Optim Eng (2011) 12:681717

DOI 10.1007/s11081-010-9132-0

Optimal topology design of industrial structures using

an evolutionary algorithm

Raj Das Rhys Jones Yi-Min Xie

Received: 25 January 2009 / Accepted: 20 December 2010 / Published online: 12 January 2011
Springer Science+Business Media, LLC 2011

Abstract The paper demonstrates the application of a modified Evolutionary Struc-

tural Optimisation (ESO) algorithm for optimal design of topologies for complex
structures. A new approach for adaptively controlling the material elimination and a
gauss point average stress is used as the ESO criterion in order to reduce the gen-
eration of checkerboard patterns in the resultant optimal topologies. Also, a conver-
gence criterion is used to examine the uniformity of strength throughout a structure.
The ESO algorithm is validated by comparing the ESO based solution with the result
obtained using another numerical optimisation method (SIMP).
The capabilities of ESO for producing an optimal design against a specified
strength constraint are illustrated using two industrial design problems, viz: a bulk-
head used in an aircraft structure and a sideframe used in a railway freight wagon.
It has been shown that topology optimisation using ESO can result in a considerable
reduction in the weight of a structure and an optimum material utilisation by generat-
ing a uniformly stressed structure. The ESO algorithm was also applied to the shape
optimisation of a local geometry of the sideframe to (locally) reduce stress levels.
The paper evaluates and establishes the ESO method as a practical tool for optimum
topology and subsequent shape design problems for complex industrial structures.

Keywords Structural optimisation Topology optimisation Evolutionary

algorithm Finite element analysis Structural analysis

R. Das ()
Centre for Advanced Composite Materials, Department of Mechanical Engineering,
University of Auckland, Symonds Street, Auckland 1010, New Zealand
e-mail: r.das@auckland.ac.nz

R. Jones
Department of Mechanical and Aerospace Engineering, Monash University, Wellington Road,
Victoria 3168, Australia

Y.-M. Xie
School of Civil and Chemical Engineering, RMIT University, Victoria 3001, Australia
682 R. Das et al.

1 Introduction

The challenge of designing efficient and economic systems has resulted in an increas-
ing trend to integrate traditional numerical and heuristic optimisation techniques into
design processes. As a result, topology optimisation has become a rapidly growing
field of structural optimisation. It focuses on determining an optimum distribution of
material within a prescribed region satisfying specified design constraints. Topology
optimisation determines the general layout of a structure for a given design specifi-
cation. It can be a very useful tool in an initial design conception stage, where the
configuration of a structure is not available, or no previous layout/design exists. In
such cases, a designer can start with an anticipated arbitrary layout and can apply an
appropriate optimisation technique to determine the optimised topology based on the
load path through the structure and the boundary constraints. Topology optimisation
can also be used to alter and improve an existing design. This is usually accomplished
by iteratively adding or removing material, by redistributing material, or by altering
material microstructure or properties. The application of topology optimisation can
result in substantial material (hence cost) savings and enhanced strength. An opti-
mised topology can be taken as a starting geometry that can subsequently be (locally)
refined using shape optimisation.
During the last two decades, researchers have developed a range of novel ideas,
commonly known as the heuristic methods. These researchers were fascinated by
the principles of nature and believed that adopting the methods followed in natural
objects and events would produce an optimum (improved) design. It should be em-
phasised that the heuristic methods cannot ensure that the solutions are truly optimal.
However, if applied with proper judgement, these methods can produce solutions,
that are close to optimum, and generally rapidly converge. The main advantage of
these methods is that they do not require gradient information. Gradient evaluation is
computationally expensive and often quite difficult in cases where the design objec-
tive function has discontinuities, or an analytical expression of the gradient function
is not easily obtainable. Hence, heuristic methods are well suited to problems where
gradient evaluation is difficult, or the objective function is non-differentiable. Also,
the heuristic methods have the ability to locate the global optimum, without becom-
ing stuck at a local optimum point even if the initial solution is far away from the
global optimum.
The primary heuristic optimisation algorithms commonly used are Simulated An-
nealing (Kirkpatrick et al. 1983; Cerny 1985), Genetic Algorithm (Holland 1975;
Goldberg 1989), the Biological methods (Schnack 1979; Arutyunyan and Drozdov
1988; Das et al. 2006; Das and Jones 2009d), and the Hard/soft kill methods (Xie
and Steven 1993, 1997). Simulated Annealing was among the first of these meth-
ods. Annealing is an iterative process used to obtain a stable crystal structural con-
figuration with minimum energy level. Kirkpatrick et al. (1983) and Cerny (1985)
independently developed the concept of applying the procedure that is followed
in an annealing process to structural optimisation problems. The optimum solu-
tion is analogous to the lowest energy state in an annealed crystal structure. Sim-
ulated Annealing provides the benefits of global as well as local search methods.
The Genetic algorithm seeks to emulate the natural law survival of the fittest. In this
Optimal topology design of industrial structures 683

method a given population is allowed to evolve through the random steps of selection,
crossover, and mutation. The continued process of selection of the fittest members to
the next generation increases the proportion of fitter members (improved solutions)
in the later generations and drives the solution towards an optimum (Holland 1975;
Goldberg 1989). A class of novel heuristic methods, called the Biological method,
was developed based on the observations in nature (Schnack 1979). It has been ob-
served that biological species, for example, plants and animals, adapt themselves in
such a way that avoids local peak stress by generating a homogeneous surface stress
distribution. These methods thus attempt to optimise structures by emulating the
growth principles. Biological methods have been developed and extended for various
applications by many researchers, such as Arutyunyan and Drozdov (1988), Mattheck
et al. (1990), and Azegami (1990), and recently by Das and Jones for fracture and
fatigue life based optimum designs (Das and Jones 2009a, 2009b, 2009c). Lastly,
the hard/soft kill optimisation methodologies have become a popular approach be-
cause of their conceptual simplicity and easy implementation with the finite element
method. These optimisation techniques are based on addition and removal of materi-
als based on a certain set of criteria. Here we will focus on a variant of the hard kill
heuristic method called Evolutionary Structural Optimisation (ESO) (Xie and Steven
1993). A comprehensive description of various heuristic optimisation techniques can
be found in the book by Pardalos and Resende (2002).
A number of optimisation methods have been developed for topology design with
stress constraints (Cheng and Jiang 1992; Duysinx and Bendse 1998; Bendse and
Sigmund 2003; Navarrina et al. 2005). One notable optimisation technique is the
Homogenisation method, extensively investigated by Bendse (1989, 1995). This
method generates an optimal topology by distributing material using a density func-
tion, and had been applied to a wide variety of topology optimisation problems. For
example, the Homogenisation method was applied to the topology optimisation with
strength constraints (Duysinx and Bendse 1998), and to the design of material mi-
crostructure for obtaining optimum elastic properties (Neves et al. 2000). Buhl (2002)
implemented a formulation for minimising compliance, in which both the topology
and the supports were included. Haber et al. (1996) proposed a modified method
called the perimeter method. This helps generate a variable topology by creating new
boundaries and adding holes, thus controlling the geometric resolution. Svanberg and
Werme (2006) used the Sequential integer programming method for topology opti-
misation based on stress constraints. In this paper, we demonstrate the potential of
the ESO method for industrial topology design problems. The aim will be to produce
a light weight structure without greatly compromising the strength.
One major difficulty encountered in the various methods for topology optimisation
of large industrial structures is the computational overhead. Pardalos and Korotkikh
(2003) have highlighted a number of difficulties in implementing the traditional opti-
misation methods to practical engineering designs. The computational effort in finite
element analysis strongly depends on the size of the structures, particularly for three-
dimensional problems. The application of topology optimisation to structural designs
has been limited due to the computational inefficiency encountered for large scale
industrial structural components. In addition, some optimisation methods, such as
SIMP, penalises the density resulting in materials with intermediate densities. In such
684 R. Das et al.

cases, large structural displacements may lead to an indefinite or negative definite

stiffness matrix in the low density elements as the topology evolves (Buhl et al. 2000;
Bruns and Tortorelli 2003). To overcome this, a number of additional measures have
been proposed, such as modifying the convergence criterion (Buhl et al. 2000) or fil-
tering the low-density elements (Bruns and Tortorelli 2003). The discrete formulation
based ESO method (Yang et al. 1999a; Querin et al. 2000) can potentially circumvent
these issues.
We previously developed a fracture based ESO algorithm (Das et al. 2005) and
evaluated its effectiveness as a practical tool for optimum shape design problems.
This algorithm was applied to damage tolerance based shape optimisation of a
stringer cutout used in many fuselages of transport aircrafts (Das et al. 2007). These
studies established the ESO method as a powerful and robust design tool for shape
optimisation of real life structures with fracture strength as the design criteria. In this
paper, we modify the traditional ESO algorithm and apply it to topology optimisation
of realistic large scale structures.

