Вы находитесь на странице: 1из 22

196 Int. J. Design Engineering, Vol. 6, No.

3, 2016

A Pareto optimisation approach for competing failure


criteria in composite design

David B. Segala and Thomas A. Wettergren*


Naval Undersea Warfare Center,
1176 Howell Street,
Newport, RI 02841 USA
Email: david.segala@navy.mil
Email: t.a.wettergren@ieee.org
*Corresponding author

Abstract: In multidimensional design optimisation, competing objectives lead


to performance trade-offs that can only be made judiciously once the trade-off
surface is understood in the context of the problem constraints. Examining the
Pareto front of the multi-objective problem provides the designer with an
understanding of which design parameter combinations lead to various
desirable trade-offs between competing objectives. We consider a bi-objective
design problem in composite materials, where we wish to maximise the energy
dissipation due to the cohesive failure between fibres and matrix material and
minimise fibre and matrix material failure of a dynamically loaded
unidirectional composite. We develop and utilise the genetic algorithm normal
boundary intersection (GANBI) method to iteratively determine the Pareto
front for this composite design framework. The utility of the approach is
demonstrated with numerical examples.

Keywords: genetic algorithms; Pareto optimisation; optimal design;


composites; finite element analysis; FEA; fibres; fracture.

Reference to this paper should be made as follows: Segala, D.B. and


Wettergren, T.A. (2016) ‘A Pareto optimisation approach for competing failure
criteria in composite design’, Int. J. Design Engineering, Vol. 6, No. 3,
pp.196–217.

Biographical notes: David B. Segala received his BS in Biomedical


Engineering and PhD in Applied Mechanics, both from the University of
Rhode Island in Kingston, RI. Following graduation, he secured a position with
the Navy working with the operational mathematics for off-board systems
group. He also holds an appointment as an Adjunct Professor in the
Engineering Department at Roger Williams University, where he teaches
special topics course on finite element modelling of physical systems. His
interest includes conducting basic and applied research in numerical modelling,
characterisation, and identification of complex physical systems in the presence
of limited information and/or computational complexity.

Thomas A. Wettergren received his BS in Electrical Engineering and PhD in


Applied Mathematics, both from the Rensselaer Polytechnic Institute. He
joined the Naval Undersea Warfare Center in Newport in 1995, where he has
served as a Research Scientist in the torpedo systems, sonar systems, and
undersea combat systems departments. He currently serves as the US Navy
Senior Technologist (ST) for Operational and Information Science, with a
concurrent title as a Senior Research Scientist at the Center. He also is an
Adjunct Professor of Mechanical Engineering at the Pennsylvania State

Copyright © 2016 Inderscience Enterprises Ltd.


A Pareto optimisation approach for competing failure criteria 197

University. His personal research interests are in planning and control of


distributed systems, applied optimisation, multi-agent systems, and search
theory. He is a senior member of the IEEE and a member of the Society for
Industrial and Applied Mathematics (SIAM).

1 Introduction

In the composites design community, multi-objective optimisation has been utilised to


satisfy specific design goals such as: minimise cost, weight, buckling and/or maximise
strength, energy absorption, resonance frequency, and stiffness in a certain direction.
Usually, one seeks to determine the optimal fibre orientation angle in a laminate. This
was first investigated by Walker (2001) who optimally determined the fibre orientation
angle of symmetrically laminated rectangle plates. The effects of bending and twisting
were investigated by maximising the biaxial buckling load and the resonance frequency.
Walker and Smith (2003) furthered this area by utilising a genetic algorithm with a finite
element analysis (FEA) model of a rectangle plate to minimise the weighted sum of the
mass and deflection of the plate to determine the fibre orientation and laminate thickness
for a set of discrete variables. The design constraint used was based on the Tsai-Wu
failure criterion (Tsai and Wu, 1971).
Further efforts in determining optimal fibre angles have been investigated using a
modified feasible direction search optimisation scheme. Topal (2009) used a laminated
cylindrical shell where the goal was to maximise the weighted sum of the fundamental
natural frequency and the critical buckling mode. Likewise, Topal and Uzman (2010)
applied a biaxial compression and uniform thermal load to a symmetrically angle-ply
square laminated plate where the multiple objectives were to maximise the buckling load
for the weighted sum of the biaxial compressive and the thermal loads. A genetic
algorithm was used (Bakar et al., 2013) to optimise woven composites by determining the
elastic constants and a particular weaving pattern. The design variables used here were
gap length, shape of the yarn section, yarn thickness, constituent materials, and fibre
volume fraction. The optimal patterns also provided high elastic properties which
minimised the fray problem due to high weave stiffness.
These early studies in multidimensional optimisation of composite design relied on a
designer combining the multiple objectives (usually via an additive weighting scheme) to
create a single objective for applying conventional optimisation techniques. Recently,
there has been interest in the utilisation of Pareto optimisation techniques to determine a
set of optimal designs that illustrates the optimal trade-off between the competing
multiple design objectives. Specifically, de Kruijf et al. (2007) and Almeida and Awruch
(2009) have both developed techniques that arrive at Pareto fronts for competing
multi-objective optimisation problems which yield a set of optimal designs. From these
optimal designs, the designer has the opportunity to chose the feasible design point which
he/she can build to. The paper by de Kruijf et al. (2007) used two conflicting design
criteria to optimise the stiffness and conduction performance of structures and materials.
Almeida and Awruch (2009) generated a Pareto front for the design of a square plate with
uniform pressure loading on its surface with the competing objectives to minimise
structural weight and deflection. Both of these studies utilised a traditional numerical
198 D.B. Segala and T.A. Wettergren

