PERGAMON Computers and Structures 68 (1998) 419±431

An evolutionary structural optimization method for sizing

problems with discrete design variables
Chu D. Nha a, Y. M. Xie a, *, G. P. Steven b
a
Department of Civil and Building Engineering, Victoria University of Technology, P.O. Box 14428, MCMC, Melbourne, Victoria
8001, Australia
b
Department of Aeronautical Engineering, University of Sydney, NSW 2006, Australia

Received 27 December 1996; accepted 6 November 1997

Abstract

This paper presents a simple evolutionary method for minimizing the weight of structures subject to displacement
constraints, where sizing design variables are discrete. Sensitivity numbers for element size reduction are derived
using optimality criteria methods. An optimal design can be obtained by gradually reducing elements which have
the lowest sensitivity numbers until one of the constrained displacements reaches its given limit. It is shown that the
proposed method can directly deal with discrete design variables. A simple smoothing technique is employed to
suppress formation of checker-board patterns. Examples are provided to demonstrate the capability of the proposed
method for discrete sizing optimization problems. # 1998 Elsevier Science Ltd. All rights reserved.

Keywords: Structural optimization; Evolutionary method; Displacement constraints; Discrete sizing variables; Checker-board
patterns

1. Introduction relaxation methods, is used to get the discrete

solution [1±5]. The values given for each discrete de-
In many practical applications of structural optimiz- sign variable are usually required to be close to each
ation, the design variables must be selected from a other for validity of converting a continuous optimal
given set of discrete values. For example, structural solution to a discrete one [3].
members may have to be selected from standard sec- Recently, a simple method for shape and layout pro-
tions or thicknesses commercially available from man- blems, called evolutionary structural optimization
ufacturers. In traditional optimization techniques, such (ESO), has been proposed by Xie and Steven [6]. The
as mathematical programming and optimality criteria ESO method is based on the concept of gradually
methods, the derivatives of the objective and constraint removing redundant material to achieve an optimal
functions with respect to design variables are needed. design. It was developed by Xie and Steven ®rst for
So these methods can only handle problems with con- shape and layout problems under stress
tinuous design variables. They require a more sophisti- consideration [6, 7] and then for frequency
cated mathematical treatment when dealing with optimization [8]. The ESO method for shape and top-
discrete design variables. Often the problem is solved ology problems with displacement constraints has been
for continuous optimal solution assuming all designs presented recently by Chu et al. [9±11]. An evolution-
variables to be continuous. Then, one of the methods, ary procedure for sizing members to increase the buck-
such as rounding-o€, branch and bound methods, ling load factor has been proposed by Manickarajah et
simulated annealing, genetic algorithm, Lagrangian al. [12]. It is found that the ESO methods are simple
and can be easily implemented into any general pur-
pose ®nite element analysis (FEA) program. Many
* Author to whom correspondence should be addressed. structural optimization solutions obtained by other

0045-7949/98/$19.00 # 1998 Elsevier Science Ltd. All rights reserved.

PII: S 0 0 4 5 - 7 9 4 9 ( 9 8 ) 0 0 0 6 2 - 5
420 C.D Nha et al. / Computers and Structures 68 (1998) 419±431

