Computers and Structures 157 (2015) 142–152

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: www.elsevier.com/locate/compstruc

Topology optimization of continuum structures under buckling

constraints
Xingjun Gao, Haitao Ma ⇑
State Key Laboratory of Subtropical Building Science, Department of Civil Engineering, South China University of Technology, Guangzhou 510640, China

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a study on topology optimization of continuum structures under buckling
Received 15 January 2015 constraints. New algorithms are developed for minimization of structural compliance considering
Accepted 19 May 2015 constraints on volume and buckling load factors. The SIMP (Solid Isotropic Material with Penalization)
Available online 8 June 2015
material model is employed and nodal relative densities are used as topology design variables. A new
approach based on the eigenvalue shift and pseudo mode identification is proposed for eliminating the
Keywords: effect of pseudo buckling modes. Two-phase optimization algorithms are also proposed for achieving bet-
Topology optimization
ter optimized designs. Numerical examples are presented to illustrate the effectiveness of the new
Buckling constraints
Pseudo buckling modes
methods.
Pseudo mode identification Ó 2015 Elsevier Ltd. All rights reserved.
Two-phase optimization algorithms

1. Introduction geometrical stiffness matrices of elements with densities and prin-

cipal stresses smaller than predefined threshold values.
Structural strength, stiffness, and stability are three of the Meanwhile, they indicated that the predefined values might have
important factors considered for assessing the design of a struc- a significant influence on optimization results. Bendsøe and
ture. Naturally, it is important to consider structural stability in Sigmund [18] pointed out that doing this might cause solution
the optimization process. Recently, buckling optimization has oscillations due to abrupt changes of objective functions and sen-
drawn more research attention. sitivities. In order to avoid the discontinuity caused by such a
For trusses, frames, and other built-up structures consisting of cut-off method, they suggested the use of different penalization
bars and beams, much work has been done to consider the stability schemes for element stiffness matrix and geometric stiffness
requirements in structural optimization, such as size optimization matrix. Currently this method appears to be a standard solution
of trusses and frames [1], shape optimization of columns, truss or for this problem and has been used by many researchers, e.g.
built-up structures [2–4] and topology optimization of truss struc- Lindgaard and Lund [19]. However, Zhou [20] showed that it might
tures [5–7]. be difficult to select an appropriate parameter value for the expres-
Neves et al. [8] and Min and Kikuchi [9] have considered struc- sion of penalization in calculating accurate buckling load factors.
tural stability in topology optimization of continuum structures. Pseudo eigenmodes may also appear in the optimization of
They investigated the reinforcement of a structure to increase its eigenfrequencies in vibration problems [21]. To eliminate these
overall stability. Neves et al.[10] extended their earlier work to pseudo modes, some methods of modifying element stiffness
the buckling optimization of periodic material micro-structures. matrix and/or mass matrix in low-density regions have been pro-
Geometrically nonlinear models have also been introduced into posed and details of these methods can be found in the research lit-
the optimization of continuum structures against buckling erature, e.g. [21–23]. A topology optimization problem considering
[11–15]. In addition, optimization of composite structures was buckling differs from the one considering vibration modes and is
considered by Lindgaard and Lund [16,17]. more complex as element geometrical stiffness matrices are
A common problem in topology optimization using the SIMP dependent on element stresses, which depend both on the struc-
material model is the appearance of pseudo buckling modes in ture itself and on the loading condition. In contrast, element mass
low-density regions. Neves et al. [8] suggested ignoring the matrices used in frequency analysis are dependent on material dis-
tribution only.
In this paper, the pseudo buckling mode problem is investigated
⇑ Corresponding author. and a new method combining eigenvalue shift and pseudo mode
E-mail addresses: gaoxj85@126.com (X. Gao), maht@scut.edu.cn (H. Ma). identification is proposed. An optimization formulation for

http://dx.doi.org/10.1016/j.compstruc.2015.05.020
0045-7949/Ó 2015 Elsevier Ltd. All rights reserved.
 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152 143

minimizing the structural compliance under material volume and 2.2. Material interpolation
buckling load factor constraints is used in the study.
This paper is organized as follows: In Section 2 the optimization It is possible to obtain continuous material distributions by
formulation and material model used are presented. In Section 3, using nodal relative densities as topology design variables [24]. It
the finite element model employed for the structural analysis is is noted that Kang and Wang [25] have presented a more general
introduced. In Section 4, expressions for the sensitivity of con- density interpolation strategy for topology optimization using
straint functions and objective function are derived. In Section 5, nodal design variables and Shepard interpolation. In this study, a
some of the existing methods for dealing with pseudo buckling more conventional interpolation scheme based on element nodal
modes are briefly discussed and a new approach is proposed. In values and shape functions is used. Within the eth element, the rel-
Section 6, new optimization algorithms are developed. In ative density distribution is expressed as
Section 7, two numerical examples are presented to demonstrate X
NN

