Composite Structures 90 (2009) 1–11

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Review

Optimum stacking sequence design of composite materials Part I: Constant

a r t i c l e i n f o a b s t r a c t

Article history: Designing an optimized composite laminate requires finding the minimum number of layers, and the best
Available online 3 February 2009 fiber orientation and thickness for each layer. To date, several optimization methods have been intro-
duced to solve this challenging problem, which is often non-linear, non-convex, multimodal, and multi-
Keywords: dimensional, and might be expressed by both discrete and continuous variables. These optimization
Composite materials techniques can be studied in two parts: constant stiffness design and variable stiffness designs. This
Optimization paper concentrates on the first part, which deals with composite laminates with uniform stacking
Stacking sequence design
sequence through their entire structure. The main optimization methods in this class are described, their
characteristic features are contrasted, and the potential areas requiring more investigation are
highlighted.
Ó 2009 Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Gradient-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1. Vanishing the function’s first gradient. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2. Steepest descent (SD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.3. Conjugate gradient (CG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.4. Quasi-Newton method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.5. Method of feasible directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.6. Approximation schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3. Direct search methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1. Partitioning methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2. Enumeration search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.3. Simplex method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.4. Random and greedy search. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.5. Simulated annealing (SA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.6. Genetic algorithm (GA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.7. Other heuristic optimization methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.7.1. Tabu search. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.7.2. Scatter search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.7.3. Particle swarm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.7.4. Ant colony. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4. Specialized algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.1. Design with lamination parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.2. Layerwise optimization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.3. Problem partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.4. Discrete material optimization (DMO). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.5. Fractal branch-and-bound method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.6. Knowledge-based methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

* Corresponding author. Tel.: +1 514 690 4151; fax: +1 514 398 7365.
E-mail address: [email protected] (H. Ghiasi).

0263-8223/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruct.2009.01.006
2 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11

5. Hybrid methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
6. Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1. Introduction 4. Hybrid methods. In this class fall methods that combine two
or more optimization techniques to benefit from the
The advantage of composite materials is that they provide strength of all constituent techniques.
excellent mechanical properties. However, using this advantage re-
quires the optimization of shape and size and the proper place- In this paper, we review optimization techniques used for con-
ment of fibers within the material, which gives a good stant stiffness design of composite laminates. The scenario of var-
opportunity to tailor the material properties; however, it increases iable stiffness design will be the focus of a future paper, which will
the complexity of the design problem. This complexity exists, not also analyze other issues, such as multi-objective optimization,
only because of numerous design variables, but also because of discretization techniques, design for manufacturing, sensitivity
having a multimodal and variable-dimensional optimization prob- analysis, and design for uncertainty. The following sections de-
lem with unattainable or costly derivatives. scribe the main optimization methods for constant stiffness design
This paper classifies and compares optimization techniques with respect to the four classes given above. This design scenario is
used for optimal lay-up selection of laminated composite materi- simpler than that for the variable stiffness design, since generally
als. The aim of the review is to provide a reference for the selection there are fewer design variables involved.
of the technique that is most appropriate to solve a given problem.
More details about the algorithms are found in Haftka and Gurdal 2. Gradient-based methods
[1], Gurdal et al. [2], and other references provided in the following
sections. These algorithms are based on the gradient of the objective and
In the literature, different classifications for optimization of the constraints, whose functions, when their mathematical closed-
composite laminate have been suggested. For example, Fang and form expression is not available, can be approximated, although it
Springer [3] identified four categories for the optimization meth- may be computationally expensive. The solutions obtained with
ods, namely: (1) analytical procedures, (2) enumeration methods, gradient-based methods are only local optima, but the advantage
(3) heuristic schemes, and (4) non-linear programming. Abrate of these methods is the faster convergence rate as opposed to that
[4], on the other hand, decided to classify optimization problems achieved by direct and heuristic methods.
with respect to their objective functions that can be either one or
a combination of in-plane properties, flexural rigidity, buckling 2.1. Vanishing the function’s first gradient
load, natural frequency, and thermal effects. Venkataraman and
Haftka [5] suggested another classification for composite panels, Setting the first gradient of the objective function to zero is the
which categorized the design methods into two groups: (1) single simplest and the most common method to find a stationary point
laminate design, and (2) stiffened plate design. of a mathematical function. Sandhu [7] used this approach to find
We follow the classification provided by Setoodeh et al. [6], in the fiber orientation for a single-layer composite laminate. The
which there are two scenarios for the design of a composite problem involved the formulation of the Tsai’s failure criterion
structure: with respect to the fiber angle, and the solution was expressed in
terms of the angle between the fibers and the stress principal
(I) Constant stiffness, in which the composite part is considered directions. This method is very fast and returns all stationary
as a single element with the same stacking sequence all over points of the objective function just in one run; however, it
the domain. The design goal is to find an optimal stacking works only for single-variable, unconstrained problems with a
sequence that is uniform for the entire structure. closed form expression for the objective function. Obstacles may
(II) Variable stiffness, in which the structure consists of multiple emerge in the formulation of a closed form failure criterion for a
elements, each of them with a different stacking sequence. general multi-layer laminate, which is the main challenge prevent-
Here, material distribution and fiber orientation might ing this method from being extensively used in composite lay-up
change over the structural domain. design.

For the first scenario, which is the focus of this paper, we exam- 2.2. Steepest descent (SD)
ine the following optimization methods:
Steepest descent (SD) is a minimization technique that per-
1. Gradient-based. These methods utilize gradient information forms, at each step, a line-search in the opposite direction of the
of the objective(s) and the constraint(s) to find the direction gradient of the objective function. To simplify the problem, SD of-
and size of the step towards the optimum solution. ten resorts to techniques that find an approximation of the ideal
2. Direct Search and Heuristic. In contrast to the previous step size, which is the one that makes the gradients of the two con-
group, these methods do not need any gradient information; secutive successive iterations mutually orthogonal. For laminate
rather they require only the values of the objective function. stacking sequence design of composite laminates, SD was used in
Heuristic and Enumeration methods are also included the past either alone [8] or in combination with other optimization
here. techniques [9]. At the onset, it can achieve large variations in the
3. Specialized techniques. These methods are developed for objective function, but as the minimum of the objective function
solving a lay-up design problem, in which some properties is reached, the convergence rate may become very slow. Besides
of composite laminates are used to simplify the problem; this limitation, the necessity of working with continuous variables
thus, they cannot be applied to a general optimization and the drawback of entrapping in local optima are other disad-
problem. vantages that limit its use for composite lay-up design.
 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11 3