more complex methods have been reproduced by the higher order term, the change in the displacement vec-
ESO methods [6±12]. The current status of ESO can be tor can be found as
found in ref. [13].
fDug ˆ ÿ‰KŠÿ1 ‰DKŠfug: 3†
This paper presents an ESO method for discrete siz-
ing optimization problems with displacement con- To ®nd the change in the speci®ed jth displacement
straints. Sensitivity numbers for element reduction are component uj, a virtual unit load vector {Fj}, in which
formulated using optimality criteria methods. A num- only corresponding jth component is equal to unity
ber of elements with the lowest sensitivity numbers will and all the others are equal to zero, is introduced.
have their sizes reduced to the next lower values from Multiplying Eq. (3) by {Fj}T, the change in the speci-
a given set of discrete values. The optimal design of ®ed displacement due to sizing the ith element is deter-
the structure is obtained by repeating the cycle of ®nite mined by
element analysis and sizing elements until one of the
displacement constraints reaches its given limit. To Duj ˆ fFj gT fDug ˆ ÿfFj gT ‰KŠÿ1 ‰DKŠfug
suppress formation of checker-board patterns, smooth- ˆ ÿfuj gT ‰DKŠfug ˆ ÿfuij gT ‰DKi Šfui g
ing of the sensitivity numbers of elements is proposed.
ˆ fuij gT ‰Ki Šfui g ÿ fuij gT ‰Ki Šnew fui g; 4†
The proposed ESO method for sizing problems proves
to be simple and capable of dealing with discrete de- where [DK] = [DKi] = [Ki]newÿ[Ki] is employed, {uj} is
sign variables directly. the solution of Eq. (1) for the virtual unit load {Fj},
{ui} and {uij} are the displacements of the ith element
due to the real load {P} and the virtual unit load {Fj},
respectively. It should be noted that Duj can be positive
2. Sensitivity numbers for element reduction or negative, which implies that uj may increase or
decrease.
Consider a problem of minimizing the weight of The value
structures subject to displacement constraints.
Thicknesses or cross-sectional dimensions of the el- aij ˆ fuij gT ‰Ki Šfui g i ˆ 1; n† 5†
ements must be assigned the values from given sets of
is known as the virtual energy of the ith element. The
discrete values. In sizing optimization problems el-
ements cannot be removed. The only way to reduce change in the element virtual energy due to only sizing
the weight of a structure is to remove under-utilised can be determined as
material from the elements by changing the element Daij ˆanew
ij ÿ aij ˆ fuij gT ‰Ki Šnew fui g ÿ fuij gT ‰Ki Šfui g
sizes (thicknesses or cross-sectional dimensions). This
i ˆ1; n†: 6†
can be done by simply assigning to their sizes the next
lower available values. Similar to the ESO method for Therefore, Eqs. (4) and (6) give
shape and topology optimization problems [9±11], the
e€ect of element size reduction on a speci®ed displace- Duj ˆ ÿfuij gT ‰DKi Šfui g ˆ ÿDaij i ˆ 1; n†; 7†
ment can be assessed using information available from
which means that, in absolute values, the change in the
a ®nite element analysis and then sensitivity numbers
speci®ed displacement is equal to the change in the vir-
for element reduction can be derived from optimality
tual energy within an element due to changing its sizes.
criteria for the general weight minimization problems.
By taking sum of aij in Eq. (5) over all elements, we
Suppose the ith element is to be sized to the next
have well known relationship
lower available dimension. This results in the change
in its weight by the value Dwi=wnew i ÿwi and the X
n
uj ˆ aij ; 8†
change in its sti€ness matrix by [DKi] = [Ki]newÿ[Ki].
iˆ1
The change in displacements can be determined by
considering equilibrium conditions before and after the where aij is also referred to as the element contri-
change. This gives bution.
By using Eq. (7), the change in the speci®ed displa-
‰KŠfug ˆ fPg 1†
cement is estimated by the change in the element vir-
and tual energy due to changing the element sizes. It is
obvious that reducing the element, whose Daij is close
‰K ‡ DKŠfu ‡ Dug ˆ fPg; 2†
to zero or vDaijv is the lowest, will result in the mini-
where [K] is the global sti€ness matrix, {u} is the glo- mum change in the displacement. When all elements
bal nodal displacement vector, and {P} is the nodal can be reduced by equal weight portions, reduction of
load vector. No change in the nodal load vector is the element with the lowest vDaijv is always the best
assumed. By subtracting (1) from (2) and ignoring the choice, because the weight of structure is reduced by
 C.D Nha et al. / Computers and Structures 68 (1998) 419±431 421

the same amount while the least change in the speci®ed valid for a general class of optimization problems with
performance is resulted in. However, when elements diminishing return on investment' [14]. It is seen that
are reduced di€erently, the eciency of reduction of the value gi is an essential measure of eciency of the
an element depends also on how much the change Dwi material within the portion to be removed from the ith
in its own weight wi. Comparing two elements whose element. By sizing elements, an uniform state where gi
reductions result in the same vDaijv, it is obvious that is equal for all elements, can be achieved. Thus, the
reduction of the element which gives a larger reduction weight of the structure is indirectly minimized. When
in weight will result in a lighter structure with an equal the design variables are continuous, various recurrence
response. This means that reduction of the element formulas are suggested based on the optimality con-
with lower ratio vDaijv/vDwiv is more ecient. ditions to derive solution to the optimum [14±17]. By
The question of whether the ratio vDaijv/vDwiv can be means of sizing elements, where size variables are dis-
used as sensitivity numbers for element reduction and crete, it is not always possible to reach such an uni-
reduction of elements with lowest values of this ratio form state. However, a more uniform gi in the
can result in a minimum weight design is addressed in resulting structure can be achieved by reducing el-
the following using Lagrange multiplier approach: ements with the lowest gi while making larger re-
duction in the weight and less change in the
constrained displacement. An iterative process is
2.1. Single constraint applied to reduce the sizes of a small number of el-
ements with the lowest gi gradually, whilst the speci®ed
The objective is to minimize the weight of a struc- displacement is slowly approaching the prescribed
ture limit, and thus the minimum weight is reached at the
X
n same time. Therefore, the sensitivity number for element
Wˆ wi 9† reduction for a single displacement constraint is de®ned
iˆ1 as
subject to a constraint imposed on a speci®ed displace- Daij j fuij gT ‰DKi Šfui g j
ment given in the form as gi ˆ gij ˆ ˆ i ˆ 1; n†: 15†
Dwi j Dwi j
j uj j ÿuj  0; 10†
This sensitivity number is easily calculated using the
where n is number of elements, wj is the weight of the results available from a ®nite element analysis of the
ith element and u*j is the prescribed limit for vujv. The static problem (1) for both the real load {P} and the
Lagrangian for the problem is virtual unit load {Fj}. It is not dicult to calculate the
change in the element sti€ness matrix and the change
L ˆ W ‡ l j uj j ÿuj †; 11† in the element weight since the new thickness or sec-
where l is a Lagrange multiplier. Taking wi (i = 1, n) tion is speci®ed as the next lower thickness or next
as design variables, the optimality conditions for the smaller section in the given sets.
problem are
2.2. Multiple constraints
@L @W @uj @uj
ˆ ÿl ˆ1ÿl ˆ0 i ˆ 1; n†: 12†
@wi @wi @wi @wi Consider a problem with multiple displacement con-
Further discussion usually needs an explicit expression straints given as
of uj in terms of design variables. However, Eq. (12) j uj j ÿuj  0 j ˆ 1; m†; 16†
can be approximated by
where m is the number of constraints, u*j is the given
Duj
1ÿl ˆ0 i ˆ 1; n†: 13† limit for vujv. The Lagrangian is
Dwi
X
m
By using Eq. (7), the optimality condition (13) LˆWÿ lj j uj j ÿuj †; 17†
jˆ1
becomes
j Daij j 1 where lj are m Lagrange multipliers. The optimality
gi ˆ ˆ ˆ constant i ˆ 1; n†: 14† conditions are
j Dwi j l
Eq. (14) represents an optimality criteria that, at the @L @W X m
@uj Xm
@uj
ˆ ÿ lj ˆ1ÿ lj ˆ0 i ˆ 1; n†;
optimum, the absolute ratio of the change in the el- @wi @wi jˆ1
@wi jˆ1
@wi
ement virtual energy and the change in the element
18†
weight is equal for all elements. This is a specialised in-
terpretation of `the correct optimality criteria statement where lj>0 for the active constraints vujv ÿ u*j =0, and
422 C.D Nha et al. / Computers and Structures 68 (1998) 419±431

