1091

A New Algorithm for Bi-Directional Evolutionary Structural

Optimization∗

Xiaodong HUANG∗∗ , Yi Min XIE∗∗ and Mark Cameron BURRY∗∗∗

In this paper, a new algorithm for bi-directional evolutionary structural optimization

(BESO) is proposed. In the new BESO method, the adding and removing of material is
controlled by a single parameter, i.e. the removal ratio of volume (or weight). The conver-
gence of the iteration is determined by a performance index of the structure. It is found that
the new BESO algorithm has many advantages over existing ESO and BESO methods in
terms of efficiency and robustness. Several 2D and 3D examples of stiffness optimization
problems are presented and discussed.

Key Words: Bi-Directional Evolutionary Structural Optimization, Stiffness Optimization,

Optimal Design, Performance Index

tionary structural optimization (AESO) method(14) . The

1. Introduction
AESO algorithm begins from the minimum domain and
Topology optimization of structures has been exten- adds material around areas of high stress. In so doing, the
sively researched in the past two decades and many op- structure that results is over-designed because the initial
timization methods have been developed(1) – (5) . The evo- domain, no matter how lowly-stressed it is, can never be
lutionary structural optimization (ESO) method has been removed. Thus, the AESO method is unable to find the
received extensive attention because it can be easily im- global optimum except for a few simple cases.
plemented and linked to existing finite element analysis By combining the AESO and ESO methods, bi-
packages. Also, the resulting optimal design provides a directional evolutionary structural optimization (BESO)
clear profile of topology (with no “grey” area) and there- method has been developed by Yang et al. and Querin et
fore easy to manufacture. al.(15), (16) In the BESO method, material of the structure
The ESO process begins with a full design domain can be either be added or removed. The amount of added
and gradually removes underutilized material to achieve and removal material in each iteration is determined by the
an optimal design. The amount of removed material rejection ratio (RR) and inclusion ratio (IR) respectively.
is controlled by two parameters: rejection ratio (RR) However, RR and IR used in the BESO method are irrele-
and evolutionary rate (ER). These two parameters must vant and the user must select the appropriate RR and IR to
be small. The theoretical basis of the ESO method obtain the corresponding optimal design. Otherwise, the
has been explained mathematically by Tanskanen(6) . To final design may not be the global optimum. Furthermore,
date, a wide range of structural optimization solutions elements for removal and those for addition are treated dif-
for stiffness optimization, frequency, buckling etc. have ferently and ranked separated, which is a bit cumbersome
been solved using the ESO method(7) – (9) . Extended and not very logical.
ESO methods and other techniques similar to the ESO In original BESO method, the volume of the struc-
method(10) – (13) have been developed in Japan recently and ture is gradually reduced when a full design domain is
successfully applied to the design of real structures(11) . used as an initial design. When BESO starts from a min-
An alternative to ESO method is the addition evolu- imum domain, the volume of the structure increases un-
∗
til far larger volume than the objective volume is obtained
Received 20th February, 2006 (No. 06-5027)
∗∗
and then decreases to the objective volume (see Figs. 4 and
School of Civil and Chemical Engineering, RMIT Univer-
sity, GPO Box 2476V, Melbourne 3001, Australia.
8 in Ref. (16)). The BESO procedure is stopped when the
E-mail: mike.xie@rmit.edu.au objective volume is reached for the first time or the sec-
∗∗∗
Spatial Information Architecture Laboratory, RMIT Uni- ond time. Obviously, the result from using this type of
versity, GPO Box 2476V, Melbourne 3001, Australia BESO procedure is problematic when a broken member

JSME International Journal Series C, Vol. 49, No. 4, 2006

1092

with no or low strain energy happens to be a part of the 1 iT

{u } [Ki ]{ui }
final topology. This is the reason that the evolutionary pa- αi = 2 (3)
rameters must be very small or certain number of itera- Vi
tions is needed to assure that the BESO method produces which is defined as the sensitivity number of the ith ele-
a reasonable solution. Thus, the computational efficiency ment which indicates the contribute of an element removal
is limited. To overcome the deficiencies of the original or addition to the mean compliance.
BESO method, a new volume control and a stop criterion Checkerboard pattern refers to the phenomena of al-
is developed in this study. ternating presence of solid and void elements ordered in
In this paper, a new algorithm for bi-directional evo- a checkerboard like fashion. It is quite common in vari-
lutionary structural optimization (BESO) method is pro- ous fixed grid finite element based structural optimization
posed. By the linear extrapolation of the sensitivity num- methods. To eliminate the checkerboard, Li et al.(17) pro-
ber in the surrounding elements, both adding and remov- posed a simple and effective smoothing algorithm. Firstly,
ing element material are realized according to a single pa- the sensitivity number at each node can be calculated by
rameter: the removal ratio of volume (or weigh) (RRV). averaging the element ones connecting to this node as
m