2.3. Conjugate gradient (CG) 2.6. Approximation schemes

Conjugate gradient (CG) methods improve the convergence rate An approximation scheme replaces the primary optimization
of the steepest descent by choosing conjugate descent directions problem with a sequence of explicit approximate sub-problems,
that are depending on both the gradient of the objective at the each expressed by a first or second-order Taylor series expansion
current step and the descent directions in the previous iteration. of the corresponding structural function [23]. The benefit is that
There are different algorithms to perform a conjugate gradient the sub-problems are solved faster than the original one, while
search; among them, Powell’s Conjugate Gradient is used by Hir- finding the solution of a set of approximate sub-problems should
ano [10] for buckling load maximization of laminated composite ideally yield to the same solution of the original problem.
plate under axial compression. The advantage of this method is Linear approximation is the simplest approximation and was
that no gradient information is required; however, unimodality used by Schmit and Farshi [24] to minimize the weight of a sym-
of the objective function is necessary. For multimodal functions, metric fibre-composite laminate. Here the non-linear program-
it is possible to compensate the effect of multi-modality by ming problem could be transformed into a sequence of linear
performing several trials from different starting points [10]. For programs because the stiffness matrix was a monotonous function
simple problems where the objective function can be expressed of the design variables (i.e. thickness of the layers with pre-as-
in closed-form, alternative CG methods, such as Powell’s Conjugate signed fiber orientations).
Directions, may also be used [11]; however, when only an approx- For non-monotonous functions, a mixed approximation was
imation of the gradient is available, techniques such as Fletcher– used by Fleury and Braibant [25]. This method, called Conlin’s
Reeves or Polak–Ribiere may perform better [12]. method, consisted of a combination of linear approximation when
the corresponding first derivatives were positive and inverse
2.4. Quasi-Newton method approximation when the first derivatives were negative. A more
robust form of Conlin’s approximation was developed by Svanberg
Due to the requirement of the second-order gradient informa- [26,27] called Method of Moving Asymptotes (MMA), in which the
tion, Newton (or Newton–Raphson) methods are rarely used for approximate function was obtained by a linearization of the corre-
composite laminate design; however, in their place, the Quasi- sponding function with respect to variables of the type 1/(xi Li) or
Newton (QN) methods are often used, for they allow determining 1/(Ui xi), depending on the sign of the derivatives of the function.
the Hessian without using second-order derivatives. In this transformation, xi is a design variable, and Ui and Li repre-
The QN method by Davidon, Fletcher, and Powell (DFP) [13] is sent the corresponding upper and lower limits for this variable
one of the most popular QN techniques used for composite lay- and are called ‘‘moving asymptotes” because they normally chan-
up design. Waddoups et al. [14] and Kicher and Chao [15] used ged between iterations. This method was used by Bendsoe et al.
the original form of DFP proposed by Fletcher and Powell [16] to [28] to maximize the buckling load factor of composite laminates.
optimize a composite cylindrical shell. A quadratic interpolation In order to achieve a better approximation for a general non-linear
of the objective function, which included strength and buckling function, Svanberg [29] included an additional non-monotonic
failure, was used in the one-dimensional minimization process. parameter into MMA formulation, which was updated at each step.
Kim and Lee [17] also used DFP for optimization of a curved actu- The modified method, called Globally Convergence MMA
ator with piezoelectric fibers. (GCMMA), was recommended to be used when the fiber orienta-
Quasi-Newton methods generally have a higher convergence tions are considered as design variables.
rate than CG methods, although their performance is problem- Bletzinger [30] and Svanberg [29] used the second-order deriv-
dependent and may change from one case to another. ative to achieve a MMA approximation that better matches the
curvature of the original function. Using a combination of method
2.5. Method of feasible directions of diagonal quadratic approximation (DQA) and a MMA by Zhang
et al. [31] is another attempt to exploit the benefits of a second-or-
Method of feasible directions (MFD) is created to solve optimi- der approximation. The use of second-order derivatives has the
zation problems with inequality constraints. Starting from a feasi- advantage of improving the reliability and the efficiency of the
ble initial point, MFD tries to find a move to a better point without optimization process, but it has the drawback of being computa-
violating any of the constraints. Since a composite lay-up design tionally expensive, especially for large scale problems. To avoid
problem usually includes several inequality constraints, MFD has calculating expensive second-order derivatives, Bruyneel and Fle-
been a good candidate for solving these problems [18]. However, ury [32] used the gradient information of the two successive iter-
like other gradient-based methods, MFD is not always able to find ations. The modified method called Generalized MMA (GMMA)
the global optimum. was employed for the optimization of composite structures where
More recently, with the presence and popularity of finite ele- both ply thicknesses and fiber orientations were considered as de-
ment methods, MFD has been adapted to be used in combination sign variables.
with finite element analyses [19,20]. In this case, the derivatives Harte et al. [33] compared the performance of Conlin’s method,
are calculated by approximation or direct sensitivity analysis. GCMMA and MDQA in optimizing the layer thicknesses and the
MFD may generate non-feasible solutions in the presence of winding angle of a glass-reinforced epoxy pipeline. When the vari-
highly non-linear constraints. In order to avoid this problem, Spal- ables were limited to only the thicknesses, Conlin’s method and
lino et al. [21] introduced a drive-away factor into the feasibility MDQA were found to perform equally. Conlin’s method required
requirement, by which the search direction was deviated from more iterations for convergence, but MDQA was computationally
the constraint boundaries towards the feasible region. Another more expensive. When the winding angle was added as a design
modification was proposed by Topal and Uzman [22], who took variable, the only algorithm capable of solving the problem was
into account not only the gradients of the objective function and the GCMMA, although the solutions required a careful examination
the retained active and/or violated constraints, but also the search since they might not be optimal solutions.
direction in the former iteration. This modified method, which Sequential quadratic programming (SQP) methods can also be
aims to increase the convergence rate, was used by author included in this section. SQP methods handle non-linear problems
researchers to optimize the buckling loads and the fundamental by constructing and solving a local quadratic program, which con-
frequencies of a composite part. sists of a quadratic model of the objective and a linear model of the
4 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11

constraints. SQP has been used for composite laminate design in the whole design space and give insight on the variables that most
several applications [34–36]. impact the performance, their use is limited to design problems
In comparison to other sequential programming techniques, governed by a small number of variables and to simple loading
Wang and Karihaloo [34] reported that SQP produced the most conditions.
precise results with fewer gradient calculations, but it requires a
larger number of functional evaluations compared to SLP (sequen- 3.3. Simplex method
tial linear programming) and SCP (sequential convex program-
ming). The comparative study was performed on minimizing the The simplex method uses the concept of a simplex, which is a
stress intensity at a crack tip, and the results cannot be generalized set of (n + 1) points in an n-dimensional design space. A simplex,
because the accuracy of the approximation depends on the shape for example, might be a line segment in one dimension, or a
of the objective function. triangle in two dimensions. This algorithm starts with an initial
The structural response function of a composite design problem simplex, which is generally improved by moving its vertices to-
is usually highly non-linear with several local optima, thus an ward better positions. As described by Nelder and Mead (NM)
approximation technique can be inaccurate, particularly when [42], the improvement is achieved by reflecting the least favour-
the fiber angles are used as design variables. If the accuracy of able point of the current simplex with respect to the other points,
the approximation is poor, inappropriate results will be obtained. expand the move if it is favourable, or contract it if it is not favour-
A higher reliability can be achieved by adopting higher order poly- able. The simplex is scaled down (shrunk) towards the most
nomials, which unfortunately are mathematically cumbersome to favourable point if no improvement is achieved. This process is ter-
handle. minated when the simplex becomes smaller than a user-defined
size.
NM method was employed by Tsau et al. [43] for optimal stack-
3. Direct search methods ing sequence design of a laminated composite loaded with tensile
forces, while the evaluation of stresses was performed by a finite
While the analytical methods are known for a fast convergence element method. It has been reported by Tsau and Liu [44] that
rate, direct search methods have the advantage of requiring no gra- the NM method is faster and more accurate than a Quasi-Newton
dient information of the objective functions and the constraints. method when it is used for lay-up selection of laminates with small
This feature is a significant advantage because in composite lami- number of layers (i.e. less than 4). The error in the simplex method
nate design derivative calculations or their approximations are of- was reported to increase with the number of design variables, as
ten costly or impossible to obtain. Direct search methods opposed to Quasi-Newton methods that have an almost constant
systematically approach the optimum solution only by using func- error. Regardless of the number of design variables, NM was
tion values from the preceding steps. As a result, several of these found to be faster than a Quasi-Newton method in computational
techniques have been revealed to be a significant tool for compos- time.
ite lay-up design, as described in the following sections. NM method is found to be practical for problems with a small
number of design variables (e.g. less than 10), but the convergence
3.1. Partitioning methods rate decreases exponentially with an increase of the number of
variables [45]. Other limitations of this method include: being a lo-
A partitioning method, such as Dichotomous search, interval- cal and unconstrained optimization method; dependency of the fi-
halving, Fibonacci and Golden section search, is a line search strat- nal solution on the initial simplex; and dependency of the
egy that can handle a single-variable optimization problem. Since convergence rate on the coefficients of expansion, contraction,
composite lay-up design generally copes with several design vari- and shrink. Despite these drawbacks, NM has been widely and effi-
ables, resorting only to a partitioning method to optimize a com- ciently used in several composite laminate problems either sepa-
posite design is rare. A sporadic example of a single variable rately or in combination with a global method, as discussed later
problem cited here is the work by Walker et al. [37], who merely in another section.
exploit the golden section method to determine the fiber angle that
best maximizes buckling resistance of a laminated cylindrical shell. 3.4. Random and greedy search
Although these methods are rarely used alone, they are effectively
used for a line search inside other optimization methods, such as A random search evaluates a number of randomly selected
conjugate gradient method and method of feasible directions [12]. points in the design space of a given optimization problem and
simply selects the best sampled point, while a track of previously
3.2. Enumeration search sampled points may be kept by the program to avoid recycling.
In contrast to the arbitrary move used in random search, a greedy
One of the first attempts in optimum design of laminated com- search evaluates a set of points around the current solution and
posite materials is Enumeration Search, consisted of trying all pos- moves one step in the direction of the best point, the last move
sible combinations of design variables and simply selecting the is retained until there is no more improvement achieved, then
best combination. Although cumbersome, this technique was used the whole procedure is repeated from the new point.
to find the lightest composite laminate during the 1970s [38,39]. Foye [46] was the first who used a random search to find the
Within this group, we include also selection chart techniques, optimum ply orientation angles of a laminated composite plate.
which aim at visualizing the impact of the variables throughout Graesser et al. [47] also used a random search, called improving
the whole design space. For instance, Park [40] plotted the change hit and run (IHR), to find a laminate with minimum number of
of the optimum fiber angle for a particular class of laminates (e.g. plies that safely sustain a given loading condition. This technique
[ h/ h]s, [ h/90/ h]s) under all possible normalized loading con- started with either a given or an arbitrary point and sought a
ditions. Weaver [41] also created selection charts, in which each new point by randomly changing one or all design variables. Once
class of laminates was displayed within an elliptical contour. Use- an improving direction was found, a one-dimensional line search
ful at the early stage of the design, these charts can be used to iden- was performed, and the search was continued from the new point.
tify a small subset of potential laminates which might be This concept is similar to the one in the Monte Carlo method used
investigated in more details. Although selection charts can map by Fang and Springer [3], in which the only difference was retain-
 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11 5