effectiveness of the proposed methods. Finally, concluding remarks qe ðx; yÞ ¼ Nk ðx; yÞqek ð2Þ
k¼1
are made.
where qek denotes nodal density value at the kth node of the ele-
ment, NN is the number of nodes in the element, and N k ðx; yÞ is
the element shape function for the kth node.
2. Problem formulation and material interpolation scheme Using the SIMP material model, the elasticity matrix at point
ðx; yÞ is expressed in terms of material relative density qe ðx; yÞ
2.1. Optimization problem formulation p
Eðx; yÞ ¼ ½qe ðx; yÞ E0 ð3Þ
The topology optimization of continuum structure may often where E0 is the elasticity matrix of the isotropic solid elastic mate-
generate designs with slender components when the allowed rial, and p P 1 is a penalization exponent number.
material volume fraction is small. If compressive stresses occur
in these structural components, structural buckling may present 3. Finite element analysis methods
serious safety concerns. Therefore, structural stability require-
ments should be considered in the optimization. The mathematical In this section, the finite element model for structural analyses
formulation of the compliance minimization problem of contin- and the computation of buckling load factors using hybrid stress
uum structures with constraints on the material volume and buck- element is briefly introduced.
ling load factors can be stated as
3.1. Finite element model
find q ¼ fq1 ; q2 ;    ; qN g
When the nodal design variable is employed, the checkerboard
min C ¼ FT U ¼ UT KU
patterns can be avoided naturally. However, a ‘‘layering’’ or ‘‘is-
s:t: KU ¼ F landing’’ phenomenon of black and white regions in the design
  ð1Þ
minkj  P k > 0 domain may appear [24]. Deng et al. [26] showed that this problem
j2J
could be effectively avoided by replacing the conventional
V ðqÞ 6 V 0 four-node displacement-based quadrilateral element with a hybrid
0 < q 6 qi 6 1 i ¼ 1; 2; . . . ; N stress element. The same approach is taken in this study, and in
this section, the basic theory and formulation of the hybrid stress
element to be used will be summarized.
where qi ði ¼ 1; 2; . . . ; NÞ are design variables of relative material
Pian and Sumihara [27] developed a four-node hybrid stress
density; N is the number of design variables; C is the structural
finite element for homogeneous plane problems. Independent ele-
compliance; U and F are the global displacement and force vectors;
ment stress and displacement fields are defined and can be
kj is the jth buckling load factor corresponding to the given load
expressed as
cases; J is a set of indices of the buckling mode considered in the  T
optimization; k denotes the lower bound of buckling load factors; r ¼ rx ; ry ; sxy ¼ Ub ð4Þ
V ðqÞ is the total material volume of the structure; V 0 is the upper  T
bound of material volume; and q is the lower bound of design vari- u ¼ ux ; uy ¼ Nd ð5Þ
ables, e.g. q ¼ 0:001. where rx ; ry and sxy are stress components, ux and uy are displace-
Through the introduction of an explicit constraint condition on ment components, U and N are interpolation matrices for element
buckling load factors, designs that fail to satisfy stability require- stress and displacement fields, respectively, b is an element stress
ments will be excluded from the feasible solution set. parameter vector, and d is the nodal displacement vector.
Theoretically, different levels of safety margins can be achieved Based on the Hellinger–Reissner variational principle, the fol-
by using different lower bound values. For example, if k ¼ 1, the lowing expressions for element stiffness matrix Ke and the stress
optimized structure will be at a critical state under normal service parameter vector b can be derived
conditions; if k > 1, the structure will be stable under normal ser-
vice conditions with a bigger safety margin for a bigger k; if
Ke ¼ GTe H1
e Ge ð6Þ
0 < k < 1, the structure may buckle under normal service condi-
tions, but cannot be a mechanism. b ¼ H1
e Ge d ð7Þ
The buckling mode index set J is introduced for two reasons. where matrices Ge and He are defined as
Firstly, when an applied load always points in the same direction, Z 1 Z 1
negative loading factors are meaningless and in this case, set J Ge ¼ UT BjJjdndg ð8Þ
should contain only the modes with positive load factors. 1 1

Secondly, when pseudo modes are among the calculated buckling Z 1 Z 1

modes, the corresponding mode indices must be excluded from jJj
He ¼ UT S0 U dndg ð9Þ
set J as these modes are not real and should be ignored. 1 1 ½qe ðn; gÞp t 0
144 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152

where J is Jacobian matrix, jJj is the determinant of J; B is the strain– @C @K XNe

@Ke
displacement matrix, S0 is the inverse of elasticity matrix E0 , i.e. ¼ UT U ¼  ðue ÞT ue ð19Þ
@ qi @ qi e¼1
@ qi
S0 ¼ E1
0 , and t 0 is the thickness of the structure.
where Ke and ue are the stiffness matrix and the displacement vec-
3.2. Element geometric stiffness matrix tor of element e, respectively, Ne is the number of elements used to
discrete the design domain.
For a plane continuum structure, the element geometric matrix From Eq. (6), we get
can be expressed as [28]
@Ke @H1
Z Z " # ¼ Ge e
Ge ð20Þ
1 1
S 0 @ qi @ qi
KGe ¼ gT gjJjdndg ð10Þ
1 1 0 S Then differentiating the identical equation He H1
e ¼ I with respect
to qi , we can get
where submatrix S and matrix g are defined as
  @H1 @He 1
rx sxy e
¼ H1 H ð21Þ
S¼ ð11Þ @ qi e
@ qi e
sxy ry
From Eq. (9), we can get the following expression
g ¼ C½ M1 M2 M3 M4  ð12Þ Z 1 Z 1
@He p  jJj @ qe ðn; gÞ
¼ UT S0 U dndg ð22Þ
in which C and Mi ði ¼ 1; 2; 3; 4Þ are matrices defined below @ qi 1 1 ½qe ðn; gÞpþ1
t0 @ qi
" #
J1 0 where
C¼ ð13Þ
0 J1
@ qe ðn; gÞ 0 if i is not adjacent to element e
¼ ð23Þ
2 3T @ qi Nk ðn; gÞ if i is the kth node of element e
@Ni @Ni
@n @g
0 0
Mi ¼ 4 @N i @N i
5 ð14Þ Then using Eqs. (19)–(22), we can obtain the sensitivity of the com-
0 0 @n @g pliance with respect to the nodal design variables.
From Eqs. (4) and (7), the stress vector can be calculated as follows
4.2. Sensitivity of buckling load factors
re ¼ UH1
e Ge d ð15Þ
Introducing an auxiliary variable jj ¼ 1=kj [29], Eq. (18) can be
Now, with a mapping function H : R3 # R44 defined as rewritten as
2
02 31 rx sxy 0 0 3  
KG ðq; UÞ  jj KðqÞ wj ¼ 0 ð24Þ
rx 6s
6 xy ry 0 0 7
HB6 7C
@4 ry 5A :¼ 6
7
7 ð16Þ Note that for the current optimization problem, both of the two glo-
4 0 0 rx sxy 5
sxy bal stiffness matrices are functions of the current design q and the
0 0 sxy ry global geometric stiffness matrix depends also on U, the nodal dis-
the element geometric matrix in Eq. (10) can be rewritten as placement vector for the given loading case.
Z Z If the eigenvalue is unimodal, the sensitivity of auxiliary vari-
1 1
KGe ¼ gT Hðre ÞgjJjdndg ð17Þ able jj with respect to variable qi can be expressed as [29]
1 1