approach of solving a sequence of single objective optimisation problems to achieve the


approximation to the Pareto front.
Numerical methods for Pareto optimisation involve forming approximations to the
Pareto front by solving a sequence of optimisation problems. The variety of
gradient-based nonlinear programming approaches (Miettinen, 1998) all involve solving
a sequence of problems, each requiring many objective evaluations. Efficiency in
reducing the number of objective evaluations has been found by developing metaheuristic
approaches of evolutionary algorithms (see Deb, 2001; Coello Coello, 1999;
VanVeldhuizen and Lamont, 2000). In these methods, each iteration evaluates a sequence
of designs, and the iterations cause the entire sequence to converge to an approximation
of the Pareto front. This provides a great computational advantage for the efficient
production of an approximate Pareto front in problems where the individual objective
evaluations have a large computational cost.
Fibre-reinforced composites such as unidirectional composites offer an advantage
over conventional materials due to their tailorability. A designer has a suite of potential
design variables to adjust in order to meet specific design goals and requirements.
Furthermore, the designer might not realise the full potential of the composite if the
optimal set of design variables is not chosen. Customisable design variables include fibre
and matrix materials (e.g., elastic, plastic, viscoelastic, high strain rate, etc.), lamina
architecture, failure mechanisms (e.g., cohesive/adhesion failure, matrix cracking, fibre
breakage, fracture, etc.), and laminate stacking schedules and orientations. These
variables are chosen in order to satisfy specific design goals such as: minimise cost,
weight, buckling and/or maximise strength, energy absorption, resonance frequency, and
stiffness in a certain direction. The large number of design variables makes determining
the optimal design of fibre-reinforced composites a very complex problem.
Complexities in the design of fibre-reinforced composites are manifested in two
ways:
1 optimal designs could be problematic to obtain due to the high dimension and
existences of parameter dependencies/coupling in the design space
2 modelling challenges due to capturing intricate geometries, failure conditions, and
loading/boundary conditions.
In the first part, it is infeasible to fully explore the design space to determine the optimal
set of designs. In the same vein, the parameter space may not be relatively large but the
parameters are coupled. In this sense, a slight change in parameter values could yield
very different dynamics. Therefore, one can use a multi-objective optimisation procedure
to determine the set of optimal designs based on a predefined objective function that
satisfies the design criterion and/or constraints. Secondly, modelling complexities of the
composites has to rely on FEA to capture the intricate geometries, failure conditions, and
loading/boundary conditions. As a result, simulating a FEA model becomes very costly –
in both simulation time and data storage. It is not feasible to run every possible design
and access which one is the most optimal for the predefined goal or requirement.
Therefore, we need a robust procedure which gives a set of optimal points that a designer
can chose a feasible solution from.
Herein, we consider the design problem of simultaneously maximising the energy
dissipation due to the cohesive failure between fibres and matrix material and minimising
the fibre and matrix material failure of a dynamically loaded unidirectional composite. To
A Pareto optimisation approach for competing failure criteria 199

achieve this, we develop a FEA-based model of this dynamic composite fibre problem.
We then present a new method of evolutionary multi-objective optimisation to generate
Pareto fronts using a genetic algorithm approach. We refer to the new method as the
genetic algorithm normal boundary intersection (GANBI) method and illustrate its
effectiveness in generating an approximation to the Pareto front for a composite design
problem using a very small number of numerical evaluations of the objectives.
In the remainder of the paper, the failure mechanisms in the composite model and the
numerical simulation of the model will be presented in Sections 2 and 3. In Section 4, the
multi-objective optimisation problem will be presented along with GANBI. The results of
GANBI will be presented in Section 5 and then the physical interpretation of the results
and what they mean for the composite designer will be presented in Section 6.

2 Energy absorption failure modes

It is commonly the goal of a designer or manufacturer to design fibre-reinforced


composites such that the cohesive bonds between the fibre and matrix interface fail
before the fibre or matrix material fails. In the composite model presented here, we will
consider material (fibre and matrix) failure and cohesive (fibre-matrix interfacial) failure.
These three failure modes are explicitly modelled using ABAQUS’s internal subroutines
(ABAQUS, 2011).

2.1 Cohesive failure


Here, we will briefly present the failure modes and the composite model. A linear elastic
traction-separation damage model is used to model the cohesive failure modes for the
cohesive surfaces defined around each fibre and the surface inside the matrix where the
fibres are nested. The cohesive failure can be modelled as a two-step procedure. First, a
maximum separation-based linear damage initiation law is used until the separation
between fibre and matrix material reaches a specified threshold value δo. Once the
separation criterion is met, the process of degradation begins by a displacement-based
linear damage evolution law.
From a mathematical perspective, the maximum separation criterion for the linear
damage initiation is expressed as
⎧ δ δ δ ⎫
max ⎨ no , so , ot ⎬ = 1, (1)
δ
⎩ n δs δt ⎭

where δn is the separation in the pure normal direction, δs is the separation in the first
shear direction, δt is the separation in the second shear direction, δno , δso , δto are the
effective separations at damage initiation, and <.> represents Macaulay brackets.
As in the initiation law, the evolution law is based on displacements as well. A scalar
damage variable D represents the overall damage at the contact point with an initial value
of zero. When a value of one is reached the bond has failed. The contact stress
components are a function of the damage variable and are defined in equations (2), (3),
and (4):
200 D.B. Segala and T.A. Wettergren

⎧(1 − D) tn , tn ≥ 0
tn = ⎨
⎩ tn , otherwise
ts = (1 − D) ts , (2)
tt = (1 − D) tt ,

where tn , ts and tt are the contact stress components without damage. The value of D is

δmf ( δmmax − δmo )


D= , (3)
δmmax ( δmf − δmo )

and where
2
δm = δn + δs2 + δt2 , (4)

is the maximum value of the effective separation attained during the loading history, and
δmf − δmo is specified as the difference between the separation at failure and initiation at a
contact point.