ing the last successful change instead of performing a line search. A large, GA is not at the risk of being stuck in a local optimum. How-
more systematic form of this technique is a greedy search method ever, finding a global solution is not necessarily guaranteed to be
used by Sargent et al. [48]. successful unless an infinite number of iterations are performed.
The advantage of a random search method is its ease of pro- Despite the high computational cost, GA has been the most popular
gramming and the freedom of selecting a priori the number of iter- method for optimizing the stacking sequence of a laminated com-
ations. However, the chance of success is strongly dependent on posite [5]. Its simple coding, which escapes gradient calculations,
the smoothness of the objective function and on the relative pro- and its flexibility of being applied to a large variety of problems
portions of satisfactory and deficient solutions. In addition, the with different types of variables and objective functions make GA
computational cost to obtain the global optimum exponentially in- particularly useful for problems with multimodal functions, dis-
creases with increase of the number of discrete variables [49]. A crete variables, and functions with costly derivatives.
greedy search is short-sighted and may never reach a better solu- Callahan and Weeks [59], Nagendra et al. [60], Le Riche and
tion if it has to go through a worse solution. Therefore, the success Haftka [61], Ball et al. [62] are among the first who adopted and
of a greedy search requires a relatively smooth design space and used GA for stacking sequence design of laminated composite
the proper selection of the step size. To overcome partially the lim- materials. GA has been used for several objective functions, such
itation of being trapped in local optima, a GRASP (greedy random- as strength [61,63], buckling loads [61,64–70], dimensional stabil-
ized adaptive search procedure), was proposed; however, it has not ity [71], strain energy absorption [72], weight (either as a con-
been used for lay-up design yet. straint or as an objective to be minimized) [73,52,74–77],
bending/twisting coupling [67], stiffness [72,78], fundamental fre-
3.5. Simulated annealing (SA) quencies [1,70,74,79], deflection [76] or finding the target lamina-
tion parameters [80].
Simulated annealing (SA), which mimics the annealing process GA has been applied to the design of a variety of composite
in metallurgy, globalizes the greedy search by permitting unfa- structures ranging from simple rectangular plates to complex
vourable solutions to be accepted with a probability related to a geometries such as sandwich plates [81], stiffened plates [82,83],
parameter called ‘‘temperature”. The temperature is initially as- bolted composite lap joints [84], laminated cylindrical panels
signed a higher value, which corresponds to more probability of [170]. GA can be often combined with finite element packages that
accepting a bad move and is gradually reduced by a user-defined analyze the stress and strain response of the composite structure
cooling schedule. Retaining the best solution is recommended in [85,76,86].
order to preserve the good solution [50]. This method is the most One of the main problems associated with GAs is the high com-
popular method after genetic algorithms (GA) for stacking se- putational intensity and the premature convergence, which may
quence optimization of laminated composite materials [48,51–53]. happen if the initial population is not appropriately selected. To in-
Generation of a sequence of points that converges to a non- crease convergence rate, reduce the risk of premature convergence,
optimal solution is one of the problems in SA. To overcome this and decrease the function evaluation time, several modifications
shortcoming, modifications of SA have been proposed such as have been proposed. Some of these attempts are summarized here:
increasing the probability of sampling points far from the current
point by Romeijn et al. [54] or using a set of points at a time instead – use of parallel computing [87–89];
of only one by Erdal and Sonmez [50]. – multi-level optimization (coarse level and fine level coding)
To increase the convergence rate, Genovese et al. [55] proposed a [87];
two-level SA including a ‘‘global annealing” where all design vari- – introduction of problem-dependent operators [90], such as
ables were perturbed simultaneously and a ‘‘local annealing” where layer addition or deletion, permutation, interlaminate swap
only one design variable was perturbed at a time. The local anneal- [83], generalized elitist [67], laminate thickness/material/
ing was performed after each iteration of the global annealing in or- fiber angle mutation [86];
der to locally improve the trial point. Its convergence speed was – recovery of previously evaluated solutions [65,75,91];
reported higher than a one-level SA and comparable to a gradi- – use of approximation methods for function evaluation
ent-based optimization method implementing a SLP method. [92,75,69], or training an artificial neural network (ANN)
In order to prevent re-sampling of the solutions, Rao and Arvind [70,79];
[56] embedded a Tabu search in SA obtaining a method called Tabu – proposing a consanguineous initial population [66], or a
embedded simulated annealing (TSA). The stacking sequence opti- hierarchical structure with aging [93].
mization of laminated composites was solved by TSA, while the
constraints were handled using a correction strategy. TSA was fas- Sometimes a combination of these methods are used; for in-
ter than classical SAs with the penalty of requiring more memory stance, Park et al. [91] used a memory approach combined with
and computation time per iteration. the permutation operator with local learning/random shuffling to
Simulated annealing is a good choice for the general case of reduce the number of function evaluation and to improve the con-
optimal lay-up selection; however, it cannot be programmed to vergence rate. They also categorized the design criteria into two
take advantage of the particular properties of a given problem; groups of layer combination dependent criteria and layer sequence
GA is more flexible in this respect, although it is often computa- dependent criteria, which recycled some of previously evaluated
tionally more time consuming [48,52,57]. It is not easy to general- candidates. However, almost exclusively all these methods are re-
ize the conclusion because there are some other researches, such as ported to provide an improvement in the GA’s numerical efficiency,
the one by Rao and Shyju [58], claiming that SA outperformed GA a comprehensive comparison among them is impossible because
both in computational performance and in finding the global opti- GA is a probabilistic method and its performance is strongly
mum for other combinatorial problems. dependent on the problem.
GA is originally developed for unconstrained optimization,
3.6. Genetic algorithm (GA) whereas composite design problems are usually constrained by
limitations in material’s strength, weight, cost or other criteria.
A genetic algorithm (GA) is an evolutionary optimization tech- Among several methods used to incorporate constraints, using a
nique using Darwin’s principal of ‘‘survival of the fittest” to penalty function strategy is the most popular one. A penalty func-
improve a population of solutions. If the population size is suitably tion is introduced to convert a constrained problem into an
6 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11

