You are on page 1of 12

Compurers & Srrvcrures Vol. 49, No. 5, pp. W-896.

1993
0 1994 Elsevier Science Ltd
Pergamon Printed in Great Britain.
0045-7949/93 $6.00 + 0.00

TECHNICAL NOTE

A SIMPLE EVOLUTIONARY PROCEDURE FOR


STRUCTURAL OPTIMIZATION
Y. M. XIE and G. P. STEVEN
Finite Element Analysis Research Centre, Engineering Faculty, University of Sydney, NSW 2006, Australia

(Received 6 July 1992)

Abstract-A simple evolutionary procedure is proposed for shape and layout optimization of structures.
During the evolution process low stressed material is progressively eliminated from the structure. Various
examples are presented to illustrate the optimum structural shapes and layouts achieved by such a
procedure. zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

1. INTRODUCITON a rejection ratio (RR) times the maximum von Mises stress
over the structure.
Structural optimization, in particular the layout optimiz- Such a finite element analysis and element elimination
ation, has been identified as the most challenging and cycle is repeated using the same RR until a steady state is
economically the most rewarding task in structural design reached. At this stage an evolution rate (ER ) is introduced
[I]. Traditionally solutions to structural optimization prob- and added to the RR. The iteration takes place again until
lems have been sought by using various mathematical a new steady state is reached.
programming techniques. However, mathematical program- This evolutionary process continues until a desired opti-
ming has been demonstrated to be inefficient and often mum is reached, for example, until all stress levels are within
unsuitable for tackling layout optimization problems and to 25% of the maximum. It might not be the absolute best
overcome these difficulties optimality criteria and optimal result but such an evolutionary optimization procedure
layout theories have been developed [2]. offers the possibility of knowing every stage of the shape and
Recently Bendsoe and Kikuchi [3] developed the so-called layout path towards the true optimum.
homogenization method in which a material model with Although the proposed procedure appears to be trivial or
infinitely many microscale voids is introduced and the primitive, we shall show in the next section the encouraging
layout optimization problem is defined by seeking the results achieved by this simple procedure. In various test
optimal porosity of such a porous medium using one of the cases, the results agree remarkably well with solutions
optimahty criteria. obtained by analytical and other mathematically rigorous
By observing the evolution of naturally occurring struc- procedures.
tures such as shells, bones and trees it becomes obvious that
the topology and shape of such structures achieve their
3. EXAMPLES
optimum over a long evolutionary period and adapt to
whatever environment they find themselves in. This paper is Example 1. Metal implant in bone
to demonstrate the possibilities of achieving similar struc-
An example of the natural evolutionary structural optim-
tural shape and layout optimization by using the finite
ization occurs in metal implants to repair bone breakage. It
element analysis and training the software to follow a
has been noticed that after the repair, if any local bone does
particular evolutionary path.
not have sufficient stress applied that part of bone dimin-
ishes as calcium is absorbed into the bloodstream.
2. A SIMPLE EVOLUTIONARY PROCEDURE
This evolutionary process is simulated with the model in
Fig. 1 using 82 x 20 bilinear quadrilateral plane stress
After finite element analysis the stress distribution elements. The ‘bone’, which is broken into two parts, is
throughout the structure is found. Often it happens that connected by the ‘metal’ in the middle. The Young’s modu-
part of the material is not effectively used. Using some lus for the metal is assumed to be 10 times as much as that
criterion for rejection, here called a rejection criterion (RC), of the bone and the system is stretched at both zyxwvutsrqponm
e nds.
such as the von Mises stress, this unneeded part of the The initial rejection ratio R& = 1% is assumed and the
material can be eliminated. For example, elements are evolution rate ER is taken as 1% Figures 2(a)(j) show the
deleted where the von Mises stress in the element is less than eroded shape of the bone at various stages. These pictures

Fig. 1. Model of metal implant in broken bone.

885
886 Technical Note

(a> RR = 1%

RR = 2%

(c>

(e)

Fig. 2(a-f).
Technical Note 887

RR = 7%

--

00 RR = 8%

(i) RR = 9%

Cj)

Fig. 2. Evolutionary history of bone erosion. (Initial rejection ratio RR, = l%, evolution rate ER = I%.)

provide a clear view of the evolutionary history of the bone


erosion. On average seven iterations are needed to achieve
a steady state for each RR increment.
Example 2. Optimum two-bar frame
As a verification of the present evolutionary procedure a
simple problem whose solution may be obtained analytically
is considered. It is a two-bar frame shown in Fig. 3. For
fixed values of the applied load F and the horizontal length
24

Fig. 3. A two-bar frame structure. Fig. 4. Design domain for the two-bar frame structure.
888 Technical Note

RR= 3% (b) RR = 6%

RR = 12%
RR = 9% (4

RR= 18%
RR = 15% (0

........................
........................
........................
........................
f ........................