Similar to the existing BESO method, the new BESO
Vi αi
algorithm begins with a minimum initial design rather than i=1
αk = (4)
a maximum. Examples in this paper show that the final m


optimal design is independent of the initial design. The Vi
i=1
similarity of the results from different initial designs will where m denotes the total number of elements connected
be discussed in details later. It shows that the new BESO to the kth node, αi the sensitivity number of the ith con-
method is more robust than traditional ESO and BESO nection element.
methods. In order to obtain the sensitivity numbers for the
In the new BESO method, a convergence criterion added elements, it is necessary to extrapolate the sensi-
based on a performance index is introduced to monitor, tivity numbers for the nodes surrounding the structure. As
and ultimately to terminate, the optimization process. The shown in Fig. 1, it is assumed that the sensitivity numbers
new BESO algorithm is very effective in significantly im- vary linearly along the coordinates, so that the unknown
proving the performance index of structures. node sensitivity number is calculated by
2. Optimization Problem and Sensitivity Number αkx− = 2αk1 − αk2 (5)
In many cases, the maximum stiffness of structures is where αkx− denotes the sensitivity number of kth node
pursued, in other words, the mean compliance of the struc- which is calculated along the negative x direction. If pos-
ture should be minimized. When a linear structure is sub- sible, this extrapolation is repeated along other coordinate
jected to the external forces {F}, the topology optimization positive and negative directions. The final sensitivity num-
problem for maximizing stiffness can be formulated with ber of the kth node is calculated by averaging all these val-
the volume constraint using elements as design variables. ues.
Finally, the smoothed sensitivity number of the candi-
1
Minimize C = {F}T {u} (1.a) date element (e) is calculated by averaging all nodal ones
2 of this element as
n


Subject to g = V ∗ − Vi xi = 0 (1.b) N


i=1 αn
α∗e =
n=1
xi ∈ {0,1} (1.c) (6)
N
where {u} are the displacement vectors. Vi is the volume of
an individual element and V ∗ the prescribed total structural
weight. The binary design variable xi declares the absence
(0) or presence (1) of an element.
When the ith element is removed or added, the change
of mean compliance is equal to the strain energy of ith
element as(5) :
1
∆Ci = {ui }T [Ki ]{ui } (2)
2
where {ui } is the nodal displacement vector of the ith ele-
ment, [Ki ] is the element stiffness matrix. Therefore, the
sensitivity number of ith element can be defined by its
strain energy dividing its volume as Fig. 1 Extrapolation of sensitivity number