unconstrained problem by assigning a penalty term to the infeasi- to its own flying experience (best solution in its individual history)
ble solutions. The penalty assigned to each solution depends on and the flying experience of the other particles (the best solution
number and intensity of violated constraints. Another method to among all particles). As such, PSO has a good potential to benefit
handle the constraints is the repair strategy [80], which transforms from parallel computing.
infeasible solutions into a feasible solution located in a close PSO was used by Suresh et al. [97] for the optimal design of a
proximity. composite box-beam of a helicopter rotor blade. It was reported
Sargent et al. [48] compared GA with some greedy algorithms that the solutions provided by PSO were closer to the optimum val-
(i.e. random search, greedy search, and simulated annealing) and ues than those given by the GA. The computational time, on the
observed that GA produced better solutions than the greedy other hand, was comparable. Kathiravan et al. [98] compared PSO
searches, which in some instances were unable to find a solution. to a gradient-based method for the maximization of the failure
Although more demanding in terms of computation time, GA re- strength of a thin walled composite box-beam, whose design vari-
vealed to be a more robust algorithm. Sivakumar et al. [94] com- ables were the ply orientation angles. The PSO was found to give
pared DFP and GA, both applied to minimize the weight of a results superior or equivalent to the gradient-based method. Fur-
laminated composite constrained by its fundamental frequencies. thermore, it did not need to start from different initial points.
It was reported that DFP converged in smaller number of iterations Chen et al. [99] introduced two improvements to PSO. First, a
when the number of constraints was small; however, finding a fea- random coefficient was added that increased the individual influ-
sible point was a difficult task when number of constraints ence and the swarm variety and improved the search ability. Sec-
increased. Also considering that the DFP could not handle discrete ond, an interference method was proposed to overcome the lack
variables, they concluded ‘‘GA seems to be the best tool to optimize of local exploration at later stages of the search, particularly when
composite laminates”. the objective function has multiple local optima. This modification
Although GA has been widely used for stacking sequence opti- was applied only if the results remained unchanged during a cer-
mization, one major shortcoming is its low convergence rate. GA tain number of consecutive iterations. The improved method,
is a population-based evolutionary algorithm, and might require which was tested on lay-up stacking sequence design, was re-
several generations before converging to a solution [58]. Each gen- ported to converge faster and to be more stable than the original
eration consists a large number of function evaluations, thus it can PSO.
be computationally time consuming and expensive.
3.7.4. Ant colony
3.7. Other heuristic optimization methods Ant colony optimization (ACO) is another heuristic search
method, which is inspired by the behaviour of the ants and their
Although GA and SA are the most popular heuristic methods ability in finding the shortest paths between their nest and the
used to optimize the stacking sequence of composite laminates, food source. Aymerich and Serra [100] illustrated the application
other heuristic methods have been also used. This section summa- of this technique to the lay-up design of laminated panels for max-
rizes the main ones. imizing the buckling load. The average performance of the ACO, as
measured by practical reliability and normalized price, was evalu-
3.7.1. Tabu search ated ‘‘comparable or even superior” to GA or TS techniques.
Tabu search (TS) is a local search method that starts from an ini-
tial point and progresses by changing the design variables, one at a
time. The entire or a part of the neighbourhood is evaluated before 4. Specialized algorithms
accepting the best solution as a starting point for the next move. TS
keeps, in a short term memory, the potential solutions that have In this category, fall methods developed explicitly for optimiz-
already been visited and marks them as ‘‘Tabu”. This strategy pre- ing laminated composite materials. These strategies exploit a num-
vents solution re-sampling. TS was implemented by Pai et al. [95] ber of properties of the composite laminates to simplify the
for discrete optimization of the stacking sequence of a composite optimization process. Often developed for a particular application,
laminate subjected to buckling and strength requirements as well these methods generally simplify the problem by restricting the
as for matrix cracking. The results were compared with a GA, design space in terms of allowable lay-up, loading condition,
which showed a comparable solution, but the computational time and/or the objective function. Since these methods are tailored to
was case dependent. a specific design problem, they lose robustness when applied to a
general optimization problem. However, if used for the particular
3.7.2. Scatter search case they are designed for they can be much faster than other
Scattered search (SS) is a strategy that generates a reference set techniques.
from a population of randomly selected solutions. From this refer-
ence set, a subset of solutions is selected and enhanced by using an 4.1. Design with lamination parameters
improvement procedure (e.g. a greedy search). The improved solu-
tions are then used to update the reference set, and the process is Using lamination parameters [101,102], which are integrated
continued with a new subset. Rao and Arvid [96] used a scatter trigonometric functions through the thickness of a laminate, in-
search for lay-up sequence optimization including thermal buck- stead of lay-up variables has the big advantage of reducing the
ling, weight, cost, fundamental frequency and buckling consider- number of parameters required to express a laminate properties
ations. Lay-up results and computational efficiency were similar to maximum of 12, regardless of number of layers [2]. Avoiding
to those obtained with GA. troublesome optimization over periodic functions of the rotation
angles and discrete number of plies is the other advantage of using
3.7.3. Particle swarm lamination parameters. The convexity of the set of lamination
Like GA and SS, particle swarm optimization (PSO) is also a pop- parameters together with the linear dependence of the stiffness
ulation-based, stochastic optimization method inspired by the on these parameters implies further simplification with respect
flocking behaviour of birds. In this method, each solution in the to the basic mathematical structure of the problem [103].
search space is called a ‘‘particle” and resembles a bird among Beside the promising advantage of using lamination parame-
others, which adjusts its position in the search space according ters, the challenge in working with these parameters is that they
 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11 7

are not independent and cannot be arbitrarily prescribed. The inate. The second examined the probability of replacing the higher
admissible range of lamination parameters are given by solving quality layers with weaker materials. Ghiasi et al. [120] used layer
the geometric relations. Several authors, such as Fukunaga and separation technique to keep the location of different layers un-
Vanderplaats [104] and Grenestedt and Gudmundson [105], have changed when a layer is added. A new layer was added to the lam-
suggested necessary conditions for different combinations of lam- inate by dividing one of the current layers into two layers. Then a
ination parameters, but the complete set of sufficient conditions scaling procedure was used to change the thicknesses.
for all twelve parameters is still unknown [103]. These methods are very fast compared to other optimization
The lamination parameters strategy requires solving the inverse techniques, because they are working only with a small portion
problem to obtain the corresponding number of plies, thicknesses of all the design variables in each step; however, they may not
and fiber orientations, which are suitable for manufacturing. Solv- reach a local optimum. This problem can be partially solved by
ing the inverse problem is not easy and the solution is not unique. revising the fiber orientations and thicknesses every time a new
Miki [106] proposed a method to visualize the admissible range of layer is introduced to the laminate. Nevertheless, the additional
lamination parameters and their corresponding lay-up parameters. computational time and cost induced by this approach might be
The method is fast and handy, although it can be applied only to excessive and of limited value since it will only assure achieving
the laminates with prescribed in-plane stiffness properties and a local optimum.
balanced-angle-ply laminates of type [±h1/±h2/ . . . /±hi]S. Just as
for the in-plane lamination diagram, the flexural lamination dia-
4.3. Problem partitioning
grams were also developed [107]. Fukunaga and Chou [108] used
a similar graphical technique for laminated cylindrical pressure
Composite lay-up design is governed by variables of different
vessels. Lipton [109] has developed an analytical method to find
nature (e.g. thickness and fiber angle), which increase the com-
the laminate configuration of a three-ply laminate under in-plane
plexity of the design problem. To simplify the design problem,
loading conditions. Autio [110], Kameyama and Fukunaga [111],
one possible approach is to split the problem into some dependent
Enrique et al. [112], and some other researchers used GA to solve
sub-problems, each described with a small number of variables of
the inverse problem.
similar nature. This section gives an overview of these techniques.
In general, optimization methods based on the lamination
Two-part methods are the most common form of problem par-
parameters are restricted to global structural responses, such as
titioning, in which the first part finds the fiber direction that pro-
stiffness, and do not include local strength constraints at the ply
vides optimum property such as stiffness, displacement, natural
level. Optimal design studies are limited to particular structural
frequency, etc., and the second part, optimizes ply thicknesses for
responses and to specific laminate configurations. Inverse problem
minimum weight [121,122]. Le Riche and Gaudin [70,71] added
must be solved in order to get the corresponding number of plies,
one more sub-problem to this technique, which includes minimi-
thicknesses and fiber orientations, which is the most challenging
zation of the total number of layers in the laminate. Farshi et al.
part of working with lamination parameters.
[123] also used three sub-problems, but embedded with a layer-
wise optimization. In this method, a new layer was added to the
4.2. Layerwise optimization
stack after optimization of fiber angles and then thicknesses, which
was repeated until the thickness of the recently added layer ap-
A layerwise optimization method optimizes the overall perfor-
proached zero.
mance of a composite laminate by sequentially optimizing one or
Partitioning the problem may also be used to improve the GA’s
some of the layers within a laminate. This method works with
performance, because when both real and discrete variables are
one layer or a subset of layers in the laminate and requires first
represented using a single string of GA the cost of the search is in-
the selection of the best initial laminate and then the addition of
creased. After partitioning, sub-problems with different type of
the layer that best improves the laminate performance, which is
variables can be represented with different strings or can be solved
usually achieved by an enumeration search [113,114].
by different optimization methods [124,125]. In general problem
Lansing et al. [9] determined the initial laminate by assuming
partition can significantly increase the convergence rate; however,
the layers at 0, 90, and ±45° carrying all the longitudinal, trans-
since the sub-problems are not independent, the final solution in
verse, and shear stresses, respectively. Massard [115] started with
certain instance might be far from the global optimum.
a one-layer laminate and found the best fiber orientation for the
single-ply laminate. Todoroki et al. [116] proposed two other ap-
proaches to find the initial laminate. They first used quasi-isotropic 4.4. Discrete material optimization (DMO)
plies to estimate the initial number of layers; the second deter-
mined the lower bond for the number of layers by using a super- Lund and Stegmann [126] used a method called discrete mate-
ply. A super-ply is a virtual layer that had the best material rial optimization (DMO) for optimum design of composite lami-
properties in all directions. nates. In this method, which can be equally applied to constant
An important issue in layerwise optimization strategy is the im- and variable stiffness design, the material stiffness is computed
pact of the newly added layer on the stress distribution of the lam- as a weighted sum of some candidate materials (fiber orientations).
inate. The addition of a new layer changes the stress field in the The discrete problem of choosing the best material (with the right
laminate, because there is a change in the position of the layers orientation) is converted to a continuous formulation where the
in the laminate and in the portion of load each layer can sustain. design variables are the weight functions on each candidate mate-
Therefore, layers inserted in previous steps are no longer optimal. rial. For each ply, the goal is to drive the influence of all but one of
Narita [117] and Narita et al. [118] tried to solve this problem by these weight functions to zero. A disadvantage of this method is
starting with a laminate with hypothetical layers with no rigidity. that it replaces each design variable (fiber angle) with several
From the outermost layer, all layers were sequentially replaced by weight factors, each corresponding to one permissible fiber angle.
an orthotropic layer and the optimum fibre orientation angle was
determined by enumeration. The first solution was then used as 4.5. Fractal branch-and-bound method
an initial approximation for the next cycle. Farshi and Rabiei
[119] proposed a method for minimum thickness design consisting Fractal branch-and-bound (FBB) method, proposed by Todoroki
of two steps. The first aimed at introducing new layers to the lam- and Terada [127,128], and Hirano and Todoroki [129,130] is based
8 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11

