Professional Documents
Culture Documents
www.elsevier.com/locate/engstruct
Abstract
This paper presents topology optimization of geometrically and materially nonlinear structures under displacement loading. A revised bidirectional evolutionary optimization (BESO) method is used. The objective of the optimization problem is to maximize the structural stiffness
within the limit of prescribed design displacement. The corresponding sensitivity number is derived using the adjoint method. The original BESO
technique has been extended and modified to improve the robustness of the method. The revised BESO method includes a filter scheme, an
improved sensitivity analysis using the sensitivity history and a new procedure for removing and adding material. The results show that the
developed BESO method provides convergent and mesh-independent solutions for linear optimization problems. When the BESO method is
applied to nonlinear structures, much improved designs can be efficiently obtained although the solution may oscillate between designs of two
different deformation modes. Detailed comparison shows that the nonlinear designs are always better than the linear ones in terms of total energy.
The optimization method proposed in this paper can be directly applied to the design of energy absorption devices and structures.
c 2008 Elsevier Ltd. All rights reserved.
Keywords: Bi-directional evolutionary structural optimization (BESO); Geometrical nonlinearity; Material nonlinearity; Nonlinear finite element analysis; Energy
absorption
1. Introduction
Topology optimization methods enable designers to find the
best structural layout for a required structural performance.
Over the last three decades, many mathematical and heuristic
optimization methods have been developed [14]. So far,
most papers dealing with the optimization method have been
concerned with the optimization of structures with linear
material and geometric deformation behaviour. However, the
linear assumptions are not always valid, such as in the cases
of compliant design and energy absorption design. Using the
sensitivity/gradient based optimization methods, a number of
papers have considered topology optimization of geometrically
nonlinear structures [58]. Topology optimization of materially
nonlinear structures has also been conducted by several
researchers [912]. Jung and Gea [13] studied topology
optimization of both geometrically and materially nonlinear
Corresponding author. Tel.: +61 3 99253320; fax: +61 3 96390138.
2058
Fig. 1. (a) Typical load-displacement curve in nonlinear analysis and objective function of the optimization problem; (b) Sensitivity analysis of nonlinear structures
under displacement loading.
from the structure at the same time as the efficient ones are
added. However, the original BESO method may sometimes
lead to non-convergent and mesh-dependent solutions. Thus, it
is less robust than other mathematical optimization methods [2].
The BESO method with perimeter control [16] is shown
to be capable of obtaining somewhat convergent and meshindependent solutions because of one extra constraint (the
perimeter length) on the topology optimization problem.
However, predicting the value of the perimeter constraint
for a new design problem is a difficult task. It follows that
the original BESO method should be modified to deal with
aforementioned problems.
Consider a typical nonlinear load-displacement curve as
depicted in Fig. 1(a). If the structure is physically loaded
under force control, the displacement at the horizontal limit
point snaps through along the dashed line. Alternatively, the
trajectory due to displacement control can follow the whole
equilibrium path. Moreover, the displacement control is more
practical in applications such as the crushing distance in energy
absorption structures. So the objective of this paper deals with
optimization problems of this type of displacementloaded
nonlinear structures using a developed BESO method. To
maximize the structural stiffness within the prescribed design
displacement as shown in Fig. 1, the external work is selected
as the objective function. The sensitivity number is formulated
with energy-based functions using the adjoint variable method.
Several examples are presented to verify the proposed method.
Maximize
f (x) = W
"
= lim
#
n
1X
T
UiT Ui1
(Fi + Fi1 )
2 i=1
Subject to g = V
M
X
Vj x j = 0
(1a)
(1b)
j=1
x j {0, 1}
(1c)
(2)
M Z
X
CeT B dv =
M
X
CeT Fe
(3)
2059
3. Sensitivity number
Consider a displacementload nonlinear structure, the
displacement is increased step by step up to the prescribed
displacement Un as shown in Fig. 1(b). The sensitivity of the
objective function with respect to a change in design variable x
is
"
n
dF
d f (x)
1X
dFi1
i
T
UiT Ui1
= lim
+
n 2
dx
dx
dx
i=1
!
#
T
n
dUi1
dUiT
1X
+
(Fi + Fi1 ) .
(4)
2 i=1 dx
dx
It is noted that the second term must be zero because the
dUT
dUT
variation of displacement dxi = 0 and dxi1 = 0 at controlled
freedoms and the external forces Fi = 0 and Fi1 = 0 at
other freedoms. An adjoint equation is introduced by adding a
series of vectors of Lagrangian multipliers i into the objective
function as
n h
1X
T
f (x) = lim
(Fi + Fi1 )
UiT Ui1
n 2
i=1
i
iT (Ri + Ri1 ) .