2.2 Fibre breakage and matrix cracking


The same approach as was described for the cohesive failure was taken to model both
fibre breakage and matrix shearing – except for how damage initiation is defined.
Damage initiation is based on a shear failure strain or an equivalent plastic strain. Once
the equivalent plastic strain value is reached for damage initiation, the same
displacement-based evolution law, as described previously for the cohesive damage
evolution, is used.

3 Ballistic impact of unidirectional composite model

For this study, we use a model of a quarter-symmetric (0/90) UD fibrous composite


laminate subjected to an impact by a fragment simulating projectile (FSP) (see Segala
and Cavallaro, 2014). Quarter-symmetric representations were permissible provided that
the material properties, displacement boundary conditions, and applied loads were
collectively quarter-symmetric. In Figure 1, the quarter-symmetry planes are the YZ- and
YX-planes. The boundary conditions were such that no translations normal to the planes
were admissible. The fibre and matrix material were all modelled with quad elements
(ABAQUS type Q4). The characteristic length of an element is 0.0001 in or 2.54 μm. The
size and spacing between the fibres are depicted in Figure 2.
The models shown in the present analyses were constructed using quarter-symmetric
representations that were permissible provided that the material properties, displacement
boundary conditions, and applied loads were collectively quarter-symmetric. In Figure 1,
the quarter-symmetry planes are the YZ- and YX-planes. The boundary conditions were
such that no translations normal to the planes were admissible. The key advantage to
using quarter symmetric representations is reduced computational time with no loss of
solution robustness. Because of the relatively small dimensions of the fibre
A Pareto optimisation approach for competing failure criteria 201

cross-sections, ply thicknesses, and the level of mesh discretisation applied, the solution
time increment was on the order of 4E-11 second.

Figure 1 (a) Surface view of the two-ply model of the polymer matrix (green) and the UHMWPE
fibres (red) (b) Subset view of the model depicting the mesh detail adjacent to two
fibres (see online version for colours)

(a) (b)

Figure 2 Size and spacing between the fibres in the matrix material model

The isotropic polymer matrix material properties are listed in the first row of Table 1. The
matrix was reinforced with ultra-high-molecular-weight polyethylene (UHMWPE) fibres
whose material properties are listed in the second row of Table 1. General contact
surfaces and surface interactions (hard overpressure) were established between the matrix
and fibres and between the fibres with each other.
202 D.B. Segala and T.A. Wettergren

Figure 3 The FSP just prior to impacting the composite laminate (see online version for colours)

Table 1 Isotropic material properties for polymer matrix (Row 1) and UHMWPE fibres
(Row 2)

E (psi) ρ (lb/in3) ν Ultimate strength (psi) ε p (in/in)


435,113.0 1.68E-4 0.4 13,900 2.5
17.1E7 1.3E-4 0.3 5.8E5 0.36
Source: Grujicic et al. (2008)
For the ballistic impact case, contact was defined for the matrix and FSP as well as the
fibres and FSP. It was necessary to establish contact between the fibres and FSP because,
as the matrix material degraded due to failure, the FSP came into contact with the fibres.
If fibre-to-FSP contact was not established, the FSP would simply pass through the
fibres. A nominal coefficient of friction, μ = 0.1, was used between the fibres, matrix, and
FSP materials.
When designing composite laminates that are subjected to severe loadings such as
ballistic impact, it is most desirable to arrest the ballistic without compromising the
structural integrity of the composite laminate. This task can be achieved by designing the
composite laminate such that it experiences cohesive failure along the fibre/matrix
interface. When the cohesive bond between the fibre and matrix breaks due to cohesive
failure, as a result of the incoming energy transferred from the ballistic to the laminate,
strain energy is released. This release of strain energy serves as a vehicle for energy
dissipation which will eventually lead to arresting the ballistic. The energy release due to
fibre breakage and matrix cracking will also mitigate energy from the ballistic. However,
the designer does not want to break or fracture the fibre and/or the matrix material
because the laminate would then result in structural decay.
In order to arrest the ballistic or mitigate the imparted energy on the composite due to
the ballistic impact, the two competing objectives will be to minimise material damage
and maximise cohesive failure. These two objectives will ensure the greatest survivability
of the composite while ensuring the least amount of energy transfer into or through the
composite. The damage due to fibre failure, matrix failure, and total failure are output
A Pareto optimisation approach for competing failure criteria 203

variables in ABAQUS which are reported as energy dissipation due to the specific
damage variables. From here, we define material failure (JMAT) as fibre failure plus matrix
failure and cohesive failure (JCOH) as the total failure minus material failure. One should
note that there are other energy loss terms but they are negligible as compared to these
values used here.
The design space that we will consider here are the displacements in the traction vs.
separation failure laws for the cohesive failure. We assume that the designer has
determined (either through experience or through some optimal design means) the fibre
and matrix materials to use. The two displacements are the effective separation damage
initiation (δo) and effective separation at failure (δf). Figure 4 depicts the design space for
the effective separations in the traction vs. separation cohesive failure law where δo and δf
are defined by:

δ o = δmin
o
+ A ⎣⎡δmax
o
− δmin
o
⎦⎤ (5)

f
δ f = δ o + δmin + B ⎡⎣δmax
f f
− δ o − δmin ⎤⎦ , (6)

where 0 ≤ A ≤ 1, 0 ≤ B ≤ 1. The range of damage initiation displacements are between


o o
δmin = 1 μm and δmax = 5 μm. Physically, this is equivalent to the fibre displacing up to
18% of its diameter before damage has fully set in and the bond starts to break. The range
f
of damage evolution displacements are between δmin = 2 μm + δ0 and δmax f
= 20 μm.
Physically, this is equivalent to the fibre displacing up to 71% of its diameter before the
bond completely fails and releases energy. At this value, the fibre is almost in the same
starting area of its neighbouring fibre. The evolution displacement’s dependence on the
minimum distance acts as a threshold just higher than that of the initiation displacement.
These displacements are graphically illustrated in Figure 5 where the original location of
the fibre (black circle) is displacement to its maximum (blue circle) at the initiation point
and then to its maximum (red circle) at the evolution point.

Figure 4 Design space for the displacements in the traction vs. separation cohesive failure law
(see online version for colours)
204 D.B. Segala and T.A. Wettergren

Figure 5 Graphical representation of the fibre separation for both damage initiation (blue) and
evolution (red) with respect to the original location of the fibre (black) (see online
version for colours)

4 Multi-objective optimisation

To solve multi-objective optimisation problems, the problem is formally represented as a


mathematical optimisation of a vector-valued function In particular, the k-objective
multi-objective optimisation problem (MOP) has as a goal to find the design(s) x that
minimise the set of objectives {f1(x), f2(x), …, fk(x)} subject to the constrained design
space D. A feasible design x* ∈ D is called Pareto optimal for a multi-objective problem
if all other designs x ∈ D have a higher value for at least one of the objective functions fi,
with i = 1, 2, ∙∙∙, k.
To develop mathematical techniques with multi-objective problems, it is convenient
to work with the individual minima of each objective. In particular, the individual
minima fi* of the ith objective fi(x) is given by

fi* = fi ( xi* ) , where xi* = { x ∈ D : fi ( x) ≤ fi ( y ), ∀y ∈ D} . (7)

This is the minimum value that the objective fi(x) obtains over the entire design space D.
Clearly, there is an individual minima for each objective that in some way bound the
achievable designs in objective space.
The convex hull of individual minima (CHIM) is a set that connects the individual
minima, and acts as a crude approximation of the Pareto front. Mathematically, it is given
by

{
C * = ΦB + F * : B = [b1 , b2 , ..., bk ] ,
T
∑ b = 1, b ≥ 0}.
i i (8)
A Pareto optimisation approach for competing failure criteria 205

where F* is the vector of individual minima given by F * = [ f1* , f 2* , ..., f k* ]T and Φ is the
k × k matrix whose ith column is given by Fi* − F * , where Fi* = F ( xi* ) is the vector
evaluation of the multi-objective performance of the specific design xi* that achieves the
minimum for the ith objective fi(x). Details of the construction of CHIM are found in Das
and Dennis (1998). An example of a CHIM for a two-objective minimisation problem is
shown in Figure 6.

Figure 6 Example of the CHIM for a two-objective minimisation problem (see online version
for colours)

Note: The Pareto front is shown in red, the individual minima are shown as stars (*), and
the CHIM is given by the black line segment.

4.1 Conventional approaches to multi-objective optimisation


Practical design optimisation requires more calculation than that provided by simple
scoping tools and thus require numerical optimisation tools. There are many approaches
for multi-objective design optimisation, which we separate into two primary categories:
gradient search methods and evolutionary optimisation methods. The gradient search
methods are based on functional relationships between design parameters and objectives.
Unfortunately, many of the problems encountered in single-objective gradient search
problems (convergence to local minima, poor estimates of gradient values, etc.) become
even more troublesome when the size of the parameter space (the dimension of x)
increases, as is often the case in practical systems design. These problems are further
exacerbated when the size of the objective space (the dimension k of F) is increased.
Examples of gradient search techniques for multi-objective optimisation are shown by
Miettinen (1998) and Voronin and Ziatdinov (2014). There are practical issues with the
use of these gradient search techniques that have been well-documented (Das and Dennis,
1997).
Evolutionary methods of optimisation are based on forming large sets of potential
designs x, evaluating the relative performance of the designs within each set, and using
some evolutionary principle to combine traits of members of each set to generate a new
206 D.B. Segala and T.A. Wettergren

set of candidate designs. This process is repeated until the iterations stabilise, at which
time the solution is presumed to have converged. These evolutionary methods include
techniques such as genetic algorithms, simulated annealing, particle swarm methods, and
ant colony optimisation. We restrict our interest in evolutionary methods to genetic
algorithms, as they are readily applicable to design problems. Commonly used genetic
algorithm approaches to multi-objective optimisation include the vector evaluated genetic
algorithm (VEGA) (Schaffer, 1984), the niched Pareto genetic algorithm (NPGA) (Horn
et al., 1994), and the non-dominated sorting genetic algorithm (NSGA) (Srinivas and
Deb, 1995). The problem with these techniques is that they tend to have a poor spread of
points along the Pareto front in practical problems, especially when evaluating only a
small number of sample designs.
To handle the problems involved with the genetic algorithm-based solvers, while also
avoiding the issues inherent in gradient approaches, we have developed a new
preprocessor for genetic multi-objective optimisation solvers. This preprocessor provides
good spread along the Pareto set in poorly-scaled multi-objective settings, while
maintaining the desirable features of the existing genetic algorithm. Our new
preprocessor-based approach is referred to as the GANBI method.

