DOI 10.1007/s0015801005876
RESEARCH PAPER
Nonuniqueness and symmetry of optimal topology of a shell
for minimum compliance
Ryo Watada · Makoto Ohsaki · Yoshihiro Kanno
Received: 23 December 2009 / Revised: 7 October 2010 / Accepted: 8 October 2010
c SpringerVerlag 2010
Abstract Uniqueness and symmetry of solution are inves
tigated for topology optimization of a symmetric continuum
structure subjected to symmetrically distributed loads. The
structure is discretized into finite elements, and the compli
ance is minimized under constraint on the structural volume.
The design variables are the densities of materials of ele
ments, and intermediate densities are penalized to prevent
convergence to a gray solution. A path of solution satisfying
conditions for local optimality is traced using the continua
tion method with respect to the penalization parameter. It is
shown that the rate form of the solution path can be formu
lated fromthe optimality conditions, and the uniqueness and
bifurcation of the path are related to eigenvalues and eigen
vectors of the Jacobian of the governing equations. This
A part of this paper was presented at the 8th World Congress of
Structural and Multidisciplinary Optimization (WCSMO8).
R. Watada · M. Ohsaki
Department of Architecture and Architectural Engineering,
Kyoto University, Kyoto, Japan
Y. Kanno
Department of Mathematical Informatics, University of Tokyo,
Tokyo, 1138656, Japan
email: kanno@mist.i.utokyo.ac.jp
Present Address:
R. Watada
Takenaka Corporation, Osaka, Japan
email: watada.ryou@takenaka.co.jp
Present Address:
M. Ohsaki (B)
Department of Architecture, Hiroshima University, 141,
Kagamiyama, HigashiHiroshima, 8398527, Japan
email: ohsaki@hiroshimau.ac.jp
way, local uniqueness and symmetry breaking process of the
solution are rigorously investigated through the bifurcation
of a solution path.
Keywords Topology optimization · Uniqueness ·
Symmetry · Shell · Continuation
1 Introduction
There have been numerous studies on topology optimiza
tion of twodimensional continua discretized into finite
elements. In most of those studies, the compliance, which
is equivalent to the external work against the specified static
loads, is minimized under constraint on the total structural
volume. The topology optimization problem is regarded as
a combinatorial optimization problem using 0–1 variables
indicating existence/nonexistence of elements, which are
relaxed to continuous variables between 0 and 1. However,
simple solution of the relaxed problem leads to a so called
gray solution, in which the variables may have intermediate
values between 0 and 1.
In order to prevent gray solutions, mainly two approaches
have been developed for optimization of plates or sheets;
namely, homogenization approach (Bendsøe and Kikuchi
1988; Suzuki and Kikuchi 1991) and density approach with
penalization (Bendsøe 1989). The former approach has
rigorous mathematical background, but is rather difficult
to implement. Therefore, the density approach has been
recently preferred especially for structures with nonlin
ear properties. Because simple application of the density
approach in conjunction with a nonlinear programming
algorithm results in a gray solution, the intermediate den
sity is penalized to have small stiffness using, e.g., the
SIMP (solid isotropic microstructure with penalty or solid
R. Watada et al.
isotropic material with penalization) approach (Rozvany
et al. 1992; Bendsøe and Sigmund 2003; Rozvany 2009),
or penalized to have artificially large structural volume
(Bruns 2005).
Another difficulty in the SIMP approach is the exis
tence of checkerboard solution, which is avoided using the
penalization in perimeter length (Petersson 1999b) or in the
gradient of the density (Petersson and Sigmund 1998). A
smoothing filter of the density (Bruns 2005, 2007) can also
be used for regularizing the problem. Optimal topology can
be controlled by varying the perimeter length or the parame
ter value of the filter function. Rietz (2007) obtained various
topologies for different values of the gradient parameter.
Recently, the level set approach has been developed for
simultaneous shapetopology optimization of plates and
sheets (Allaire et al. 2002, 2004; Wang et al. 2003; Sethian
and Wiegmann 2000; Yamasaki et al. 2010). Yamada et al.
(2009) utilized the phase field model in conjunction with
the level set approach to control the complexity (number of
holes) of optimal topology. However, in these approaches,
the intrinsic properties of optimal topology cannot be inves
tigated, because the complexity of the solution is artificially
controlled through the penalty terms. Therefore, in this
paper, we investigate the properties of the optimal solu
tion using the SIMP approach without additional penalty
on the perimeter length or the gradient (slope) of density
distribution. Note that it is important for optimization of a
symmetric structure that the symmetry property of the solu
tion should be investigated before assigning the types and
parameter values for regularization.
In the SIMP approach, a large penalization parameter
leads to a 0–1 solution; however, an inappropriate param
eter value will restrict symmetry properties and result in
a convergence to a local optimal solution which is simply
denoted by KKT point satisfying the KarushKuhnTucker
(KKT) conditions. Therefore, a KKT point is usually traced
gradually increasing the penalization parameter utilizing the
so called continuation method (Allgower and Georg 1993)
to select an appropriate value of the penalization parame
ter as well as the parameters for density filter or penalty
of gradients.
It has been shown that the process of gradually increas
ing the penalization parameter converges to an approxi
mate local optimal solution for the original 0–1 problem
(Martínez 2005) and an exact local solution under certain
conditions (Rietz 2001). In the predictorcorrector contin
uation method using the Euler predictor, the governing
equations are differentiated, and the solution path is traced
along the path (Mittelmann and Roose 1990). This pro
cess is basically the same as the parametric programming
approach (Gal 1979; Fiacco 1983) or the homotopy method
(Watson and Haftka 1989) for tracing KKT points corre
sponding to the various parameter values (Nakamura and
Ohsaki 1988). However, in most of the continuation meth
ods for the plate (sheet) topology optimization problems
(Rozvany et al. 1994), the governing equations are not
differentiated, and the solutions are found consecutively
with increasing value of the penalization parameter. Stolpe
and Svanberg (2001) investigated the trajectory of the opti
mal solution with respect to the penalization parameter for
a problem for minimizing the worst value of compliances
under multiple loading conditions.
There have been several papers on topology optimiza
tion of shells (Belblidia and Bulman 2002) including the
early paper by Maute and Ramm (1997). Moses et al.
(2003) investigated symmetric optimal topologies of cir
cular plates assigning the symmetry conditions. However,
to the authors’ knowledge, there has been no investiga
tion on the mechanism of symmetrybraking process of the
optimal topology of an axisymmetric shell. If the penaliza
tion parameter is small and intermediate density is allowed,
the optimal solution of an axisymmetric shell subjected to
symmetric loads is highly likely to be axisymmetric. How
ever, if the intermediate density is penalized, then some ribs
should be generated to result in a solution with reduced
symmetry. Although the solution with intermediate den
sity has no practical meaning, it is important to investigate
the symmetrybreaking process of the KKT point in view
of bifurcation of solution paths. In the same manner as
bifurcation theory (Ikeda and Murota 2010), the unique
ness of the solution is strongly related to the symmetry of
the solution. Uniqueness and stability of a local optimal
solution have been studied by many researchers. Jog and
Haber (1996) derived the conditions of stability using incre
mental form of the variational problem. Petersson (1999a)
investigated convergence of the solution with respect to the
mesh size for simple loading conditions. Kanno et al. (2001)
investigated symmetry properties of optimal solutions of
semidefinite programming using group theoretic approach,
and showed that optimal crosssectional areas of trusses
for maximum eigenvalues are symmetric if the geometry
is fixed and symmetric. Stolpe (2010) showed that optimal
truss with symmetric geometry and loading conditions may
not be symmetric if crosssectional areas are discrete vari
ables. Rozvany (2011) showed that the solution is usually
unique and at least one solution must be symmetric for a
symmetric problem with continuous variables.
The purpose of this paper is to investigate the symmetry
reduction process of the local optimal solution. A solution
path is defined with respect to the penalization parame
ter, and a bifurcation point is detected as a singular point
of the solution path. This process is carried out without
using any filter or penalty for the gradient in order to inves
tigate the intrinsic symmetryreduction process. We first
define local nonuniqueness of the KKT point as a bifur
cation of the solution path with respect to the penalization
Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance
parameter. The formulation for numerical continuation with
respect to the penalization parameter is rigorously derived
by differentiating the KKT conditions and the stiffness
(equilibrium) equations. Then, a condition for local unique
ness of the solution is derived as the singularity of the
Jacobian of governing equations (Ohsaki 2006; Ikeda and
Murota 2010; Ohsaki and Ikeda 2007). In the numerical
examples, the symmetryreduction process of the optimal
solution as a function of the penalization parameter is stud
ied in details. It is shown that a ribbed shell with reduced
symmetry is generated through a bifurcation process of the
solution path.
2 Conditions for bifurcation of solution path
and stability of solution of nonlinear equations
Suppose a set of variables is found as a solution of nonlinear
equations defined with a parameter. Then a path or a curve
of the solution is defined parametrically in the space of the
variables and parameter. In this section, a basic method
ology of bifurcation analysis of solution path of nonlinear
equations is summarized for the completeness of the paper
and to clarify the relation between the stability of itera
tive solution process and the bifurcation (nonuniqueness) of
solution path (Golubitsky and Schaegger 1979).
Consider a problem of finding the solution x ∈ R
q
of the
nonlinear equations
F(x, t ) = 0 (1)
where t is the problem parameter, and F(x, t ) ∈ R
q
is con
tinuously differentiable with respect to x and t . Since the
solution is defined by (1) for each specified value of t , it
is regarded as a function of t , and is denoted by ˜ x(t ). Sup
pose we have a solution ˜ x(t
0
) for t = t
0
. Then the solution
for t = t
0
+ t with a specified increment t of t can be
estimated using the following linear approximation of the
governing equations (1):
J
¸
˜ x(t
0
+t ) − ˜ x(t
0
)
¸
+
∂F
∂t
t = 0 (2)
where J ∈ R
q×q
is the Jacobian of F, for which the (i, j )
component J
i j
is defined as
J
i j
=
∂ F
i
∂x
j
, (i, j = 1, . . . , q) (3)
If J is nonsingular, then the solution path can be uniquely
found using an incrementaliterative method for the gov
erning equations (1) in a similar manner as the arclength
method for geometrically nonlinear equilibrium analysis. In
contrast, if J is singular, then bifurcation of the solution
path exists and the solution cannot be determined uniquely
at the bifurcation point (Ikeda and Murota 2010; Ohsaki and
Ikeda 2007).