@ jj ~ T @KG 1 @K ~ @K
¼ wj  wj  v Tj U ð25Þ
@ qi @ qi kj @ qi @ qi
3.3. Determination of buckling load factors
where w ~ j is the eigenvector, and v j is the adjoint displacement vec-
The linear buckling load factors can be calculated from the fol- tor. Adjoint displacement vectors are determined by solving the fol-
lowing equation lowing equation
 
K þ kj KG wj ¼ 0 ð18Þ Kvj ¼ Pj ð26Þ
where K is the global stiffness matrix of the structure, KG is the glo- where
bal geometric stiffness matrix, and kj and wj are the jth buckling 8 @K 9
> ~T G w
w ~ >
load factor and the corresponding buckling mode vector. >
> j @u1 j >
>
< =
Normally the buckling modes are ordered according to the mag- ~ T @KG w
Pj ¼ w ~j ¼ .. ð27Þ
j
@U > . >
nitudes of buckling load factors and k1 will be the smallest. >
> ~ T @KG ~ >
>
: wj @u wj ;
d

4. Sensitivity analysis in which d is the number of degrees of freedom of the structure.

Note that the eigenvectors must satisfy the orthonormalization con-
In this section, the sensitivities of compliance, buckling load ~ k ¼ djk ; djk is Kronecker’s delta.
~ T Kw
dition, i.e. w j
factors and volume with respect to nodal design variables are
By differentiating Eq. (17), we get
derived.
Z 1 Z 1
@KGe @ Hðre Þ
4.1. Sensitivity of compliance ¼ gT gjJjdndg ð28Þ
@ qi 1 1 @ qi

When the applied loads are independent of design variables, the Since mapping H is linear, the derivatives of Hðre Þ with respect to
sensitivity of structural compliance can be expressed as qi can be expressed as
 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152 145

 !
@ H UH1 where @ qe ðn; gÞ=@ qi is given in Eq. (23).
@ Hðre Þ e Ge ue @H1
¼ ¼ H U e Ge ue ð29Þ
@ qi @ qi @ qi
5. Methods for dealing with pseudo buckling modes
The sensitivity of the global geometrical stiffness matrix can then be
given as
Pseudo buckling modes usually appear in low-density regions
X
Ne Ne Z
X 1 Z 1 during the optimization process. In this section, some existing
@KG @KGe @ Hðre Þ
¼ ¼ gT gjJjdndg methods for dealing with this problem are first investigated
@ qi @ qi
e¼1 e¼1 1 1
@ qi
! through a simple example, and then a new approach is proposed.
XNe Z 1 Z 1
@H1
¼ gT H U e Ge ue gjJjdndg ð30Þ
e¼1 1 1
@ qi
5.1. A discussion on existing methods for dealing with pseudo buckling
Using Eqs. (21) and (30), we can calculate @KG =@ qi . modes
Following the same procedure, the sensitivity of the elemental
geometrical stiffness matrix with respect to displacement compo- For an investigation of existing methods for dealing with local
nent uj ðj ¼ 1; 2; . . . ; dÞ can be calculated pseudo buckling modes, the 2D portal frame structure shown in
Z Z Fig. 1 is considered. Its first buckling load factor is to be calculated.
1 1
@KGe @ Hðre Þ The frame with a thickness of 10 mm is clamped at the bottom, and
¼ gT gjJjdndg
@uj 1 1 @uj is modeled with a regular 40  20 mesh of four-node plane stress
Z 1 Z 1
 elements. A downward concentrated force F is applied at the center
@ue
¼ gT H UH1 e G e gjJjdndg on the top edge. The outer frame with a thickness of 5 mm has fully
1 1 @uj
Z Z 1 dense material q ¼ 1:0 and the inner region has a density
1 
gT H UH1 q ¼ 0:001.
¼ e Ge I gjJjdndg ð31Þ
1 1 For a real structural design, the low-density region should be
where I is an 8  1 vector. If uj is not a nodal displacement of the eth ignored and only the buckling modes occurring in the solid parts
or the regions with a high relative density value should be consid-
element, then I is a zero vector. Otherwise I is a unit vector and the
ered. In order to distinguish these modes from the pseudo modes,
kth component is one if the uj is the kth nodal displacement compo-
we call them ‘‘real’’ buckling modes, and their corresponding
nent of the eth element.
eigenvalues ‘‘real’’ buckling load factors. Thus, the first real buck-
Once the sensitivities of the introduced auxiliary variables
ling load factor is calculated by using a model with elements of
jj ¼ 1=kj are obtained, we can use the chain rule to calculate
the low density excluded, and a buckling load factor of 2.41 is
the sensitivities of eigenvalues kj as follows
obtained. This is taken as a reference solution.
@kj @kj @ jj 1 @ jj @ jj For most of the existing methods for suppressing pseudo
¼ ¼ ¼ k2j ð32Þ
@ qi @ jj @ qi j2j @ qi @ qi modes, different schemes are used for choosing modulus for stiff-
ness matrix ðEKe Þ and geometric stiffness matrix ðEKG
e Þ. In this study,
If the eigenvalue is multimodal with multiplicity m greater than
the three schemes in Table 1 are considered.
one, individual eigenvalues may no longer be differentiable func-
The results are summarized in Table 2. Quite different results
tions of the design variables. In such situations, the method pro-
have been obtained. Even with the same method, we may still
posed by Gravesen et al. [30] can be used. If j1 ¼ j2 , we can
obtain very different results when different parameter values are
calculate sensitivities of functions j1 þ j2 and j1 j2 using the fol-
used.
lowing expressions