4.2 The GANBI method


The use of normal boundary intersection techniques as a method to obtain an estimate of
the Pareto front was developed for gradient search techniques by Das and Dennis (1998).
While theoretically effective, those methods do not apply well to practical engineering
problems as the numerical gradients that are required are computationally prohibitive and
often of questionable numerical accuracy. We extend the work on normal boundary
intersection techniques to make them appropriate for use as a preprocessor for
general-purpose multi-objective solvers.
The GANBI method is based on the concept of building up an approximate Pareto
surface by advancing this coarse approximation of the CHIM as far as possible at
regularly spaced intervals while staying within the constraints of the design space D. This
concept is the basis of the NBI approach of Das and Dennis (1998) as developed for use
in gradient search optimisation techniques. To illustrate the Pareto approximation used by
GANBI, consider the bi-objective problem from Figure 6. If we pick four uniformly
placed normals to the CHIM as shown in Figure 7, then the new multi-objective problem
of maximising the distance from the CHIM while simultaneously minimising the distance
to each of the normals becomes a multi-objective problem that explicitly seeks to
maintain a good spread of points along the Pareto front. Because the spread along the
Pareto front is now achieved explicitly, rather than through a sheer number of points, this
makes the GANBI approach useful for design problems where objective evaluations are
very computationally costly. To formally develop the objectives used in the GANBI
approach, we begin with m evenly-spaced points Fi along the CHIM.
The vector normal to the CHIM that goes through a given point Fi on the CHIM is
given by
Ni (ξ ) = Fi − ξ ΦT U , (9)

where U represents a vector of all ones, i.e., U = [1, 1, …, 1]T. Note that the normal Ni(ξ)
in equation (9) is parameterised by ξ ≥ 0, where ξ = 0 represents the location on the
A Pareto optimisation approach for competing failure criteria 207

CHIM and increasing values of ξ correspond to moving further from the CHIM. Thus, the
goal of GANBI is to find designs along Ni(ξ) that maximise the value of ξ for each Ni(ξ)
while staying within the constraints of the design space D. These objectives are obviously
in conflict, and the multi-objective solution of this optimisation problem provides an
estimate of the Pareto surface that tends to lead to a good spread of points along the
Pareto front for a small number of objective evaluations.

Figure 7 Example of the concept of the GANBI method applied to the problem in Figure 6
(see online version for colours)

Note: In this example, four normals are chosen as locations to ‘push out’ the CHIM to
form an approximation to the Pareto front, given by the black segmented line.
The individual objective of maximising distance along the line Ni(ξ) is, in itself, a
multiple objective. One goal is to stay close to the line, and a second is to move far along
it. Rather than consider these goals as further dimensions of the objective space, we
develop a single objective for each normal Ni(ξ) that balances these goals. To form the
objective for a given design x, we find the point along Ni(ξ) that is closest to F(x) in a
least-squares Euclidean sense. This value of ξ is given by ξF(x) as

ξ F ( x ) = ( ( ΦT U ) ΦT U ) ( ΦT U ) Gi ( F ( x) ) ,
T −1 T
(10)

where
Gi ( F ( x) ) = F * − F ( x) − ΦT Bi , (11)

for the normal Ni(ξ).


Then, the objective we use for the GANBI method for each normal Ni(ξ) is to
minimise the function hi(x) given by

hi ( x) = d ( Ni ( ξ F ( x ) ) , F ( x) ) − 2ξ F ( x ) (12)

where d(·, ·) is the Euclidean distance measure in the objective space. We note that the
objective hi(x) in equation (12) is only dependent on the objective space evaluation of the
design x, not explicitly on the design parameters themselves. The value of two found in
208 D.B. Segala and T.A. Wettergren

the weighting between the components of the objective in equation (12) has been found
to be very good for convergence. In using GANBI, the objective functions hi(x) are
evaluated for each i = 1, …, m for each given design x, and the resulting vector-valued
objectives H(x) = [h1(x), h2(x), …, hm(x)]T are the result of the GANBI preprocessor that
is used in a standard genetic multi-objective solver. Figure 8 shows a flowchart of the
procedure used to apply the GANBI method to an engineering design problem using
FEM analysis. Note that the shaded box in the middle of the flowchart is the only step of
the entire process that requires the FEM analysis of the engineering design to be
optimised. It is this efficiency in greatly reducing the number of evaluations required that
makes GANBI particularly beneficial for systems that require very computationally
complex FEM analysis evaluations.

Figure 8 Flowchart describing the GANBI method for Pareto optimisation

For the examples in this paper, we utilise the NSGA algorithm (Srinivas and Deb, 1995)
as our standard solver to use with GANBI. The approach shown in GANBI has been
previously shown (Wettergren, 2006) to provide rapid convergence to well-spread
A Pareto optimisation approach for competing failure criteria 209

representations of the Pareto front in analytically tractable two-objective optimisation


problems. We also note that those same analytical solutions were obtained with m = 4
normals, which we use as a guideline for our analysis.