Series C, Vol. 49, No. 4, 2006 JSME International Journal

 1093

where N is the total node number of the candidate element. Using the finite element analysis, the performance in-
Thus, the sensitivity numbers for all elements within dex of the structure can infinitely approach the theoret-
the structure as well as possible added elements surround- ical one through the adjustment of the elements. When
ing the structure are used to compare. Elements in the the structure satisfies the convergence criterion defined as
current structure can therefore be removed if they satisfy Eq. (10), an optimum for the present removal ratio of vol-
Eq. (7.a) ume is assumed to be obtained.
|PIi − PIi−1 |
α∗e ≤ αth (7.a) errori = ≤ error (10)
PIi
For these elements surrounding the current structure will where errori is the defined performance error for the ith
be added if they satisfy Eq. (7.b) iteration and error is the maximum allowable error which
α∗e > αth (7.b) is specified by the user.
Sometimes, an oscillation state occurs when elements
where αth is the threshold of the sensitivity number which are added in an iteration and the same elements are re-
is determined by the current removal ratio of volume moved in the subsequent iteration. Thus, these elements
(RRVi ). For example, if there are 1 000 elements in de- will be added and removed in successive iterations. Nor-
sign domain and α1 > α2 ··· > α1 000 and RRVi corresponds mally, the occurrence of the oscillation state is caused by
to a design with 800 elements then αth = α800 . This new a low specified value of the maximum allowable error.
element removal and addition scheme ranks all elements When an oscillation occurs, the current design is assumed
(void and solid) together, while in the original BESO to be an optimum for the present removal ratio of volume
methods(14), (15) elements for removal and those for addi- although the convergence criterion has not been satisfied
tion are treated differently and ranked separated, which is yet.
a bit cumbersome and not very logical. The proposed iterative procedure is similar to the
The cycle of finite element analysis and element re- conventional optimality criteria approach where if only
moval and addition is repeated using the same value of marginal improvement in compliance over last design, the
RRVi until the solution is converged. Then an evolution- iterative procedure would be stopped. However, the de-
ary rate (ER) is introduced and added to the removal ratio signs of the conventional optimality criteria approach con-
of volume, i.e. tain elements of different densities (or thicknesses) rather
RRVi+1 = RRVi + ER, i = 0, 1, 2, 3··· (8) than elements of the same density in the BESO method.
With this increased removal ratio of volume, the cycle of 4. Bi-Directional Evolutionary Structural Optimiza-
finite element analysis and element removal and addition tion Procedure
takes place again until another optimum is obtained. The
To explain better how the bi-directional evolutionary
BESO procedure would be stopped when both the conver-
structural optimization can be implemented into a com-
gent criterion and final objective volume are satisfied.
puter program, a step-by-step algorithm is given as fol-
3. Performance Index (PI) and Convergence Crite- lows:
rion Step 1: The maximum design domain is discretized with
a densely finite element mesh.
Performance index (PI) can be introduced to identify Step 2: Define all boundary constraints, the design loads
the performance of various optimal designs. Thus, the and nonlinear material properties.
definition of the performance index should be directly re- Step 3: Carry out a linear finite element analysis of the
lated to the optimization objective. In the present case, structure.
the stiffness per unit volume or weight denotes the usage Step 4: Output the results of analysis into BESO pro-
efficiency of the material. So the performance index is de- gram.
fined by Step 5: Calculate the sensitivity number for all existed
1 elements as Eq. (3).
PI = (9)
CV Step 6: Calculate the sensitivity number for all existed
Where C and V are the mean compliance and volume nodes as Eq. (4).
(or weight) of the current design. While designs have a Step 7: Calculate the sensitivity number for all candidate
same volume, the one with a highest stiffness is the op- nodes surrounding the structure as Eq. (5).
timum which has the highest performance index. Thus, Step 8: Calculate the sensitivity number for all elements
for a given removal ratio of volume, the structure should as Eq. (6)
be gradually evolved to an optimum with a highest per- Step 9: Remove and add elements which satisfy
formance index by several removing and adding elements Eqs. (7.a) and (7.b).
process. Step 10: If the convergence condition Eq. (10) is satis-

JSME International Journal Series C, Vol. 49, No. 4, 2006

1094

fied, an optimum is reached and the optimiza- nodes. All topologies shown in the following sections are
tion process skips Step 11. Otherwise, go to next smoothed.
step.
5. Examples and Discussion
Step 11: If an oscillatory state is reached when a group
of elements are removed and added back to the 5. 1 Example 1: Michell type structure
structure in the successive iterations. The opti- The classic Michell type structure is fixed at both sup-
mum is reached. Otherwise, repeats step (3) to ports as shown in Fig. 3. The dimensions of the rectangu-
(11) until an optimum is reached. lar domain are 0.2 m by 0.1 m. The thickness of the plate
Step 12: For the optimization process to continue, the re- is 0.001 m. A 100 N concentrated force is applied at the
moval ratio of volume is increased by Eq. (10). center of bottom edge. The linear material is used with
Repeat step (3) to (10) until another optimum is Young’s modulus E = 210 GPa and Poisson’s ratio ν = 0.3.
reached. Because of symmetry, only half the structure is modeled
Figure 2 shows the flow chart which defines the logical with 100×100 four node elements. The BESO parameters
steps for the new BESO method. are: RRV0 = 0.01, ER = 0.01 and error = 0.001.
In the above procedure, the convergent optimums cor-
responding to various intermediate volumes are found. If
only one optimum corresponding to the final volume, V ∗ ,
is requested, the convergence condition may be checked
only after the objective volume V ∗ is satisfied in step 10.
In this way, the computation time can be saved signifi-
cantly.
In order to present the final optimal topology with
manufacturable boundary, an intuitive smoothing tech-
nique is applied. The coordinates of every node on the
boundary are averaged by the coordinates of neighbor