 Method (a) uses a material model without penalization on
@ ðj1 þ j2 Þ ~ T @KG 1 @K ~ @K intermediate density. For this problem, it give a result 156.3%
¼ w1  w1  v T1 U
@ qi @ qi k1 @ qi @ qi higher than the reference solution. The significantly higher buck-

 ling load factor is due to the contribution of the inner region,
þw ~ T @KG  1 @K w ~ 2  v T @K U ð33Þ which may act as a thin membrane. As the material modulus
2 2
@ qi k2 @ qi @ qi
for the low density region is not scaled down with a small factor,


  say using penalization exponent number p > 1, pseudo buckling
@ ðj1 j2 Þ ~ T @KG  1 @K w ~ 1  v T @K U
¼ j2 w 1 1 modes can be avoided. However, this simple treatment may lead
@ qi @ qi k1 @ qi @ qi


  to optimized designs with many gray regions because no penal-
~T @K G 1 @K ~ 2  v T @K U ization is applied. The existence of these gray regions may lead
þ j1 w 2  w 2 ð34Þ
@ qi k2 @ qi @ qi to the over-estimation of buckling load factors, as shown in this
Other methods for dealing with multiple eigenvalues can be example (see Zhou [20]). Therefore, much care should be taken
found in the research literature [8,10,23,31]. when using this method.

4.3. Sensitivity of total material volume

F
The total volume of material used in the structure can be calcu-
lated as
X
Ne Ne Z
X 1 Z 1
H

V¼ V e ¼ t0 qe ðn; gÞjJjdndg ð35Þ

e¼1 e¼1 1 1

Its sensitivity can then be expressed as

X
Ne XNe Z 1 Z 1 W
@V @V e @ qe ðn; gÞ
¼ ¼ t0 jJjdndg ð36Þ
@ qi e¼1 @ qi e¼1 1 1
@ qi
Fig. 1. Portal frame.
146 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152

Table 1 (1) Some pseudo buckling modes have eigenvalues close to zero,
Three methods used for calculating the first buckling load factor of the portal frame. or more accurately, much smaller than those of the real
Method Scheme for modulus calculation Reference modes. As there may be many such pseudo modes, a large
number of eigenvalues have to be determined to include
(a) EKe ¼ EKe G¼ q e E0 Zhou [20]
some of the real modes.
(b) 0 if qe < ql Neves et al. [8]
EKe ¼ qpe E0 ; EKe G ¼
qpe E0 if qe P ql (2) The deformation mainly occurs in low-density regions, and
 
(c) EKe ¼ ql þ ð1  ql Þqpe E0 ; EKe G ¼ qpe E0 Bendsøe and Sigmund [18] as a result, the modal strain energy in low-density regions
makes a major contribution to the total modal strain energy.
Note: ql is a predefined parameter.

Based on these two observations, a new strategy is proposed.

Although method (b) produces a similar result as the reference

solution, it may cause solution oscillations in an optimization pro- 5.2.1. Calculation of candidate buckling modes
cess [18]. Besides parameter ql , this method requires an additional In linear buckling analysis, it is a common practice to calculate
parameter of critical normalized stress value rc . Geometrical stiff- only the first several eigenmodes to reduce the computational cost.
ness matrices of low-density elements with normal stress less than If a model has many pseudo buckling modes, it will be necessary to
rc will be ignored in the optimization process [8,10]. Generally calculate a large number of eigenvalues. Otherwise, it is possible
speaking, the critical normalized stress rc is difficult to predefine, that all those calculated are pseudo modes, which are useless.
and an improper value may result in different topology results [8]. Thus, the existence of pseudo modes may cause a dramatic cost
Method (c) can avoid the discontinuity problem of method (b). increase in eigenvalue calculation.
However, as the introduced parameter ql plays a critical role in However, if we know the range of real buckling load factors, this
avoiding the appearance of pseudo modes, using an improper value issue could be easily resolved by applying the eigenvalue-shift
may lead to erroneous results [19,20]. A large value of ql may technique [32]. By applying an appropriate shift value, real
result in unrealistically high stiffness of void elements as shown low-order buckling modes can be calculated at a reasonable cost
in this example (e.g. ql P 105 ). On the other hand, a small value as pseudo models with very small eigenvalues are excluded from
of ql , such as ql 6 109 for this example, is insufficient for avoiding the calculation.
the pseudo buckling mode. Hence, the difficulty of this method is When a uniform material distribution is used as the initial
to choose a proper value ql to calculate accurate real buckling load design, no pseudo buckling modes will appear at the early stage
factors. Although the lower bound of relative density (say, 0.001) is of the optimization, because there are no low-density regions. In
often used for ql , such a value may still be too large and as a result, this case, the range of real buckling load factors is easily available.
the structural stability can be overestimated. When a non-uniform initial design is used, the range of real eigen-
On the other hand, if buckling load factors of the portal frame values can still be determined, for example, by using the method
are calculated by using the standard SIMP model with p ¼ 3 and proposed by Neves et al. [8]. One can also calculate a relatively
including all the elements in the low-density region, the first 148 large number of buckling load factors in the first iteration and find
buckling modes obtained are pseudo modes with very small eigen- real low-order buckling modes using the identification technique
values and the 149th mode with a load factor of 2.41 is the first real presented in Section 5.2.2. Therefore, it is always possible to deter-
mode. Therefore, it is important to find reliable methods for deal- mine the real low-order buckling modes and calculate the corre-
ing with these pseudo modes. sponding load factors for the initial design.
As similar situations may occur for other topology optimization Let kðk1Þ be the first real buckling load factor for the last design.
problems, the development of effective methods for dealing with Assuming that the design will not change dramatically, we can use
pseudo modes is an important research topic. A new approach is kðk1Þ as the shift to calculate M buckling modes of a new, modified
proposed in the following sections. design, and then pick the smallest eigenvalue of the real modes.
Obviously, it is more likely that we can find the real first buckling
mode if a larger M is used. However, considering that it requires
5.2. Methods for eliminating the effects of pseudo modes
more computer resources and time to compute more eigenvalues
than necessary, a smaller number is preferred. Based on the
An investigation into the occurrence and characteristics of
assumption that optimized designs change only moderately
pseudo buckling modes reveals that pseudo modes have the fol-
between iterations, it is possible to choose a reasonably small
lowing two important features:
number for M. In this study, M ¼ 50 is used for all examples, which
means that only 50 eigenmodes are determined in the eigenvalue
extraction.
Table 2 Therefore, by applying an eigenvalue shift, possible low-order
Results for the first buckling load factor (p ¼ 3).
pseudo modes can be excluded from the eigenvalue extraction
Method ql k1 Error (%) computation.
(a) – 6.19 156.3
(b) P 103 2.41 0.0
5.2.2. Pseudo buckling modes identification
104 0.99 58.9
Aside from low-order pseudo models with very small eigenval-
(c) 103 34.37 1323.8 ues, other pseudo modes among the M eigenmodes may still exist.
104 6.56 171.9 These high-order pseudo modes must be identified to avoid a situ-
105 2.73 13.3
ation where a wrong constraint is introduced. We can make use of
106 2.45 1.3
the second feature of pseudo buckling modes to determine
107 2.42 0.1
whether a mode is real or not. As low-density elements (regions)
108 2.41 0.0
make the major contribution to the total modal strain energy of
109 1.98 17.8
a pseudo mode, we can divide the modal strain energy of a mode
1010 1.09 54.8
into two parts, one from the low-density regions and the other
 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152 147