We next consider a process of finding the solution of
(1) for a fixed value of t at t
0
from an arbitrary initial
solution using Newton iteration. Let F
0
denote the current
value of F in the iterative process. Then the equation for
finding the increment x of x for satisfying (1) with linear
approximation is written as
F
0
+Jx = 0 (4)
Therefore, if J is singular, then the solution of (1) cannot be
determined uniquely using (4). Furthermore, if F is the gov
erning equation of a reciprocal system that has a potential
function, then F is defined as the gradient of the potential,
and J is a symmetric matrix. Therefore, for a reciprocal sys
tem, the Newton iteration using (4) diverges if J is singular.
Hence, the solution process is unstable at the bifurcation
point of the solution path.
Jog and Haber (1996) derived the conditions of stability
using an incremental form of the variational problem for a
minmax optimization problem. However, our approach for
investigation of uniqueness of the optimal solution based on
continuation method is applicable to any type of nonlinear
programming problem.
3 Optimization problem and optimality conditions
Consider a symmetric plate or shell discretized into finite
elements. The number of elements and the number of
degrees of freedom are denoted by m and n, respectively.
Let d
j
denote the density of the j th element, for which the
upper and lower bounds are assigned as
0 ≤ d
j
≤ 1, ( j = 1, . . . , m) (5)
The design variable vector is given as d = (d
1
, . . . , d
m
)
.
Let P ∈ R
n
denote the specified nodal load vector. The
stiffness matrix is denoted by K ∈ R
n×n
. Then the nodal
displacement vector U ∈ R
n
is obtained from the following
stiffness (equilibrium) equation:
K(d)U(d) = P (6)
The objective function to be minimized is the compliance
W(d) defined as
W(d) = P
U(d) (7)
R. Watada et al.
We use the SIMP approach to prevent gray solutions
by penalizing the stiffness of an element with intermediate
density. The stiffness matrix K is defined with the matrix
K
i
∈ R
n×n
corresponding to the unit density of the i th
element and the penalization parameter p as
K =
m
¸
i =1
¸
ε +(1 −ε)d
p
i
¸
K
i
(8)
where ε is a sufficiently small positive value for prevent
ing numerical instability. The volume of the i th element is
expressed as a product of the area A
i
, thickness h
i
, and
density d
i
. Then the problem of minimizing the compli
ance under a constraint on the total structural volume is
formulated with respect to the variable vector d as
Minimize W(d) = P
U(d) (9a)
subject to
m
¸
i =1
A
i
h
i
[ε +(1 −ε)d
i
] −
¯
V ≤ 0 (9b)
0 ≤ d
i
≤ 1, (i = 1, . . . , m) (9c)
where
¯
V is the specified upper bound of the total structural
volume.
The Lagrangian of problem (9) is formulated as
L(d, λ, μ
U
, μ
L
)
= W(d) +λ
m
¸
i =1
A
i
h
i
[ε +(1 −ε)d
i
] −
¯
V
+
m
¸
i =1
μ
U
i
(d
i
−1) +
m
¸
i =1
μ
L
i
(−d
i
) (10)
where λ, μ
U
i
, and μ
L
i
are the Lagrange multipliers
that have nonnegative values at the optimal solution. By
differentiating (10) with respect to d
i
and using (6–8), we
have
∂L
∂d
i
=−(1 −ε) pd
p−1
i
U
K
i
U
+ λA
i
h
i
(1 −ε) +μ
U
i
−μ
L
i
(11)
where the standard approach of sensitivity analysis of
compliance has been used (Choi and Kim 2004). Define
G
i
(d) as
G
i
(d) = −(1 −ε) pd
p−1
i
U
K
i
U +λA
i
h
i
(1 −ε) (12)
Then the firstorder optimality conditions (KKT condi
tions) are derived as
G
i
(d)
⎧
⎨
⎩
= 0 for 0 < d
i
< 1
≤ 0 for d
i
= 1
≥ 0 for d
i
= 0
(13)
The set of indices of elements satisfying 0 < d
i
< 1, and
accordingly G
i
(d) = 0, is denoted by I; i.e.,
I = {i  0 < d
i
< 1} (14)
and we define s by s = I.
4 Sensitivity of optimal solution with respect
to penalization parameter
Equations for computing the sensitivity coefficients with
respect to p, which are called parametric sensitivity
coefficients for brevity, of the optimal solutions are derived
below, where (·)
indicates differentiation with respect to
p. For this purpose, the vector of state variables U and the
Lagrange multiplier λ are also regarded as functions of p.
By differentiating (6) with respect to p and using (8),
we have
−KU
−
m
¸
i =1
(1 −ε) pd
p−1
i
d
i
K
i
U
= (1 −ε)
m
¸
i =1
d
p
i
ln d
i
K
i
U (15)
By differentiating the volume constraint (9b) and multiply
ing 1/2, we obtain
1
2
m
¸
i =1
(1 −ε)A
i
h
i
d
i
= 0 (16)
Suppose the active side constraints remain active at the
KKT point corresponding to the parameter value in the
neighborhood of the current value; i.e., μ
U
i
> 0 and μ
L
i
> 0
are satisfied for d
i
= 1 and d
i
= 0, respectively. Further
more, transition of an inactive side constraint to be active
is not considered, because we increase the parameter dis
cretely and find the solution for each specified value of the
parameter. Therefore, we can assume that the solution is
generally a regular KKT point. Bifurcation at a degenerate
point may be investigated similarly using directional deriva
tive of the solution. However, it is shown in the numerical
examples that this simple continuation method is practi
cally effective for obtaining a symmetric optimal topology.
Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance
Hence, for the elements i ∈ I, differentiation of G
i
(d) = 0
with respect to p leads to
−(1 −ε) pd
p−1
i
U
K
i
U
−
1
2
(1 −ε) p( p −1)d
p−2
i
d
i
U
K
i
U
+
1
2
(1 −ε)A
i
h
i
λ
=
1
2
(1 −ε) pd
p−1
i
ln d
i
U
K
i
U
+
1
2
(1 −ε)d
p−1
i
U
K
i
U, (i ∈ I) (17)
For i / ∈ I, we have d
i
= 0. Therefore, there are n + s + 1
linear equations (15), (16), and (17) for n +s +1 variables
U
, d
i
(i ∈ I), and λ
.
The indices of elements are rearranged so that I =
{1, . . . , s}, and define d
0
= (d
1
, . . . , d
s
)
. Then the
linear equations for computing the parametric sensitivity
coefficients of the KKT point are written in the following
form using the notations defined below:
⎛
⎜
⎜
⎝
−K B
12
0
B
12
B
22
B
23
0
B
23
0
⎞
⎟
⎟
⎠
⎛
⎝
U
d
0
λ
⎞
⎠
=
⎛
⎝
b
1
b
2
0
⎞
⎠
(18)
which is simply written as
BX
= b (19)
Let H
i
= U
K
i
U and b
2
= (b
2
1
, . . . , b
2
s
)
. The (i, j )
components of B
12
∈ R
n×s
, B
22
∈ R
s×s
, and the i th
component of B
23
∈ R
s
are denoted by B
12
i j
, B
22
i j
, and
B
23
i
, respectively. The j th column of K
i
is denoted by
k
i j
∈ R
n
. Then the components of the symmetric matrix
B ∈ R
(n+s+1)×(n+s+1)
and the constant vector b ∈ R
n+s+1
are given as
B
12
i j
= −(1 −ε) pd
p−1
j
k
j i
U (20a)
B
22
i i
= −
1
2
(1 −ε) p( p −1)d
p−2
i
H
i
,
B
22
i j
= 0 for i = j (20b)
B
23
i
=
1
2
(1 −ε)A
i
h
i
(20c)
b
1
= (1 −ε)
m
¸
i =1
d
p
i
ln d
i
K
i
U (20d)
b
2
i
=
1
2
(1 −ε)(1 + p ln d
i
)d
p−1
i
H
i
(20e)
The path of the optimal solutions can be traced succes
sively solving (19) (Ohsaki and Nakamura 1996). Since B
is symmetric, the stability of solution is detected from the
eigenvalues or the condition number of the matrix.
5 Uniqueness of local optimal solution for specified
penalization parameter
We can solve the first equation of (18) for U
as
U
= −K
−1
b
1
+K
−1
B
12
d
0
(21)
which is incorporated into the second and third equations of
(18) to obtain
B
22∗
B
23
B
23
0
d
0
λ
=
b
2∗
0
(22)
b
2∗
= b
2
+B
12
K
−1
b
1
,
B
22∗
= B
22
+B
12
K
−1
B
12
(23)
Note that the elements of vector B
23
have nonzero values as
seen in (20c).
We can see from (20b) that B
22
is a diagonal negative
definite matrix for p > 1, and a null matrix for p = 1.
In contrast, the matrix B
12
K
−1
B
12
is positive definite,
because K
−1
is positive definite and B
12
has full column
rank as discussed below. Therefore, the symmetric matrix
B
22∗
is positive definite at p = 1, and its lowest eigenvalue
may decrease to 0 as p is increased. When B
22∗
is positive
definite, the first equation of (22) can be solved for d
0
as
d
0
=
B
22∗
−1
b
2∗
−B
23
λ
(24)
By incorporating this into the second equation of (22) and
using positive definiteness of (B
22∗
)
−1
, we obtain
λ
=
B
23
B
22∗
−1
b
2∗
B
23
B
22∗
−1
B
23
(25)
Hence, λ
is uniquely determined, and d
0
is also uniquely
found using (24) and (25). Therefore, the parametric sensi
tivity coefficients of a KKT point are uniquely determined;
accordingly, no bifurcation occurs along a solution path if
B
22∗
is nonsingular. This way, the conditions for unique
ness of the solution have been derived using a simple and
rigorous manner based on the uniqueness of a solution path.