Fig. 3 Design domain of a Michell type structure

(a) (b)

(c) (d)

(e) (f)

Fig. 4 Optimal designs with various remained material ratio

and performance index: (a) V f = 80% and PI =
2.93 × 10−4 N−1 mm−4 ; (b) V f = 60%, PI = 3.51 ×
10−4 N−1 mm−4 ; (c) V f = 40% and PI = 4.26 ×
10−4 N−1 mm−4 ; (d) V f = 30% and PI = 4.61 ×
10−4 N−1 mm−4 ; (e) V f = 20% and PI = 5.08 ×
10−4 N−1 mm−4 ; (f) V f = 10% and PI = 5.31 ×
Fig. 2 Flowchart depicting logical steps of the BESO method 10−4 N−1 mm−4

Series C, Vol. 49, No. 4, 2006 JSME International Journal

 1095

(a) (b)

(c) (d)

Fig. 6 Evolutionary history of the structural topology starting

from the initial guess design with V f = 40%: (a) initial
guess design; (b) design after iteration 20; (c) design
Fig. 5 Evolutionary history of the performance indexes for all after iteration 60; (d) final optimal design
optimal designs

The evolution history of the structure is shown in

Fig. 4 (a) – (f) with various remained material volume, V f .
All structures consist of an arch and several spokes be-
tween the load and the top of the arch. However, the num-
ber and arrangement of spokes would be adjusted with the
remained material volume. Figure 5 shows the history
of the performance index of optimums during the whole
evolution processes. The performance index increases
continuously until the final remained volume/weight (say
V f = 12%) arrives. The highest performance index of the
design is 5.39 × 10−4 N−1 mm−4 . Thus, the performance of
the design improves significantly comparing with the ini- Fig. 7 Evolutionary history of the performance index and
structure volume/weight
tial full design with PI = 2.39 × 10−4 N−1 mm−4 .
In this way, all optimal designs with various remained
material ratios are obtained. Sometimes, only one op-
timum with a certain volume/weight constraint, V ∗ , is
(a) (b)
needed. The convergence condition may be checked only
after the objective volume V ∗ is satisfied. In this way, the
computation time can be saved significantly. To further
save the computation time, we may start the BESO method
from the initial guess design. In the following sections,
(c) (d)
we try two initial guess designs: one design is satisfied
the volume/weight constraint and another is a minimum Fig. 8 Evolutionary history of the structural topology starting
possible design. from a minimum design space: (a) initial minimum
The initial guess design with only V f = 40% ma- design; (b) design after iteration 15; (c) design after
iteration 60; (d) final optimal design
terial of the design domain is shown in Fig. 6 (a). The
BESO parameters are set to RRV0 = 0.60, ER = 0.0 and
error = 0.001. Thus, the location of elements is ad- and its results are more accurate than the BESO method
justed step by step and Fig. 6 (a) – (d) shows the history starting from an initial guess design whose topology is to-
of the evolving process. The topology of the optimum tally different from the optimum. It can be believed that
(Fig. 6 (d)) is very close to the Fig. 4 (c). Its performance much closer topology and performance index would be
index is 4.23 × 10−4 N−1 mm−4 which is a bit lower than obtain if a smaller convergence error is specified in the
the performance index of Fig. 4 (c), 4.26×10−4 N−1 mm−4 . later cases. Figure 7 shows the evolutionary histories of
These small differences can be attributed to the assigned the performance index and structural volume/weight. It
convergence error in the later case. In the BESO starting can be seen that the performance index increases steadily
from the full initial design, the topology after every itera- and the final optimum possesses the highest performance
tion is very close to the next optimum. Thus its calculated index in the whole optimization process.
convergence error is much lower than the assigned value Next we conduct the BESO method starting from the