(5)
Because Ri and Ri1 are equal to zero, the modified objective
function (5) is same as the original objective function (1a).
Thus, the sensitivity of the modified objective function is
n
dF
1X
dFi1
d f (x)
i
T
T
= lim
+
Ui Ui1
n 2
dx
dx
dx
i=1
dRi1
dRi
+
.
(6)
iT
dx
dx
Within a small loading step, a linear forcedisplacement
relationship can be assumed:
Fi Fi1 = Kit (Ui Ui1 ).
dFi
dFi1
Fi
Fi1
+
=
+
.
dx
dx
x
x
Similarly, for the residual force we have
(8b)
dRi
dRi1
Ri
Ri1
+
=
+
.
(9b)
dx
dx
x
x
Substituting the above equations into Eq. (6) and utilizing Eq.
(2), Eq. (6) can be further simplified as
"
n
F
d f (x)
1X
Fi1
i
T
T
= lim
Ui Ui1
+
n 2
dx
x
x
i=1
(10)
+ Fi1
x , adjoint variable
(11)
To verify the above sensitivity analysis equation, a simple onedimensional system is examined in the Appendix.
In the evolutionary structural optimization method, a
structure can be optimized by removing and adding elements.
That is to say that, the element itself is treated as the design
variable. Thus, when one element is totally removed from the
system, the variation of the objective function is
#
"
n
1X
int
T
T
int
Ui Ui1 1Fi + 1Fi1 .
1 f (x) = lim
n 2
i=1
(13)
From Eqs. (2) and (3), the variation of internal force due to
removing one element can be written as
1Fiint = CeT Fie .
(14)
n
X
e
E ie E i1
= E ne
(15)
i=1
(9a)
(7)
(8a)
Fi
x
!#
i = Ui Ui1 .
= lim
Fint
Fint
Fi
Fi1
i +
i1
x
x
x
x
iT
(16)
2060
i =
N
P
(17)
(ri j )
J =1
(i = 1, 2, 3 )
(20a)
(20b)
(21a)
(ri j )i
j=1
th
i > add
(21b)
2061
(22)
N
P C
Wi j+1
j=1
Fig. 2. Design domain, initial guess design and support conditions for a beam.
6. Examples
6.1. Examples for a small design displacement
A beam with 400 mm long and 100 mm high is simply
supported at bottom corners as shown in Fig. 2. A small
displacement d = 5 mm is applied at the center of bottom
edge. It is assumed that the available material can only cover
30% volume of the design domain, and material has Youngs
modulus E = 200 GPa, Poissons ratio v = 0.3, yield stress
y = 300 MPa and plastic hardening modulus E p = 0.3E.
To save computation time, BESO starts from the initial guess
design with 30% material of the design domain which is also
shown in Fig. 2. BESO parameters are E R = 0, A Rmax = 2%,
filter radius r = 10 mm, = 0.01% and i max = 150. Thus,
only 30% of elements in the design domain are involved in finite
element analysis in each iteration.
To verify the proposed BESO method, the topology
optimization using linear finite element analysis was first
carried out to find the linear design. Fig. 3 shows the
evolutionary histories of the external work and topology. It
can be seen that the external work (non-dimensionized by the
external work of the initial guess design) gradually increases
until a convergent value is obtained in about 120 iterations.
Also, the topology also gradually develops to a stable and
mesh-independent design as shown in the last iteration. The
resultant topology is well-known and can be verified by other
mathematical optimization methods [2,22].
2062
Fig. 5. Comparison of the applied load and external work between the linear
and nonlinear design for d = 5 mm.
2063
Fig. 10. Comparison of external work against displacement for various designs.
2064
Fig. 11. Comparison of applied load against displacement for various designs.
Fig. 12. Design domain, initial guess design and support conditions for a 3D structure.
2065
Fig. 13. Comparison of topologies between linear and nonlinear design: (a)
linear design; (b) nonlinear design.
Fig. 14. Comparison of applied load and external work between the linear and
nonlinear design.
2066
EA
u 1
L
EA
(u 2 u 1 ).
F2 =
L
F1 =
1
u2.
1+
1
1
F1A u 1A + F2A (u 2A u 1A ).