2 Evolutionary Structural Optimisation

Evolutionary Structural Optimisation (ESO) is a gradient-less, heuristic optimisa-

tion method that mimics the Darwinian principles of evolution in naturally occurring
structures. This method was proposed by one of the authors (by Xie and Steven 1993,
1997). ESO works by imitating biological structures in nature. It has been observed
that naturally occurring species tend to achieve shapes that are close to fully stressed
configurations, as this leads to an optimal material utilisation.
The basic principle of ESO can be expressed as: if a portion of material in a struc-
ture does not contribute effectively to the functioning of the structure with respect to
its design objective(s), then it can be removed from that region. This gradual removal
process leads to a structure that meets the design objective(s) subject to constraints
(Xie and Steven 1997). Since ESO removes material completely without distributing
the material locally in a specified region or altering its density, it is also called a hard
kill or a (0, 1) method.
The ESO method removes inefficient material from a structure based on a set of
predefined criteria. Here the term inefficient means that the material is not taking
part, or contributing effectively, to the overall performance of the structure. Mater-
ial elimination is accomplished by progressively removing elements from the finite
element (FE) model. ESO can be used with various design objective functions. In
general, the objective function may be the weight of a structure, or a relevant per-
formance criterion. A factor or index (fitness number) is defined for each element,
which measures the contribution of that element towards the overall structural behav-
iour. The formulation of the fitness number or contribution factor for the elements is
an important step. Given an initial design domain, loads, and constraints, ESO will
remove those elements whose contribution factors are less than a reference value. The
updated structure is then re-analysed. Depending on the response of this new struc-
ture, the algorithm again identifies elements with a low fitness number and eliminates
Optimal topology design of industrial structures 685

them from the structure. This process is continued until the resulting structure satis-
fies some convergence criteria. The history of optimisation (i.e. intermediate struc-
tural topologies) is recorded so that a designer has the option to choose the most
suitable topology.
One advantage of ESO is that it is a relatively robust method. It follows the same
basic principle irrespective of the field of application in that it requires an FE model of
the structure, and a fitness or contribution factor (sensitivity index) for each element
to decide whether an element should be removed. Due to this universal formulation,
variants of the ESO method have been effectively applied to other fields, viz: dynam-
ics and vibration (Xie and Steven 1996; Zhao et al. 1996, 1997; Yang et al. 1999b),
buckling (Manickarajah et al. 1998, 2000), heat conduction and thermo-elasticity (Li
et al. 1999a, 1999b, 2000), and recently in damage tolerance based shape optimisa-
tion (Das and Jones 2004a; Das et al. 2007).
ESO has many other advantages. It uses the finite element method (FEM) to per-
form structural analysis. This enables the evaluation of the response of a complex
structural geometry with realistic loading and boundary conditions. Also in gen-
eral, as the topological configuration of a structure changes during optimisation, it
needs to be continuously remeshed. This may lead to FE mesh distortion and con-
sequent degradation in the solution accuracy. In this regard ESO uses a fixed mesh,
i.e. the initial FE model throughout the optimisation process, thus avoiding remesh-
ing and the associated mesh distortion problems. Moreover, the optimiser can be
externally interfaced with the FE software without requiring access to the source
code.

3 Topology optimisation using modified ESO

Topology design using ESO may be adopted from two different design perspectives.
The most common application is in the initial product or design conception stage. In
this case an arbitrary block of material is chosen and a set of given loads and supports
are specified. ESO takes away inefficient material from the structure and identifies
an optimum material distribution. This leads to an optimal initial design layout.
However, ESO performs equally well when modifying an existing design. In this case
the relatively inefficient portions of the structure are identified and removed. This
algorithm can highlight the parts that are not performing effectively. Accordingly,
one can either remove those portions, or redesign them to achieve a lighter design
and/or a more uniformly stressed structure.
The choice of the ESO parameter depends on the objective function. In stress
based optimisation the use of von Mises, or maximum principal stress, is common.
For the traditional ESO algorithm, the stress for each element is usually determined
by taking the stress value at the element centroid.
We have modified some aspects of the conventional ESO algorithm to reduce or
suppress some well-known demerits, such as the generation of checkerboard pat-
terns, mesh-dependence of topologies, and generation of unrealistic designs (Zhou
and Rozvany 2001). The ESO algorithm used here incorporates the following:
686 R. Das et al.

The stress for an element used in the material removal criterion is evaluated by
weighted averaging the stresses at the gauss points used in finite element integra-
tion for that corresponding element. This weighted average stress for an element
will be termed as the representative stress in this paper.
An adaptive variation approach is used for determining the elimination factor that
controls the rate of material removal.
A convergence function is introduced to monitor the uniformity in stress distrib-
ution as the topology of the structure evolves. This function measures the overall
material utilisation in the structure.

3.1 ESO criteria

The stress based ESO works by removing elements from lowly stressed regions. In
this study, a representative von Mises stress for each element is chosen as the ESO
criterion. It has been observed that some structural responses, when used as the ESO
design criteria, may lead to the generation of checkerboard patterns and impractical
topologies. Details regarding the formation of checkerboard patterns and procedures
to avoid them can be found in Li et al. (2001). It is found that if the stress at the
centroid of an element is used as the ESO criterion (as commonly used), it may lead
to the generation of checkerboard patterns. Here we adopted a weighted average
stress for each element derived from the corresponding gauss point stresses, which
were calculated using the displacement field. The values of the weights were depen-
dent on the locations of the gauss points with respect to the element centroid. This
approach uses a smoothed stress field when determining the ESO criterion, and this
was found to reduce the generation of checkerboard patterns. At every iteration, the
elements with a lower von Mises stress are removed. At a given stage (iteration), the
ith element is removed if:
v,i v,ref (1)
where v,i is the representative von Mises stress for the ith element calculated by
averaging the corresponding stresses at the gauss points of the ith element, v,ref is
the reference von Mises stress.
In the present case, the reference stress (v,ref ) is taken to be:

v,ref = EF v,max (2)

where v,max is the maximum von Mises stress in the structure, and EF is the elimi-
nation factor that determines the cut-off reference stress.
For stress optimisation, the weight of the structure itself can be considered as
the objective function, whereby the stresses are treated as constraints. The constraint
function is formed as a function of suitable stresses at various predefined locations of
the structure, and has a general form:

F = f (1 , 2 , . . . , n ) (3)

where 1 , 2 , . . . , n are some suitable design stresses (e.g. von Mises or maximum
principal stress) at n specified locations of the structure. The functional form of F ,
Optimal topology design of industrial structures 687

the nature of the stress, and the locations depend on the optimisation class (sizing,
topology, shape, etc.) and the specific problem category. For example, general topol-
ogy optimisation using ESO will use a representative elemental stress (as described
before) as the design stress.
For the topology optimisation problems presented, the weight of the structure will
be considered as the objective function, and the maximum von Mises stress will be
used as the constraint function. The convergence of the optimum solution will be
assessed by monitoring the variability of stress distribution in the structure through
a convergence/variance function. Here we propose the following convergence func-
tion F :
   
ne
ne 2 
i=1 (i max )
1 ( )2 1
i av
F = C1 i=1
+ C2 (C1 + C2 )
av ne 1 max ne 1
(4)
where i is the representative stress for the ith element, av and max are the aver-
age and the maximum stress for all of the elements, and ne is the total number of
elements currently existing in the structure at a given iteration. For topology opti-
misation, i = v,i , C1 and C2 are user specified constants (weighting factors). In
the present examples, we took C1 = C2 = 1. The function F is a linear combination
of the standard deviations with respect to both the average stress and the maximum
stress, and it thus provides an overall variability index of the stress in the structure.