Note that the matrix K may be nearly singular and the
condition number of K
−1
may be very large, if a small value
is assigned for ε. However, singularity of K leads only to
nonuniqueness of the displacements that may not have any
R. Watada et al.
1
41
81
121
161
Fig. 1 A spherical shell model
effect on nonuniqueness of the density variables (Koˇ cvara
and Outrata 2006). In fact, the large eigenvalue of K
−1
, if
exists, corresponds to the eigenmode that has large defor
mation at the nodes connected by the elements with low
density only. Let f
i
∈ R
n
denote the vector of equivalent
nodal loads of the i th element with d
i
= 1 corresponding to
the displacement vector U. Using (20a), the i th column of
B
12
, denoted by R
i
∈ R
n
, is given as
R
i
= −(1 −ε) pd
p−1
i
f
i
(26)
(a) p = 1.00, W = 1.4528 × 10
−5
(b) p = 1.78, W = 1.7819 × 10
−5
(c) p = 1.80, W = 1.7935 × 10
−5
(d) p = 1.82, W = 1.7947 × 10
−5
(e) p = 2.94, W = 2.2391 × 10
−5
(f) p = 2.96, W = 2.2148 × 10
−5
Fig. 2 Optimal solutions for various values of parameter p
Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance
Hence, R
i
(i = 1, . . . , s) are generally independent, and
B
12
has full column rank.
The matrix B
12
K
−1
B
12
in (23) is denoted by S ∈ R
s×s
.
Then, the (i, j )component S
i j
of S is written as follows
using the displacement vector U
j
∈ R
n
against F
j
:
S
i j
= U
j
R
i
(27)
Therefore, the matrix S = B
12
K
−1
B
12
is bounded if the
work in the j th element ( j ∈ I) done by the nodal loads that
is proportional to the equivalent nodal loads of the i th ele
ment (i ∈ I) is bounded. This condition is usually satisfied
because d
i
has moderately large value for the elements in I.
Suppose there exists a singular point where the lowest
eigenvalue of B
22∗
vanishes. The rate form (22) is converted
to an incremental formas follows for the increments δd
0
and
δλ of d
0
and λ, respectively, corresponding to the increment
δp of the parameter:
B
22∗
B
23
B
23
0
δd
0
δλ
= δp
b
2∗
0
(28)
As seen from (20c), the elements of vector B
23
has nonzero
values, and B
23
has the same symmetry property as the
solution with uniform density. Let denote the eigenvec
tor corresponding to the zero eigenvalue of B
22∗
; i.e., the
kernel of the matrix B
22∗
is given as c with arbitrary
coefficient c. If has a lower symmetry property than B
23
,
then
B
23
= 0 and B
23
is included in the range of B
22∗
.
In this case, for the constant parameter value, i.e., δp = 0,
there exists a solution δd
0
= β and δλ = 0 with β being
an arbitrary nonzero value. It is easily observed from (20d),
(20e), and (23) that b
2∗
has the same symmetry property as
0
0.5
1.0
1.5
2.0
2.5
3.0
1 1.5 2 2.5 3
( 10 )
5
(a) 1 p 3
0
0.5
1.0
1.5
2.0
1.5 1.6 1.7 1.8 1.9 2
( 10 )
6
(b) 1.5 p 2
2.5 2.6 2.7 2.8 2.9 3
0
0.5
1.0
1.5
2.0
( 10 )
6
(c) 2.5 p 3
Fig. 3 Eigenvalues of the matrix B
22∗
R. Watada et al.
the current optimal solution. Therefore, if
b
2∗
= 0, i.e.,
if b
2∗
is included in the range of B
22∗
, then there exists a par
ticular solution that may have nonzero value of δp. Hence,
bifurcation of solution occurs when B
22∗
becomes singular.
Jog and Haber (1996) derived the conditions of unique
ness and stability of the optimal solution based on incremen
tal form of variational problem. They first used continuum
formulation that is reduced to a finite element formula
tion. In the following, we summarize their results in a
matrixvector form.
The optimization problem is formulated as a maxmin
problem:
Find max
d
min
U
1
2
U
KU −P
U
−λ
m
¸
i =1
d
i
−
¯
V
−γ
m
¸
i =1
d
i
(1 −d
i
)
(29a)
subject to 0 ≤ d
j
≤ 1, ( j = 1, . . . , m) (29b)
where γ is the penalty parameter. From the stationary con
ditions of the Lagrangian of problem (29), we obtain the
equilibrium equation (6) and the optimality conditions (13)
for the elements in I, where G
i
(d) in (13) is to be replaced
by
G
i
(d) = pd
p−1
i
U
K
i
U +2λ −2γ (1 −2d
i
) (30)
The incremental equations are derived in the following
form for solving the equilibrium equations and optimality
conditions using Newton iteration:
K B
12
B
12
B
22
δU
δd
0
=
b
1
b
2
(31)
with appropriate modification of matrices and vectors due
to incorporation of the penalty term γ d
i
(1 − d
i
) and the
differences in the signs of the strain energy in the objective
function. Note that the increment of λ is not considered in
(31), i.e., λ is fixed for some reason in the iterative process.
(a) p = 1.78
(b1) p = 2.94 (b2) p = 2.94 (b3) p = 2.94
Fig. 4 Eigenvectors of B
22∗
for p = 1.78 and 2.94
Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance
Then, the conditions of uniqueness of the solution are: K is
positive definite and B
22
is negative definite.
6 Numerical examples
As a numerical example, uniqueness and symmetry of opti
mal solution is investigated for an axisymmetric shell. The
symmetry of the solution is indicated using the standard
notations of group theory (Kettle 2007). If the solution does
not change after application of reflection with respect to
one of n different planes containing the axis of symme
try, and also with respect to rotation of one of n different
angles around the axis of symmetry, then it has the dihedral
symmetry D
n
.
Consider a spherical shell as shown in Fig. 1 subjected
to the vertical concentrated load P at each node on the top
ring. The design domain is discretized to 10 and 20 elements
in longitudinal and circumferential directions, respectively,
i.e., the total number of elements is 400. The geometrical
parameters are R = 40.0 m, H = 11.5 m, θ = π/3,
and α = π/12. Young’s modulus is 2.10 × 10
8
kN/m
2
,
Poisson’s ratio is 0.3, and P = 1.0 kN. In the following,
the units of length and force are m and kN, respectively,
which are omitted for brevity. However, the objective func
tion is multiplied by 10
9
to prevent numerical difficulty. The
values in the figures are those without scaling. The small
density parameter for ε for preventing numerical instabil
ity is 1.0 × 10
−6
, the maximum thickness is h
i
= 0.5 m,
and the upperbound volume
¯
V is 50% of the value of the
solution with d
i
= 1 for all elements. Optimization is
carried out using SNOPT Ver. 7 (Gill et al. 2002), where
the sequential quadratic programming (SQP) is used. The
default values are used for the parameters except the strict
tolerance 10
−12
for feasibility. The tolerance of optimality
is also small enough so that optimization is carried out until
no improvement is achieved.
0.01
0.1
1
1 1.5 2 2.5 3
(a) B
22
1 1.5 2 2.5 3
1
100
1000
10
(b) B
22
10
100
1000
1 1.5 2 2.5 3
(c) B
12
B
12
1 1.5 2 2.5 3
*
10
8
10
4
10
5
10
6
10
7
(d) K
Fig. 5 Variations of condition numbers
R. Watada et al.
The parametric sensitivity coefficients of optimal solu
tions are first verified. For p = 1.0, the optimal densities of
the elements 1, 41, 81, 121, and 161 indicated in Fig. 1 are
d
1
= d
41
= 1.0000, d
81
= 0.58805, d
121
= 0.33348, and
d
161
= 0.21647. The sensitivity coefficients are obtained
as d
1
= d
41
= 0, d
81
= −0.3232, d
121
= 0.05138, and
d
161
= 0.06652. The sensitivity coefficients obtained by
the forward finite difference approach with p = 0.01
are d
1
= d
41
= 0, d
81
= −0.3226, d
121
= 0.05133,
and d
161
= 0.06640, which have good agreement with the
analytical results.
Local optimal solutions are found for the parameters
between p = 1.0 and 3.0 with the increment p = 0.02, by
tracing a solution path assigning the solution of p − p
as the initial solution for the SQP algorithm. Figure 2
shows the distributions of d
i
of optimal solutions for various
parameter values in gray scale. The values of the objective
function W are also shown. The same axisymmetric solu
tion is found from any initial solution for p = 1.0, because
the objective function is convex and the volume constraint
is linear with respect to d. Optimization in the parameter
range 1.00 ≤ p ≤ 1.78 leads to axisymmetric solutions as
shown in Fig. 2a and b, which have D
20
symmetry.
For each parameter value, the eigenvalues of B
22∗
of
the KKT point are plotted in Fig. 3. As seen in Fig. 3b,
the matrix B
22∗
becomes singular with a zero eigenvalue
between p = 1.78 and 1.80. Accordingly, the symmetry of
solution for p = 1.80 in Fig. 2c is reduced to D
10
, and the
difference between the two solutions for p = 1.78 and 1.80
is proportional to the eigenvector corresponding to the zero
eigenvalue of the optimal solution for p = 1.78, which has
D
10
symmetry as shown in Fig. 4a.