lj=0 for the passive constraints vujv ÿ u*j <0. The !1=b
optimality conditions (18) can be approximated by j uj j
lnew
j ˆ lold
j j ˆ 1; m† 25†
uj
X
m
Duj
1ÿ lj ˆ0 i ˆ 1; n†: 19†
jˆ1
Dwi where b is a step control parameter. The use of the
recurrence formula (25) requires that the initial values
Using Eq. (7), the optimality condition (19) becomes of the Lagrange multipliers have to be assumed. The
X
m value of a Lagrange multiplier corresponding to a less
Daij
1ÿ lj ˆ0 i ˆ 1; n† 20† potentially active constraint is reduced, which scales
jˆ1
Dwi down the contribution from the corresponding energy
density of an element to the weighted sensitivity num-
or ber de®ned by Eq. (22). It has been found that because
X
m
Daij the recurrence formula (25) successively reduces the
Zi ˆ lj ˆ1 i ˆ 1; n†: 21† Lagrange multiplier corresponding to a potentially less
jˆ1
Dwi
active constraint, it may not be able to pick up a
Eq. (21) represents an optimality criterion for multiple change when the constraint becomes the potentially
constraints that, at the optimum, the weighted sum of most active in a later iteration. The Lagrange multi-
plier corresponding to this constraint may still remain
the absolute ratio of the change in the element virtual
smaller than the others.
energy and the change in the element weight is equal
To avoid the drawback of the recurrence formula
to unity for all elements, where the weighting par-
(25), we propose a simple formula for Lagrange multi-
ameters are the Lagrange multipliers. Reducing the el-
pliers given in the form as
ements with the lowest values of Zi will result in more !1=b
uniform Zi among remained elements, hence a more j uj j
ecient design. The value Zi can be viewed as a lj ˆ j ˆ 1; m†; 26†
uj
measure of the eciency of reduction of an element in
the case of multiple constraints. Therefore, the sensi- where Lagrange multipliers are determined only by the
tivity number for element reduction for multiple displace- ratios of the current values of the constrained displace-
ment constraints is de®ned as ments to their given limits. A special case of the for-
mula (26) where b = 1 has been used by Chu et al. [9±
X
m
aij
Zi ˆ lj i ˆ 1; n† 22† 11]. Similar to Eq. (25), formula (26) also reduces the
jˆ1
Dwi contribution from a potentially less active constraint to
the sensitivity numbers de®ned by Eq. (22). The advan-
Using Eq. (15), the above equation can be rewritten as tage of the formula (26) is its simplicity and ability to
X
m pick up any change in the relative order of a con-
Zi ˆ lj gij 1 ˆ 1; n† 23† straint. The Lagrange multiplier corresponding to the
jˆ1 current potentially most active constraint always has
the largest value, so that the contribution from this
which implies that the sensitivity number of an element
constraint to the weighted sensitivity number is domi-
for multiple constraints is the weighted sum of its sen- nant. The advantage of both formulas (25) and (26) is
sitivity numbers corresponding to each constraint. that an equivalent single virtual load vector
To calculate the sensitivity numbers in Eq. (22) or {F} = Slj{Fj} can be used to determine the sensitivity
(23), the values for Lagrange multipliers are needed. number of an element without the necessity of deter-
The method of evaluation of the Lagrange multipliers mining individual values aij. This reduces the compu-
plays an important role in any optimization procedure tation time substantially when the number of
using an optimality criterion approach. Various for- constraints increases.
mulas, with their advantages and disadvantages, for The Lagrange multipliers can also be determined by
determining Lagrange multipliers, have been given in linear equations which are obtained by combining the
literature [14±17]. A simple way is to assume all the optimality criterion (21) and the constraint Eq. (24),
constraints in Eq. (16) are equality constraints, which where uj is determined by Eq. (8). The linear equations
gives have the form
!
j uj jˆ uj j ˆ 1; m† : 24† X m X n
Daip †aij
lp ˆ uj j ˆ 1; m†: 27†
pˆ1 iˆ1
Dwi
Multiplying both sides of this relation by lbj and taking
bth root, a recurrence formula can be written as Although Eq. (27) is valid only at the optimum, it can
 C.D Nha et al. / Computers and Structures 68 (1998) 419±431 423