Fig. 5(a-f ).
Technical Note 889

RR = 30%
RR = 21% zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
(J)
(9

tt .......................
.......................

Fig. 5. Evolutionary path towards the optimum two-bar frame structure. (Initial rejection ratio RR, = I %,
evolution rate ER = 1%. Figure S(j) gives H = 2L, which agrees exactly with the analytical solution.)

L, the optimal height is obtained as H = 2L, if this structure RR, = I%, taking the evolution rate ER as 1%. Figures
is assumed to be a truss. 5(a)< j) show the different stages of the evolution. After the
To get such an optimum structural layout, a model shown RR reaches 30% a perfect two-bar system is achieved. The
in Fig. 4 is used where the rectangular ‘design domain’ is final solution in Fig. 5(j) gives H = 2L, which is exactly the
larger than the size L x 2L (L = 10 m). It is discretized using same as the analytical solution.
25 x 60 bilinear quadrilateral plane stress elements. Young’s On average eight iterations are needed to converge to a
modulus E = 100 GPa and Poisson’s ratio v = 0.3 are as- steady state at each RR increment.
sumed. The same problem has been modelled by Suzuki and
Kikuchi [4] using the homogenization method with 40 x 96
elements.
We start the evolution with the initial rejection ratio L=2H d
t-

_d- F
1F
A-
Fig. 7. Design domain for the Michell type structure.
H

Fig. 6. A Michell type structure.


890

RR = 2.5%
(a)

..............
..............
..............

RR = 5%

............... ...
pu

RR = 7.5%
(c>

............ .........................

..............

RR = 10%

RR = 12.5%

Fig. 8(a e)
Technical Note 891

RR = 15%

................
. ... ... . . .... ...... ....... . .
...............

......................... ................
................................................ ....
...................................... .......
.,.,........,.....
..................................... ............
................
. ...............

fg) RR = 17.5%

.................
.................

.................

.................

RR = 20%

0) RR = 22.5%

Cj> RR = 25%

Fig. 8. Evolutionary path towards the optimum Michell type structure. (Initial rejection ratio RR, = 1%,
evolution rate ER = 0.5%. Figure 8(j) gives B = x/4, which agrees exactly with the analytical solution.)
892 Technical Note zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPO

E.xample 3. A Michell type structure

The second verification of the present evolutionary pro-


cedure of layout optimization is whether it can reproduce a
Michell type structure shown in Fig. 6, which has been
solved analytically as a typical problem of structural layout
optimization [5].
A design domain of the size 2H x H (H = 5 m) as shown
in Fig. 7 is discretized into 50 x 25 bilinear quadrilateral
plane stress elements. The two corners at the bottom are
assumed fixed and the load is applied at the middle of the
bottom. Young’s modulus E = 100 GPa and Poisson’s ratio
v = 0.3 are assumed. The initial rejection ratio RR, is equal
to 1% and the evolution rate ER is 0.5%.
F
Figures 8(a)(j) show the evolutionary history of this
Fig. 9. Change of support at the bottom right corner.
model from a rectangular plate into the final truss type

(4 RR = 2.5%

RR = 5%

RR = 7.5%
Cc)

RR = 10%

Fig. lO(a4).
(e) zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
RR = 12.5%

RR = 15%

w RR = 17.5%

RR = 20%

...................................................

(9 RR = 22.5%

............
............ ............

...................................................
.................
...............

RR = 25%
0)

RR, = I%,
Fig. 10. A different e volutionary history due to changing support. (Initia ~1 r,ejection ratio
evolution rate ER = O.S%.)
893
Technical Note

structure, which is similar to the analytical solution in


Fig. 3(a). It is worth pointing out that the angle 0 in Fig. 8(j)
is equal to 45”. which is exactly the same as in the analytical
solution.
On average four iterations are needed to achieve a steady
state for each RR increment.

E.xample 4. Chutqy of’ supporr


To examine the effect of changing support of a structure,
the model in Example 3 is now reanalysed with the support
at the bottom right changed from fixed to rolling condition
as shown in Fig. Y.
Keeping all the other parameters the same as in Example
Fig. 11. Design domain for a short cantilever. 3, we now get a diRerent evolutionary history as shown in

RR = 11%
(4

RR = 12%

(cl RR = 13%

RR = 14%

Fig. I2(a~ d)
Technical Note 895 zyxwvutsr

......................
......................
......................

RR = 17%
w

RR = 18%
O-4 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

Fig. 12. Evolutionary path to the optimum configuration. (Initial rejection ratio RR, = lo%, evolution
rate ER = l%.)