2
2
F1A
F1A du 1A
1
du 1A
+
u 1A + F1A
u 1A d
2
d
1 F2A
F2A du 1A
+
+
2
u 1A d
du 1A
1
(u 2A u 1A ) F2A
2
d
EA
E A
2
=
u
u2
2L(1 + )3 2A 2L(1 + )3 2A
E A
E A
+
u 22A +
u2
3
2L(1 + )
2L(1 + )3 2A
EA
=
u2 .
2L(1 + )2 2A
2067
Ep
Ep
(u 2A + u 2B ) + Y 1
2L
E
Ep
EEp A
=
(u 2A + u 2B )
(E p + E)2 2L
Ep
+ Y 1
(u 2B u 2A ).
E
2L 1A
EA
u2 .
=
2L(1 + )2 2A
Thus,
d f 1 ()
1 F1A
F2A
=
u 1A +
(u 2A u 1A ) .
d
2
(A.1)
EA
u 1
L
F2 = A Y + E p
u2 u1
Y
L
E
Thus,
d f 2 ()
1
=
d
2
1
(F1A + F1B )(u 1B u 1A )
2
1
+ (F2A + F2B )(u 2B u 1B u 2A + u 1A ).
2
u 1A d
u 1B d
1
du 1B
du 1A
(u 1B u 1A ) + (F1A + F1B )
2
d
d
1 F2A du 1A
F2B du 1B
+
+
2 u 1A d
u 1B d
(u 2B u 1B u 2A + u 1A )
du 1A
du 1B
1
+ (F2A + F2B )
2
d
d
2
E Ep A
Ep
=
(u 2A + u 2B )
(E p + E)3 2L
E E 2p A
Ep
+ Y 1
(u 2B u 2A )
E
(E p + E)3
Ep
Ep
(u 2A + u 2B ) + Y 1
(u 2B u 2A )
2L
E
E E 2p A
Ep
+
(u 2A + u 2B )
3
(E p + E) 2L
Ep
E2 E p A
+ Y 1
(u 2B u 2A ) +
E
(E p + E)3
(u 2B u 2A )
F2B
1 F2A
(u 2B u 2A u 1B + u 1A )
+
+
2
EA
=
(u 1A + u 1B )(u 1B u 1A ) + 0
2L
Ep
EEp A
=
(u 2A + u 2B )
2
(E p + E) 2L
Ep
+ Y 1
(u 2B u 2A ).
E
Setting F1 = F2 , gives
Ep
Ep
L
u 2 + Y 1
.
u1 =
E p + E L
E
f 2 () =
F1A
F1B
+
(u 1B u 1A ).
(A.2)
EA
u 1
L
F2 = A 2B E
u2 u1
L
1
L
(u 2 2B ).
1
E
1
(F1B + F1C )(u 1C u 1B )
2
1
+ (F2B + F2C )(u 2C u 1C u 2B + u 1B ).
2
u 1B d
u 1C d
1
du 1C
du 1B
(u 1C u 1B ) + (F1B + F1C )
2
d
d
1 F2B du 1B
F2C du 1C
+
+
(u 2C u 1C
2 u 1B d
u 1C d
1
du 1B
du 1C
u 2B + u 1B ) + (F2B + F2C )
2
d
d
2068
EA
u 2B + u 2C
L
2B (u 2C u 2B )
2
E
L(1 )3
A d B
(u 2C u 2B )
(1 )2 d
E A
L
u 2B + u 2C
+
2B
2
E
L(1 )3
u 2B + u 2C
E A
L
(u 2C u 2B )
2B
2
E
L(1 )3
A d B
(u 2C u 2B )
(u 2C u 2B ) +
(1 )2 d
u 2B + u 2C
E A
L
2B (u 2C u 2B )
2
E
L(1 )3
EA
u 2B + u 2C
L
2B (u 2C u 2B ).
=
2
E
L(1 )2
=
1 F2B
F2C
+
+
(u 2C u 2B u 1C + u 1B )
2
EA
(u 1B + u 1C )(u 1C u 1B ) + 0
=
2L
EA
L
u 2B + u 2C
=
2B (u 2C u 2B ).
2
E
L(1 )2
Thus,
d f 3 ()
1
=
d
2
F1B
F1C
+
(u 1C u 1B ).
(A.3)
References
[1] Bendse MP, Kikuchi N. Generating optimal topologies in structural
design using a homogenization method. Computer Methods in Applied
Mechanics and Engineering 1988;71:197224.
[2] Bendse MP, Sigmund O. Topology optimization: Theory, methods and
applications. Berlin, Heidelberg: Springer-Verlag; 2003.
[3] Xie YM, Steven GP. A simple evolutionary procedure for structural
optimization. Computers and Structures 1993;49:8856.