from the rest of the structure. We can then identify this mode smallest buckling load factor. It is both simple and natural to just
based on the energy contributions of these two parts. add this new constraint equation and then use an existing opti-
First, a threshold parameter ql is introduced, and all of the mization algorithm to solve the new problem. Based on this idea,
nodes can be divided into two sets based on nodal relative density the iterative solution procedure in Fig. 2 is proposed for the solu-
values, i.e. nodes with low densities N l ¼ fijqi 6 ql ; 1 6 i 6 N g and tion of the topology optimization problem in Eq. (1). To start with,
nodes with high densities N h ¼ fijqi > ql ; 1 6 i 6 N g. the design domains are discretized and the optimization parame-
Then, the degrees of freedom can be divided into two groups ters are defined. In the FE analysis of the initial design in Step 2
based on the grouping of nodes [33], and a buckling mode vector and the current design in Step 5, linear static responses, the ‘real’
can be decomposed into two vectors: buckling load factors kj and corresponding mode vectors w ~ j , the
structural compliance C and the volume of material used in the
Wj ¼ Wlj þ Whj ð37Þ
design V are all determined. In Step 3, the sensitivities of the object
where Wlj is composed of displacement components for degrees of
freedom of nodes in set N l and zero elements, and Whj is composed
of the displacement components for degrees of freedom of nodes in
set Nh and zero elements. Start
Hence, the modal strain energy ratio of low-density regions r lj ,
defined as the ratio of contribution from the nodes in N l to the total 1. Define optimization model and choose
modal strain energy, is given by
optimization parameters: V0 , λ , ρ l, MWl , p, εC, ερ , M
WTlj KWlj þ WTlj KWhj WTlj KWj
r lj ¼ ¼ ð38Þ
WTj KWj WTj KWj
2. Determine λ j , ψ j , U, C , V (ρ) by FE analysis
The pseudo buckling mode identification criterion can be stated as

r lj P MWl ð39Þ
3. Calculate the sensitivities of constraint functions and
where MWl is a predefined parameter with a value between 0 and objective function
1. If the modal strain energy ratio of low-density regions is bigger
than the value, the corresponding mode is regarded as pseudo
mode; otherwise, the mode is treated as real. 4. Update design variables using optimization solver
A region with a relative density less than ql is usually regarded (MMA)
being of low-density, and all numerical tests conducted in this
study have shown that the threshold value ql ¼ 0:1 is appropriate.
For a clear black-white design in which the design variables are 5. Determine λ j , ψ j , U, C , V (ρ) by FE analysis
equal to either one or the lower bound q, the modal strain energy
ratio for a pseudo mode can be bigger than 0.98, as demonstrated
by the example in Section 5.1. During the optimization iteration, it No
is very likely that a topology design contains gray regions. For a Converged?
pseudo mode, deformation in these regions may be small but is
not zero, causing a reduction in the modal strain energy ratio that Yes
may have an effect on a pseudo mode check. In order to enhance End
the reliability of the mode identification results, it is necessary to
decrease the value of MWl . In this study, the selected parameter Fig. 2. Flow chart of the iterative solution procedure.
values ql ¼ 0:1 and MWl 2 ½0:6; 0:7 are used and prove to be
appropriate for all of the conducted numerical tests.
The proposed identification method is used on the 150 calcu-
lated buckling modes of the example structure in Section 5.1,
and all the modes are correctly identified. The pseudo mode iden-
tification is an important component of the optimization algorithm
proposed in the next section, and numerical experiments show
that the identification method is very effective and reliable.

6. Optimization algorithm

Based on the finite element analysis and sensitivity analysis,

topology optimization problem (see Eq. (1)) can now be solved
by using a gradient-based optimization algorithm. Optimization
algorithms based on the well-known MMA optimization solver
[34] are developed. Different optimization strategies are also pro-
posed for achieving high-quality local solutions.

6.1. Computational procedure

Compared with the conventional problem of compliance mini-

mization under material volume constraint, the optimization Fig. 3. 2-D continuum structure under distributed load (a) problem description and
model in Eq. (1) has an additional buckling constraint on the (b) topology from compliance minimization subjected to volume constraints.
148 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152

 
and constraint functions with respect to the nodal design variables  ðkþ1Þ 
C  C ðkÞ =C ðkÞ 6 eC ;
are computed. In addition, we will check the multiplicity of the
(c) all of the constraint conditions are satisfied.
first eigenvalue. If the multiplicity m is bigger than one, we will
use multimodality formulation to calculate its sensitivities. In
6.2. Improved optimization strategies
Step 4, the design is modified by using optimization solver MMA.
To check the solution convergence, the following criteria are used:
While the algorithm presented earlier is applicable to the buck-
ling optimization problem, as will be shown through numerical
(a) the maximum change in the design variables between two
examples in the next section, it does not always work well. As
consecutive iterations is smaller than a predefined tolerance
  there is usually a confliction between the requirements for struc-
eq , i.e. qðkþ1Þ  qðkÞ 1 6 eq ;
tural stiffness and stability, the buckling constraint should not be
(b) the change in the objective function between two consecu-
treated just like the one in the material volume, implying that
tive iterations is smaller than a predefined tolerance eC , i.e.
refined algorithms are required to achieve more optimized designs.
It is proposed that the optimization process is separated into two
phases and at each phase, different optimization models and/or
-2
material models are employed. In the present paper, two improved
x10
8.2 algorithms are given and compared with the simple one phase
Algorithm A 8.05 algorithm shown in Fig. 2. The initial value for nodal design vari-
8.0 Algorithm B ables in all of the three algorithms is uniform with the volume frac-
Algorithm C 8.00
7.8 7.70 tion. The three algorithms investigated are as follows:
Compliance