on branch-and-bound method, whose computational cost is ing such knowledge-based systems is problem-dependent and
reduced using a response surface to approximate the objective needs a great deal of experience.
function in terms of lamination parameters. This technique per-
forms an evaluation of the objective function to prune inefficient 5. Hybrid methods
branches (laminates) by utilizing the fractal patterns of feasible re-
gion of lamination parameters in symmetrical laminates. Later on, A hybrid method combines two or more optimization methods
FBB was expanded to unsymmetrical laminates by Matsuzaki and to benefit from the advantages of all of them in order to obtain a
Todoroki [131]. It has also been adopted for complicated structures better convergence rate, to achieve a global optimum, to have a
with multiple laminates such as hat stiffened panels [132]. better accuracy, to make the optimization method more robust,
or for other reasons.
4.6. Knowledge-based methods In order to handle both continuous and discrete design vari-
ables in a general multimodal non-linear problem, Seeley and
Experimental or knowledge-based methods (KBM) aims at Chattopadhyay et al. [136] developed a combined simulated
bringing the expert’s knowledge into the computer design process annealing and sequential programming optimization techniques
by performing a screening of the concepts ill-suited to the specific (more specifically BFGS). The method was applied to the design
requirements and focusing quickly on the most viable and attrac- of a scaled airplane wing model, represented by a flat composite
tive alternatives. To produce a short list of good layups, these plate, with piezoelectric actuation to improve the aeroelastic
methods may use different tools, such as discrimination charts response.
and tables eliminating unsuitable arrangements [133] or a list of Incorporating a local optimization technique into GA [65,81] is
rules and recommendations for good lay-ups [134,135]. A KBM the most popular hybrid approach, in which a local search algo-
can be faster than a numerical method because it reduces the num- rithm is applied to some newly generated individuals to drive
ber of candidates at the onset, but it may not explore the entire them to a local optimum. These locally optimal solutions replace
range of potential layups [134]. In addition, the process of develop- the current individuals in the population to prepare the next gen-

Table 1
Properties of different optimization methods used for composite lay-up design.

Method name Method Global Continuous Discrete Constrained Derivative Deterministic Relative convergence rate
Vanishing the function gradient D0 X 2 X
Steepest descent SD X 1 X
Conjugate gradient
Linear conjugate gradient LCG X 1 X SD+*/in n steps
Powell’s method CG-P X X SD+/in n steps
Fletcher–Reeves CG-FR X 1 X Aprx. in n steps
Polak–Ribiere CG-PR X 1 X CG-FR+
Quasi-Newton
Broyden–Fletcher–Goldfarb–Shanno BFGS X 1 X Super linear
Davidon–Fletcher–Powell DFP X 1 X Super linear
Method of feasible directions MFD X 1 X
Modified feasible direction (Topal) MFD X X 1 X MFD+
Conlin’s approximation method Colin X X 1 X
Method of moving asymptotes MMA X X 1 X
Globally convergent MMA GCMMA X X 1 X
Globally convergent MMA-2 GCMMA2 X X 2 X GCMMA+
Generalized MMA GMMA X X 1 X GCMMA+
Method of diagonal quadratic approximation MDQA X X 2 X GCMMA+
Enumeration search ES X X X X
Partitioning methods X
Golden section GS X X X
Simplex method
Nelder–Mead simplex NM X X
Random search RS X X X ES**
Monte Carlo MC X X X RS+
Improving hit and run IHR X X X X MC
Greedy search GRS X X X X RS+
Greedy randomized adaptive search procedure GRASP X X X X GRS
Simulated annealing SA X X X X GRS
Improved SA ISA X X X 1 SLP/SA+
Genetic algorithm GA X X X
GA + local improvements GA+L X X X GA+
GA + special operators GA+O X X X GA+
Scatter search SS X X X X GA
Tabu search TS X X X
Particle swarm optimization PSO X X X
Ant colony optimization ACO X X X
Knowledge-based methods KBM X X X X
*
Having a better convergence rate.
**
Having almost a similar convergence rate.
 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11 9