0
2
4
6
8
10
12
14
16
18
20
22
1 1.5 2 2.5 3
(a)  d
0

0
2
4
6
8
10
12
1 1.5 2 2.5 3
7
( 10 )
(b)
0
50
100
150
200
1 1.5 2 2.5 3
(c) s
1.4
1.5
1.6
1.7
1.8
1.9
2.0
2.1
2.2
2.3
1 1.5 2 2.5 3
5
( 10 )
(d) W
Fig. 6 Variations of the norm of parametric sensitivity vector d
0
, Lagrange multiplier λ, cardinality s of I, and objective function W
Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance
The condition numbers of B
22
, B
22∗
, B
12
B
12
, and K are
plotted in logscale in Fig. 5. The norm of parametric sensi
tivity vector, Lagrange multiplier, cardinality s of I, and
objective function are plotted in Fig. 6. As is seen, the
condition number of B
22∗
and the norm d
0
 of paramet
ric sensitivity vector drastically increase around p = 1.78
at which B
22∗
has very small positive eigenvalue. It is also
seen from Fig. 5 that the n × s matrix B
12
has full col
umn rank s (< n). The condition number of K jumps to
a very large value at the singular point of B
22∗
; however,
B
22∗
has moderately small condition number in the range
p ≥ 1.80, although the condition number of K is very large.
For p = 2.0, the condition number of K is 6.913 × 10
7
,
which is very large; i.e., U is nearly nonunique. By con
trast, the condition numbers of B
22
and B
∗
22
are 0.2469 and
9.413, respectively, which are sufficiently small. Therefore,
(a)
(b)
Fig. 7 Optimal topology of 15×40 model with p = 3; a continuation
approach, b direct optimization
instability of the structure, or nonuniqueness of the displace
ment vector U is not related to the bifurcation of the optimal
solution.
More symmetrybreaking singular points can be found
by further increasing the parameter p. The solution for
p = 1.82 has D
5
symmetry as shown in Fig. 2d, and the
solutions in the parameter range 1.82 ≤ p ≤ 2.94 have
the same symmetry property, because B
22∗
remains to be
positive definite. The lowest eigenvalue of B
22∗
becomes
zero between the parameter range 2.94 ≤ p ≤ 2.96, and
the symmetry of the solution reduces from D
5
to D
2
as
shown in Fig. 2d and e. For p = 2.94, the lowest eigen
values of B
22∗
are threefold, for which the eigenvectors
are shown in Fig. 4, where (b1) and (b3) have D
5
and
D
1
symmetry, respectively, while (b2) does not have any
symmetry property.
There also exist other points at which the lowest eigen
value of the matrix B
22∗
and the condition numbers of
the matrix B
22
and B
22∗
jump discontinuously due to the
variation of the size s of these matrices.
Since the shape is fixed and only the topology is varied,
the mechanism of resisting external loads distributed around
the top ring did not change as an result of increasing the
penalization parameter.
Finally, optimal topologies are found for a finer mesh
with 15 and 40 elements in longitudinal and circumferential
directions, respectively, i.e., the total number of elements is
600. Starting with the uniform initial design d
i
= 0.5 for all
elements with p = 1.0, the parameter p is increased to 3.0
with the increment 0.01. The optimal topology obtained in
this continuation approach is shown in Fig. 7a. As is seen,
a topology with D
10
symmetry has been obtained; however,
some gray elements exist near the boundary. In contrast, if
we optimize directly for p = 3.0 from the uniform initial
Fig. 8 Optimal topology of 15 × 40 model with p = 4; direct
optimization
R. Watada et al.
solution, then the topology as shown in Fig. 7b is obtained.
As is seen, no clear symmetry is observed.
The optimal topology for p = 4.0 using continuation
approach is shown in Fig. 8a. An almost clear 0–1 solu
tion has been found with d
i
= 1.0 for 341 elements,
0.966 ≤ d
i
≤ 0.968 for 10 elements, and d
i
= 0.0 for
249 elements. The number of elements with d
i
1.0 is not
equal to the half of the total number elements, because the
elements around the boundary have larger area than those
around the center. If we optimize directly for p = 4.0 from
the uniform initial solution, then the topology is obtained as
shown in Fig. 8b, which has no clear symmetry.
7 Conclusions
A simple formulation has been presented for investigating
the path of the firstorder optimal solution of an axisym
metric shell that minimizes the compliance under specified
load and the total structural volume. The conditions of
nonuniqueness of the solution are derived based on a bifur
cation of the solution path with respect to the penaliza
tion parameter of the SIMP approach. The formulation
for numerical continuation with respect to the penalization
parameter is rigorously derived by differentiating the KKT
conditions, stiffness (equilibrium) equations, and volume
constraint.
The main results are summarized as follows:
1. The symmetry reduction process of the KKT point
can be defined as an bifurcation of solution path with
respect to the penalization parameter of SIMP approach.
2. Convergence of optimization algorithm and uniqueness
of solution path are defined by the same condition of
the singularity of a matrix similarly to geometrically
nonlinear analysis of structures.
3. Nonuniqueness of displacement does not have any
effect on nonuniqueness of the KKT point.
4. It has been shown quantitatively that symmetry of KKT
point reduces through bifurcation to the critical eigen
vector of the Jacobian of the governing equations in
the continuation process of KKT point using SIMP
approach.
5. A procedure has been developed for detecting bifur
cation of the solution and the direction of reduc
tion of symmetry using eigenvalue analysis of the
Jacobian of the governing equations. A procedure
has also been developed for condensing Jacobian to
remove the derivatives of displacements and sepa
rate the nonuniqueness properties of displacements and
design variables.
Thus, a unified approach has been presented for inves
tigation of uniqueness and bifurcation (symmetrybreaking
process) of KKT point. A local optimal solution with appro
priate symmetry can be found through investigation of
symmetry using continuation approach. Note that the sym
metry properties without filter can be used for appropriately
setting the parameters for filters and other regularization.
The numerical example of an axisymmetric shell shows
that a solution path of an axisymmetric shell has a bifurca
tion point where the Jacobian of the governing equations
is singular. The KKT point is nonunique at the bifurca
tion point, and a symmetrybreaking bifurcation path exists
in the direction of the eigenvector corresponding to the
zero eigenvalue analyses of the Jacobian. This way, the
symmetryreduction process of the KKT point is character
ized as a bifurcation process of a solution path with respect
to the penalization parameter. Although the solution with
intermediate density has no practical meaning, it is impor
tant to investigate the symmetry breaking process of the
KKT point on the basis of bifurcation of a solution path.
The condition numbers of the matrix for computing the
sensitivity coefficients of the KKT points, and the corre
sponding submatrices, have been computed to show that
nearly singular stiffness matrix does not lead to nonunique
ness of the KKT point.
Note that the proposed method is effective for any prob
lem with symmetry including a circular disc subjected to a
twisting force at the center circle, which is often used as
an example of Michell structure. We used a shell structure,
because it has high symmetry. Furthermore, the method pro
posed in this paper can also be used for the case where a
filter is used, and it can be applied to a largescale prob
lem without any difficulty, because the eigenvalue analysis
should be carried out for the matrix of the size equal to the
variables that have intermediate values.
References
Allaire G, Jouve F, Toader AM (2002) A level set method for shape
optimization. C R Acad Sci 334:1125–1130
Allaire G, Jouve F, Toader AM (2004) Structural optimization using
sensitivity analysis and a levelset method. J Comput Phys
104:363–393
Allgower EL, Georg K (1993) Continuation and path following. Acta
Numer 2:1–64
Belblidia F, Bulman S (2002) A hybrid topology optimization algo
rithm for static and vibrating shell structures. Int J Numer
Methods Eng 24:835–852
Bendsøe MP (1989) Optimal shape design as a material distribution
problem. Struct Optim 1:193–202
Bendsøe MP, Kikuchi N(1988) Generating optimal topologies in struc
tural design using a homogenization method. Comput Methods
Appl Mech Eng 71:197–224
Bendsøe MP, Sigmund O (2003) Topology optimization: theory, meth
ods and applications. Springer, Berlin
Bruns TE (2005) A reevaluation of the SIMP method with filtering and
an alternative formulation for solid–void topology optimization.
Struct Multidisc Optim 30:428–436
Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance
Bruns TE (2007) Topology optimization by penalty (TOP) method.
Comput Methods Appl Mech Eng 196:4430–4443
Choi KK, KimNH(2004) Structural sensitivity analysis and optimiza
tion, vol 1. Linear systems. Springer, New York
Fiacco AV (1983) Introduction to sensitivity and stability analysis in
nonlinear programming. Academic
Gal T (1979) Postoptimal analyses, parametric programming, and
related topics. McGraw Hill
Gill PE, Murray W, Saunders MA (2002) SNOPT: an SQP algo
rithm for largescale constrained optimization. SIAM J Optim 12:
979–1006
Golubitsky M, Schaegger D (1979) A theory for imperfect bifurcation
via singularity theory. Commun Pure Appl Math 32:21–98
Ikeda K, Murota K (2010) Imperfect bifurcation in structures and
materials—engineering use of grouptheoretic bifurcation theory,
2nd edn. Applied mathematical sciences, vol 149. Springer, New
York
Jog CS, Haber RB (1996) Stability of finite element models for
distributedparameter optimization and topology design. Comput
Methods Appl Mech Eng 130:203–226
Kanno Y, Ohsaki M, Murota K, Katoh N (2001) Group symmetry
in interiorpoint methods for semidefinite program. Optim Eng
2:293–320
Kettle SFA (2007) Symmetry and structure: readable group theory for
chemists. Wiley, Chichester
Koˇ cvara M, Outrata JV (2006) Effective reformulation of the truss
topology design problem. Optim Eng 7:201–219
Martínez JM (2005) A note on the theoretical convergence properties
of the SIMP method. Struct Multidisc Optim 29:319–323
Maute K, Ramm E (1997) Adaptive topology optimization of shell
structures. AIAA J 35:1767–1773
Mittelmann HD, Roose D (eds) (1990) Continuation techniques and
bifurcation problems. Birkhäuser, Basel
Moses E, Fuchs MB, Ryvkin M (2003) Topology design of mod
ular structures under arbitrary loading. Struct Multidisc Optim
24:407–417
Nakamura T, Ohsaki M (1988) Sequential optimal truss generator
for frequency ranges. Comput Methods Appl Mech Eng 67(2):
189–209
Ohsaki M (2006) Local and global searches of approximate optimal
designs of regular frames. Int J Numer Methods Eng 67:132–147
Ohsaki M, Nakamura T (1996) Minimum constraint perturbation
method for topology optimization of systems. Eng Opt 26:
171–186
Ohsaki M, Ikeda K (2007) Stability and optimization of structures—
generalized sensitivity analysis. Mechanical engineering series.