7.6
7.46 (a) Algorithm A: solve optimization problem (Eq. (1)) by follow-
7.4 7.32 7.50 ing the procedure shown in Fig. 2 with no changes.
7.25 7.38
7.17 (b) Algorithm B: separate the optimization process into two
7.2 7.28
7.02 7.21 phases and use different optimization models. In the first
7.08
7.0 phase, solve the conventional problem of compliance mini-
6.8
mization under the volume constraint by ignoring the buck-
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 ling constraints. In the second phase, solve the original
Lower bound of buckling load factor problem (Eq. (1)) with the solution from the first phase as
the initial design.
Fig. 4. Compliance vs the lower bound of buckling load factor.

Table 3
Optimized topologies for different buckling constraints.

Algorithm k ¼ 0:6 k ¼ 0:8 k ¼ 1:0 k ¼ 1:2 k ¼ 1:4 k ¼ 1:6

A 40 40 40 40 40 40

35 35 35 35 35 35

30 30 30 30 30 30

25 25 25 25 25 25

20 20 20 20 20 20

15 15 15 15 15 15

10 10 10 10 10 10

5 5 5 5 5 5

0 0 0 0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20

B 40 40 40 40 40 40

35 35 35 35 35 35

30 30 30 30 30 30

25 25 25 25 25 25

20 20 20 20 20 20

15 15 15 15 15 15

10 10 10 10 10 10

5 5 5 5 5 5

0 0 0 0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20

C 40 40 40 40 40 40

35 35 35 35 35 35

30 30 30 30 30 30

25 25 25 25 25 25

20 20 20 20 20 20

15 15 15 15 15 15

10 10 10 10 10 10

5 5 5 5 5 5

0 0 0 0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152 149

(c) Algorithm C: separate the optimization process into two The iteration history curves in Fig. 5 show that algorithm A
phases and use different material penalization for calculat- requires considerably more iterations than algorithms B and C
ing buckling load factors. In the first phase, use normal for convergence. From the topological changes shown in Fig. 6, it
material penalization for calculating compliance and its sen- can be found that with algorithm A, modulus penalization makes
sitivity, but do not penalize the material modulus for calcu- the buckling constraint harder to be satisfied and causes the mate-
lating the buckling load factor (i.e. use p ¼ 1). In the second rial to distribute over a much larger region than required at the
phase, use the normal material penalization to solve the early stage of the optimization. At the later stage, the optimization
original problem (Eq. (1)) with the solution from the first algorithm will guide the design to change and produce better
phase as the initial design. designs. However, the intermediate designs could be so different
from the final design that the solution requires a large number of
It is noteworthy that, for both algorithms B and C, a ‘pre-solve’
iterations to converge. In Fig. 5(b), we can see that in iterations
phase is added and the second phase is the same as the solution of
167–480, the objective function increases first and then decreases
the original problem by using algorithm A, but with a different ini-
until the solution terminates. This is because different constraints
tial design from the first phase. In the first phase, the buckling con-
are considered in the two phases of algorithm B. In the first phase,
straints are either ignored as in algorithm B or considered but in a
the buckling constraint is not considered at all, thus the optimized
modified form as in algorithm C. In contrast, the initial design for
topology structure has high stiffness but the buckling constraints
algorithm A is a uniform distribution in the design domain.
may not be satisfied. In the second phase, the optimization algo-
The numerical examples in the next section will show that by
rithm will steer the design to improve structural stability and
separating the optimization procedure into two phases, the solu-
may cause a reduction in stiffness. However, once the buckling
tion may converge in fewer iterations and better performance of
constraints are satisfied, compliance will begin to decrease as opti-
the optimized structures can be achieved.
mization proceeds. Fig. 5(c) shows a drop in first buckling load fac-
tor at iteration 300. This sudden change is due to the switch of the
7. Numerical examples solution phase. From this iteration onwards, the normal material
penalization will be employed for buckling analysis. It can be seen
Two examples are considered in this section and the designs from Fig. 6 that for algorithm B and C, at the first stage, the compli-
obtained by using the three algorithms are compared. ance of topological designs is relatively small and material is

7.1. Example 1 a
The first example is the optimization of a 2D continuum
column-like structure under distributed loads. The load and sup-
port conditions are shown in Fig. 3(a). The design domain is a rect-
angular area of unit thickness with height H = 40 and width W = 20.
A distributed load q = 0.05 is applied at the top edge with a width
d = 2/3. The design domain is discretized into 60  120
equally-sized square four-node elements. The material constants
used are Young’s modulus E ¼ 1:0 and Poisson’s ratio m ¼ 0:3.
The prescribed material volume fraction number is set to 0.35.
The initial value of all nodal design variables are set to the vol-
ume fraction. The optimized design of a conventional minimum
compliance problem is shown in Fig. 3(b) and its corresponding b
first buckling load factor is 0.45. When considering buckling con-
straints, the parameters for convergence criterion are ec ¼ 0:01
and eq ¼ 0:001 and the move limit on design variables is
mq ¼ 0:003.
This optimization problem has been solved using the three
algorithms for different lower bounds of buckling load factors.
The curves of compliance versus the lower bound of buckling load
factor are presented in Fig. 4, while the obtained topologies are
shown in Table 3. It can be seen that the compliance increases as
the lower bound of buckling load factor increases. This means that
with a fixed amount of material, the improvement of structural
stability can be achieved only by a reduction in structural stiffness. c
A comparison of the designs obtained with buckling constraints
and those obtained with the volume constraint only reveals that
the requirement on the structural stability tends to distribute the
available material over a larger area. In contrast, the maximization
of structural stiffness causes the material to distribute along the
load transfer path.
From Fig. 4, it is found that the compliance values for the three
solutions are very close for all of the considered lower bound val-
ues. However, the topologies obtained have some minor differ-
ences. This indicates that very likely, the solutions are just local
optima, and for this particular problem, different local solutions Fig. 5. Iteration history of compliance and first buckling load factor with k ¼ 1:0 for
have similar topologies and compliance values. three algorithms. (a) Algorithm A; (b) Algorithm B; and (c) Algorithm C.
150 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152

Table 4
The first three buckling modes of optimized designs at k ¼ 1:0.