eration. This combination utilizes the advantage of GA in finding have low rate of convergence, which is also a problem dependent
the global optimum and the advantage of a local optimization factor. Consequently, comparing stochastic methods is unfeasible
method in quickly arriving to an optimum solution. This hybrid because their heuristic nature make them strongly dependent on
optimization technique needs less number of function evaluation the problem. Until now, Genetic Algorithm has been the most pop-
compared to pure GA. ular method with simulated annealing ranked the second.
Another common form of hybridization is to restart a local opti- The next group of methods called ‘‘specialized methods” in-
mization technique, several times. A good example is the global- cludes a variety of techniques developed for the design of a specific
ized bounded Nelder–Mead (GBNM) by Luersen and Le Riche type of laminates, loading conditions, and/or objective. These
[137]. This method was particularly adapted to tackle multimodal, methods are usually fast, but are limited to certain types of prob-
discontinuous optimization problems, for which it is uncertain that lems and can find only a local minimum.
a global optimization can be afforded. The algorithm consists of Finally, hybrid methods benefit from the advantages of all their
several restarts of a local search, performed by the Nelder–Mead constituent methods. The review of the main hybrid methods in
simplex method, each initiated from a simplex likely far from pre- this paper has highlighted their promising future.
viously sampled points. The proximity to previously sampled point Optimization methods for variable stiffness design will be the
was controlled by a multi-dimensional probability function, which focus of a second paper, which will review also other issues,
has been later simplified to a one-dimensional function by Ghiasi such as multi-objective optimization, discretization techniques,
et al. [138] in order to reduce the computational time and improve design for manufacturing, sensitivity analysis, and design for
the convergence rate. The local-global search has been shown to uncertainty.
converge faster than an evolutionary method, such as GA when
the number of iterations is small; however the two methods per-
form similarly, when the number of iterations increases. References
Oh and Lee [139] used the same technique, but replacing the
[1] Haftka RT, Gurdal Z. Elements of structural optimization. 3rd ed. Kluwer;
probability function with a domain elimination technique, which 1992.
consisted of making a set of eliminated regions, and terminating [2] Gurdal Z, Haftka RT, Hajela. Design and optimization of laminated composite
the local search when the current search is adjacent to an elimi- materials. Wiley-Interscience; 1999.
[3] Fang C, Springer GS. Design of composite laminates by a Monte Carlo method.
nated region. J Compos Mater 1993;27(7):721–53.
Rao and Shyju [58] integrated SA and TS into an algorithm [4] Abrate S. Optimal design of laminated plates and shells. Compos Struct
called multiple start guided neighbourhood search (MSGNS). The 1994;29:269–86.
[5] Venkataraman S, Haftka RT. Optimization of composite panels – a review. In:
multi-start scheme helps overcome the sensitivity of SA on the
Proceedings of the 14th annual technical conference of the American society
cooling schedule, while the Tabu search prevents recycling the pre- of composites, Dayton, OH, September 27–29; 1999.
vious solutions. For stacking sequence optimization of laminated [6] Setoodeh S, Abdalla MM, Gurdal Z. Design of variable – stiffness laminates
using lamination parameters. Compos B: Eng 2006;37:301–9.
composites, MSGNS was found to be as effective as GA in obtaining
[7] Sandhu RS. Parametric study of optimum fiber orientation for filamentary
multiple near optimal solutions, and superior in terms of computa- sheet. Aerospace structures information and analysis center AFFDL/FBR
tional performance to algorithms like GA, SA, and SS. WRAFB, TM-FBC-71-1, Ohio; 1971.
[8] Cairo RP. Optimum design of boron epoxy laminates. TR AC-SM-8089,
Grumman Aircraft Engineering Corporation, Bethpage, New York;
November 1970.
6. Concluding remarks [9] Lansing W, Dwyer W, Emerton R, Analli E. Application of fully stressed design
procedures to wing and empennage structures. J Aircraft 1971;8(9):683–8.
[10] Hirano Y. Optimum design of laminated plates under axial compression. AIAA
The most popular methods for optimizing the stacking se- J 1979;17(9):1017–9.
quence of laminated composites with uniform stiffness have been [11] Moh JS, Hwu C. Optimization for buckling of composite sandwich plates. AIAA
reviewed in this paper. The optimization techniques are classified J 1997;35(5):863–8.
[12] Nocedal J, Wright SJ. Numerical optimization. 2nd ed. New York: Springer-
into four categories: gradient-based methods, direct search and Verlag; 2006.
heuristic methods, specialized techniques, and hybrid methods. A [13] Davidon WC. Variable metric method for minimization. SIAM J Opt
brief list of all these methods and their properties is shown in Table 1991;1(17).
[14] Waddoups ME, McCullers LA, Olsen FO, Ashton JE. Structural synthesis of
1. The last column in this table shows the relative convergence rate
anisotropic plates. In: AIAA/ASME 11th Struct., Struct. Dyn. and Math Conf.,
of each method compared to others, when such data are available Denver, Colorado; 1970.
in literature. [15] Kicher TP, Chao TL. Minimum weight design of stiffened fibre composite
Gradient-based methods are found to be generally faster than cylinders. J Aircraft 1971;8:562–8.
[16] Fletcher R, Powell MJD. A rapidly convergent descent method for
other techniques and can find a local minimum in a small number minimization. Comput J 1963;6:163–8.
of iterations. These methods, however, are limited to problems [17] Kim C, Lee DY. Design optimization of a curved actuator with piezoelectric
with continuous design variables and first or second derivatives fibers. Int J Mod Phys B (IJMPB) 2003;17(8–9):1971–5.
[18] Saravanos DA, Chamis CC. An integrated methodology for optimizing the
(either exact or approximate). Since they reach local optimum, passive damping of composite structures. Polymer Compos 1990;11(6):
the final solution depends on the initial point. If the problem has 328–36.
a small number of design variables and the objective function is [19] Ha SK, Kima DJ, Sung TH. Optimum design of multi-ring composite flywheel
rotor using a modified generalized plane strain assumption. Int J Mech Sci
smooth, then gradient-based methods are the best choice, espe- 2001;43:993–1007.
cially if several restarts from different initial points can be affor- [20] Thompson MD, Eamon CD, Rais-Rohani M. Reliability-based optimization of
ded. These methods do not find large application to composite fiber-reinforced polymer composite bridge deck panels. J Struct Eng 2006;
132(12):1898–906.
lay-up design because gradient information is often either unavail- [21] Spallino R, Giambanco G, Rizzo S. A design algorithm for the optimization of
able or very expensive to determine. laminated composite structures. Eng Comput 1999;16(3):302–15.
In contrast, direct search methods are very popular because no [22] Topal U, Uzman U. Maximization of buckling load of laminated composite
plates with central circular holes using MFD method. Struct Multidisc Optim
derivatives need to be calculated. These methods can be divided
2008;35:131–9.
into deterministic and stochastic. Stochastic methods are more [23] Schmit LA, Fleury C. Structural synthesis by combining approximation
appropriate for composite lay-up design, because of their capabil- concepts and dual methods. AIAA J 1980;18(10):1252–60.
ities of handling a mixture of continuous and discrete variables, [24] Schmit LA, Farshi B. Optimum laminate design for strength and stiffness. Int J
Num Meth Eng 1973;7:519–36.
finding the global optimum of a multi-modal objective function, [25] Fleury C, Braibant V. Structural optimisation: a new dual method using mixed
and working with a population of solutions. These methods usually variables. Int J Numer Meth Eng 1986;23(3):409–28.
10 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11