be used to estimate the Lagrange multipliers. The ution. So the original limits can be used in formulas
incremental procedure based on Eq. (27) to update lj (25)±(27) to determine the Lagrange multipliers.
is as follows
!
Xm X n
Daip †aij
lp ˆ 1 ‡ r†uj ÿ ruj j ˆ 1; m†; 3. Optimization procedure
pˆ1 iˆ1
Dwi
28† Using the constraint limit scaling, the evolutionary
procedure for size selection to minimize the weight of
where r is a step control parameter. The Lagrange structures subject to displacement constraints can be
multipliers determined by Eqs. (27) and (28) take into stated as follows:
account the interdependence of di€erent constraints by Step 1: Model the structure by ®nite elements with
including coupling terms. However, the computation maximum available sizes.
e€ort to determine every aij and Daij, to assemble the Step 2: Analyse the structure for the given loads and
coecients for the Lagrange multipliers and to solve virtual unit loads.
the set of simultaneous equations increases substan- Step 3: If any constraint is violated, go to Step 7.
tially with increase in the number of constraints. Otherwise, go to Step 4.
Eqs. (25)±(28) are valid under the assumption that Step 4: Calculate the sensitivity numbers gij for each
one or more constraints are active. The use of these element corresponding to each constraint. For multiple
equations to determine Lagrange multipliers requires a constraints, scale the constraint limits, select the active
procedure to make one or more constraints active. A set of constraints, calculate the Lagrange multipliers
simple procedure widely used in continuous sizing op- for the active set and the weighted sensitivity numbers
timization problems is scaling design variables by a Zi for each element.
scalar. In the proposed ESO method for discrete sizing Step 5: If the sensitivity numbers are uniform, go to
optimization where elements are allowed to have only Step 7. Otherwise, go to Step 6.
the sizes from the given sets, an alternative procedure Step 6: Select lower sizes for a number of elements
to make constraints active is to scale all the given dis- which have the lowest sensitivity numbers. Go back to
placement limits, so that the potentially most active Step 2.
constraint becomes active. The activeness of a con- Step 7: Terminate the iterative process. Scale actual
straint in the form (16), is characterised by the value weights to get the objective (scaled) weights for the
fj=vujv/u*j (j =1,m). The closer the value of fj to speci®ed limit and select the optimum solution among
unity, the more potentially active is the constraint. The the available designs.
scaling factor for the displacement limits is determined Initially, all elements are assigned their given maxi-
as mum thicknesses or sections. If after analysis this in-
! itially chosen design violates the displacement
j uj j constraints, then the optimization process is terminated
f ˆ maxjˆ1;m fj ˆ maxjˆ1;m : 29†
uj because no further feasible design can be obtained by
removing material from the structure. Sensitivity num-
Therefore, the given limits u*j is scaled to fu*j when bers are calculated assuming that the element thickness
Eqs. (25)±(28) are used to determine Lagrange multi- will change to the next lower value. During the optim-
pliers. It is easy to check that the potentially most ization process elements whose thicknesses or sections
active constraint has the current value of the corre- have been reduced to the minimum thickness or sec-
sponding displacement as its limit, hence becoming tion, are kept unchanged. The optimization process is
active. Other potentially less active constraints are less terminated when the displacements reach the pre-
active or passive. As displacements increase due to re- scribed limits. The iterative process is also terminated
duction of elements, f increases to unity and the when sensitivity numbers become uniform where
scaled limits approach the given limits. By using the optimality conditions are satis®ed.
scaled limits, we try to minimize the weight of the In the proposed procedure, the number of elements
structure while keeping displacements at the levels gov- subjected to size reduction needs to be speci®ed, which
erned by the potentially most active constraint. When can be prescribed by one of the following options:
the Lagrange multipliers are determined using linear
equations, any negative value will be temporarily set to 1. by an integer number;
zero, which implies that the corresponding constraint 2. by the ratio of the number of elements subjected to
is passive. It is seen from formulas (25)±(27) that scal- size reduction to the total number of elements,
ing the given limits simply scales all the Lagrange mul- called as the element changing ratio (ECR); or
tipliers by a constant, either fÿ(1/b) or f. As it will be 3. by the ratio of the weight (or material) to be
shown in the next section, this does not a€ect the sol- removed in each iteration to the total weight (or
424 C.D Nha et al. / Computers and Structures 68 (1998) 419±431