3.2 Elimination factor

The elimination factor plays an important role in controlling the optimisation process.
A high value will lead to a rapid convergence, but may cause instability. The instabil-
ity may not allow capturing some of the topologies near the optimal region and may
also drive the solution away from an optimum. In contrast, a very low value will re-
quire a large number of iterations to reach an optimum, and can dramatically increase
the solution time. Therefore, it is instructive to establish an optimum value by initial
numerical tests.
In the present work, a new approach for calculating the elimination factor is con-
sidered. The approach commonly used in literature is a fixed increment approach (Xie
and Steven 1997), in which an initial value of the elimination factor is chosen and it
is increased by a constant factor (IF) whenever the volume change, or the number
of elements removed at a given iteration, is less than a user specified value. In this
approach the elimination factor (EF) is given by:

EF = EF 0 + IF (5)

where EF 0 is the initial value of EF and IF is the increment factor. This method does
not take into account the current state of the structure. As a result, it may remove more
elements from an optimal structure leading to instability, or remove less elements in
the initial or intermediate stages resulting in slow convergence. It may also lead to an
increase in the checkerboard pattern of the topology.
688 R. Das et al.

To overcome this, a new adaptive variation approach is proposed here. In this

approach, the elimination factor is calculated depending on the current maximum
and minimum value of the ESO criteria. It is given by:

EF = SR + FR (1 SR) (6)

where SR is the ratio of the minimum to maximum value of the ESO criterion. In
the present stress based case, SR = min /max , and FR is a control factor. With this
formula the reference von Mises stress becomes (using (2)):

ref = min + FR (max min ) (7)

Equations (6) and (7) determine the level of material elimination depending on the
current state of the structure, e.g. relative variation in the stress levels. In the initial
stages there is usually a wide variation in stress levels throughout the structure. The
stress range (difference between the maximum and minimum stress) remains large
that results in a higher volume removal per iteration. As the structure approaches
the optimum point, the ESO criteria (in the present case von Mises stress) tends to
be more uniform throughout the structure. In this situation the difference between
the reference stress and the elemental stress gradually reduces and the number of
elements removed per iteration slowly decreases. This gradual and stable change in
the topology provides a better scope to record a number of closely spaced topologies
near the optimum region. As such, it furnishes the analyst with several design options
to choose from. Therefore, the present work will use the adaptive variation approach
for the selection of the elimination factor.

3.3 Initial design domain

The specification of the initial design domain depends on the loading and constraints.
Sufficient material must be present between load points and constraints so that the
algorithm can capture the near optimal topology. On the other hand, if the design
domain is too large, the computational effort will increase accordingly. Hence, a suit-
able balance based on experience and intuition is often required in choosing a guessed
starting topology. A good guess of the initial design domain can significantly reduce
the computer time, yet can produce an improved design. However, if the design do-
main is too restricted, some improved designs may not be realised.

3.4 Analysis methodology

We summarise below the steps required for stress based topology optimisation using
the ESO method.
(i) A geometric model of the structure was constructed using a pre-processor.
FEMAP was used for developing the initial solid models in the present study.
(ii) The optimisation category, design constraints, design criteria, and ESO para-
meters were specified as the inputs to the optimiser.
(iii) Finite element analysis was performed to evaluate the stress distribution for
this geometry.
Optimal topology design of industrial structures 689

(iv) The elemental representative stresses were then computed by averaging the
stresses at the gauss points of the corresponding elements.
(v) For each element, the elemental von Mises stress was compared with the refer-
ence von Mises stress calculated using (7).
(vi) An element was removed from the structure if the elemental von Mises stress
was below the reference value. In the present work this was achieved by re-
ducing the stiffness of the element to a very low value so that it (effectively)
took no load. The values of the Youngs modulus and Poissons ratio for the
removed material were taken as 0.1 GPa and 0.3, respectively.
(vii) The resultant new structure was converted to an FE model for re-analysis. This
completed the first iteration.
(viii) Steps iivii were repeated until convergence was achieved, i.e. the objective
function attained a minimum value, or did not appreciably change between
successive iterations, or satisfied an appropriate convergence criteria.

3.5 Software development

An in-house optimisation software NASESO has been developed that implements the
ESO algorithm for both topology and shape optimisation. The software is interfaced
with the FEMAP pre/post-processor and the finite element code NE-NASTRAN to
form an Integrated Design Optimisation System. Modelling of the initial geometry or
design domain can be performed in FEMAP. The appropriate geometric constraints
are also specified interactively within FEMAP. The design objective, response con-
straints, optimisation parameters, ESO criterion, and optimisation category (shape or
topology) are specified as input. The initial FE analysis model exported from FEMAP
acts as the input to this software system. The optimisation using the ESO method is
then performed iteratively until convergence. A detailed description of the software
can be found in NASESO User Manual (Das and Jones 2004b).
We will first demonstrate the ESO algorithm for simple examples. It will then be
applied to topology optimisation of complex real life structures. Two case studies
will be considered, viz: a bulkhead used in an aircraft structure, and a sideframe
used in a railway freight wagon. In each case, the ESO algorithm will act upon an
existing design to generate an improved (topological) layout with appropriate limits
on the maximum tolerable stress. The effectiveness of the ESO algorithm for the
shape optimisation of a local geometry of the sideframe to reduce stress level will
also be illustrated.

4 Illustrative examples

The basic purpose here is to perform a preliminary study and the validation of the
ESO method used by applying it to simple problems. The problems of the design of
two cantilever beams were considered.

4.1 Validation example: Optimisation of a 2D short cantilever beam

This example serves as a validation problem for the present study. This was analysed
in the literature using a different optimisation method (SIMP) (Sigmund 2001). The
690 R. Das et al.

Fig. 1 Initial design domain of

the short cantilever beam

Fig. 2 Initial von Mises stress distribution of the short-cantilever beam

problem consists of a short cantilever beam loaded by a vertical force of 100 N acting
at the mid point of the right end of the structure, and the opposite (left) end was rigidly
fixed, see Fig. 1. The width and height of the beam were 160 and 100 mm respec-
tively, and the thickness was 1 mm. The structure was meshed regularly with 1440
quadrilateral plate elements. The material was assumed to be steel with a Youngs
modulus of 207 GPa and a Poissons ratio of 0.3. The initial von Mises stress distrib-
ution is shown in Fig. 2. The initial control factor for the adaptive variation approach
was 0.01. In this example, material was allowed to be removed from everywhere of
the structure, i.e. no constrained (fixed) region was specified.
The ESO algorithm started removing material from the low stress regions. Some of
the intermediate topologies obtained in the evolutionary process are shown in Fig. 3.
The optimised topology is presented in Fig. 4 along with the associated von Mises
stress distribution. The final structure was nearly uniformly stressed within the limi-
tation of elemental resolution. By this we mean that protruding element corners had
very low stresses and high stress concentrations occurred at sharp internal element
corners, even though the major portions of the structure were equally stressed.
Optimal topology design of industrial structures 691

Fig. 3 Topology evolution

history of the short-cantilever
beam

Fig. 4 Optimised topology of the short-cantilever beam with von Mises stress distribution

The convergence function (given by (4)) was monitored throughout the optimi-
sation process. The convergence function decreased up to iteration 147, at which
it attained a minimum value of 0.412. It then increased slightly until the structure
completely collapsed at iteration 155. Therefore, the structural configuration corre-
sponding to the 147th iteration represented the most uniformly stressed structure that
692 R. Das et al.

Fig. 5 Change in weight of the

short-cantilever beam

Fig. 6 Variation in the

maximum stress to weight ratio
of the short-cantilever beam

can be taken as the optimal topology. However, the topologies in the iteration range
145155 were very close. Thus, selecting any of them could have made practically
no difference. The topologies obtained here were similar to those obtained via a pe-
nalization method (SIMP) presented in Sigmund (2001). This established that the
optimal solution predicted by the ESO algorithm agreed with that obtained using a
different topology optimisation method.
As the optimisation progressed, the change in weight of the structure is shown in
Fig. 5. Figure 6 shows the corresponding variation of the maximum stress to weight
ratio. The weight of the structure reduced by 73%, whereas the maximum von Mises
stress increased by 15%, from an initial value of 19.7 to 23.3 MPa. As a result, the
maximum stress to weight ratio increased six times that of the original structure, i.e.
from 140 to 845 MPa/Kg. Hence, the reduction in weight clearly outweighs the in-
crease in the maximum stress. It can be inferred that the optimum design provides a
good compromise between the strength and the weight of the structure. The strength
to weight ratio of a structure can provide useful information at the design conception
stage. One can select a minimum weight design for a given constraint on the maxi-
mum acceptable stress limit. The corresponding structural configuration can then be
chosen from the evolution history for the design feasibility study.