Figs lO(aHj). There are two extra horizontal bars present and extra material is needed at joints to support
at the bottom which obviously are in tension and bending moments.
pulling the bottom right corner from moving further to the
right.
Example 5. A short ~~niil~er
On average four iterations are needed to converge to a
steady state at each RR increment. Next let us apply the optimization procedure to a short
It is noticed from results of both Example 3 and Example cantilever subject to the vertical force at the free end, see
4 that there are a cluster of elements at the location where Fig. 11. The rectangular design domain is discretized into
the load is applied, which are in contrast to solutions from 64 x 40 quadrilateral bilinear plane stress elements. Young’s
truss theory in which the pin-joint condition is assumed. modulus E = 100 GPa and Poisson’s ratio Y = 0.3 are
This is consistent with the modelling whereby rigid joints are assumed.
896 Technical Note

Figures lZ(aHh) show the evolutionary path towards the iteration cycles automatically. At the end of each
the optimum configuration. The final solution shown iteration or each RR increment, a picture of the remaining
in Fig. 12(h) is very similar to what Suzuki and elements is stored. When the whole solution is finished, these
Kikuchi [4] obtained using a totally different approach of pictures are shown in a sequence, typically for 2 seconds
the homogenization method. each frame. This has resulted in a collection of interesting
On average ten iterations are needed to achieve a steady videos of the evolutionary history of various structures.
state for each RR increment. To speed up this optimization procedure two possible
improvements are currently under investigation. One is to
use the similarity of the element stiffness matrices when
4. DISCUSSION identical quadrilateral or brick elements are zyxwvutsrqponml
use d to dis-
cretized the whole or most part of the design domain. The
The examples shown here are all plane stress problems. other is to employ reanalysis techniques, see for example 171,
However, it is obvious that the proposed approach applies to avoid assemble and decompose the global stiffness matrix
equally to general two- and three-dimensional problems. at every iteration.
The only change in the existing finite element analysis codes It has been shown that the present evolutionary procedure
is adding a few lines regarding comparing the element von is able to give the design engineer at least a rough guide to
Mises stress with the maximum von Mises stress and then the optimum layout and shape of a structure under certain
rejecting the low stressed elements. loading conditions. The final design of a structure can be
During the evolutionary process, it is not necessary to improved by using existing shape optimization codes which,
generate a new mesh. Instead we assign the material prop- although very limited in the capability of changing topology
erty number of the rejected elements to zero and ignore these of a structure, can provide details of the shape and size of
elements when the globle stiffness matrix is assembled in the a structure.
following solutions. Thus as more and more elements are
eliminated, the number of equations becomes less and less. REFERENCES
Typically it takes 5-8 iterations to reach a steady state at
each value of the rejection ratio. We have found that the I. G. I. N. Rozvany, Historical review of layout optimiz-
initial rejection ratio RR,, must be small, say 1% and that ation. In Structural Optimization, Proceedings of the
the evolution rate must also be small, again 1% is typical. IVTAM Symposium on Structural Optimization (Edited
If the evolution rate ER is too high then over rejection by G. I. N. Rozvany and B. L. Karihaloo), Melbourne,
occurs and the structure becomes singular. When this 9-13 Feb. 1988. Kluwer, Dordrecht (1988).
happens, it is necessary that the software steps back and 2. G. I. N. Rozvany, Structural Design via Optimality
starts off the current evolutionary cycle with a halved ER to Criteria. Kluwer, Dordrecht (1988).
try again. 3. M. P. Bendsae and N. Kikuchi, Generating optimal
It is noted that most of the examples have reached close topologies in structural design using a homogenization
to their optimum configurations with the RR round about method. Comput. Meth. appl. Mech. Engng. 71,197-224
25%. This indicates that there is still a wide range of stress (1988).
between the lowest and the highest stressed elements using 4. K. Suzuki and N. Kikuchi, A homogenization method
the element von Mises stress compared with the maximum for shape and topology optimization. Comput. Meth.
von Mises stress as the rejection criterion. Because of the appl. Mech. Engng. 93, 291-381 (1991).
presence of rigid joints and fixed support joints there can be 5. W. S. Hemp, Michell’s structural continua. In Opti-
significant stress raises at these locations which can explain mum Structures, Chap. 4. Clarendon Press, Oxford
the low terminal RR value. It is a possible further develop- (1973).
ment to compare the element von Mises stress with its 6. STRAND6 Reference Manual and User Guide. G + D
average over all the elements as an alternative rejection Computing, Sydney (1991).
criterion. 7. B. H. V. Topping, Exact and approximate static struc-
It can take nearly 100 iteration to reach a suitable final tural re-analysis. In CISM Advanced Course on Shape
optimum configuration. We currently use STRAND6 [6] and Lay-out Optimization of Structural Systems. Udine
software which allows the user to set up batch files to handle (1990).

You might also like