JSME International Journal Series C, Vol. 49, No. 4, 2006

1096

Fig. 10 Design domain and support conditions of a 3D

cantilever beam

Fig. 9 Evolutionary history of the performance index and

structure volume/weight

initial minimum design as shown in Fig. 8 (a). The vol-

ume/weight constraint is set to the 20% of the full design
domain. The BESO parameters are RRV0 = 0.80, ER = 0.0
and error = 0.001. Figure 8 (a) – (d) shows the evolv-
(a) (b)
ing topology history using the present BESO method.
Figure 9 shows the histories of the performance index and
structural volume/weight. It can be seen that elements
are added to the structure from 1st iteration to 16th iter-
ation. Then, locations of elements are adjusted because
the volume/weight constraint is satisfied. The structure is
evolved to the optimum (95th iteration) when the specified
convergence error is satisfied. The topology of the opti- (c) (d)
mum (Fig. 8 (d)) is similar to the optimum obtained by the
BESO starting from the initial full design (Fig. 4 (e)). The
performance index of the optimum is 5.07×10−4 N−1 mm−4
which is also close to the performance index of Fig. 4 (e),
5.07 × 10−4 N−1 mm−4 .
This well-known topology design problem has been
widely reported in the literature. The above optimal
topologies show comparable shapes with those gener- (e) (f)
ated with the original ESO method (Fig. 2.8 in Ref. (5)) Fig. 11 Optimal designs with various remained material ratio
and BESO method (Fig. 9 in Ref. (16)) as well as those and performance index: (a) V f = 80% and PI =
proposed by classical methods like Michell’s layout the- 4.44 × 10−7 N−1 mm−4 ; (b) V f = 60%, PI = 5.87 ×
ory(18) . However, the present BESO method starting from 10−7 N−1 mm−4 ; (c) V f = 40% and PI = 48.63 ×
initial guess design is more efficient than the original 10−7 N−1 mm−4 ; (d) V f = 30% and PI = 1.12 ×
10−4 N−1 mm−6 ; (e) V f = 20% and PI = 1.59 ×
BESO method because only small portion of elements (no
10−6 N−1 mm−4 ; (f) V f = 10% and PI = 2.66 ×
more than the objective volume due to ER = 0.0) is calcu- 10−6 N−1 mm−4
late in the finite element analysis.
5. 2 Example 2: a 3D cantilever
Figure 10 shows the dimensions of the initial design mained volume/weight ratios are shown in Fig. 11 (a) –
domain and supports for a 3D cantilever. The applied (f). Figure 12 shows the performance index of all opti-
force is 100 N downward at the middle of the free end. mal designs. The performance index increases gradually
The mechanical properties of the material are Young mod- as the material is removed from the design domain step by
ulus E = 1 GPa and Poisson ratio v = 0.33. The maximum step. The performance index of the optimal design with
design domain is meshed with 64 000(80 × 40 × 20) eight V f = 10% is about 7.5 times than that of the initial full
node solid elements. design.
Firstly, we start BESO from the full design. The As explained in the Example 1, the BESO method
BESO parameters are set to be RRV0 = 0.01, ER = 0.01 may starts from any initial guess design with any mate-
and error = 0.001. Optimal designs with various re- rial in the design domain (without singularity). For exam-

Series C, Vol. 49, No. 4, 2006 JSME International Journal

 1097

Fig. 14 Evolutionary history of the performance index and

structure volume/weight
Fig. 12 Evolutionary history of the performance indexes for all
optimal designs