4.2 Optimisation of a 3D long cantilever beam

Next we considered a 3D cantilever beam with a square cross-section and an as-

pect ratio of 8:1. The length of the beam was 160 mm and its cross-section was
20 20 mm2 . The beam was made of steel with a Youngs modulus of 200 GPa, a
Optimal topology design of industrial structures 693

Fig. 7 Geometry and initial

design domain of the 3D
cantilever beam

Fig. 8 Initial (top) and

optimised (bottom) topologies
with associated von Mises stress
distribution

Poissons ratio of 0.33, and a density of 8000 Kg/m3 . A uniformly distributed load
of 25 N/mm, acting vertically downwards, was applied along the upper edge of one
of the ends, whilst the other end was rigidly fixed, see Fig. 7. The beam was meshed
with 8000 21-noded hexahedral elements and the model had 37,301 nodes. The von
Mises stress distributions for the initial FE model and the optimal topology are shown
in Fig. 8.
The optimal topology was similar to an I-beam with internal holes. We observe
localised high stresses due to the presence of sharp edges and corners in the vicin-
694 R. Das et al.

ity of the regions from where elements were removed, which is physically intuitive.
This requires the final optimised design be smoothed to eliminate the local stress
concentrations.
The optimised structure had a locally uniform stress distribution. Although the
maximum von Mises stress increased due to the generation of stress raisers, the
weight savings resulted in an improved maximum stress to weight ratio, leading to a
nearly fully stressed structure.

4.3 Checkerboard pattern and mesh dependency of optimal topologies

One major concern with topology optimisation is the generation of checkerboard type
patterns in the optimised structures. The main reason for this is that the discretised
structure develops localised artificial stiffness in the regions of element removal,
which contributes to the generation of unrealistic topologies (Bendse and Sigmund
2003). The pattern and location of iterative element removal appear to affect the dis-
tribution of this artificial stiffness. Another important aspect is to examine the sensi-
tivity of optimum topologies with mesh attributes. Differences can often arise due to
the change in the locations of the element centroids with mesh configuration, which
causes a shift in the stress evaluation points. For example, a finer mesh increases the
number of locations/points considered that may affect the material elimination pat-
tern somewhat. Thus, the optimal topologies are bound to vary slightly with mesh at-
tributes. The same argument is true for different element shapes, as these also change
the element centroid positions.
In this study, we used a weighted average stress based on the stresses at the gauss
points to calculate the representative stress for a given element. This smoothens the
stress field over an element. The modified ESO criterion (determined based on the
smoothed stresses) and the adaptive variation of the elimination factor (controlling
the pattern of material removal) were found to be adequate to suppress the genera-
tion of checkerboard patterns and reduce the mesh dependency of topologies in the
present study. The optimisation results were reasonably stable with variations in the
mesh density and element shape. The optimised topologies for different mesh pat-
terns were also found to be similar. It is noteworthy that the structural layout obtained
with ESO is an initial estimate only, which should be subjected to (local) redesigns
and refinements. The optimal topology obtained will be manufacturable provided the
jagged contours are smoothed using a CAD program. Hence, a minor variation in the
ESO produced optimal topology may not cast a significant influence on the final
smoothed design.

5 Topology optimisation of a bulkhead

In the last example, ESO was used for the initial design stage. A primary strength of
ESO is that it is a versatile tool that can be used for redesigning an existing structure.
The need to lighten structural components without compromising their functionalities
was first recognised in the aircraft industry. ESO is particularly suitable for accom-
Optimal topology design of industrial structures 695

Fig. 9 (a) A set of full F/A-18 bulkheads (courtesy: DSTO, Melbourne), (b) Half solid model (3D) of the
initial bulkhead

plishing weight reduction under a given set of loads and constraints. Here we will
demonstrate this by applying ESO to obtain an improved topology for an F/A-18 air-
craft bulkhead. The aim here is to produce a lighter bulkhead structure satisfying the
prescribed geometric, functional, and strength requirements.
The bulkhead under consideration forms a key component of the F/A-18 aircraft.
This aircraft, inducted into Royal Australian Air Force (RAAF) in 1983, is well-
known for its manoeuvrability and high performance. Weight reduction and life ex-
tension programs are two important areas of active research. In the context of topol-
ogy optimisation we will focus on the former aspect. The aircraft as a whole is an
extremely complex structure. There are many components that can be potentially op-
timised to reduce their weights. One specific region of the F/A-18 aircraft that has
potential for further weight reduction is the structural components associated with
the centre barrel. The F/A-18 centre barrel section and its associated bulkheads can
be seen in Fig. 9a.
A half-symmetric 3D model of the bulkhead, shown in Fig. 9b, was analysed here.
It may be noted that in the 3D solid model some of the complicated features of the
actual structure were simplified provided their effects on the main analysis results
of interest were negligible. The major portion of the bulkhead was made of an alu-
minium alloy (Al7050-T7451). The Youngs modulus (E), Shear modulus (G), and
Poissons ratio () were taken as 71,015.7 MPa, 26,889.4 MPa, and 0.33 respectively.
Beryllium-copper alloy was used in a small portion of the wing attachment lug with
material properties: E = 127,552 MPa, G = 50,193.6 MPa, and = 0.27. A linear-
elastic finite element analysis was performed throughout.
696 R. Das et al.

Fig. 10 (a) Schematic of the bulkhead at a test rig with loads applied through the attachment holes (Dixon
and Molent 2003), (b) Loads and constraints used in the analysis
Optimal topology design of industrial structures 697

5.1 Finite element modelling

Depending on the state of the aircraft such as take-off, landing, and various flight
conditions, the bulkheads are subjected to a range of load cases. Figure 10a shows a
schematic of the bulkhead in a test rig where the loads were applied through attach-
ment forks using an actuator system. In the analysis, the loads and constraints were as
indicated in the xz view of the model shown in Fig. 10b. The model was symmetric
about the yz plane and accordingly appropriate symmetry constraints were imposed
on the relevant planes. In this study we considered the most critical load case, techni-
cally known as the Combined Set 55 (Dixon and Molent 2003), shown in Fig. 10b.
In this load case, the loads at the wing attachment holes act in the horizontal direc-
tion under the in-service loading condition of the aircraft. The loads were modelled
as bearing loads and applied on the semi-cylindrical surfaces of the holes. The resul-
tant load in the upper hole was 100.869 KN acting horizontally towards left, and the
same in the lower hole was 101.708 KN acting horizontally towards right.
The 3D solid model was meshed with 34,306 (21 noded) brick elements and it had
169,333 nodes. The finite element mesh of the initial bulkhead is shown in Fig. 11a.
In this study the von Mises stress was taken as the ESO criterion. The von Mises
stress distribution is shown in Figs. 11 and 12. The maximum von Mises stress for
the initial structure was 86.4 MPa. Highly stressed regions can be observed around

Fig. 11 (a) Finite element mesh of the 3D bulkhead, (b) von Mises stress distribution throughout the
structure
698 R. Das et al.

Fig. 12 (a) Locations of the

holes in the bulkhead along with
the von Mises stress distribution
around them, (b) Geometric
constraints imposed on the hole
boundaries showing the regions
of the bulkhead kept unaltered
during the topology optimisation

hole D, and some portions around the flange hole. The loads from holes A and B
were transmitted to the region around hole D and the left end of the structure along
the upper and lower sections of the central flange hole. This resulted in some regions,
such as the lower protruded portion below hole C, a part of the upper left region, and
some portions around the flange hole boundary, being lowly stressed. These areas
of the structure represented potential inefficient utilisation of material in sharing the
loads.
A number of geometric constraints were imposed to meet the functionality re-
quirements. The holes in the bulkhead serve different functions. For example, holes
A and B (Fig. 12a) are used to transmit loads. The large central flange hole holds
the bulkhead flange. Other holes are also used as various connection points. Hence,
it was a design requirement that the hole boundaries be retained in the final topol-
Optimal topology design of industrial structures 699

ogy. This was accomplished by dividing the structure into a restricted domain and
a design domain using the FEMAP pre-processor. The restricted domain consisted
of the regions around the hole boundaries (Fig. 12b), whereas the rest of the structure
constituted the design domain. Alteration of the topology was permitted only in the
design domain, and the restricted domain was kept intact throughout the optimisation
process.