[26] Svanberg K. The method of moving asymptotes – a new method for structural [58] Rao ARM, Shyju PP. Development of a hybrid meta-heuristic algorithm for
optimization. Int J Numer Meth Eng 1987;24:359–73. combinatorial optimisation and its application for optimal design of
[27] Svanberg K. The method of moving asymptotes (MMA) with some extensions. laminated composite cylindrical skirt. Comput Struct 2008;86(7–8):796–815.
Optimization of large structural systems, GIN Rozvany. Dordrecht: Kluwer [59] Callahan KJ, Weeks GE. Optimum design of composite laminates using
Academic Publishers; 1993. p. 555–66. genetic algorithm. Compos Eng 1992;2(3):149–60.
[28] Bendsøe MP, Olhoff N, Taylor JE. A variational formulation for multicriteria [60] Nagendra S, Haftka RT, Gurdal Z. Stacking sequence optimization of simply
structural optimization. J Struct Mech 1983;11:523–44. supported laminates with stability and strain constraints. AIAA J
[29] Svanberg K. A globally convergent version of MMA without linesearch. In: 1992;30(8):2132–7.
Olhoff N, editor. First world congress of structural and multidisciplinary [61] Le Riche R, Haftka RT. Optimization of laminate stacking sequence for
optimization, Rozvany GIN; 1995. p. 9–16. buckling load maximization by generic algorithm. AIAA J 1993;31:951–69.
[30] Bletzinger KU. Extended method of moving asymptotes based on second- [62] Ball NR, Sargent PM, Ige DO. Genetic algorithm representations for laminate
order information. Struct Optim 1993;5(3):175–83. layups. Artif Intell Eng 1993;8(2):99–108.
[31] Zhang WH, Domaszewski M, Fleury C. New mixed convex approximation [63] Park JH, Hwang JH, Lee CS, Hwang W. Stacking sequence design of composite
method with applications for truss configuration optimization. Struct Optim laminates for maximum strength using genetic algorithms. Compos Struct
1995;15(3–4):237–41. 2001;52(2):217–31.
[32] Bruyneel M, Fleury C. Composite structures optimization using sequential [64] Gurdal Z, Haftka RT, Nagendra S. Genetic algorithm for the design of
convex programming. Adv Eng Software 2002;33:697–711. laminated composite panels. SAMPE J 1994;30:207–18.
[33] Harte AM, McNamara JF, Roddy I. Evaluation of optimisation techniques in [65] Kogiso N, Watson LT, Gurdal Z, Haftka RT. Genetic Algorithm with local
the design of composite pipelines. J Mater Process Tech 2001;118(1– improvement for composite laminate design + D2. Struct Optim 1994;7(4):
3):478–84. 207–18.
[34] Wang J, Karihaloo BL. Cracked composite laminates least prone to [66] Todoroki A, Haftka RT. Stacking sequence matching by a two stage genetic
delamination. Proce Math Phys Sci 1994;444(1920):17–35. algorithm with consanguineous initial population. In: Proceedings of the 38th
[35] Liu B. Two-level optimization of composite wing structures based on panel AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics and materials
genetic optimization. PhD thesis, Aerospace Engineering, Mechanics, and conference part 2; 1997. p. 1297–302.
Engineering Science, University of Florida; 2001. [67] Soremekun G, Gurdal Z, Haftka RT, Watson LT. Composite laminate design
[36] George TJ, Herman Shen MH, Huybrechts SM, Meink TE, Wegner PM. Optimal optimization by genetic algorithm with generalized elitist selection. Compos
design of composite chambercore structures. Compos Struct 2002;52(3– Struct 2001;79:131–43.
4):277–86. [68] Messager T, Pyrz M, Gineste B, Chauchot P. Optimal laminations of thin
[37] Walker M, Reiss T, Adali S, Weaver PM. Application of MATHEMATICA to the underwater composite cylindrical vessels. Compos Struct 2002;58(4):
optimal design of composite shells for improved buckling strength. Eng 529–37.
Comput 1998;15(2):260–7. [69] Todoroki A, Ishikawa T. Design of experiments for stacking sequence
[38] Waddoups ME. Structural airframe application of advanced composite optimizations with genetic algorithm using response surface
materials-analytical methods. Report AFML-TR-69-101, vol. VI Air Force approximation. Compos Struct 2004;64(3–4):349–57.
Materials Laboratory, Wright-Patterson Air Force Base, Ohio; 1969. [70] Abouhamze M, Shakeri M. Multi-objective stacking sequence optimization of
[39] Verette RM. Stiffness, strength and stability optimization of laminated laminated cylindrical panels using a genetic algorithm and neural networks.
composites. Report NOR-70-138, Northrop Aircraft Co., Hawthorne, Compos Struct 2007;81(2):253–63.
California; 1970. [71] Le Riche R, Gaudin J. Design of dimensionally stable composites by
[40] Park WJ. An optimal design of simple symmetric laminates under the first ply evolutionary optimization. Compos Struct 1998;41:97–111.
failure criterion. J Compos Mater 1982;16:341–55. [72] Potgeiter E, Stander N. The genetic algorithm applied to stiffness
[41] Weaver PM. Designing composite structures: lay-up selection. In: maximization of laminated plates: review and comparison. Struct Optim
Proceedings of the institution of mechanical engineers – part G. J Aerospace 1998;15:221–9.
Eng, vol. 216; 2002. p. 105–16. [73] Rajendran I, Vijayarangan S. Optimal design of a composite leaf spring using
[42] Nelder JA, Mead R. Simplex method for function minimization. Comput J genetic algorithms. Compos Struct 2001;79:1121–229.
1965;7(4):308–13. [74] Sivakumar K, Iyengar NGR, Deb K. Optimum design of laminated composite
[43] Tsau LR, Chang YH, Tsao FL. Design of optimal stacking sequence for plates with cut-outs using a genetic algorithm. Compos Struct
laminated FRP plates with inplane loading. Comput Struct 1995;55(4): 1998;42(3):265–79.
565–80. [75] Gantovnik VB, Gurdal Z, Watson LT. A genetic algorithm with memory for
[44] Tsau LR, Liu CH. A comparison between two optimization methods on the optimal design of laminated sandwich composite panels. Compos Struct
stacking sequence of fiber-reinforced composite laminate. Comput Struct 2002;58(4):513–20.
1995;55(3):515–25. [76] Walker M, Smith R. A technique for the multiobjective optimization of
[45] Han L, Neumann M. Effect of dimensionality on the Nelder–Mead simplex laminated composite structures using genetic algorithms and finite element
method. Optim Meth Software 2006;21(1):1–16. analysis. Compos Struct 2003;62(1):123–8.
[46] Foye R. Advanced design concepts for advanced composite air frames. AFML- [77] Naik GN, Gopalakrishnan S, Ganguli R. Design optimization of composites
TR-68-91, Air Force Materials Laboratory, Wright-Patterson Air Force Base, using genetic algorithms and failure mechanism based failure criterion.
vol. 1–11, Ohio; 1968. Compos Struct 2008;83:354–67.
[47] Graesser DL, Zabinsky ZB, Tuttle ME, Kim GI. Designing laminated composites [78] Todoroki A, Watanabe K, Kobayashi H. Application of genetic algorithms to
using random search techniques. Compos Struct 1991;18:311–25. stiffness optimization of laminated composite plates with stress-
[48] Sargent PM, Ige DO, Ball RN. Design of laminate composite layups using concentrated open holes. JSME Int J A: Mech Mater Eng 1995;38(4):458–64.
genetic algorithms. Eng Comput 1995;11(2):59–69. [79] Apalak MK, Yildirim M, Ekici R. Layer optimisation for maximum
[49] Savic V, Tuttle M, Zabinsky Z. Optimization of composite I-sections using fiber fundamental frequency of laminated composite plates for different edge
angles as design variables. Compos Struct 2001;53:267–77. conditions. Compos Sci Technol 2008;68:537–50.
[50] Erdal O, Sonmez FO. Optimum design of composite laminates for maximum [80] Todoroki A, Haftka RT. Stacking sequence optimization by a genetic algorithm
buckling load capacity using simulated annealing. Compos Struct with a new recessive gene like repair strategy. Compos B: Eng 1998;29:
2005;71(1):45–52. 277–85.
[51] Lombardi M, Haftka RT, Cinquini C. Optimization of composite plates for [81] Lin CC, Lee YJ. Stacking sequence optimization of laminated composite
buckling using simulated annealing. In: Proceedings of the AIAA/ASME/ASCE/ structures using genetic algorithm with local improvement. Compos Struct
AHS/ASC 33rd structures, structural dynamics and materials conference, Part 2004;63:339–45.
5; Dallas, Texas; April 13–15, 1992. p. 2552–63. [82] Nagendra S, Haftka RT, Gurdal Z. Design of blade stiffened composite panels
[52] Sadagopan D, Pitchumani R. Application of genetic algorithms to the by a genetic algorithm approach. In: Proceedings of the AIAA/ASME/AHS 34th
optimal tailoring of composite materials. Compos Sci Technol 1998;58: structures, structural dynamics and materials conference AIAA, Washington
571–89. DC; 1993. p. 2418–36.
[53] Bakis CE, Arvin AC. Optimal design of press-fitted filament wound composite [83] Nagendra S, Jestin D, Gurdal Z, Haftka RT, Watson LT. Improved genetic
flywheel rotors. Compos Struct 2006;72(1):47–57. algorithm for the design of stiffened composite panels. Comput Struct
[54] Romeijn HE, Zabinsky ZB, Graesser DL, Neogi S. New reflection generator for 1996;58(3):543–55.
simulated annealing in mixed-integer/continuous global optimization. J [84] Madenci E, Kradinov V, Ambur DR. Application of genetic algorithm for
Optim Theory Appl 1999;101(2):403–27. optimum design of bolted composite lap joints. Compos Struct 2007;77(2):
[55] Genovese K, Lamberti L, Pappalettere C. Improved global-local simulated 148–59.
annealing formulation for solving non-smooth engineering optimization [85] Muc A, Gurba W. Genetic algorithms and finite element analysis in
problems. Int J Solid Struct 2005;42(1):203–37. optimization of composite structures. Compos Struct 2001;54(2–3):275–81.
[56] Rao ARM, Arvind N. Optimal stacking sequence design of laminate composite [86] Deka DJ, Sandeep G, Chakraborty D, Dutta A. Multiobjective optimization of
structures using tabu embedded simulated annealing. Struct Eng Mech laminated composites using finite element method and genetic algorithm. J
2007;25(2):239–68. Reinf Plast Compos 2005;24(3):273–85.
[57] Di Sciuva M, Gherlone M, Lomario D. Multiconstrained optimization of [87] Punch WF, Averill RC, Goodman ED, Lin SC, Ding Y, Yip YC. Optimal design of
laminated and sandwich plates using evolutionary algorithms and higher- laminated composite structures using coarse-grain parallel genetic
order plate theories. Compos Struct 2003;59(1):149–54. algorithms. Comput Syst Eng 1994;5:415–23.
 H. Ghiasi et al. / Composite Structures 90 (2009) 1–11 11