Fig. 15 Design domain for a bridge, where the solid section

(a) (b) denotes the non-designable material

However, the BESO starting from the initial guess design

saved the computation time significantly comparing with
the BESO starting from the full design, although the per-
formance index of the present design is little lower than
that of the later method.
5. 3 Example 3: an optimal bridge
(c) (d)
Optimization for bridge type structure having the road
Fig. 13 Evolutionary history of the structural topology starting on its upper part has been studied by Cui et al.(10) The di-
from the initial guess design: (a) initial guess design; mensions of the design domain, the loading and support-
(b) design after iteration 10; (c) design after iteration
ing conditions are given in Fig. 15. The mechanical prop-
15; (d) final optimal design
erties of the material are assumed to be E = 210 GPa, and
v = 0.3. In order to assure that the pressure is applied on
ple, the initial design is used as shown in Fig. 13 (a). The the full top surface, a non-designable layer is defined with
BESO parameters are defined with RRV0 = 0.8, ER = 0.0 a thickness 1.5 m as shown in Fig. 15 with solid section.
and error = 0.000 1. Figure 13 shows the evolutionary his- The design objective is to maximum the stiffness subject
tory of the structural topology. The final optimal design to the remained material ratio V f = 20% comparing with
with V f = 20% is shown in Fig. 13 (d). The evolutionary the designable domain.
histories of the material volume and performance index Due to symmetry, only quarter of the structure is
are shown in Fig. 14. The material is added to the struc- modeled with 140×10×40 eight node solid elements. The
ture until the specified V f = 20% is satisfied. Then loca- BESO method starts from the minimum design as shown
tions of elements are adjusted step by step and the perfor- in Fig. 16 (a). The BESO parameters are RRV0 = 0.8, ER =
mance index is improved gradually. Eventually, the struc- 0.0 and error = 0.000 1. After 121 iterations, the conver-
ture evolves to an optimum with the highest performance gence criterion is satisfied and the evolutionary history of
index. the structural topology is shown in Fig. 16 and Fig. 16 (d)
Comparing with two results using different initial de- and (e) shows the final optimal topology. Figure 17 shows
signs, the optimal topologies are globally similar to each the history of the performance index. It can be seen that
other (Figs. 11 (e) and 13 (d)). Moreover, their perfor- the performance index has a big drop when one of sup-
mance index are very close (PI = 1.59×10−6 N−1 mm−4 for ports breaks abruptly. Then a new structural topology is
Fig. 11 (e) and PI = 1.58 × 10−6 N−1 mm−4 for Fig. 13 (d)). formed and the performance index increases steadily until

JSME International Journal Series C, Vol. 49, No. 4, 2006

1098

(a) (a)

(b) (b)

(c) (c)

(d) (d)

(e) (e)
Fig. 16 Evolutionary history of the structural topology starting Fig. 19 Evolutionary history of the structural topology starting
from the initial guess design: (a) initial design; (b) from the initial guess design: (a) design after iteration
design after iteration 40; (c) design after iteration 80; 20; (b) design after iteration 40; (c) design after
(d) final optimal design; (e) rendered optimal design iteration 60; (d) final optimal design; (e) rendered
optimal design

Fig. 17 Evolutionary history of the performance index and

structure volume/weight
Fig. 20 Evolutionary history of the performance index and
structure volume/weight

as shown in Fig. 18, the BESO method starting from the

minimum design as shown in Fig. 16 (a) is applied. The
BESO parameters are RRV0 = 0.8, ER = 0.0 and error =
0.000 1 which are identical to the above case. Figure 19
Fig. 18 Design domain for a revised bridge shows the evolutionary history of the structural topology.
After 74 iterations, the final optimum is obtained as shown
the next support breaks. After the several adjustments of in Fig. 19 (d) and (e) which differs from the Fig. 16 (d) and
the supports, the final optimum is obtained and possesses (e). Figure 20 shows the performance index history during
the highest the performance index, 5.9 × 10−4 N−1 m−4 . the optimization process. The performance index of the
If we apply a constraint on the initial design domain final optimum is 3.51 × 10−4 N−1 m−4 which is lower than