material), called as the material removal ratio in one or more cross-sectional dimensions will be
(MRR). taken into account by using di€erent section proper-
ties. Each member and each group of members can
Options (1) and (2) are always equivalent. All three have their own set of properties. Initially all elements
options are equivalent only when equal weight por- are assigned the properties with maximum thicknesses
tions can be removed from elements and the initial or sections. After calculation of sensitivity numbers, el-
weight is considered as the total weight. When the cur- ements with the lowest sensitivity numbers are reduced
rent weight is considered as the total weight and equal by changing their properties to the next properties
weight portions can be removed from elements, the with lower thicknesses or smaller sections. The proper-
number of elements subjected to sizing decreases as the ties with minimum thicknesses or smallest sections are
weight of structure reduces. In general, when elements identi®ed as non-design properties. Elements are kept
are changing di€erently and the material removal ratio unchanged when they reach their minimum thicknesses
(MRR) is prescribed, the number of elements subjected or sections.
to reduction varies.
To get an accurate solution either the number of el-
ements subjected to sizing, the value of ECR or MRR
should be small enough to ensure a smooth change in 4. Suppression of checker-board patterns
sti€ness of the structure and consequently, a smooth
change in displacements. It is observed that the level of It is observed that checker-board patterns often
change in displacements depends not only on the num- appear in solutions derived by the proposed method
ber of elements subjected to size reduction (de®ned by for two-dimensional continuum structures. This is
ECR), but also on how much the change in the sizes is quite typical in topology optimization using ®nite el-
or how big the step size is. For bar or plate elements, ements. With checker-board patterns the obtained
the decrement in cross-section areas or in thicknesses, shapes may become practically unacceptable. The ori-
respectively, is considered as a step size. For beam el- gin of checker-board patterns is still not fully under-
ements more than one step size may be needed. The stood but it is likely to be related to the ®nite element
change in the weight of elements is also an important approximation as a numerical phenomenon [18]. DiaÂz
factor, which re¯ects any change in any cross-sectional and Sigmund [19] pointed out that material in checker-
dimension of elements. So the decrement in the el- board arrangements in 4-node element meshes appears
ement weight should be considered as a common step to be locally sti€er than any real material built. It is
size to avoid speci®cation of any particular physical desirable to suppress the formation of checker-board
dimension. As the step size increases, more material is patterns. One way is to use higher-order elements.
removed from elements and the total change in displa- Rodrigues and Fernandes [20] showed that the
cements will be bigger despite the same ECR is used. checker-board patterns which appear by using 4-node
In this case, a smaller number of elements subjected to elements, can be avoided by using 9-node elements.
sizing (or smaller ECR) should be used. It is found Another way is to use a special algorithm to control
that the material removal ratio (MRR) more closely formation of checker-board patterns as proposed by
relates to the level of change in displacements between Bendsùe et al. [18]. A simpler algorithm for density re-
two adjacent designs. In the case where equal weight distribution was suggested by Youn and Park [21] to
portions are removed from the structure, the number suppress formation of checker-board patterns.
of elements subjected to reduction (or ECR) can be In the problem with displacement constraints, the
determined from the prescribed value of the MRR and sensitivity numbers de®ned by Eq. (15) can be
the step size. It is noted that in the ESO method for expressed by the following formula
shape and topology optimization, the element removal ei
ratio ERR of 1%±4% are used. When regular ®nite el- gi ˆ i ˆ 1; n† 30†
ri
ement meshes and an equal step size are used, the el-
ement removal ratio ERR and the material removal where ri is mass density and
ratio MRR are equivalent. So these values are also Daij
adopted for MRR in sizing problems. ei ˆ i ˆ 1; n† 31†
Dvi
The discrete sizes are prescribed by using di€erent
element properties. They are input either in ascending is the virtual energy density in the part of the element
or descending order as the property number increases. to be removed, Dvi is the volume of this part. It is
The number of properties for a plate element is equal obvious that in continuum structures, energy density is
to the number of given thicknesses. The number of a continuous function. However, the energy density ei
properties for a beam element is equal to the number determined by Eq. (31) is only the average value and
of available sections. For beam elements, any change therefore becomes a stepwise function, experiencing
 C.D Nha et al. / Computers and Structures 68 (1998) 419±431 425

discontinuity at the nodes. To provide continuity of

the energy density at the nodes, smoothing technique
is needed. The simplest procedure for smoothing
energy density is as follows:

1. Calculate the virtual energy density at each node by

averaging the virtual energy densities of elements
connecting to the node.
2. Calculate the smoothed virtual energy density for
an element by averaging the energies densities of its
nodes.

The smoothed sensitivity numbers are obtained from

the smoothed energy density. It will be seen through
examples that the optimization procedure based on the
smoothed sensitivity numbers is able to suppress for-
mation of checker-board patterns.