Springer, New York
Petersson J (1999a) A finite element analysis of optimal variable
thickness sheets. SIAM J Numer Anal 36(6):1759–1778
Petersson J (1999b) Some convergence results of perimetercontrolled
topology optimization. Comput Methods Appl Mech Eng 171:
121–140
Petersson J, Sigmund O (1998) Slope constrained topology optimiza
tion. Int J Numer Methods Eng 41:1417–1434
Rietz A (2001) Sufficiency of a finite exponent in SIMP (power law)
methods. Struct Multidisc Optim 21:159–163
Rietz A (2007) Modified conditions for minimum of compliance.
Comput Methods Appl Mech Eng 196:4413–4418
Rozvany GIN (2009) A critical review of established methods of
structural topology optimization. Struct Multidisc Optim 37:
217–237
Rozvany GIN (2011) On symmetry and nonuniqueness in exact
topology optimization. Struct Multidisc Optim. Published online.
doi:10.1007/s0015801005640
Rozvany GIN, Zhou M, Birker T (1992) Generalized shape optimiza
tion without homogenization. Struct Optim 4:250–252
Rozvany GIN, Zhou M, Sigmund O (1994) Topology optimization.
In: Adeli H (ed) Advances in design optimization, Chapter 10.
Chapman and Hall, London
Sethian JA, Wiegmann A (2000) Structural boundary design via level
set and immersed interface methods. J Comput Phys 163(2):
489–528
Stolpe M (2010) On some fundamental properties of structural topol
ogy optimization problems. Struct Multidisc Optim 41:661–670
Stolpe M, Svanberg K (2001) On the trajectories of the epsilon
relaxation approach for stressconstrained truss topology opti
mization. Struct Multidisc Optim 21:140–151
Suzuki K, Kikuchi N (1991) A homogenization method for shape
and topology optimization. Comput Methods Appl Mech Eng 93:
291–318
Wang MY, Wang XM, Guo DM (2003) A level set method for struc
tural topology optimization. Comput Methods Appl Mech Eng
192:227–246
Watson LT, Haftka RT (1989) Modern homotopy methods in optimiza
tion. Comput Methods Appl Mech Eng 74(3):289–305
Yamada T, Nishiwaki S, Izui K, Yoshimura M, Takezawa A (2009)
A new level set based topology optimization method using a
fictitious interface energy model based on phase field method
concepts. In: Proc. 8th world congress of structural and multi
disciplinary optimization (WCSMO8), paper no. 1136. Lisbon
Yamasaki S, Nomura T, Kawamoto A, Sato K, Izui K, Nishiwaki S
(2010) A level set based topology optimization method using the
discretized signed distance function as the design variables. Struct
Multidisc Optim 41:685–698
Although the solution with intermediate density has no practical meaning. because the complexity of the solution is artificially controlled through the penalty terms. Moses et al. it is important to investigate the symmetrybreaking process of the KKT point in view of bifurcation of solution paths. Fiacco 1983) or the homotopy method (Watson and Haftka 1989) for tracing KKT points corresponding to the various parameter values (Nakamura and Ohsaki 1988). if the intermediate density is penalized. the optimal solution of an axisymmetric shell subjected to symmetric loads is highly likely to be axisymmetric. an inappropriate parameter value will restrict symmetry properties and result in a convergence to a local optimal solution which is simply denoted by KKT point satisfying the KarushKuhnTucker (KKT) conditions. A smoothing filter of the density (Bruns 2005. (2003) investigated symmetric optimal topologies of circular plates assigning the symmetry conditions. Recently. This process is carried out without using any filter or penalty for the gradient in order to investigate the intrinsic symmetryreduction process. and a bifurcation point is detected as a singular point of the solution path. However. a large penalization parameter leads to a 0–1 solution. Rozvany (2011) showed that the solution is usually unique and at least one solution must be symmetric for a symmetric problem with continuous variables. which is avoided using the penalization in perimeter length (Petersson 1999b) or in the gradient of the density (Petersson and Sigmund 1998). This process is basically the same as the parametric programming approach (Gal 1979. However. There have been several papers on topology optimization of shells (Belblidia and Bulman 2002) including the early paper by Maute and Ramm (1997). Bendsøe and Sigmund 2003. In the same manner as bifurcation theory (Ikeda and Murota 2010). Petersson (1999a) investigated convergence of the solution with respect to the mesh size for simple loading conditions. 2003. 1992. In the predictorcorrector continuation method using the Euler predictor. and showed that optimal crosssectional areas of trusses for maximum eigenvalues are symmetric if the geometry is fixed and symmetric. Watada et al. Uniqueness and stability of a local optimal solution have been studied by many researchers. (2001) investigated symmetry properties of optimal solutions of semidefinite programming using group theoretic approach.R. Jog and Haber (1996) derived the conditions of stability using incremental form of the variational problem. to the authors’ knowledge. The purpose of this paper is to investigate the symmetryreduction process of the local optimal solution. Sethian and Wiegmann 2000. If the penalization parameter is small and intermediate density is allowed. Wang et al. 2007) can also be used for regularizing the problem. in this paper. Another difficulty in the SIMP approach is the existence of checkerboard solution. Kanno et al. Stolpe and Svanberg (2001) investigated the trajectory of the optimal solution with respect to the penalization parameter for a problem for minimizing the worst value of compliances under multiple loading conditions. the intrinsic properties of optimal topology cannot be investigated. Stolpe (2010) showed that optimal truss with symmetric geometry and loading conditions may not be symmetric if crosssectional areas are discrete variables. 1994). 2004. however. in these approaches. It has been shown that the process of gradually increasing the penalization parameter converges to an approximate local optimal solution for the original 0–1 problem (Martínez 2005) and an exact local solution under certain conditions (Rietz 2001). then some ribs should be generated to result in a solution with reduced symmetry. Rozvany 2009). we investigate the properties of the optimal solution using the SIMP approach without additional penalty on the perimeter length or the gradient (slope) of density distribution. Therefore. Therefore. the uniqueness of the solution is strongly related to the symmetry of the solution. 2010). A solution path is defined with respect to the penalization parameter. the governing equations are differentiated. in most of the continuation methods for the plate (sheet) topology optimization problems (Rozvany et al. In the SIMP approach. Rietz (2007) obtained various topologies for different values of the gradient parameter. Note that it is important for optimization of a symmetric structure that the symmetry property of the solution should be investigated before assigning the types and parameter values for regularization. Optimal topology can be controlled by varying the perimeter length or the parameter value of the filter function. We first define local nonuniqueness of the KKT point as a bifurcation of the solution path with respect to the penalization . a KKT point is usually traced gradually increasing the penalization parameter utilizing the so called continuation method (Allgower and Georg 1993) to select an appropriate value of the penalization parameter as well as the parameters for density filter or penalty of gradients. (2009) utilized the phase field model in conjunction with the level set approach to control the complexity (number of holes) of optimal topology. there has been no investigation on the mechanism of symmetrybraking process of the optimal topology of an axisymmetric shell. and the solution path is traced along the path (Mittelmann and Roose 1990). the level set approach has been developed for simultaneous shapetopology optimization of plates and sheets (Allaire et al. Yamasaki et al. or penalized to have artificially large structural volume (Bruns 2005). Yamada et al. isotropic material with penalization) approach (Rozvany et al. and the solutions are found consecutively with increasing value of the penalization parameter. the governing equations are not differentiated. However. 2002. However.
. the solution process is unstable at the bifurcation point of the solution path. . then bifurcation of the solution path exists and the solution cannot be determined uniquely at the bifurcation point (Ikeda and Murota 2010. it ˜ is regarded as a function of t. Ohsaki and Ikeda 2007). The number of elements and the number of degrees of freedom are denoted by m and n. t) = 0 Rq Therefore. Let P ∈ Rn denote the specified nodal load vector. a condition for local uniqueness of the solution is derived as the singularity of the Jacobian of governing equations (Ohsaki 2006. . Since the solution is defined by (1) for each specified value of t. our approach for investigation of uniqueness of the optimal solution based on continuation method is applicable to any type of nonlinear programming problem. . j = 1. In this section. then the solution of (1) cannot be determined uniquely using (4). Ohsaki and Ikeda 2007). In contrast. respectively. . Hence. Then the nodal displacement vector U ∈ Rn is obtained from the following stiffness (equilibrium) equation: K(d)U(d) = P (6) If J is nonsingular. and is denoted by x(t). We next consider a process of finding the solution of (1) for a fixed value of t at t 0 from an arbitrary initial solution using Newton iteration. then the solution path can be uniquely found using an incrementaliterative method for the governing equations (1) in a similar manner as the arclength The objective function to be minimized is the compliance W (d) defined as W (d) = P U(d) (7) . In the numerical examples. then F is defined as the gradient of the potential. method for geometrically nonlinear equilibrium analysis. . if J is singular. Let d j denote the density of the jth element. Ikeda and Murota 2010. Then a path or a curve of the solution is defined parametrically in the space of the variables and parameter. (i. . Jog and Haber (1996) derived the conditions of stability using an incremental form of the variational problem for a minmax optimization problem. dm ) . However. if F is the governing equation of a reciprocal system that has a potential function. It is shown that a ribbed shell with reduced symmetry is generated through a bifurcation process of the solution path. . Then the equation for finding the increment x of x for satisfying (1) with linear approximation is written as F0 + J x = 0 (4) 2 Conditions for bifurcation of solution path and stability of solution of nonlinear equations Suppose a set of variables is found as a solution of nonlinear equations defined with a parameter. and J is a symmetric matrix. ( j = 1. (1) 3 Optimization problem and optimality conditions where t is the problem parameter. j)component Ji j is defined as Ji j = ∂ Fi . for which the (i. for a reciprocal system. Sup˜ pose we have a solution x(t 0 ) for t = t 0 . Furthermore. if J is singular.Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance parameter. . and F(x. a basic methodology of bifurcation analysis of solution path of nonlinear equations is summarized for the completeness of the paper and to clarify the relation between the stability of iterative solution process and the bifurcation (nonuniqueness) of solution path (Golubitsky and Schaegger 1979). for which the upper and lower bounds are assigned as 0 ≤ d j ≤ 1. the symmetryreduction process of the optimal solution as a function of the penalization parameter is studied in details. . . m) (5) where J ∈ Rq×q is the Jacobian of F. . Therefore. The formulation for numerical continuation with respect to the penalization parameter is rigorously derived by differentiating the KKT conditions and the stiffness (equilibrium) equations. Then. The stiffness matrix is denoted by K ∈ Rn×n . the Newton iteration using (4) diverges if J is singular. t) ∈ is continuously differentiable with respect to x and t. Then the solution for t = t 0 + t with a specified increment t of t can be estimated using the following linear approximation of the governing equations (1): ˜ J x(t 0 + ˜ t) − x(t 0 ) + ∂F t =0 ∂t (2) Consider a symmetric plate or shell discretized into finite elements. q) ∂x j (3) The design variable vector is given as d = (d1 . Let F0 denote the current value of F in the iterative process. Consider a problem of finding the solution x ∈ Rq of the nonlinear equations F(x.