5 Results

To develop an approximation to the Pareto front with a small number of total objective
evaluations, we ran the GANBI method for m = 4 normals using a genetic population of
eight members for each generation. A population size of 8 was chosen as it is the minimal
number required to exercise the GANBI method, and we have a desire to minimise the
number of design evaluations considered in this analysis due to their computational
complexity. Each design is represented by an ordered pair x = (δ0, δf) and ABAQUS runs
were made to obtain the two objective values of F(x) = (JCOH, JMAT), where δ0 and δf are
the effective separations at damage initiation for the cohesive bond and the effective
separations at failure of the cohesive bonds, respectively, and JCOH and JMAT are the
energy dissipation due to the failure of the cohesive bonds and the energy dissipation due
to the failure of the material (both fibre and matrix), respectively. The initial population
was chosen by randomly selecting eight values of x within the constrained design space
shown in Figure 4. Those initial designs were each evaluated to obtain the objective
results shown in Figure 9. Since the goal of this problem is to minimise JMAT and
maximise JCOH, we show the non-dominated points as those in the lower right of the
figure, marked by red x’s. We similarly show the corresponding mapping of the space
back into the design (parameter) space in Figure 10. Note that in this design space view
there are not enough points to have a clear delineation of regions of the design space that
correspond to non-dominated designs.

Figure 9 Objective space view of non-dominated (red x’s) and dominated (blue o’s) designs
found from the first iteration (see online version for colours)

Note: Values shown are in micro-Joules (μJ).


210 D.B. Segala and T.A. Wettergren

Figure 10 Parameter space view of non-dominated (red x’s) and dominated (blue o’s) designs
found from the first iteration (see online version for colours)

Note: Values shown are in microns (μm).

Figure 11 Objective space view of non-dominated (red x’s) and dominated (blue o’s) designs
found from the second iteration (see online version for colours)

Note: Values shown are in micro-Joules (μJ).


Taking the resulting objective values F(x) from the first iteration, we executed the
GANBI procedure to generate a new generation of eight designs x to improve the Pareto
approximation. These new designs were evaluated to find another set of F(x) objectives
that were aggregated with the first set to have 16 total designs to evaluate
domination/non-domination to determine a new Pareto approximation. The results are
shown in Figure 11 along with the corresponding design (parameter) values in Figure 12.
A Pareto optimisation approach for competing failure criteria 211

From these plots it is clear that there is more distinction in the Pareto front. Also note that
the Pareto front, while only containing six points, extends across the objective space of
Figure 11. We also can see that there are distinct regions of non-dominated designs
appearing in the design space of Figure 12.

Figure 12 Parameter space view of non-dominated (red x’s) and dominated (blue o’s) designs
found from the second iteration (see online version for colours)

Note: Values shown are in microns (μm).

Figure 13 Objective space view of non-dominated (red x’s) and dominated (blue o’s) designs
found from the third iteration (see online version for colours)

Note: Values shown are in micro-Joules (μJ).


212 D.B. Segala and T.A. Wettergren

Figure 14 Parameter space view of non-dominated (red x’s) and dominated (blue o’s) designs
found from the third iteration (see online version for colours)

Note: Values shown are in microns (μm).


We continued the GANBI procedure to develop a third generation of designs, and again
computed the corresponding objective results. The resulting non-dominated solutions that
provide an approximate Pareto front for the entire 24 designs that were evaluated are
shown in Figure 13. The corresponding design (parameter) values are given in Figure 14.
We note that there appears to be a clear view of a Pareto front, even though we only
evaluated a total of 24 designs, wherein these 24 designs were chosen through a directed
random search using the GANBI multi-objective solver.
To show the effectiveness of the Pareto approximations shown for each generation,
we use some common measures of Pareto algorithm performance (Sayin, 2000). The first
measure we use is the number of non-dominated solutions in a generation. For a set of P
total population members, let the subset Pd ⊂ P represent the dominated members and
Pnd ⊂ P represent the non-dominated members, where Pd ∪ Pnd = P . The cardinality of
Pnd , given by | Pnd |, is a simple measure of quality of the Pareto approximation. This
measure is shown in the second column of Table 2 and is also given by the number of x’s
shown in the objective space plots for each generation. We expect that the cardinality will
increase with each generation for an effective algorithm as there is a larger number of
total population members created with each generation.
Another popular measure is the uniformity of the Pareto approximation (Sayin, 2000).
The uniformity δ is defined as the smallest distance between any two members of the
Pareto front. Since the two objectives in this application have different scales, we scale
the distance measure by taking the relative distance in each objective (relative to the
extremal points of the entire population in that objective) and use the Euclidean distance
in that scaled space. Then, formally we have
A Pareto optimisation approach for competing failure criteria 213

δ = min d ( xi , x j ) , ( xi , x j ) ∈ Pnd , (13)


i≠ j

as the measure of uniformity of the Pareto approximation. We show those values for each
generation as the third column of Table 2. Note that the first and second generation have
the same value since they share common designs within the non-dominated set. Clearly,
it is desirable to have a larger measure of uniformity to spread points out along the Pareto
front. Unfortunately, the uniformity measure suffers from a bias based upon the
cardinality (more points are more likely to have closer distances) and also is biased by
any single pair of neighbouring designs. To alleviate these issues, a measure of spread
has been proposed (Deb, 2001) that averages the uniformity across the entire set of
non-dominated designs, and normalises by the cardinality. Specifically, the measure of
spread Δ is given by