Series C, Vol. 49, No. 4, 2006 JSME International Journal

 1099

that of the previous case. It is reasonable because there is (5) Xie, Y.M. and Steven, G.P., Evolutionary Structural
more constraint on the design domain in the later case. Optimization, (1997), Springer, London.
Above two examples also demonstrate the efficiency (6) Tanskanen, P., The Evolutionary Structural Optimiza-
of the present BESO method because the elements in- tion Method: Theoretical Aspects, Comput. Methods
Appl. Mech. Engrg., Vol.191 (2002), pp.5485–5498.
volved in finite element analysis is no more than 20% of
(7) Chu, D.N., Xie, Y.M., Hira, A. and Steven, G.P.,
total elements in the design domain. The convergence cri- Evolutionary Structural Optimization for Problems
terion used here is very strict, ensuring that the BESO al- with Stiffness Constraints, Finite Element Anal. Des.,
gorithm really has converged to an optimum. The designs Vol.21 (1996), pp.239–251.
with the performance index just a few percent above the (8) Xie, Y.M. and Steven, G.P., Evolutionary Structural
“optimal” one may be obtained using a large error such Optimization for Dynamic Problems, Comput. Struct.,
as 0.001 in approximately 50 iterations. Vol.58 (1996), pp.1067–1073.
(9) Manickarajah, D., Xie, Y.M. and Steven, G.P., An
6. Conclusions Evolutionary Method for Optimization of Plate Buck-
ling Resistance, Finite Elements in Anal. Des., Vol.29,
In this paper, a new algorithm for the bi-directional No.3–4 (1998), pp.205–230.
evolutionary structural optimization (BESO) has been de- (10) Cui, C., Ohmori, H. and Sasaki, M., Computational
veloped. Several 2D and 3D examples have demonstrated Morphogenesis of 3D Structures by Extended ESO
the capability of the new BESO method to optimize a Method, Journal of the International Association for
structure by adding and removing material. The devel- Shell and Spatial Structures: IASS, Vol.44 (2003),
oped BESO method has several advantages over tradi- pp.51–61.
tional ESO and BESO methods: (11) Ohmori, H., Futai, H., Iijima, T., Muto, A. and
Hasegawa, Y., Application of Computational Morpho-
1. The adding and removing of material is controlled
genesis to Structural Design, Proceeding of Computa-
by a single parameter: the removal ratio of volume (or tional Science Symposium, Nagoya University, Japan,
weight ratio). October 11–13, (2005), pp.45–52.
2. The final design is independent of the initial de- (12) Mitsui, K. and Sogabe, H., Heuristic Method for Topol-
sign, making the optimization algorithm more robust. The ogy and Shape Optimization of Structure Changing Pe-
final topology and performance index are very close even riodically, Proceeding of Computational Science Sym-
when the initial designs are totally different. posium, Nagoya University, Japan, October 11–13,
3. The performance index provides a clear indication (2005), pp.53–58.
(13) Kita, E. and Toyoda, T., Structural Design Using Cellu-
of structural efficiency of the resulting topologies and an
lar Automata, Structural Optimization, Vol.19 (2000),
effective termination criterion for the optimization proc- pp.64–73.
ess. (14) Querin, O.M., Steven, G.P. and Xie, Y.M., Evolution-
4. The proposed BESO method starting from initial ary Structural Optimisation by an Additive Algorithm,
guess design is more efficient than the original BESO Finite Elements Anal. Design, Vol.34 (2000), pp.291–
method because only small portion of elements which is 308.
no more than the objective volume is calculated in the fi- (15) Yang, X.Y., Xie, Y.M., Steven, G.P. and Querin,
nite element analysis. O.M., Bidirectional Evolutionary Method for Stiffness
Optimization, AIAA Journal, Vol.37, No.11 (1999),
References pp.1483–1488.
(16) Querin, O.M., Young, V., Steven, G.P. and Xie,
(1) Bendsøe, M.P. and Kikuchi, N., Generating Optimal Y.M., Computational Efficiency and Validation of
Topologies in Structural Design Using a Homogeniza- Bi-Directional Evolutionary Structural Optimisation,
tion Method, Comput. Methods Appl. Mech. Eng., Comput. Methods Appl. Mech. Engrg., Vol.189 (2000),
Vol.71 (1988), pp.197–224. pp.559–573.
(2) Bendsøe, M.P. and Sigmund, O., Topology Opti- (17) Li, Q., Steven, G.P. and Xie, Y.M., A Simple Checker-
mization: Theory, Methods and Applications, (2003), board Suppression Algorithm for Evolutionary Struc-
Springer-Verlag, Berlin Heidelberg. tural Optimization, Struct. Multidisc. Optim., Vol.22
(3) Sigmund, O., A 99 Line Topology Optimization Code (2001), pp.230–239.
Written in MATLAB, Struct. Multidisc. Optim., Vol.21 (18) Michell, A.G.M., The Limits of Economy of Material
(2001), pp.120–127. in Frame Structures, Phil. Mag., Vol.8 (1904), pp.589–
(4) Xie, Y.M. and Steven, G.P., A Simple Evolution- 597.
ary Procedure for Structural Optimization, Comput.
Struct., Vol.49 (1993), pp.885–896.

JSME International Journal Series C, Vol. 49, No. 4, 2006