Fig. 1. Optimal design for a quarter of the plate with two

5. Examples thicknesses, MRR = 1% and limit of 2.5 mm.

We shall illustrate the capability of the proposed Using the smoothing technique the optimal designs are
method for sizing optimization problems with discrete obtained for each set of thicknesses as shown in
design variables. All examples are solved using a Figs. 1±3(b) where no checker-board patterns are
486DX2/66 MHz personal computer. Typically, one observed. It is seen from Figs. 1±3(b) that the thickness
example takes a few hours. distributions are similar. The areas with minimum
thickness in all cases are located along the elastic hinge
5.1. Plate in bending lines, which have been reported in the literature [22].
The use of original sensitivity numbers leads to opti-
Consider an example of least weight design for a
mal solutions with checker-board patterns. Fig. 3a
plate in bending with discrete thicknesses. A simply
supported square plate with sides of 8 m is carrying at
the centre a point load P = 100 kN normal to its
plane. The Young's modulus is E = 30 GPa and the
Poison's ratio n = 0.2. The plate is designed for di€er-
ent sets of discrete thicknesses ranged from the mini-
mum 0.2 m to the maximum 0.4 m. Because of the
symmetry, only a quarter of the plate is analysed using
400 four-node quadrilateral plate elements. Initially, all
elements are assigned the maximum thickness 0.4 m
and the initial out-of-plane displacement at the centre
is 1.78 mm. The limit 2.5 mm is imposed on the displa-
cement at the centre. Using MRR = 1%, the solutions
are obtained for the following cases:

(a) Design for two thicknesses: Dt = 0.2 m,

ECR = 2% (i.e. 8 elements subjected to thickness
reduction in each iteration).
(b) Design for three thicknesses: Dt = 0.1 m,
ECR = 4% (i.e. 16 elements subjected to thickness
reduction in each iteration).
(c) Design for ®ve thicknesses: Dt = 0.05 m,
ECR = 8% (i.e. 32 elements subjected to thickness Fig. 2. Optimal design for a quarter of the plate with three
reduction at each iteration). thicknesses, MRR = 1% and limit of 2.5 mm.
426 C.D Nha et al. / Computers and Structures 68 (1998) 419±431

The eciency of the proposed ESO method can be

easily pointed out by scaling the obtained designs to
keep the speci®ed displacement at the limit and follow-
ing the change in the scaled weight during the optimiz-
ation process. The scaled weight, which is also referred
to as the objective weight, for this example can be
obtained as follows. It is known that the sti€ness
matrix of the simply supported plate in bending is a
function of the cube of the thickness. When the thick-
ness is scaled by a factor a, the sti€ness matrix will be
scaled by a3, and the displacement will be reduced by
1/(a3). Conversely, to get to the limit u*, the actual dis-
placement uact should be scaled by the ratio u*/uact,
which can be done by scaling the thickness by the
cubic root of the reciprocal uact/u*. Therefore, the
objective weight for any ESO topology is obtained as
 act 1=3
u
W obj ˆ W act  : 32†
u
Similarly the initial objective weight is calculated from
the initial actual weight W act
0 and the initial displace-
ment uact
0 as
 act 1=3
u
W0obj ˆ W0act  0 : 33†
u
Thus, the relative objective weight is
 act 1=3  act 1=3
obj W obj W act u act u
Wrel ˆ obj ˆ act  act ˆ Wrel  act ;
W0 W0 u0 u0
34†
where W act
rel is the relative actual weight. It is seen in
Eq. (33) that the limit u* is eliminated, which means
that the pattern of the relative change in the objective
weight does not depend on the speci®ed value for the
limit in this case.
The change in the objective weights in all three cases
are very similar as shown in Fig. 4. The objective
weights reach the minimum values when the displace-
ment reaches the limit. The minimum objective weight
for two thicknesses is 83.61%, for three thicknesses is
Fig. 3. Optimal designs for a quarter of the plate with ®ve
83.51% and for ®ve thicknesses is 83.54% of the initial
thicknesses, MRR = 1% and limit of 2.5 mm: (a) using orig-
inal sensitivity numbers; (b) using smoothed sensitivity num-
value, which are almost the same. All optimal designs
bers. obtained by the proposed method are more ecient in
comparison with the initial uniform design. It is worth
noting that the obtained solutions are optimal for the
speci®ed set of ®xed thicknesses and the prescribed
shows the optimal design obtained for ®ve thicknesses limit of the displacement.
where checker-board patterns spread over a large area. In the case where the thicknesses are allowed to
It is seen in Fig. 3(a, b) that the overall thickness dis- change proportionally, i.e. they can be scaled by the
tributions obtained with and without using smoothing same factor, the optimum thickness distribution may
technique are similar. This particular example shows be obtained beyond the speci®ed limit. In this case, a
that the proposed simple smoothing technique can larger limit for the displacement should be used to
e€ectively suppress formation of checker-board pat- continue the optimization process until no further
terns making the solution more practically acceptable. improvement can be achieved. A close look at these
 C.D Nha et al. / Computers and Structures 68 (1998) 419±431 427

Fig. 4. Optimization history of the plate in bending, MRR = 1%, limit of 2.5 mm.

cases reveals that the objective weight will reach a 3.0 mm is imposed on displacements at three loaded
minimum value at a certain step and then start rising. points.
This is due to the fact that the process starts from the Using MRR = 1% and employing the smoothing
uniform maximum thickness design. As more material technique, solutions are obtained for the following
is removed later from the design, more elements are cases:
assigned the minimum thickness. Sooner or later a uni-
form design with the minimum thickness will be (a) Design for two thicknesses: Dt = 0.1 m,
obtained by which the objective weight will get back to ECR = 2% (i.e. 8 elements subjected to thickness
its initial value. reduction at each iteration), the optimal design is
It is worth pointing out from Fig. 4 that the use of shown in Fig. 6.
same value 1% for MRR in all three cases gives very (b) Design for three thicknesses: Dt = 0.05 m,
similar change in the objective weight despite large ECR = 4% (i.e. 16 elements subjected to thickness
di€erences in Dt (0.20, 0.10 and 0.05 m) and big di€er- reduction at each iteration), the optimal design is
ences in ECR (2, 4 and 8%). This illustrates that the shown in Fig. 7.
material removal ratio is much more important than (c) Design for ®ve thicknesses: Dt = 0.025 m,
the element changing ratio in the evolutionary method ECR = 8% (i.e. 32 elements subjected to thickness
for sizing optimization. reduction at each iteration), the optimal design is
shown in Fig. 8.