Algorithm A B C
1st mode

Iter=200 Iter=400 Iter=600

λ 1 = 1.00, C = 8.58 × 10−2 λ 1 = 1.00, C = 8.07 × 10−2 λ 1 = 1.00, C = 7.86 × 10−2

2nd mode

Iter=800 Iter=1000 3rd mode

Iter=1200
λ1 = 1.00, C = 7.57 × 10−2 λ1 = 1.00, C = 7.38 × 10−2 λ1 = 1.00, C = 7.32 × 10−2
(a) Topological changes during optimization for Algorithm A

The first three buckling modes of optimized structures for

k ¼ 1:0 are shown in Table 4. The first real buckling load factors
Iter=250 Iter=350 Iter=450 for the three algorithms are all bimodal. This validates that the
λ 1 = 1.00, C = 7.49 × 10−2 λ 1 = 1.00, C = 7.33 × 10−2 λ 1 = 1.00, C = 7.32 × 10−2 multimode sensitivity analysis methods presented at the end of
(b) Topological changes during optimization for Algorithm B Section 4.2 is effective.

7.2. Example 2

The short cantilever beam shown in Fig. 7(a) is considered. The

design domain is a rectangular area of unit thickness with height
H = 2 and width W = 1. A concentrated load F ¼ 0:005 is applied
to the center of the right edge. The design domain is discretized
into 40  80 equally-sized square four-node elements. The mate-
rial constants used are Young’s modulus E ¼ 1:0 and Poisson’s ratio
m ¼ 0:3. The prescribed material volume fraction is set to 0.15.
The initial value of all nodal design variables are set to the vol-
Iter=300 Iter=500 Iter=700
ume fraction. The optimized design of a conventional minimum
λ 1 = 0.61, C = 7.47 × 10−2 λ 1 = 0.85, C = 8.17 × 10−2 λ 1 = 1.00, C = 7.55 × 10−2 compliance problem is shown in Fig. 7(b) and the corresponding
(c) Topological changes during optimization for Algorithm C first buckling load factor is equal to 0.22.
The optimized designs obtained by employing three algorithms
Fig. 6. Topological changes during optimization with a low bound of buckling load for different lower bounds of buckling load factors are shown in
factor of k ¼ 1:0. Table 5. It can be clearly seen that as the lower bound of buckling
load factor increases, the structural member in compression
becomes shorter and wider, resulting in stability improvement.
mainly distributed along the load transfer path with a relatively At the same time, some gray regions appear. The appearance of
low stability. At the later stage, some material is moved away from gray region means that it is impossible to obtain a clear
the load transfer path to improve structural stability. black-white design to satisfy the constraint.
 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152 151

-4
x10
9.0
8.53
8.5 Algorithm A
Algorithm B 7.95
8.0 Algorithm C

Compliance
7.5
7.02 7.04
7.0 6.76 6.78

6.5 6.24
5.96 6.01
5.84
H

6.0 5.64
5.48
5.31
5.5 5.16
4.97 5.07
4.84 5.45 5.58
F 5.0 5.25 5.27 5.42
4.80 5.02
4.5
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Lower bound of buckling load factor
W Fig. 8. Compliance vs the lower bound of buckling load using three algorithms.

(a) (b) one-phase algorithm and algorithm C is the best of the three. In
addition, this example again validates the effectiveness of the
Fig. 7. Short cantilever beam under concentrated load (a) problem description and
new approach combining pseudo buckling mode identification
(b) topology of compliance minimization subjected to volume constraint.
and eigenvalue shift in dealing with pseudo buckling modes.

8. Conclusions
Table 5
Optimized topology configuration subjected to different constraints using three
algorithms. The compliance minimization problem under volume and sta-
bility constraints is considered. A new approach has been proposed
Algorithm k ¼ 0:4 k ¼ 0:6 k ¼ 0:8 k ¼ 1:0
for dealing with the well-known pseudo buckling mode problem.
A
In addition, in consideration of the non-linear and non-convex nat-
ure of the problem, two-phase algorithms are suggested for achiev-
ing better local optimization solutions. Numerical examples are
presented to show the effectiveness of the new algorithms.
It should be pointed out that the proposed approach for dealing
with the pseudo buckling mode problem can be applied to vibra-
tion optimization problems [35]. Further, the two-phase optimiza-
B tion strategies are effective for improving the performance of
optimized designs. The same idea could be useful for other engi-
neering optimization problems, which may also be highly
non-linear and non-convex.

Acknowledgement

The authors would like to acknowledge the support from State

C
Key Laboratory of Structural Analysis for Industrial Equipment,
Dalian University of Technology, Dalian 116024, P.R. China
(Project No. GZ1305).