is denoted by I . λ. μ ) U L m − KU − i=1 (1 − ε) pdi m p p−1 di Ki U = W (d) + λ i=1 m ¯ Ai h i [ε + (1 − ε)di ] − V m = (1 − ε) i=1 di ln di Ki U (15) + i=1 U μi (di − 1) + i=1 L μi (−di ) (10) By differentiating the volume constraint (9b) and multiplying 1/2. . . Therefore. Define G i (d) as G i (d) = −(1 − ε) pdi p−1 U Ki U + λAi h i (1 − ε) (12) Suppose the active side constraints remain active at the KKT point corresponding to the parameter value in the U L neighborhood of the current value. (i = 1. i. We use the SIMP approach to prevent gray solutions by penalizing the stiffness of an element with intermediate density. By differentiating (10) with respect to di and using (6–8). Bifurcation at a degenerate point may be investigated similarly using directional derivative of the solution. and μi are the Lagrange multipliers that have nonnegative values at the optimal solution. we have (1 − ε)Ai h i di = 0 i=1 (16) ∂L p−1 = − (1 − ε) pdi U Ki U ∂di U L + λAi h i (1 − ε) + μi − μi (11) where the standard approach of sensitivity analysis of compliance has been used (Choi and Kim 2004). the vector of state variables U and the Lagrange multiplier λ are also regarded as functions of p. transition of an inactive side constraint to be active is not considered. because we increase the parameter discretely and find the solution for each specified value of the parameter.. μ .R. By differentiating (6) with respect to p and using (8). m) (9c) ¯ where V is the specified upper bound of the total structural volume. μi . we can assume that the solution is generally a regular KKT point. For this purpose. The volume of the ith element is expressed as a product of the area Ai . Then the problem of minimizing the compliance under a constraint on the total structural volume is formulated with respect to the variable vector d as Minimize W (d) = P U(d) m (14) and we define s by s = I . of the optimal solutions are derived below. . Furthermore. thickness h i . we have m (9a) (9b) subject to i=1 ¯ Ai h i [ε + (1 − ε)di ] − V ≤ 0 0 ≤ di ≤ 1. where (·) indicates differentiation with respect to p. i. which are called parametric sensitivity coefficients for brevity. I = {i  0 < di < 1} where ε is a sufficiently small positive value for preventing numerical instability. . . Watada et al. respectively. The Lagrangian of problem (9) is formulated as L(d. and density di . it is shown in the numerical examples that this simple continuation method is practically effective for obtaining a symmetric optimal topology. μi > 0 and μi > 0 are satisfied for di = 1 and di = 0.. and accordingly G i (d) = 0.e. 4 Sensitivity of optimal solution with respect to penalization parameter Equations for computing the sensitivity coefficients with respect to p.e. we obtain 1 2 m U L where λ. The stiffness matrix K is defined with the matrix Ki ∈ Rn×n corresponding to the unit density of the ith element and the penalization parameter p as m Then the firstorder optimality conditions (KKT conditions) are derived as ⎧ ⎨ = 0 for 0 < di < 1 G i (d) ≤ 0 for di = 1 ⎩ ≥ 0 for di = 0 (13) K= i=1 ε + (1 − ε)di Ki p (8) The set of indices of elements satisfying 0 < di < 1. However.
the parametric sensitivity coefficients of a KKT point are uniquely determined. respectively. for the elements i ∈ I . we have di = 0. bs ) . Then the linear equations for computing the parametric sensitivity coefficients of the KKT point are written in the following form using the notations defined below: ⎛ ⎞ −K B12 0 ⎛ ⎞ ⎛ 1⎞ b ⎜ 12 ⎟ U 22 23 ⎟ ⎝d ⎠ = ⎝b2 ⎠ ⎜B (18) B B ⎠ 0 ⎝ λ 0 23 0 B 0 which is simply written as BX = b (19) which is incorporated into the second and third equations of (18) to obtain B22∗ B23 B23 0 d0 λ = b2∗ 0 (22) b2∗ = b2 + B12 K−1 b1 . Then the components of the symmetric matrix B ∈ R(n+s+1)×(n+s+1) and the constant vector b ∈ Rn+s+1 are given as Note that the elements of vector B23 have nonzero values as seen in (20c). the first equation of (22) can be solved for d0 as d0 = B22∗ −1 b2∗ − B23 λ (24) By incorporating this into the second equation of (22) and using positive definiteness of (B22∗ )−1 . and the ith components of B component of B23 ∈ Rs are denoted by Bi12 . and a null matrix for p = 1. Note that the matrix K may be nearly singular and the condition number of K−1 may be very large. . Therefore. no bifurcation occurs along a solution path if B22∗ is nonsingular. we obtain λ = B23 B23 B22∗ B22∗ −1 2∗ b −1 23 B Bi12 j = p−1 −(1 − ε) pd j k ji U (20a) (25) 1 p−2 22 Bii = − (1 − ε) p( p − 1)di Hi . because K−1 is positive definite and B12 has full column rank as discussed below. Therefore. . . In contrast. accordingly. . s}. 5 Uniqueness of local optimal solution for specified penalization parameter We can solve the first equation of (18) for U as U = −K−1 b1 + K−1 B12 d0 (21) 1 p−2 − (1 − ε) p( p − 1)di di U Ki U 2 1 + (1 − ε)Ai h i λ 2 1 p−1 = (1 − ε) pdi ln di U Ki U 2 1 p−1 + (1 − ε)di U Ki U. . . differentiation of G i (d) = 0 with respect to p leads to − (1 − ε) pdi p−1 U Ki U The path of the optimal solutions can be traced successively solving (19) (Ohsaki and Nakamura 1996). When B22∗ is positive definite. . We can see from (20b) that B22 is a diagonal negative definite matrix for p > 1. Bi22 . if a small value is assigned for ε. Since B is symmetric. (i ∈ I ) 2 (17) For i ∈ I . and λ . λ is uniquely determined. The indices of elements are rearranged so that I = {1. ds ) . . and (17) for n + s + 1 variables U . Therefore. there are n + s + 1 / linear equations (15). j)12 ∈ Rn×s . and define d0 = (d1 . di (i ∈ I ). . and its lowest eigenvalue may decrease to 0 as p is increased. the conditions for uniqueness of the solution have been derived using a simple and rigorous manner based on the uniqueness of a solution path. B22 ∈ Rs×s . and d0 is also uniquely found using (24) and (25). 2 Bi22 = 0 for i = j j Bi23 = 1 (1 − ε)Ai h i 2 m (20b) (20c) (20d) b1 = (1 − ε) i=1 di ln di Ki U p bi2 = 1 p−1 (1 − ε)(1 + p ln di )di Hi 2 (20e) Hence. . The (i. B22∗ = B22 + B12 K−1 B12 (23) 2 2 Let Hi = U Ki U and b2 = (b1 . . the stability of solution is detected from the eigenvalues or the condition number of the matrix. and j j Bi23 . This way. singularity of K leads only to nonuniqueness of the displacements that may not have any . . The jth column of Ki is denoted by ki j ∈ Rn . the matrix B12 K−1 B12 is positive definite.Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance Hence. (16). the symmetric matrix B22∗ is positive definite at p = 1. However.