The recurrence formulae (25) and the ratio formulae

5.2. Plate under multiple constraints (26) are used to calculate the Lagrange values, where

This serves as an example of structures subject to

multiple displacement constraints. A rectangular plate
with dimensions Lx=4 m and Ly=2 m is clamped at
one long edge. Three point loads P = 20 kN, normal
to the plate, are acting at the free corners and in the
middle of the free edge (see Fig. 5). The Young's mod-
ulus is E = 30 GPa and the Poison's ratio is n = 0.2.
The plate is being designed for di€erent sets of thick-
nesses with the minimum of 0.1 m and the maximum
of 0.2 m. Because of the symmetry, only a half of the
plate is analysed using 400 four-node quadrilateral
plate elements. Initially, all elements are assigned the
maximum thickness t1=0.2 m and the initial maximum Fig. 5. Design domain for the plate under multiple con-
out-of-plane displacement is 2.22 mm. A limit of straints.
428 C.D Nha et al. / Computers and Structures 68 (1998) 419±431

Fig. 6. Optimal design for the plate under multiple constraints with two thicknesses, MRR = 1% and limit of 3.0 mm.

the initial values are assumed to be unity in Eq. (25) reduces to the minimum value when the active displa-
and b = 1 in both cases. Both types of formulae give cement reaches the limit. The minimum weight for two
the same solutions for each case. It is observed that thicknesses is 86.75%, for three thicknesses is 86.65%
the thickness distributions in the optimal designs in all and for ®ve thicknesses is 86.73% of the initial value.
cases are very similar. No formation of checker-board It is observed that the maximum displacement in all
patterns has been observed. cases have very small changes in the ®rst several iter-
The change in the objective weights are shown in ations and then gradually increases in the following
Figs. 9 and 12. The objective weight continuously steps while an equal amount of material is removed in

Fig. 7. Optimal design for the plate under multiple constraints with three thicknesses, MRR = 1% and limit of 3.0 mm.
 C.D Nha et al. / Computers and Structures 68 (1998) 419±431 429

Fig. 8. Optimal design for the plate under multiple constraints with ®ve thicknesses, MRR = 1% and limit 3.0 mm.

each step. This also illustrates the e€ectiveness of the is designed for the set of 10 thicknesses ranged from
formulated sensitivity numbers for multiple displace- the minimum 0.1 mm to the maximum 1 mm with the
ment constraints. step size Dt = 0.1 mm. The plate is modelled by 800
four-node quadrilateral plate elements. All elements
5.3. Plate under torsional loading are initially given the maximum thickness. The initial
vertical displacement at the corners is 0.80 mm. The
A plate of dimensions 0.20  0.10 m is clamped limit of 1.5 mm is imposed on the displacements at the
along one short edge and is under the in¯uence of a corners.
torsional loading, which is created by applying two For MRR = 1% and the step size Dt = 0.1 mm, the
loads of the same magnitude P = 1 N, but opposite number of elements subjected to thickness reduction is
direction at the corner nodes of the free boundary (see equal to 10% of the total elements (i.e. 80 elements
Fig. 10). The Young's modulus is E = 90 GPa and the subject to thickness reduction at each iteration). Both
Poison's ratio is n = 0.3. A similar example was con- formulae (25) and (26) are used which give the same
sidered by Tenek and Hagiwara in ref. [22]. The plate results. The use of smoothing technique leads to the

Fig. 9. Optimization history for the plate under multiple constraints, MRR = 1%, limit of 3.0 mm.
430 C.D Nha et al. / Computers and Structures 68 (1998) 419±431

optimal design without checker-board patterns as

shown in Fig. 11. The objective weight continuously
reduces to the minimum value of 76.47% of the initial
value.

6. Conclusions

This paper has shown that the proposed ESO

Fig. 10. Design domain for the plate under torsional loading. method can directly deal with discrete design variables

Fig. 11. Optimal design for the plate under torsional loading, MRR = 1%, limit of 1.5 mm.