5.2 Optimisation using ESO

The ESO algorithm was used to reduce the weight of the structure by removing inef-
ficient material. At each optimisation cycle a finite element analysis was performed
using NE-NASTRAN. The optimisation algorithm was then employed to modify the
topology based on the von Mises stress field. At each iteration, a reference von Mises
stress was calculated using (7). An element was then removed from the structure if
its von Mises stress was lower than the reference stress.
The gradual evolution of the structural topology and the associated variation in
the stress pattern, as the optimisation progressed, is illustrated by a few intermediate
topologies in Fig. 13. The extent of material removal at various stages of optimisa-
tion was primarily governed by the relative stress levels among various portions of
the structure. Topology evolution history is useful in exploring alternative designs.
Sometimes a specific topology may be better in terms of satisfying other design cri-
teria, although it may not necessarily represent an optimum design. This history also
helps in identifying innovative designs, which may not be easily realised using con-
ventional design guidelines or common experience.
The change in weight of the bulkhead as material being removed is shown in
Fig. 14, and the corresponding maximum von Mises stress generated in each topology
is shown in Fig. 15. The optimum point can be located on the basis of the weight and
maximum stress histories. The aim of the present study was to reduce the weight pro-
vided the maximum stress did not exceed an allowable value. This maximum allow-
able design stress is usually determined based on failure criteria, material properties,
operating conditions, and an adequate factor of safety. Here the maximum accept-
able design stress (a ) was set to be 95 MPa, which is 10% higher than the initial
maximum stress. It was required to find a minimum weight structure for which the
maximum stress was just below a .
To this end, a constant stress line representing a = 95 MPa was drawn in Fig. 15.
The design topology just preceding the point of intersection of this horizontal line
with the stress history curve corresponded to a structure that had a maximum von
Mises stress just lower than the permissible value. This point was taken as the final
design. In this problem the optimised structure was obtained at the 86th iteration.
The maximum von Mises stress, under the given loads and constraints, associated
with this (optimal) topology was 94.9 MPa (at the next iteration the maximum stress
was 96.1 MPa and exceeded the maximum limit). For this problem the loss of mater-
ial and the resulting change in the load path led to a gradual increase in the maximum
von Mises stress. After the optimal point, a sharp rise in the maximum stress for sub-
sequent iterations is noticed. From the weight history, the weight of this optimal
structure is found to be 75% of that of the original bulkhead. Therefore, approx-
imately 25% weight reduction has been achieved with an increase in the maximum
700 R. Das et al.

Fig. 13 Evolution history of the bulkhead topology and the associated von Mises stress distribution
Optimal topology design of industrial structures 701

Fig. 14 Change in weight of

the bulkhead with change in
topology

Fig. 15 Variation of the

maximum von Mises stress of
the bulkhead with change in
topology

Fig. 16 Variation of the

convergence function of the
bulkhead with change in
topology

von Mises stress by 8.5 MPa only. This weight saving will have a major impact on the
design of the entire aircraft structure as it usually incorporates a number of bulkheads.
This shows that redesigning or optimising the bulkheads can significantly contribute
to the overall weight reduction program of the aircraft.
The convergence function, given by (4), was used to examine the convergence of
the adopted optimal solution. The variation of the convergence function is shown in
Fig. 16. The optimum point was found to lie in a uniform region indicating the con-
702 R. Das et al.

Fig. 17 Variation of the ratio of

the maximum stress to weight of
the bulkhead with change in
topology

vergence of the solution. The convergence function decreased from an initial value
of 0.69 to 0.62 (in the optimal structure). This shows that the optimal structure has
a more uniform stress distribution than the original design. This is a typical feature
of the ESO method, which tends to reduce the variability in stress levels within a
structure. Another important parameter of interest is the ratio of the maximum stress
to weight shown in Fig. 17. This is sometimes used as a design performance char-
acteristic index, particularly for layout optimisation, to assess the efficiency of ma-
terial utilisation. In the present case the stress to weight ratio increased from 4.00 to
5.85 MPa/Kg. This confirms that the optimal structure has an improved utilisation of
material.
The resultant optimal topology is shown in Fig. 18. It can be observed that the
material was removed from the lowly stressed regions. The hole boundaries were
kept as before by enforcing the geometric constraints. The optimal topology has the
following features:
(i) The region between holes A and B was lowly stressed in the original structure. In
the improved (optimal) topology, a small amount of material was removed from
this region, and the stress field became more uniform, see Figs. 12a and 18.
(ii) A major part of the load from holes A and B is now transmitted along load path
1 to the upper left end support. This is manifested by a slightly higher and a
more uniform stress distribution in the top section (above the flange hole) of the
bulkhead. Load is also transferred through load path 2, which passes through the
section near point X in zone 2 (Fig. 18), to eventually reach the region surround-
ing hole D. However, this section does not have sharp features that could have
resulted in significant stress concentrations. Although in the original structure
the left portion of the flange hole (zone 1 in Fig. 18) did carry load, we notice
that in the optimal layout very little load is transmitted through this section, i.e.
along load path 3.
(iii) From the optimal structure it can be noticed that most of the material was re-
moved from the regions marked as zones 1, 2 and 3 in Fig. 18. In the original
structure zone 3 had very low stresses as it was not in the primary load paths,
see Fig. 12a. Hence, considerable material was removed from this zone. Al-
though zone 1 carried some load in the original structure, it was still inefficiently
Optimal topology design of industrial structures 703

Fig. 18 Optimal topology of the bulkhead

utilised, and this led to the topology being altered in this zone. The lower left
portion of the flange hole (in zone 2) did not effectively take part in sharing
loads, and as a result material was taken away from this zone as well.
The topology obtained using ESO is a conceptual design and needs post-processing,
which may include incorporation of additional design features or constraints, local
modification of the structural geometry, shape optimisation of local features, etc. For
example, we notice some partially floating hole boundaries in the final topology.
This is because of the imposed geometric constraints that the boundaries of all the
holes in the original structure must remain unaltered to enable them to serve as at-
tachment points. It may be noted that all the three zones indicated in Fig. 18, including
the regions around the holes, need redesigning to generate an acceptable final design.

6 Topology optimisation of a sideframe

Similar to the requirements of aircraft designs, many other capital intensive indus-
tries also emphasise weight reduction as part of the design criteria to reduce cost
and sustain commercial competitiveness by enhancing performance. For example,
the importance of weight reduction without compromising performance cannot be
overstated for components used in rail industries. Reducing weight of carriage parts
can lead to an increased tonnage capacity. The AusLink-White-Paper (2004) high-
lighted the need to double Australias freight capacity in a decade. To realise this,
one measure being considered is to make the carriages and associated components
lighter and more durable.
704 R. Das et al.

