vari-
ables, the size of the designs space (SDS) is 55,944. A distribution
of weight and cost of all the feasible designs in the problem is
shown in Fig. 9. Points A to F in this figure are the designs that form
the pareto-optimal set, which must be obtained by the GA. Details
of the points of the pareto-optimal set are presented in Table 4.
Due to the simultaneous minimization of cost and weight, the
multiobjective approach is used in the formulation of the objective
function, given in Eq. (3), which contains both objectives. The
weighting factor a is introduced to allow a shifting on the empha-
sis given to each of the objectives, driving the GA to converge to
one of the points in the pareto-optimal set, according to the value
attributed to this factor. The constraints of the problem are also
considered in the objective function. In this example, a feasible de-
sign is determined by the safety factor k*, which is given by the
minimum value between kb and kf. A penalization is applied to Fig. 9. Weight and cost of feasible designs.
450 F.S. Almeida, A.M. Awruch / Composite Structures 88 (2009) 443–454
Table 4 Table 5
Designs of the pareto-optimal set Optimization results with GA for the plate with in-plane load
Fig. 10. Shallow shell geometry, boundary conditions and material properties.
452 F.S. Almeida, A.M. Awruch / Composite Structures 88 (2009) 443–454
Fig. 12. Central displacement load level for the optimal design and for a low fitness value.
parameter ki. If the adopted value is too big, the nonlinear solution
is inaccurate or even may not converge. In the other hand, if the va-
lue is smaller than the necessary, the analysis process may became
very slow, which is very undesirable when an optimization by GA
is used. The challenge in this case is to find a value for ki that pro-
vide accurate analyses for all the designs with the minimum com-
putational cost.
In this example a strategy for the automatic refinement of the
parameter ki is implemented to reduce the cost of the analyses
and to guarantee the convergence and the accuracy of the nonlin-
ear solution. The first step in the implementation is carried out
monitoring the structural response during the nonlinear solution
and identifying anomalous behaviors that can be interpreted as
solution instability or error. When this situation occurs, the pro-
cess is stopped and restarted using a smaller value for ki. The
way the structural response is monitored and the factor of reduc-
tion of the parameter ki are problem dependent and can be imple-
mented in different forms. In the present example the error is
identified when the increment in the displacement of the central Fig. 14. Load level central displacement for the best design obtained in optimi-
point of the shell in a specific load step is greater than 10 times zations 1, 2, 3 and 4.
References [6] Yang Y, Shieh M. Solution method for nonlinear problems with multiple
critical points. AIAA J 1990;28(12):2110–6.
[7] Daniel IM, Ishai O. Engineering mechanics of composite materials. Oxford
[1] Walker M, Smith RE. A technique for the multiobjective optimization of
University Press; 1994.
laminated composite structures using genetic algorithms and finite element
[8] Soremekun GAE. Genetic algorithms for composite laminate design and
analysis. Compos Struct 2003;62:123–8.
optimization. M.Sc. thesis (Master of Science in Engineering Mechanics).
[2] Gürdal Z, Haftka RT, Hajela P. Design and optimization of laminated composite
Polytechnic Institute and State University, Blacksbourg, Virginia, USA; 1997.
materials. Wiley & Sons; 1999.
[9] Nagendra S, Jestin D, Gürdal Z, Haftka RT, Watson LT. Improved genetic
[3] Soremekun GAE, Gürdal Z, Haftka RT, Watson LT. Composite laminate design
algorithm for the design of stiffened composite panels. Comput Struct
optimization by genetic algorithm with generalized elitist selection. Comput
1996;58(3):543–55.
Struct 2001;79:131–43.
[10] Almeida FS. Laminated composite material structures optimization with
[4] Goldberg DE. Genetic algorithms in search, optimization, and machine
genetic algorithms. M.Sc. thesis – PPGEC/UFRGS, Porto Alegre, Rio Grande do
learning. Addison-Wesley Publishing Company; 1989.
Sul, Brazil; 2006 [in Portuguese].
[5] Bathe KJ, Ho L. A simple and effective element for analysis of general shell
structures. Comput Struct 1981;13:673–81.