[88] Henderson JL. Laminated plate design using genetic algorithms and parallel [115] Massard TN. Computer sizing of composite laminates for strength. J Reinf
processing. Comput Syst Eng 1994;5:441–52. Plast Compos 1984;3:300–45.
[89] Kere P, Lento J. Design optimization of laminated composite structures using [116] Todoroki A, Sasada N, Miki M. Object-oriented approach to optimize
distributed grid resources. Compos Struct 2005;71(3–4):435–8. composite laminated plate stiffness with discrete ply angles. J Compos
[90] Le Riche R, Haftka RT. Improved genetic algorithm for minimum thickness Mater 1996;30(9):1020–41.
composite laminate design. Compos Eng 1995;5:143–61. [117] Narita Y. Layerwise optimization for the maximum fundamental frequency of
[91] Park CH, Lee Wl, Han WS, Vautrin A. Improved genetic algorithm for laminated composite plates. J Sound Vib 2003;263:1005–16.
multidisciplinary optimization of composite laminates. Comput Struct [118] Narita Y, Hodgkinson JM. Layerwise optimisation for maximising the
2008;86(19–20):1894–903. fundamental frequencies of point-supported rectangular laminated
[92] Liu B, Haftka RT, Akgun M. Composite wing structural optimization using composite plates. Compos Struct 2005;69:127–35.
genetic algorithms and response surfaces. AIAA J 1998; 98: paper 98: 4854. [119] Farshi B, Rabiei R. Optimum design of composite laminates for frequency
[93] Antonio CAC. A hierarchical genetic algorithm with age structure for constraints. Compos Struct 2007;81(4):587–97.
multimodal optimal design of hybrid composites. Struct Multidisc Optim [120] Ghiasi H, Pasini, D Lessard L. Layer separation for optimization of composite
2006;31:280–94. laminates. In: Proceedings of the ASME 2008 international design
[94] Sivakumar K, Iyengar NGR, Deb K. Optimization of composite laminates with engineering technology conference and computers and information in
cutouts using genetic algorithm. Eng Optim Variable Metric Complex Search engineering conference IDETC/CIE, Brooklyn, New York; August 3–6, 2008.
Meth 2000;32(5). [121] Soares CMM, Franco Correia V, Mateus H, Herskovits JA. Optimization of
[95] Pai N, Kawa A, Weng M. Optimization of laminate stacking sequence for laminated composite structures using a bilevel strategy. Compos Struct
failure load maximization using Tabu search. Compos B: Eng 2003;34(4): 1995;33:147–57.
405–13. [122] Kim TU, Sin HC. Optimal design of composite laminated plates with the
[96] Rao ARM, Arvind N. A scatter search algorithm for stacking sequence discreteness in ply angles and uncertainty in material properties considered.
optimisation of laminate composites. Compos Struct 2005;70:383–402. Comput Struct 2001;79(29–30):2501–9.
[97] Suresh S, Sujit PB, Rao AK. Particle swarm optimization approach for [123] Farshi B, Herasati S. Optimum weight design of fiber composite plates in
multi-objective composite box-beam design. Compos Struct 2007;81(4): flexure based on a two level strategy. Compos Struct 2006;73:495–504.
598–605. [124] Antonio CA. Multilevel genetic algorithm for optimization of geometrically
[98] Kathiravan R, Ganguli R. Strength design of composite beam using gradient non-linear stiffened composite structures. Struct Multidisc Optim
and particle swarm optimization. Compos Struct 2007;81:471–9. 2001;24:372–86.
[99] Chen J, Ge R, Wei J. Probabilistic optimal design of laminates using improved [125] Antonio CA, Torres Marques A, Soeiro AV. Optimization of laminated
particle swarm optimization. Eng Optim 2008;40(8):695–708. composite structures using a bi-level strategy. Compos Struct 1995;33:
[100] Aymerich F, Serra M. Optimization of laminate stacking sequence for 193–200.
maximum buckling load using the ant colony optimization (ACO) [126] Lund E, Stegmann J. On structural optimization of composite shell structures
metaheuristic. Compos A: Appl Sci Manuf 2008;39(2):262–72. using a discrete constitutive parametrization. Wind Energy
[101] Miki M. Material design of composite laminates with required in-plane elastic 2005;8(1):109–24.
properties. In: Hayashi T, Kawata K, Umekawa S. editors. Progress in science [127] Terada Y, Todoroki A, Shimamura Y. Stacking sequence optimizations using
and engineering of composites, ICCM-IV, Tokyo; 1982. p. 1725–31. fractal branch and bound method for laminated composites. JSME Int J A
[102] Tsai SW. Theory of composite design. Think composites, Dayton, OH, ISBN:0- 2001;44(4):490–8.
9618090-3-5; 1992. [128] Todoroki A, Terada Y. Improved fractal branch and bound method for
[103] Hammer VB, Bendsoe MP, Pedersen P. Parametrization in laminate design for stacking-sequence optimizations of laminates. AIAA J 2004;42(1):141–8.
optimal compliance. Int J Solid Struct 1997;34(4):415–34. [129] Hirano Y, Todoroki A. Stacking sequence optimizations for composite
[104] Fukunaga H, Vanderplaats GN. Stiffness optimization of orthotropic laminates using fractal branch and bound method (application for
laminated composites using lamination parameters. AIAA J 1991;29:641–6. supersonic panel flutter problem with buckling load condition). Adv
[105] Grenestedt JL, Gudmundson P. Layup optimization of composite material Compos Mater 2004;13(2):89–107.
structures. In: Pedersen P, editor. Optimal design with advanced [130] Hirano Y, Todoroki A. Stacking-sequence optimization of composite delta
materials. Amsterdam: Elsevier Science; 1993. p. 311–36. wing to improve flutter limit using fractal branch and bound method. JSME
[106] Miki M. Material design of fibrous laminated composites with required Int J A 2005;48(2):65–72.
flexural Stiffness. In: Mechanical behavior of materials-IV, proceeding of the [131] Matsuzaki R, Todoroki A. Stacking-sequence optimization using fractal
fourth international conference; 1985. p. 387–400. branch-and-bound method for unsymmetrical laminates. Compos Struct
[107] Miki M, Sugiyama Y. Optimum design of laminated composite plates using 2007;78:357–550.
lamination parameters. AIAA J 1993;31(5):921–2. [132] Todoroki A, Sekishiro M. New iteration fractal branch and bound method for
[108] Fukunaga H, Chou TW. Simplified design techniques for laminated cylindrical stacking sequence optimizations of multiple laminates. Compos Struct
pressure vessels under stiffness and strength constraints. J Compos Mater 2007;81:419–26.
1998;22:1156–69. [133] Karbhari VM, Wilkins DJ. Use of decision support systems for the efficient
[109] Lipton R. On optimal reinforcement of plates and choice of design selection and design of composites and their products. Int J Mater Prod Tech
parameters. Contr Cybernet 1994;23:481–93. 1992;7(2):125–49.
[110] Autio M. Determining the real lay-up of a laminate corresponding to optimal [134] Hochard Ch. Optimum design of laminated composite structures. Compos
lamination parameters by genetic search. Struct Multidisc Optim Struct 2004;63:159–65.
2000;20(4):301–10. [135] Cao Y, Li M Xu, Y Liu G. Parameter optimization of the dominant anisotropic
[111] Kameyama M, Fukunaga H. Optimum design of composite plate wings for layer of alternately laminated damped structures. Compos Struct
aeroelastic characteristics using lamination parameters. Comput Struct 2007;77(1):104–10.
2007;85(3–4):213–24. [136] Chattopadhyay A, Seeley CE, Jha R. Aeroelastic tailoring using piezoelectric
[112] Enrique HJ, Weavers PM, Friswell MI. Optimization of long anisotropic actuation and hybrid optimization. Smart Mater Struct 1999;8(1):83–91.
laminated fiber composite panels with T-shaped stiffeners. AIAA J [137] Luersen MA, Le Riche R. Globalized Nelder–Mead method for engineering
2007;45(10):2497–509. optimization. Comput Struct 2004;82(23–26):2251–60.
[113] Kere P, Koskli J. Multicriterion optimization of composite laminates for [138] Ghiasi H, Pasini D, Lessard L. Constrained globalized Nelder–Mead method for
maximum failure margins with an interactive descent algorithm. Struct simultaneous structural and manufacturing optimization of a composite
Multidisc Optim 2002;23:436–47. bracket. J Compos Mater 2008;42(7):717–36.
[114] Kere P, Lyly M, Koski J. Using multicriterion optimization for strength design [139] Oh SH, Lee BC. Development of an efficient algorithm for global optimization
of composite laminates. Compos Struct 2003;62:329–33. by simplex elimination. Eng Optim 2003;35(6):607–25.