Pnd
dj −d
j =1
Δ= (14)
Pnd d

where dj is the distance from the jth non-dominated design to its nearest neighbour
(amongst non-dominated designs in the objective space) and d is the mean value of the
dj’s. Pareto front approximations with a larger spread have more evenly spread apart
non-dominated points that span the extent of the extremal values of the design space. The
spread values for our problem are shown in the rightmost column of Table 2. Clearly,
they show that the Pareto approximations improve with each generation.
Table 2 Quality measures for the Pareto approximations obtained

Generation Total designs | P | Cardinality | Pnd | Uniformity δ Spread Δ


1 8 4 0.115 0.366
2 16 6 0.115 0.393
3 24 8 0.040 0.653

6 Discussion

In a design optimisation framework, the optimal set of design parameters or


corresponding values to those parameters is usually left up to the designer who
sometimes bases the decision on experience alone. Here, we offer a methodology which
takes the designer out-of-the-loop and determines the optimal set of values for the design
parameters which are chosen a priori. The methodology utilises a genetic algorithm
(GANBI) to determine the Pareto front for a given set of competing objective functions.
Now, the decision on what is ‘optimal’ and what are ‘optimal’ values is solely based on
objective functions that are provided from the subject matter experts.
214 D.B. Segala and T.A. Wettergren

Figure 15 Objective space view of non-dominated designs with regions corresponding to separate
sections of the Pareto space (see online version for colours)

Notes: Regions are separated by different marker shapes. Values shown are in
micro-Joules (μJ).

Figure 16 Parameter space view of non-dominated (red x’s) designs with regions corresponding
to those found in the Pareto space of Figure 15 (see online version for colours)

Notes: Regions are separated by different marker shapes corresponding to those in


Figure 15. Values shown are in microns (μm).
In this work, we sought out to determine the optimal separation displacements in a linear
damage initiation and evolution law to describe the cohesive failure between the fibre and
matrix material of a unidirectional composite. The two competing objective functions
were to minimise material damage (the aggregate of fibre and matrix damage) while
A Pareto optimisation approach for competing failure criteria 215

maximising the cohesive failure. These two objectives will ensure the greatest
survivability of the composite while ensuring the least amount of energy transfer into or
through the composite due to the imparted energy of the ballistic impact.
A desirable trait of Pareto optimisation approaches for design optimisation is to be
able to identify regions of the design space (parameter space) that correspond to specific
portions of the Pareto front. To do this for the composites design problem, we have taken
the plots from Figure 13 and Figure 14 and removed the dominated points (the blue
circles). From the resulting non-dominated-only points, we have grouped the points by
their relative proximity to one another in the objective space. Those groupings are shown
in Figure 15 and their corresponding designs are shown in parameter space in Figure 16.
Figure 16 clearly demonstrates the difficulty in choosing the optimal values of the
two separation displacements that satisfy the objective functions from only the parameter
space. From a build-test-build perspective, it would be a prohibitive task in both cost and
material to determine these displacements or groupings. However, the inclusion of
information about the objective space can steer the designer to the groupings in parameter
space.
From a design standpoint, we can interpret what the composite design would
resemble for the given groupings in objective space, as shown in Figure 15. First,
consider the light blue (Δ) region that is able to completely maximise the cohesive failure
but poorly minimises the material failure. As a grouping, it actually performs the best in
the cohesive failure objective but performs the worst in the material failure objective.
Mapping to the parameter space results in a very small region where the two designs are
almost overlapping. In both of these designs, the maximum values of both separation
displacements are found. Physically, these designs would chose the maximum the fibre
can displace in both the damage initiation and evolution. Next, if we consider the
grouping with the black circle (×), there is only one design. This was not grouped in with
either of the green (◊) or purple (□) groupings due to the distance apart in the parameter
space. Although it does moderately well in both objectives, it would be very challenging
from a design perspective to build this design as the cohesive bond would have to have
the exact properties.
The green group results in the least amount of material failure as compared to the
other groupings and does moderately well at maximising the cohesive failure. By
mapping this region to parameter space, the grouping has constant value for the damage
initiation separation but a large spread, at almost 50% of the design space, in values for
the damage evolution separation. Next, the purple group is roughly in the middle of the
material failure objective and adequately maximises the cohesive failure. In the parameter
space, the same conclusion can be drawn for the green grouping as did the purple. The
relatively constant value of damage evolution separation is needed but there can be some
variability in the damage initiation separation.
Interestingly, referring to the green and purple group, a unique design trade-off based
on the available materials and technologies the designer may have can be seen. If the
designer has the ability to design the cohesive bond to specific damage initiation values
much easier than specific damage evolution values, then the best choice would be to
choose the green grouping. In the opposite case, if the damage evolution value is easier
than the purple group would be the optimal choice.
216 D.B. Segala and T.A. Wettergren

7 Conclusions

In multi-objective design optimisation, optimal designs can be established by determining


the Pareto front for a set of competing objective functions. Here, we present a general
method to determine the Pareto front based on competing objectives and demonstrate the
results through a numerical example of the design of a composite laminate which is
subjected to a ballistic impact. We iteratively determine the Pareto front using the
GANBI method to determine the optimal set of designs to minimise material damage (the
aggregate of fibre and matrix damage) while maximising the cohesive failure. With only
24 numerically evaluated designs, the derived Pareto front yields four distinct groups in
objective space which are mapped to corresponding regions in parameter space. A
designer can choose the appropriate designs (separation displacements) for the materials
and technologies that exist based on the identified regions in parameter space. Two of the
regions in parameter space give the designer more flexibility in cohesive bond damage
initiation or evolution displacements values which could reduce the need for a unique or
specialised material.