Fig. 19 A typical sideframe in

a test rig at Monash University

Next we consider an optimisation example of a rail component. Many components

used in the rail industry had been (historically) designed based on a combination of
traditional design techniques together with trial and error processes. This emphasises
the need for optimisation of these components to improve performance. In this study
we will focus on the topology optimisation of a sideframe. The sideframe is one of the
crucial load carrying members that transfers the entire carriage load from the bogie
to the wheels, see Fig. 19. The terminologies used for various parts of the sideframe
are also shown (Harder 2000).
The sideframe is symmetric about its middle vertical plane. This enabled a half
model of the sideframe to be analysed. Figure 20 shows the half model of the side-
frame along with loads and constraints. The carriage load is transmitted through the
bolster to the sideframe. The load exerted on the right portion of the half model is
schematically shown in Fig. 20b. Here we considered the standard design load case
with the applied load being 193.65 KN. The reaction from the wheels (through bear-
ings) on one half of the sideframe was also applied. There were two constraints; the
symmetry about the mid-plane (yz) was imposed on the right surfaces (symmetry
planes), and the two points on the left (A and B in Fig. 20a), where the sideframe is
attached to the wheel assembly, were kept fixed.
Figure 21 shows the finite element model of the side frame. Topology optimisa-
tion needs a fine mesh to provide enough resolution so as to iteratively allow gradual
evolution of the topology. A fine mesh also allows capturing of relatively closely
spaced topologies, which may lead to an improved optimal solution. In practice, a
trade off between the mesh size and the computational time needs to be considered.
The current finite element model had 310,647 (both 21 noded hexahedral and 8 noded
tetrahedral) elements and 488,795 nodes. The loads were applied as distributed pres-
sures over the respective surfaces. The sideframe was assumed to be a steel casting
with E = 200,000 MPa, G = 76,923.1 MPa, and = 0.3. The problem was analysed
on the VPAC Linux cluster with 3 GHz P-IV processors.
The aim of the work was to produce a reduced weight structure without exceeding
a prescribed maximum (design) stress limit and meeting all the design and perfor-
mance criteria. Here the maximum design stress was set to be 388 MPa. The first
step was to identify the potential locations where material could be removed in order
to lighten the structure. The von Mises stress of each element was taken to be the
Optimal topology design of industrial structures 705

Fig. 20 (a) Initial geometric

model of the sideframe,
(b) Loads and constraints on the
sideframe

ESO design criterion. Figure 22 shows the von Mises stress distribution for the initial
sideframe. The maximum von Mises stress in the initial structure was 356.1 MPa.
The primary load flows were along the inclined middle member (relative to the half
model) and the straight upper part, see Fig. 22. It is of interest to note that several
parts of the structure were ineffective in terms of sharing the load, as evidenced by the
large lowly stressed regions especially outside the primary load paths. However, all
these inefficient portions could not be removed from the structure as many of the fea-
tures were needed to satisfy the desired performance or functionality requirements.
The geometric constraints were imposed in those portions that were required to
be present in the final design. For this purpose, the entire structure was divided into
design and non-design (restricted) domains. The portions of the sideframe that were
kept unaltered throughout the optimisation are shown in Fig. 23. A portion of the left
part, i.e. the pedestal, was kept fixed irrespective of its stress level. Other retained
areas consisted of the regions near the connection zones between the wheel bearing
assembly and the sideframe, including the pedestal roof that connects the sideframe
706 R. Das et al.

Fig. 21 Finite element mesh and boundary conditions used for optimisation

with the wheel bearing assembly through an adapter. The material removal was also
restricted for the symmetry planes and the surfaces that carry distributed loads, e.g.
the spring seat (shown in Fig. 19) on the right of the half model. These geometric
constraints influence the intermediate topologies and load paths, and therefore affect
the final optimised topology.
The ESO algorithm was then applied to generate a more efficient topology. The
initial finite element model of the sideframe was input to the in-house optimisation
system NASESO. The change in weight of the structure as the optimisation pro-
gressed is shown in Fig. 24. In the initial stages, the rate of material removal was
high due to a large variability in stress distribution and the existence of large portions
of lowly stressed material. As a result, the weight of the sideframe showed a steep
decrease reducing to 72.5% of the original weight by 60 iterations, after which the
structural configuration did not appreciably change as the weight reduced only by ad-
ditional 6% up to iteration 201, at which the optimisation process was terminated.
The variation of the maximum von Mises stress is shown in Fig. 25. It can be ob-
served that although a significant weight reduction (33.6%) has been achieved, the
maximum stress remained essentially constant throughout the optimisation process.
This is useful in that a considerable weight reduction can be achieved without signif-
icantly compromising the strength.
One advantage of ESO is that the evolution history of the structure can be
recorded. This furnishes useful information about structural behaviour with change
in load paths and configuration. The topologies of the structure obtained at 25th and
49th iterations are shown in Fig. 26. In the left portion (pedestal) of the sideframe and
some other zones such as the column and spring seat (shown in Fig. 26a), the mater-
ial was relatively lightly loaded. Hence, at the beginning material was removed from
Optimal topology design of industrial structures 707

Fig. 22 von Mises stress

distribution for the initial
structure of the sideframe

these regions, see Fig. 26b. The load transfer through the structure, shown schemati-
cally in Fig. 26b, also substantiates such a pattern of material removal.
A common practice to terminate a topology optimisation process using ESO is to
set a permissible level of the design stress. The structural configuration at which the
maximum stress is just below the allowable value is taken as the optimum design
topology, as followed previously in the case of bulkhead optimisation. In the present
case, the maximum stress for the initial structure occurred at the junction of the side-
frame pedestal roof and the wheel bearing adapter attachment (Fig. 22a). We have
noticed that this maximum stress was not greatly affected by the alteration of the
load paths due to change in the structural layout. Hence, we could have continued to
iteratively remove material until the maximum stress exceeded the permissible value
of 388 MPa, which is 9% more than the initial maximum stress.
However, one particular region of the sideframe that was of specific concern was
the junction of the pedestal roof and the inclined middle member, see the encircled
region in Fig. 22a. This area is known to be prone to failure by fracture, and is termed
as the critical zone in this study. This is also supported by localised high stresses
708 R. Das et al.

Fig. 23 The portions of the sideframe FE model that were geometrically constrained

Fig. 24 Weight reduction

history of the sideframe: the
weight of the sideframe is
shown corresponding to 9% (")
and 5% (Q) stress limits over the
initial peak maximum principal
stress in the critical zone

Fig. 25 Variation of the

maximum von Mises stress of
the sideframe
Optimal topology design of industrial structures 709

Fig. 26 Intermediate topologies: (a) Iteration No. 25 (W/W0 = 85.6%), (b) Iteration No. 49
(W/W0 = 75.9%)

on the surface in this zone. Taking this constraint into account for the topology opti-
misation, we set a limit on the maximum principal stress in this critical region. For
this local shape, the maximum principal stress governs the criticality of flaws on the
710 R. Das et al.

Fig. 27 Variation of the peak

maximum principal stress (p )
and maximum von Mises stress
(v ) in the critical zone showing
the design points corresponding
to 9% (") and 5% (Q) stress
limits over the initial peak
maximum principal stress
((p )initial )

Fig. 28 Variation of the

convergence function for the
sideframe

boundary surface. For this reason, the peak maximum principal stress in the critical
zone was chosen as the ESO design criterion for this problem.
The allowable peak maximum principal stress was set to be 175 MPa, i.e. 9%
higher than that of the original design (160.1 MPa). The variations in the peak max-
imum principal stress and the maximum von Mises stress in the critical zone are
shown in Fig. 27. It was found that the maximum principal stress in this region of the
structure obtained at iteration 201 was 174.6 MPa, which was just below the allow-
able value. This structure thus corresponded to the terminal point of the optimisation
process and was taken as the final design. It is to be noted that this stress at the criti-
cal zone can be reduced further by a localised shape optimisation, if desired. This
will be studied in the next section.
It is noteworthy that if a lower design stress limit is imposed, the resultant weight
reduction for the optimal structure would be less, see Figs. 24 and 25. For example,
a design stress limit of 5% higher than the initial peak maximum principal stress (i.e.
168.9 MPa) would result in a topology (corresponding to iteration 117) with a 28.8%
weight reduction, whereas the current limit resulted in a weight reduction of 33.6%.
Figure 28 shows that the convergence function (4) steadily decreased from a value
of 0.798 to 0.684 indicating a more local uniformity in the stress distribution. This
can be observed by comparing the initial, intermediate, and final topologies of the
sideframe. The change in the stress pattern in the encircled areas of the topology
Optimal topology design of industrial structures 711

Fig. 29 Variation of the ratio of

the maximum stress to weight of
the sideframe

obtained at iteration 49 is an example of the relative uniformity in the stress field

at different regions of the sideframe, see Fig. 26b. The regions around the column
and the spring seat were lowly stressed in the initial structure (Fig. 22b). The stress
level in these regions became more uniform as the optimisation progressed. This can
be observed in the topologies corresponding to the 25th and 49th iterations, i.e. by
comparing the encircled zones in Figs. 26a and 26b.
One important parameter that measures the load carrying efficiency of a struc-
tural member is the relative change in the maximum stress with weight, as shown
in Fig. 29. The stress to weight ratio increased from an initial value of 2.14 to
3.21 MPa/Kg at the design point. The increase in the stress to weight ratio highlights
the enhanced structural efficiency of the final topology.
The final optimised topology is shown in Fig. 30. As expected, the ESO algorithm
identified the regions of the sideframe that did not effectively contribute to the overall
structural performance, and consequently removed inefficient material from those
regions. The various areas of the sideframe where material was mainly removed are
marked in Fig. 30. The primary locations were the pedestal, the column, and some
regions in the upper straight member and around the spring seat. It is envisaged that
redesigning these portions could result in a considerably lighter structure with an
improved material utilisation.