Fig. 12. Optimization history of the plate under torsional loading, MRR = 1%, limit of 1.5 mm.
 C.D Nha et al. / Computers and Structures 68 (1998) 419±431 431

in sizing problems. The optimal solution is obtained by [7] Xie YM, Steven GP. Optimal design of multiple load
simply repeating the cycle of analysis and element re- case structures using an evolutionary procedure. Engng
duction. The thicknesses of elements are gradually Comput 1994;11:295±302.
reduced to the next lower values from the given sets. [8] Xie YM, Steven GP. Evolutionary structural optimization
for dynamic problems. Comput Struct 1996;58:1067±73.
Compared to other existing solution methods, where
[9] Chu DN, Xie YM, Hira A, Steven GP. An evolutionary
repeated analysis is needed to ®nd a continuous sol- procedure for structural optimization with displacement
ution and the problem is re-solved many times to get a constraints. In: Building for the 21st century,
discrete solution, the proposed evolutionary optimiz- Proceedings of the Fifth East Asia±Paci®c Conference on
ation method is extremely simple and ecient. Only Structural Engineering and Construction, 1995;Vol.
one ®nite element analysis is required for each design 2:1091±6.
step. Examples given in this paper, where the number [10] Chu DN, Xie YM, Hira A, Steven GP. Evolutionary
of design variables is in the range from four to eight structural optimization for problems with sti€ness con-
hundreds, illustrate the capability of the proposed straints. Finite Elements Anal Des 1996;21:239±51.
method for sizing problems with discrete design vari- [11] Chu DN, Xie YM, Hira A, Steven GP. On various
aspects of evolutionary structural optimization for pro-
ables. The proposed simple smoothing technique can
blems with sti€ness constraints. Finite Elements Anal
e€ectively suppress formation of checker-board pat- Des 1997;24:197±212.
terns in the solution. Although the examples provided [12] Manickarajah D, Xie YM, Steven GP. A simple method
in this paper are only two-dimensional continuum for the optimization of columns, frames and plates
structures, the method can also be applied to other against buckling. In: Kitipornchai S, Hancock GJ,
types of structures such as trusses and frames. For Bradford MA, editors. Structural stability and design.
beam elements, one can choose a range of sections Rotterdam, Brook®eld: A. A. Balkema Publishers,
commercially available from manufacturers as a set of 1995:175±80.
discrete sections for sizing. [13] Xie YM, Steven GP. Evolutionary structural optimiz-
It is found that in the proposed discrete sizing ation. London: Springer, 1997.
[14] Berke L, Khot N. Performance characteristics of optim-
method, the material removal ratio (MRR) is an im-
ality criteria methods. In: Rozvany, Karihaloo , editors.
portant parameter determining the magnitude of Structural optimization. Dordrecht: Kluwer Academic
change between two consecutive designs. It is obvious Publishers, 1988:31±7.
that the smaller the value for the material removal [15] Venkayya VB, Khot NS, Berke L. Application of optim-
ratio is used, the more accurate the solution will be, ality criteria approaches to automated design of large
but at a higher computational cost. Although the value practical structures. AGARD Second Symposium on
of 1% for the material removal ratio is used, which Structural Optimization, Milan, Italy, 1973.
gives good results, the in¯uence of this parameter on [16] Morris AJ. Foundation of structural optimization: a uni-
the ®nal designs needs further investigations. ®ed approach. New York: Wiley, 1982.
[17] Haftka RT, GuÈrdal Z, Kamat MP. Elements of struc-
tural optimization, 2nd edn. Dordrecht: Kluwer
Academic Publishers, 1990.
[18] Bendsùe MP, DiaÂz A, Kikuchi N. Topology and general-
References ized layout optimization of elastic structures. In: Bendsùe
MP, Mota Soares CA, editors , editors. Topology design
[1] Huang MW, Arora JS. Engineering optimization with of structures. Amsterdam: Kluwer Academic Publishers,
discrete variables. A collection of technical papers, 36th 1993.
AIAA/ASME/ASCE/AHS/ASC Structures, Structural [19] DiaÂz A, Sigmund O. Checkerboard patterns in layout
Dynamics, and Material Conference and AIAA/ASME optimization. Struct Optimiz 1995;10:40±5.
Adaptive Structures Forum, New Orleans, U.S.A., 1995, [20] Rodrigues HC, Fernandes PA. Topology optimization of
Part 1, 1995:1475±83. linear elastic structures subjected to thermal loads. In:
[2] Olsen GR, Vanderplaats GN. Methods for nonlinear op- Bendsùe MP, Mota Soares CA, editors , editors.
timization with discrete design variables. AIAA J Topology design of structures. Amsterdam: Kluwer
1989;27(11):1584±89. Academic Publishers, 1993:437±50.
[3] Ringerts UT. On methods for discrete structural optimiz- [21] Youn SK, Park SH. A study on the shape and topology
ation. Engng Optim 1988;13:47±64. optimization of structures. Computational Mechanics
[4] Sandgren E. Nonlinear integer and discrete programming '95ÐTheory and Applications. In: Proceedings of the
in mechanical design optimization. J Mech Des ASME International Conference on Computational Engineering
1990;112:223±9. Science, July 1995, Hawaii, U.S.A. vol. 1. Springer,
[5] Schmit L, Fleury C. Discrete-continuous variables struc- 1995:183±8.
tural synthesis using dual methods. AIAA J [22] Tenek LH, Hagiwara I. Optimal rectangular plate and
1980;18:1515±24. shallow shell topologies using thickness distribution or
[6] Xie YM, Steven GP. A simple evolutionary procedure for homogenization. Comput Meth Appl Mech Engng
structural optimization. Comput Struct 1993;49:885±6. 1994;115:111±24.