Acknowledgements

This work has been supported by the Section 219 internal investment program of the
Naval Undersea Warfare Center.

References
ABAQUS/Explicit (2011) Version 6.11, ABAQUS Inc.
Almeida, F. and Awruch, A. (2009) ‘Design optimization of composite laminated structures
using genetic algorithms and finite element analysis’, Composite Structures, Vol. 88, No. 3,
pp.443–454.
Bakar, I.A.A., Kramer, O., Bordas, S. and Rabczuk, T. (2013) ‘Optimization of elastic properties
and weaving patterns of woven composites’, Composite Structures, Vol. 100, pp.575–591.
Coello Coello, C.A. (1999) ‘An updated survey of evolutionary multiobjective optimization
techniques: state of the art and future trends’, in Proceedings of the Congress on Evolutionary
Computation, Vol. 1, pp.3–13.
Das, I. and Dennis, J.E. (1997) ‘A closer look at drawbacks of minimizing weighted sums of
objectives for Pareto set generation in multicriteria optimization problems’, Structural
Optimization, Vol. 14, No. 1, pp.63–69.
Das, I. and Dennis, J.E. (1998) ‘Normal-boundary intersection: a new method for generating the
Pareto surface in nonlinear multicriteria optimization problems’, SIAM Journal of
Optimization, Vol. 8, No. 3, pp.631–657.
de Kruijf, N., Zhou, S., Li, Q. and Mai, Y-W. (2007) ‘Topological design of structures and
composite materials with multiobjectives’, International Journal of Solids and Structures,
Vol. 44, No. 22, pp.7092–7109.
Deb, K. (2001) Multiobjective Optimization Using Evolutionary Algorithms, Wiley, Chichester,
UK.
Grujicic, M., Arakere, G., He, T., Bell, W.C., Cheeseman, B.A., Yen, C.F. and Scott, B. (2008) ‘A
ballistic material model for cross-plied unidirectional ultra-high molecular weight
polyethylene fiber-reinforced armor-grade composites’, Materials Science and Engineering A,
Vol. 498, Nos. 1–2, pp.231–241.
A Pareto optimisation approach for competing failure criteria 217

Horn, J., Nafpliotis, N. and Goldberg, D.E. (1994) ‘A niched Pareto genetic algorithm for
multiobjective optimization’, in Proceedings of the First IEEE Conference on Evolutionary
Computation (ICEC 94), Vol. 1, pp.82–87.
Miettinen, K. (1998) Nonlinear Multiobjective Optimization, Kluwer, Norwell, MA.
Sayin, S. (2000) ‘Measuring the quality of discrete representations of efficient sets in multiple
objective mathematical programming’, Mathematical Programming Series A, Vol. 87, No. 3,
pp.543–560.
Schaffer, J.D. (1984) Some Experiments in Machine Learning Using Vector Evaluated Genetic
Algorithms, PhD dissertation, Vanderbilt University.
Segala, D.B. and Cavallaro, P.V. (2014) ‘Numerical investigation of energy absorption
mechanisms in unidirectional composites subjected to dynamic loading events’,
Computational Materials Science, Vol. 81, pp.303–312.
Srinivas, N. and Deb, K. (1995) ‘Multiobjective function optimization using nondominated sorting
genetic algorithms’, Evolutionary Computation, Vol. 2, No. 3, pp.221–248.
Topal, U. (2009) ‘Multiobjective optimization of laminated composite cylindrical shells for
maximum frequency and buckling load’, Materials & Design, Vol. 30, No. 7, pp.2584–2594.
Topal, U. and Uzman, I. (2010) ‘Multiobjective optimization of angle-ply laminated plates’, Finite
Elem. Anal. Des., Vol. 46, No. 3, pp.273–279.
Tsai, S. and Wu, E. (1971) ‘A general theory of strength for anisotropic materials’, Journal of
Composite Materials, Vol. 5, No. 1, pp.58–80.
VanVeldhuizen, D.A. and Lamont, G.B. (2000) ‘Multiobjective evolutionary algorithms: analyzing
the state-of-the-art’, Evolutionary Computation, Vol. 8, No. 2, pp.125–147.
Voronin, A.N. and Ziatdinov, Y.K. (2014) ‘Hierarchical structure of multicriteria problems’,
Intelligent Control and Applications, Vol. 5, No. 1, pp.12–18.
Walker, M. (2001) ‘Multiobjective design of laminated plates for maximum stability using the
finite element method’, Composite Structures, Vol. 54, No. 2, pp.389–393.
Walker, M. and Smith, R. (2003) ‘A technique for the multiobjective optimisation of laminated
composite structures using genetic algorithms and finite element analysis’, Composite
Structures, Vol. 62, No. 1, pp.123–128.
Wettergren, T.A. (2006) The Genetic-Algorithm-Based Normal Boundary Intersection (GANBI)
Method: An Efficient Approach to Pareto Multiobjective Optimization for Engineering
Design, NUWC-NPT Technical Report 11,741, Naval Undersea Warfare Center Division
Newport [online] http://www.dtic.mil/dtic/tr/fulltext/u2/a455270.pdf (accessed 16 February
2016).

Вам также может понравиться