6.1 Shape optimisation of the potential critical region of the sideframe

Having determined the improved (optimal) design layout (topology), it is instructive

to perform design modifications locally to reduce stress levels in specific regions of
interest. To illustrate this, we will use the ESO based shape optimisation technique to
reduce the maximum principal stress in the potential critical zone shown in Fig. 22a
(the encircled area). Figure 31a shows a close-up view of the critical zone. In this
case the aim was to reduce the peak stress in this region by modifying the shape of
the surface locally. The alteration of the region was accomplished by taking away
material from the curved surface only, see Fig. 31. The rest of the structure was
kept unaffected throughout the optimisation.
The initial maximum principal stress distribution in the critical zone is shown in
Fig. 31a. The peak maximum principal stress in the original shape was 160.1 MPa.
712 R. Das et al.

Fig. 30 Final design topology

of the sideframe with 9% peak
maximum principal stress
constraint: the encircled areas
represent the regions where
material was primarily removed

The ESO algorithm removed some portions of the relatively lowly stressed material
from the middle of the surface. As a result, the surface stress field became more uni-
form with a peak value of 138.4 MPa, which represents a 13.6% reduction in the peak
stress. The modified surface profile is shown in Fig. 31b. This demonstrates the abil-
ity of this ESO based approach to locally redesign various portions of a (previously)
topology optimised structure to reduce local peak stresses.
Optimal topology design of industrial structures 713

Fig. 31 Maximum principal

stress distribution in the
(a) initial shape, and in the
(b) improved shape of the
critical zone

7 On constraint implementation in ESO

Geometric constraints are often necessary to meet desired performance requirements

and to ensure manufacturability within available resources. To implement this with
ESO, a structure is divided into different domains with prescribed attributes associ-
ated with them. One common constraint is to retain certain original design features
or regions in the optimal design. This can be accomplished by constraining speci-
fied regions of the structure from which material should not be removed irrespective
of the stress state. However, it is desirable to allocate as much of the total structure
as possible for material removal. This provides the optimiser a greater flexibility to
drive the solution towards an improved (optimum) design. The best possible topology
can evidently be achieved with no geometric constraints at all, although in practice a
variety of functional and/or manufacturing constraints usually need to be imposed.
The most important thing to note is that ESO generates a conceptual design and
identifies potentially under-performing regions, irrespective of whether the input to
714 R. Das et al.

the algorithm is an existing design or a guess design. Hence, the topology pro-
duced by ESO generally needs to be post-processed. The post-processing may in-
volve some, or all, of the following procedures, viz: (a) smoothing the topology to
produce a manufacturable layout by eliminating jagged surfaces, (b) inclusion of de-
sirable features and constraints, (c) redesigning certain parts of the structure, and
(d) reanalysing the final topology to check if it meets all the design/performance cri-
teria.
In general, two approaches can be followed to incorporate desirable geometric
characteristics, and/or to retain some of the features present in the original design,
viz: incorporation of the constraints in the initial design (as followed here), and inclu-
sion of them in the final design by modifying the ESO produced (optimal) topology.
In the first approach, the required geometric features are to be included in the starting
structure and appropriately constrained so that the optimisation algorithm can nei-
ther remove nor distort these features. The advantages of this approach are: (a) the
design specifications are guaranteed to be part of the final optimal topology, (b) the
optimisation algorithm takes into account the presence of the constraints throughout,
and produces an optimal topology under the given circumstances, and (c) the finally
obtained design topology will usually require minor modifications, as most of the de-
sign requirements would have been already incorporated. However, restricting a large
portion of the structure will considerably reduce the design space, and may limit the
capability of the optimiser to locate innovative and improved topologies. Thus, this
may restrain the natural evolution of the structural topology towards an optimum.
An alternative approach to this is to allocate most of the structure for topology evo-
lution and incorporate the design features later. This leads to a better exploration of
the design space. However, this may also pose difficulties as the locations where the
desired design features are required in the final structure may be considerably altered
by the optimiser, whereby a major alteration of the final topology will be required.
This may degrade the performance of the resulting structure from that of the ESO
produced optimal topology. Thus, the final (altered) design may no longer remain
optimal. On the other hand, if an essential geometric feature is removed by the opti-
miser, it merely points out that the component or part is not being effectively utilised,
and has the potential to be redesigned to reduce weight or to enhance performance of
the primary structure. So both the approaches have relative merits and demerits, and
the choice should be based on the nature of a specific problem. This also suggests
that adopting a combined approach may be beneficial in some cases.
Implementation of stress constraints in ESO involves monitoring the variation in
the maximum stress with reduction in weight of the structure. The nature of this
variation is expectedly found to be problem dependent. For example, in the bulk-
head problem the maximum stress showed a steep rise from the 70th iteration to the
optimum point (iteration 86), whereas the change of volume (per iteration) did not
appreciably change from that being followed previously, see Figs. 14 and 15. On the
other hand, in the case of the sideframe, the maximum stress remained essentially the
same throughout the optimisation process (Fig. 25), whilst the weight of the structure
considerably reduced (Fig. 24). For complex structures, the maximum stress can be
significantly influenced by changes in the local geometry as well as in the primary
load paths due to the removal of material. The effect of former is particularly preva-
lent in the case of ESO produced topologies due to the presence of uneven boundaries
Optimal topology design of industrial structures 715

and surfaces. For the sideframe, the local geometry around the regions of initial max-
imum stress was not modified by the optimiser, and the stresses in the other altered
parts of the structure did not exceed this initial maximum stress, thus keeping the
location and value of the maximum stress nearly the same throughout the optimisa-
tion.
To ensure a stable convergence, the objective and/or convergence functions need
to be monitored, and the material removal should be accordingly controlled. In this
work, the change in the weight of the structure in a given optimisation cycle was
restricted to be within a permissible limit. For example, in the case of bulkhead prob-
lem, the maximum permissible amount of material removal per iteration was limited
to 4% of the original weight of the structure, although throughout the majority of
the optimisation process the material removal was within 1%, see Fig. 14. This
prevented an abrupt change in the topology and ensured the stability of the process.

8 Conclusions

This paper has investigated the Evolutionary Structural Optimisation algorithm for
topology generation of complex industrial structures. The traditional ESO algorithm
has been extended incorporating a number of features, such as an improved ESO
criterion, an adaptive variation of the material elimination factor, and a convergence
function. The modified ESO algorithm developed was validated against another op-
timisation method, was found to converge well, and did not produce any checker-
board patterns or unrealistic topologies. The capabilities of the ESO algorithm were
demonstrated by considering two topology design problems encompassing different
industrial applications, viz: layout design of a bulkhead of an F/A-18 aircraft and a
sideframe of a rail bogie. In both the cases, the ESO algorithm acted upon existing
designs to generate improved (topological) layouts. The current designs of the two
example problems were analysed and the regions that were not effectively perform-
ing were identified. Gradual material elimination from various parts of the structure
led to the evolution of an improved topology.
The ESO algorithm is found to be a reliable, robust, efficient, and practical tool
for topology optimisation of real life complex structures. It can be used to produce
conceptual (approximate) structural topologies for given loading and boundary con-
ditions. The ESO based techniques are relatively inexpensive, easy to implement and
can serve as a good starting point for further design improvement by subsequently
applying other (shape) optimisation techniques. However, this method generates ir-
regular boundaries and surfaces that need post-processing to produce smooth and
manufacturable designs. The geometry of the ESO generated structures can be fur-
ther locally modified using (ESO based) shape optimisation to reduce the local peak
stresses in the resulting structures.
A significant weight reduction of the structure can often be achieved without
greatly compromising the strength below an acceptable limit. This requires moni-
toring the (generated) topologies and the corresponding maximum stress values. The
structure with the maximum stress just below the allowable limit can be taken as the
lightest design under the given stress constraint.
716 R. Das et al.