1 A spherical shell model effect on nonuniqueness of the density variables (Koˇ vara c and Outrata 2006). 1 41 81 121 161 Fig.80. is given as Ri = −(1 − ε) pdi p−1 fi (26) (a) p = 1.R. In fact.2148 × 10 −5 Fig. the large eigenvalue of K−1 . W = 1.94.7819 × 10 −5 (c) p = 1. Watada et al.96. W = 2. W = 2. 2 Optimal solutions for various values of parameter p .4528 × 10 −5 (b) p = 1. corresponds to the eigenmode that has large deformation at the nodes connected by the elements with low density only. Let fi ∈ Rn denote the vector of equivalent nodal loads of the ith element with di = 1 corresponding to the displacement vector U. denoted by Ri ∈ Rn . W = 1.78.00.82.7935 × 10 −5 (d) p = 1. W = 1.2391 × 10 −5 (f) p = 2. if exists. W = 1. the ith column of B12 . Using (20a).7947 × 10 −5 (e) p = 2.
then B23 = 0 and B23 is included in the range of B22∗ .0 0 1.5 2 2. . i.5 p 2 ( 10 6) 2.Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance Hence. Suppose there exists a singular point where the lowest eigenvalue of B22∗ vanishes. the (i. i.0 0.5 1.0 0. .8 2. The matrix B12 K−1 B12 in (23) is denoted by S ∈ Rs×s . Ri (i = 1. and (23) that b2∗ has the same symmetry property as 3. Then. and B12 has full column rank.0 1.6 1. and B23 has the same symmetry property as the solution with uniform density.8 1.0 2.. Let denote the eigenvector corresponding to the zero eigenvalue of B22∗ .9 2 (a) 1 p 3 (b) 1.0 1. .6 2. the kernel of the matrix B22∗ is given as c with arbitrary coefficient c. for the constant parameter value. corresponding to the increment δp of the parameter: B22∗ B23 B23 0 δd0 δλ = δp b2∗ 0 (28) Therefore.5 0.5 0 1 1.5 1.0 1. there exists a solution δd0 = β and δλ = 0 with β being an arbitrary nonzero value.9 3 (c) 2. In this case. .5 0 2.7 2.5 1.5 Fig.. j)component Si j of S is written as follows using the displacement vector U j ∈ Rn against F j : Si j = U j Ri (27) δλ of d0 and λ.5 2. The rate form (22) is converted to an incremental form as follows for the increments δd0 and As seen from (20c).5 3 1. respectively. It is easily observed from (20d).e.e. δp = 0.5 ( 10 5) ( 10 6) 2. (20e). the matrix S = B12 K−1 B12 is bounded if the work in the jth element ( j ∈ I ) done by the nodal loads that is proportional to the equivalent nodal loads of the ith element (i ∈ I ) is bounded. This condition is usually satisfied because di has moderately large value for the elements in I . 3 Eigenvalues of the matrix B22∗ p 3 . s) are generally independent. If has a lower symmetry property than B23 .5 2. the elements of vector B23 has nonzero values.7 1.
then there exists a parif b ticular solution that may have nonzero value of δp.. From the stationary conditions of the Lagrangian of problem (29).78 (b1) p = 2. 2∗ is included in the range of B22∗ . The optimization problem is formulated as a maxmin problem: 1 Find max min U KU − P U U 2 d m where γ is the penalty parameter.94 (b3) p = 2. Therefore. i. ( j = 1. . where G i (d) in (13) is to be replaced by G i (d) = pdi p−1 U Ki U + 2λ − 2γ (1 − 2di ) (30) The incremental equations are derived in the following form for solving the equilibrium equations and optimality conditions using Newton iteration: K B12 B12 B22 δU δd0 = b1 b2 (31) −λ i=1 ¯ di − V m −γ i=1 di (1 − di ) (29a) (29b) subject to 0 ≤ d j ≤ 1. 4 Eigenvectors of B22∗ for p = 1. m) with appropriate modification of matrices and vectors due to incorporation of the penalty term γ di (1 − di ) and the differences in the signs of the strain energy in the objective function.94 . i. . . if b2∗ = 0. λ is fixed for some reason in the iterative process.78 and 2. Hence.94 Fig.R. . (a) p = 1. we summarize their results in a matrixvector form. In the following. the current optimal solution.e. They first used continuum formulation that is reduced to a finite element formulation..e. we obtain the equilibrium equation (6) and the optimality conditions (13) for the elements in I .94 (b2) p = 2. bifurcation of solution occurs when B22∗ becomes singular. Note that the increment of λ is not considered in (31). Watada et al. Jog and Haber (1996) derived the conditions of uniqueness and stability of the optimal solution based on incremental form of variational problem.
The geometrical parameters are R = 40. 5 Variations of condition numbers (d) K . Optimization is carried out using SNOPT Ver.3.5 3 1 1.01 1 1.5 2 2. uniqueness and symmetry of optimal solution is investigated for an axisymmetric shell.0 kN. the maximum thickness is h i = 0. H = 11. The small density parameter for ε for preventing numerical instability is 1.5 2 2. The design domain is discretized to 10 and 20 elements in longitudinal and circumferential directions. θ = π/3. i. The tolerance of optimality is also small enough so that optimization is carried out until no improvement is achieved.5 2 2. However. and also with respect to rotation of one of n different angles around the axis of symmetry. and P = 1. The values in the figures are those without scaling. 7 (Gill et al.. then it has the dihedral symmetry Dn . 1 subjected to the vertical concentrated load P at each node on the top ring.5 3 (a) B 1000 22 (b) B 10 8 22* 10 7 100 10 6 10 5 10 1 10 4 1. which are omitted for brevity. 1 1000 100 0.5 2 2. The default values are used for the parameters except the strict tolerance 10−12 for feasibility. where the sequential quadratic programming (SQP) is used. ¯ and the upperbound volume V is 50% of the value of the solution with di = 1 for all elements. respectively.e. Poisson’s ratio is 0. In the following. the objective function is multiplied by 109 to prevent numerical difficulty.Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance Then. the units of length and force are m and kN. 2002).5 m.5 m.5 3 1 1 1. the conditions of uniqueness of the solution are: K is positive definite and B22 is negative definite. Consider a spherical shell as shown in Fig. Young’s modulus is 2. 6 Numerical examples As a numerical example. The symmetry of the solution is indicated using the standard notations of group theory (Kettle 2007).0 × 10−6 .5 3 (c) B 12 B 12 Fig. the total number of elements is 400.10 × 108 kN/m2 . respectively. and α = π/12. If the solution does not change after application of reflection with respect to one of n different planes containing the axis of symmetry.1 10 0.0 m.
Accordingly. The values of the objective function W are also shown. because the objective function is convex and the volume constraint is linear with respect to d. 41.5 3 1.78 and 1. 4a.0 and 3. 2a and b.0000.0 with the increment p = 0.78 and 1.05133.0. the optimal densities of the elements 1. and d161 = 0. The sensitivity coefficients are obtained as d1 = d41 = 0. 121.1 2. Optimization in the parameter range 1. d81 = 0.0 100 1.01 are d1 = d41 = 0.80 is proportional to the eigenvector corresponding to the zero eigenvalue of the optimal solution for p = 1. and objective function W .58805. 3.5 2 2. For p = 1.33348. d121 = 0.21647.7 50 1. Figure 2 shows the distributions of di of optimal solutions for various parameter values in gray scale.78 leads to axisymmetric solutions as shown in Fig.3226. 81. 1 are d1 = d41 = 1. Lagrange multiplier λ.78.8 1.6 1. As seen in Fig. The parametric sensitivity coefficients of optimal solutions are first verified. d121 = 0. d81 = −0.3232. which have D20 symmetry. and d161 = 0.5 3 (a) d0  200 2.0.R. d121 = 0. The sensitivity coefficients obtained by the forward finite difference approach with p = 0. which have good agreement with the analytical results. by tracing a solution path assigning the solution of p − p as the initial solution for the SQP algorithm.80 in Fig. the eigenvalues of B22∗ of the KKT point are plotted in Fig.5 0 1 1. and d161 = 0. Watada et al. Local optimal solutions are found for the parameters between p = 1.9 1. cardinality s of I . d81 = −0. the symmetry of solution for p = 1.5 3 (c) s (d) W Fig.5 3 ( 10 7 ) 12 10 8 6 4 2 0 1 ( 10 5 ) 1. 3b.06640.80. 22 20 18 16 14 12 10 8 6 4 2 0 1 1. 2c is reduced to D10 .06652.5 2 2. 6 Variations of the norm of parametric sensitivity vector d0 . and the difference between the two solutions for p = 1.05138. For each parameter value.5 2 2.2 (b) 150 2. which has D10 symmetry as shown in Fig.4 1 1.02. and 161 indicated in Fig.3 2.00 ≤ p ≤ 1. The same axisymmetric solution is found from any initial solution for p = 1. the matrix B22∗ becomes singular with a zero eigenvalue between p = 1.5 2 2.
however. and K are plotted in logscale in Fig. 4. 2d and e.01. the condition numbers of B22 and B∗ are 0. 6. the total number of elements is 600.78 at which B22∗ has very small positive eigenvalue. although the condition number of K is very large.0.82 ≤ p ≤ 2. and the solutions in the parameter range 1. By contrast. B22∗ .0. Since the shape is fixed and only the topology is varied. The condition number of K jumps to a very large value at the singular point of B22∗ . while (b2) does not have any symmetry property. It is also seen from Fig. or nonuniqueness of the displacement vector U is not related to the bifurcation of the optimal solution. where (b1) and (b3) have D5 and D1 symmetry. respectively. respectively. respectively. the parameter p is increased to 3.82 has D5 symmetry as shown in Fig. B12 B12 . the mechanism of resisting external loads distributed around the top ring did not change as an result of increasing the penalization parameter. 5 that the n × s matrix B12 has full column rank s (< n). direct optimization . As is seen. For p = 2.0 from the uniform initial (b) Fig. and the symmetry of the solution reduces from D5 to D2 as shown in Fig.94 ≤ p ≤ 2. if we optimize directly for p = 3. 7a. some gray elements exist near the boundary. 7 Optimal topology of 15 × 40 model with p = 3. Lagrange multiplier.94. The optimal topology obtained in this continuation approach is shown in Fig. however. which is very large.e. Finally.Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance The condition numbers of B22 . More symmetrybreaking singular points can be found by further increasing the parameter p.2469 and 22 9. i. The lowest eigenvalue of B22∗ becomes zero between the parameter range 2.5 for all elements with p = 1. 2d. B22∗ has moderately small condition number in the range p ≥ 1. For p = 2. for which the eigenvectors are shown in Fig.96. i. which are sufficiently small.94 have the same symmetry property.0 with the increment 0. the condition number of K is 6.913 × 107 . 5. U is nearly nonunique. the lowest eigenvalues of B22∗ are threefold. a topology with D10 symmetry has been obtained. The solution for p = 1. cardinality s of I . a continuation approach. (a) instability of the structure. and objective function are plotted in Fig.e. because B22∗ remains to be positive definite. b direct optimization Fig.. optimal topologies are found for a finer mesh with 15 and 40 elements in longitudinal and circumferential directions. The norm of parametric sensitivity vector. In contrast. the condition number of B22∗ and the norm d0  of parametric sensitivity vector drastically increase around p = 1. Starting with the uniform initial design di = 0.. 8 Optimal topology of 15 × 40 model with p = 4. As is seen. There also exist other points at which the lowest eigenvalue of the matrix B22∗ and the condition numbers of the matrix B22 and B22∗ jump discontinuously due to the variation of the size s of these matrices. Therefore.80.413.