References

[1] Khot N, Venkayya V, Berke L. Optimum structural design with stability

constraints. Int J Numer Methods Eng 1976;10:1097–114.
[2] Olhoff N, Rasmussen SH. On single and bimodal optimum buckling loads of
clamped columns. Int J Solids Struct 1977;13:604–14.
[3] Pedersen NL, Nielsen AK. Optimization of practical trusses with constraints on
eigenfrequencies, displacements, stresses, and buckling. Struct Multidisc
Compliance values for different optimized designs are shown in Optim 2003;25(5–6):436–45.
Fig. 8. From these results, one can see that the stiffness of the opti- [4] Gu YX, Zhao GZ, Zhang HW, Kang Z, Grandhi RV. Buckling design optimization
mized structure reduces in order to meet a higher stability require- of complex built-up structures with shape and size variables. Struct Multidisc
Optim 2000;19:183–91.
ment. This is the same as the observations from the last example. [5] Bojczuk D, Mroz Z. Optimal topology and configuration design of trusses with
On the other hand, the results obtained by using different methods stress and buckling constraints. Struct Optim 1999;17(1):25–35.
are quite different in the example. It is easily seen that, for all the [6] Guo X, Liu W, Li HY. Simultaneous shape and topology optimization of truss
under local and global stability constraints. Acta Mech Solida Sin
lower bounds of buckling load factors considered, the two-phase 2003;16(2):95–101.
optimization algorithms B and C produce designs with much smal- [7] Guo X, Cheng GD, Olhoff N. Optimum design of truss topology under buckling
ler compliance values. This is a significant improvement due to the constraints. Struct Multidisc Optim 2005;30(3):169–80.
[8] Neves MM, Rodrigues H, Guedes JM. Generalized topology design of structures
introduction of a first phase in the optimization algorithm. It can
with a buckling load criterion. Struct Optim 1995;10:71–8.
also be seen that algorithms B and C produce similar designs only [9] Min SJ, Kikuchi N. Optimal reinforcement design of structures under the
when k is less than 0.6. When k is bigger than 0.6, algorithms C pro- buckling load using the homogenization design method. Struct Eng Mech
duce better designs. These observations clearly show that the three 1997;5:565–76.
[10] Neves MM, Sigmund O, Bendsøe MP. Topology optimization of periodic
algorithms may produce significantly different solutions. For this microstructures with a penalization of highly localized buckling modes. Int J
problem, the proposed two-phase algorithms are better than the Numer Methods Eng 2002;54(6):809–34.
152 X. Gao, H. Ma / Computers and Structures 157 (2015) 142–152

[11] Buhl T, Pedersen C, Sigmund O. Stiffness design of geometrically nonlinear [24] Rahmatalla SF, Swan CC. A Q4/Q4 continuum structural topology optimization
structures using topology optimization. Struct Multidisc Optim 2000;19(2): implementation. Struct Multidisc Optim 2004;27(1-2):130–5.
93–104. [25] Kang Z, Wang Y. Structural topology optimization based on non-local Shepard
[12] Sekimoto T, Noguchi H. Homologous topology optimization in large interpolation of density field. Comput Methods Appl Mech Eng 2011;200(49-
displacement and buckling problems. JSME Int J A-Solid M 2001;44:616–22. 52):3515–25.
[13] Bruns TE, Sigmund O, Tortorelli DA. Numerical methods for the topology [26] Deng X, Wei P, Ma H. Topology optimization of 2D continuum using nodal
optimization of structures that exhibit snap-through. Int J Numer Methods Eng design variables. In: Proceedings of the Seventh China–Japan–Korea Joint
2002;55(10):1215–37. Symposium on Optimization of Structural and Mechanical Systems.
[14] Bruns TE, Sigmund O. Toward the topology design of mechanisms that exhibit HuangShan (China); 2012.
snap-through behavior. Comput Methods Appl Mech Eng 2004;193(36– [27] Pian THH, Sumihara K. Rational approach for assumed stress finite elements.
38):3973–4000. Int J Numer Methods Eng 1984;20:1685–95.
[15] Kemmler R, Lipka A, Ramm E. Large deformations and stability in topology [28] Crisfield MA, Remmers JJ, Verhoosel CV. Nonlinear finite element analysis of
optimization. Struct Multidisc Optim 2005;30(6):459–76. solids and structures. 2nd ed. John Wiley & Sons; 2012.
[16] Lindgaard E, Lund E. Nonlinear buckling optimization of composite structures. [29] Rodrigues HC, Guedes JM, Bendsøe MP. Necessary conditions for optimal
Comput Methods Appl Mech Eng 2010;199(37–40):2319–30. design of structures with a non-smooth eigenvalue based criterion. Struct
[17] Lindgaard E, Lund E. A unified approach to nonlinear buckling optimization of Optim 1995;9:52–6.
composite structures. Comput Struct 2011;89(3–4):357–70. [30] Gravesen J, Evgrafov A, Nguyen DM. On the sensitivities of multiple
[18] Bendsøe MP, Sigmund O. Topology optimization: theory, methods and eigenvalues. Struct Multidisc Optim 2011;44(4):583–7.
applications. 2nd ed. Berlin Heidelberg (New York): Springer; 2003. [31] Seyranian AP, Lund E, Olhoff N. Multiple eigenvalues in structural optimization
[19] Lindgaard E, Dahl J. On compliance and buckling objective functions in problems. Struct Optim 1994;8:207–27.
topology optimization of snap-through problems. Struct Multidisc Optim [32] Bathe KJ, Ramaswamy S. An accelerated subspace iteration method. Comput
2013;47(3):409–21. Methods Appl Mech Eng 1980;23:313–31.
[20] Zhou M. Topology optimization for shell structures with linear buckling [33] Brehm M, Zabel V, Bucher C. An automatic mode pairing strategy using an
responses. In: WCCM, VI. BeiJing, China; 2004. p. 795–800. enhanced modal assurance criterion based on modal strain energies. J Sound
[21] Pedersen NL. Maximization of eigenvalues using topology optimization. Struct Vib 2010;329(25):5375–92.
Multidisc Optim 2000;20(1):2–11. [34] Svanberg K. The method of moving asymptotes – a new method for structural
[22] Tcherniak D. Topology optimization of resonating structures using SIMP optimization. Int J Numer Methods Eng 1987;24:359–73.
method. Int J Numer Methods Eng 2002;54(11):1605–22. [35] Gao X, Ma H. A new method for dealing with pseudo modes in topology
[23] Du JB, Olhoff N. Topological design of freely vibrating continuum structures for optimization of continua for free vibration. Chin J Theoret Appl Mech
maximum values of simple and multiple eigenfrequencies and frequency gaps. 2014;46(5):739–46.
Struct Multidisc Optim 2007;34(2):91–110.