The ESO technique can also be applied to obtain a structure with an improved
material utilisation. This method attempts to reduce the variation in stress levels and
produces a relatively uniformly stressed structure, thus leading to a more optimum
material utilisation. In this case, the convergence function involving the ESO parame-
ter needs to be monitored throughout the optimisation process. The minimum value
of this function will be indicative of a design with an approximately uniform stress
distribution. In this ESO formulation, it is also simple to include multiple design pa-
rameters in setting the material removal criteria. Note that here we only focused on
structural performance and efficiency. The economic feasibility, considering the cost
savings achieved from an optimal structure and the possible increase in the cost of
fabrication for implementing an optimal topology, needs to be considered to assess
the viability of implementation of a topology optimised structure.

Acknowledgements The authors wish to acknowledge the support of Noran Engineering, the distributor
of NE-NASTRAN, in integrating NE-NASTRAN into the present optimisation system. The authors also
wish to thank DSTO, Bradken Rail, and Rail CRC for providing the initial designs presented here.

References

Arutyunyan NK, Drozdov AD (1988) Optimization problems in the mechanics of growing solids. Mech
Compos Mater 24(3):359369
AusLink-White-Paper, http://www.dotars.gov.au/auslink/white_paper/index.aspx (accessed 20 January,
2004)
Azegami H (1990) Proposal of shape optimisation using a constitutive equation of growth. JSME Int J
33(1):6471
Bendse MP (1989) Optimal shape design as a material distribution problem. Struct Optim 1:193202
Bendse MP (1995) Optimization of structural topology, shape, and material. Springer, Heidelberg
Bendse M, Sigmund O (2003) Topology optimizationtheory, methods and applications. Springer, New
York
Bruns TE, Tortorelli DA (2003) An element removal and reintroduction strategy for the topology opti-
mization of structures and compliant mechanisms. Int J Numer Methods Eng 57(10):14131430
Buhl T (2002) Simultaneous topology optimization of structure and supports. Struct Multidiscipl Optim
23(5):336346
Buhl T, Pedersen CBW, Sigmund O (2000) Stiffness design of geometrically nonlinear structures using
topology optimization. Struct Multidiscipl Optim 19(2):93104
Cerny V (1985) Thermodynamical approach to the travelling salesman problem: an efficient simulation
algorithm. J Optim Theory Appl 45:4151
Cheng G, Jiang Z (1992) Study on topology optimization with stress constraints. Eng Optim 20(2):129
148
Das R, Jones R (2004a) Designing structures for optimum fracture strength. In: Proceedings of the inter-
national conference on failure analysis and maintenance technologies, Brisbane, Australia
Das R, Jones R (2004b) NASESO user manual v 2.0. Monash University, Australia
Das R, Jones R (2009a) Designing cutouts for optimum residual strength in plane structural elements. Int
J Fract 156(2):129153
Das R, Jones R (2009b) Fatigue life enhancement of structures using shape optimisation. Theor Appl Fract
Mech 52:165179
Das R, Jones R (2009c) Development of a 3D biological method for fatigue life based optimisation and its
application to structural shape design. Int J Fatigue 31(2):309321
Das R, Jones R (2009d) Damage tolerance based design optimisation of a Fuel Flow Vent Hole in an
aircraft structure. Struct Multidiscipl Optim 38(3):245265
Das R, Jones R, Xie YM (2005) Design of structures for optimal static strength using ESO. Eng Fail Anal
12(1):6180
Das R, Jones R, Peng D (2006) Optimisation of damage tolerant structures using a 3D biological algorithm.
Eng Fail Anal 13(3):362379
Optimal topology design of industrial structures 717

Das R, Jones R, Chandra S (2007) Damage tolerance based shape design of a stringer cutout using evolu-
tionary structural optimisation. Eng Fail Anal 14(1):118137
Dixon B, Molent L (2003) Ex-service F/A-18 centre barrel fatigue flaw identification test plan. DSTO,
Melbourne, Australia. Report No. DSTO-TR-1426
Duysinx P, Bendse MP (1998) Topology optimization of continuum structures with local stress con-
straints. Int J Numer Methods Eng 43(8):14531478
Goldberg DE (1989) Genetic algorithms, in search, optimization and machine learning. Addison-Wesley,
Reading
Haber RB, Jog CS, Bendse MP (1996) New approach to variable-topology shape design using a constraint
on perimeter. Struct Optim 11(12):112
Harder RF (2000) Dynamic modelling and simulation of three-piece North American freight vehicle sus-
pensions with non-linear frictional behaviour using ADAMS/Rail. In: 5th ADAMS/Rail users con-
ference, Haarlem
Holland JH (1975) Adaptation in natural and artificial systems. MIT Press, Cambridge
Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by simulated annealing. Science
220(4598):671680
Li Q, Steven GP, Querin OM, Xie YM (1999a) Shape and topology design for heat conduction by evolu-
tionary structural optimization. Int J Heat Mass Transfer 42(17):33613371
Li Q, Steven GP, Querin OM, Xie YM (1999b) Evolutionary topology optimisation for thermoelastic
structures. ASME Trans J Mech Des 121
Li Q, Steven GP, Querin OM, Xie YM (2000) Structural topology design with multiple thermal criteria.
Eng Comput 17(6):715734
Li Q, Steven GP, Xie YM (2001) A simple checkerboard suppression algorithm for evolutionary structural
optimization. Struct Multidiscipl Optim 22(3):230239
Manickarajah D, Xie YM, Steven GP (1998) Evolutionary method for optimization of plate buckling
resistance. Finite Elem Anal Des 29(34):205230
Manickarajah D, Xie YM, Steven GP (2000) Optimization of columns and frames against buckling. Com-
put Struct 75(1):4554
Mattheck C, Baumgartner A, Walther F (1990) Design and growth rules for biological structures and their
application to engineering. Fatigue Fract Eng Mater Struct 13(5):535550
Navarrina F, Muios I, Colominas I, Casteleiro M (2005) Topology optimization of structures: A minimum
weight approach with stress constraints. Adv Eng Softw 36(9):599606
Neves MM, Rodrigues H, Guedes JM (2000) Optimal design of periodic linear elastic microstructures.
Comput Struct 76:421429
Pardalos PM, Korotkikh V (2003) Optimization and industry: new frontiers. Kluwer Academic, Dordrecht
Pardalos PM, Resende MGC (2002) Handbook of Applied Optimization. Oxford University Press, London
Querin OM, Young V, Steven GP, Xie YM (2000) Computational efficiency and validation of bi-directional
evolutionary structural optimization. Comput Methods Appl Mech Eng 189(2):559573
Schnack E (1979) An optimization procedure for stress concentrations by the finite element technique. Int
J Numer Methods Eng 14:124151
Sigmund O (2001) A 99 line topology optimization code written in Matlab. Struct Multidiscipl Optim
21:120127
Svanberg K, Werme M (2006) Sequential integer programming methods for stress-constrained shape and
topology optimization. In: III European conference on computational mechanics, p 498
Xie YM, Steven GP (1993) Simple evolutionary procedure for structural optimization. Comput Struct
49(5):885896
Xie YM, Steven GP (1996) Evolutionary structural optimization for dynamic problems. Comput Struct
58(6):10671073
Xie YM, Steven GP (1997) Evolutionary structural optimization. Springer, Heidelberg
Yang XY, Xie YM, Steven GP, Querin OM (1999a) Bidirectional evolutionary method for stiffness opti-
mization. AIAA J 37(11):14831488
Yang XY, Xie YM, Steven GP, Querin OM (1999b) Topology optimisation for frequencies using an evo-
lutionary method. J Struct Eng Am Soc Civ Eng 125(12):14321438
Zhao C, Steven GP, Xie YM (1996) Evolutionary natural frequency optimization of thin plate bending
vibration problems. Struct Optim 11(34):244251
Zhao C, Steven GP, Xie YM (1997) Evolutionary optimization of maximizing the difference between two
natural frequencies of a vibrating structure. Struct Optim 13(23):148154
Zhou M, Rozvany GIN (2001) On the validity of ESO type methods in topology optimization. Struct
Multidiscipl Optim 21(1):8083