Springer. Jouve F. Convergence of optimization algorithm and uniqueness of solution path are defined by the same condition of the singularity of a matrix similarly to geometrically nonlinear analysis of structures. then the topology is obtained as shown in Fig. Sigmund O (2003) Topology optimization: theory. stiffness (equilibrium) equations. Int J Numer Methods Eng 24:835–852 Bendsøe MP (1989) Optimal shape design as a material distribution problem. J Comput Phys 104:363–393 Allgower EL. 8a. We used a shell structure. Note that the proposed method is effective for any problem with symmetry including a circular disc subjected to a twisting force at the center circle.0 for 341 elements. 7 Conclusions A simple formulation has been presented for investigating the path of the firstorder optimal solution of an axisymmetric shell that minimizes the compliance under specified load and the total structural volume. 5. A procedure has also been developed for condensing Jacobian to remove the derivatives of displacements and separate the nonuniqueness properties of displacements and design variables. because the eigenvalue analysis should be carried out for the matrix of the size equal to the variables that have intermediate values. 4. Watada et al.R. 7b is obtained. Acta Numer 2:1–64 Belblidia F. which is often used as an example of Michell structure. The symmetry reduction process of the KKT point can be defined as an bifurcation of solution path with respect to the penalization parameter of SIMP approach. no clear symmetry is observed. It has been shown quantitatively that symmetry of KKT point reduces through bifurcation to the critical eigenvector of the Jacobian of the governing equations in the continuation process of KKT point using SIMP approach. have been computed to show that nearly singular stiffness matrix does not lead to nonuniqueness of the KKT point. As is seen. The numerical example of an axisymmetric shell shows that a solution path of an axisymmetric shell has a bifurcation point where the Jacobian of the governing equations is singular. Note that the symmetry properties without filter can be used for appropriately setting the parameters for filters and other regularization. because the elements around the boundary have larger area than those around the center. C R Acad Sci 334:1125–1130 Allaire G. solution. A procedure has been developed for detecting bifurcation of the solution and the direction of reduction of symmetry using eigenvalue analysis of the Jacobian of the governing equations. then the topology as shown in Fig. The KKT point is nonunique at the bifurcation point.0 for 249 elements. References Allaire G. Jouve F. An almost clear 0–1 solution has been found with di = 1. Although the solution with intermediate density has no practical meaning. The conditions of nonuniqueness of the solution are derived based on a bifurcation of the solution path with respect to the penalization parameter of the SIMP approach. Comput Methods Appl Mech Eng 71:197–224 Bendsøe MP. Toader AM (2004) Structural optimization using sensitivity analysis and a levelset method. and a symmetrybreaking bifurcation path exists in the direction of the eigenvector corresponding to the zero eigenvalue analyses of the Jacobian. the symmetryreduction process of the KKT point is characterized as a bifurcation process of a solution path with respect to the penalization parameter. and volume constraint. The formulation for numerical continuation with respect to the penalization parameter is rigorously derived by differentiating the KKT conditions. Georg K (1993) Continuation and path following. The condition numbers of the matrix for computing the sensitivity coefficients of the KKT points. The optimal topology for p = 4. 2. Toader AM (2002) A level set method for shape optimization. 8b. Kikuchi N (1988) Generating optimal topologies in structural design using a homogenization method.0 is not equal to the half of the total number elements. The main results are summarized as follows: 1. Berlin Bruns TE (2005) A reevaluation of the SIMP method with filtering and an alternative formulation for solid–void topology optimization. a unified approach has been presented for investigation of uniqueness and bifurcation (symmetrybreaking process) of KKT point.0 from the uniform initial solution. and the corresponding submatrices. and it can be applied to a largescale problem without any difficulty.968 for 10 elements. because it has high symmetry.0 using continuation approach is shown in Fig. Nonuniqueness of displacement does not have any effect on nonuniqueness of the KKT point. Struct Multidisc Optim 30:428–436 . Thus. Bulman S (2002) A hybrid topology optimization algorithm for static and vibrating shell structures.966 ≤ di ≤ 0. Struct Optim 1:193–202 Bendsøe MP. and di = 0. 0. methods and applications. A local optimal solution with appropriate symmetry can be found through investigation of symmetry using continuation approach. This way. 3. If we optimize directly for p = 4. Furthermore. the method proposed in this paper can also be used for the case where a filter is used. which has no clear symmetry. it is important to investigate the symmetry breaking process of the KKT point on the basis of bifurcation of a solution path. The number of elements with di 1.
Struct Multidisc Optim 29:319–323 Maute K. Ikeda K (2007) Stability and optimization of structures— generalized sensitivity analysis. Published online. Sigmund O (1994) Topology optimization. Eng Opt 26: 171–186 Ohsaki M.1007/s0015801005640 Rozvany GIN. Chichester Koˇ vara M. Wiegmann A (2000) Structural boundary design via levelset and immersed interface methods. AIAA J 35:1767–1773 Mittelmann HD. Comput Methods Appl Mech Eng 171: 121–140 Petersson J. Sato K. Yoshimura M. Nakamura T (1996) Minimum constraint perturbation method for topology optimization of systems. Ohsaki M. SIAM J Numer Anal 36(6):1759–1778 Petersson J (1999b) Some convergence results of perimetercontrolled topology optimization. Int J Numer Methods Eng 67:132–147 Ohsaki M. New York Petersson J (1999a) A finite element analysis of optimal variable thickness sheets. Haftka RT (1989) Modern homotopy methods in optimization. New York Jog CS. In: Proc. Schaegger D (1979) A theory for imperfect bifurcation via singularity theory. McGraw Hill Gill PE. Optim Eng 7:201–219 Martínez JM (2005) A note on the theoretical convergence properties of the SIMP method. Roose D (eds) (1990) Continuation techniques and bifurcation problems. Lisbon Yamasaki S. Saunders MA (2002) SNOPT: an SQP algorithm for largescale constrained optimization. Birker T (1992) Generalized shape optimization without homogenization. Izui K. parametric programming. Comput Methods Appl Mech Eng 196:4430–4443 Choi KK. Nishiwaki S. Birkhäuser. Ohsaki M (1988) Sequential optimal truss generator for frequency ranges. Murota K (2010) Imperfect bifurcation in structures and materials—engineering use of grouptheoretic bifurcation theory. Comput Methods Appl Mech Eng 93: 291–318 Wang MY. doi:10. Optim Eng 2:293–320 Kettle SFA (2007) Symmetry and structure: readable group theory for chemists. Springer. SIAM J Optim 12: 979–1006 Golubitsky M. Int J Numer Methods Eng 41:1417–1434 Rietz A (2001) Sufficiency of a finite exponent in SIMP (power law) methods. Kikuchi N (1991) A homogenization method for shape and topology optimization. Struct Multidisc Optim 41:685–698 . Academic Gal T (1979) Postoptimal analyses. Struct Multidisc Optim. Kim NH (2004) Structural sensitivity analysis and optimization. Struct Multidisc Optim 24:407–417 Nakamura T. Murota K. Murray W. Struct Optim 4:250–252 Rozvany GIN. paper no. Chapman and Hall. and related topics. Wang XM. Ryvkin M (2003) Topology design of modular structures under arbitrary loading. Comput Methods Appl Mech Eng 74(3):289–305 Yamada T. Haber RB (1996) Stability of finite element models for distributedparameter optimization and topology design. 8th world congress of structural and multidisciplinary optimization (WCSMO8). Fuchs MB. Svanberg K (2001) On the trajectories of the epsilonrelaxation approach for stressconstrained truss topology optimization. Ramm E (1997) Adaptive topology optimization of shell structures. Sigmund O (1998) Slope constrained topology optimization. 1136. Comput Methods Appl Mech Eng 67(2): 189–209 Ohsaki M (2006) Local and global searches of approximate optimal designs of regular frames. London Sethian JA. Mechanical engineering series. Springer. Linear systems. Comput Methods Appl Mech Eng 192:227–246 Watson LT. Zhou M. Guo DM (2003) A level set method for structural topology optimization. Zhou M. Applied mathematical sciences.Nonuniqueness and symmetry of optimal topology of a shell for minimum compliance Bruns TE (2007) Topology optimization by penalty (TOP) method. Katoh N (2001) Group symmetry in interiorpoint methods for semidefinite program. Commun Pure Appl Math 32:21–98 Ikeda K. 2nd edn. In: Adeli H (ed) Advances in design optimization. Comput Methods Appl Mech Eng 130:203–226 Kanno Y. vol 149. Struct Multidisc Optim 21:159–163 Rietz A (2007) Modified conditions for minimum of compliance. Springer. Kawamoto A. Nishiwaki S (2010) A level set based topology optimization method using the discretized signed distance function as the design variables. Nomura T. Takezawa A (2009) A new level set based topology optimization method using a fictitious interface energy model based on phase field method concepts. Chapter 10. Izui K. Struct Multidisc Optim 41:661–670 Stolpe M. J Comput Phys 163(2): 489–528 Stolpe M (2010) On some fundamental properties of structural topology optimization problems. Struct Multidisc Optim 37: 217–237 Rozvany GIN (2011) On symmetry and nonuniqueness in exact topology optimization. Basel Moses E. Comput Methods Appl Mech Eng 196:4413–4418 Rozvany GIN (2009) A critical review of established methods of structural topology optimization. Wiley. Struct Multidisc Optim 21:140–151 Suzuki K. New York Fiacco AV (1983) Introduction to sensitivity and stability analysis in nonlinear programming. vol 1. Outrata JV (2006) Effective reformulation of the truss c topology design problem.