Struct Multidisc Optim (2015) 51:11591172

DOI 10.1007/s00158-014-1190-z

EDUCATIONAL ARTICLE

Matlab code for a level set-based topology optimization

method using a reaction diffusion equation
Masaki Otomori Takayuki Yamada Kazuhiro Izui
Shinji Nishiwaki

Received: 10 March 2014 / Revised: 3 October 2014 / Accepted: 21 October 2014 / Published online: 19 November 2014
Springer-Verlag Berlin Heidelberg 2014

Abstract This paper presents a simple Matlab implementation for a level set-based topology optimization method in
which the level set function is updated using a reaction diffusion equation, which is different from conventional level
set-based approaches (Allaire et al. 2002, 2004; Wang et
al. 2003) that use the Hamilton-Jacobi equation to update
the level set function. With this method, the geometrical
complexity of optimized configurations can be easily controlled by appropriately setting a regularization parameter.
We explain the code in detail, and also the derivation of
the topological derivative that is used in the level set-based
topology optimization. Numerical results for stiffness maximization problems are provided to facilitate the readers
understanding. The presented code is intended for educational purposes only. This paper was inspired by previously published papers presenting Matlab code for a SIMP
method (Sigmund 2001; Andreassen et al. 2011), a level
set-based method (Challis 2010), and FreeFem++ code for
a structural optimization method (Allaire and Pantz 2006).
Readers can investigate results provided by these different
methods and discover the prominent aspects of each particular method. The code presented here can be downloaded
from http://www.osdel.me.kyoto-u.ac.jp/members/yamada/
codes.html.
Keywords Education Matlab code Topology
optimization Level set method Topological derivative
Reaction diffusion equation

M. Otomori T. Yamada () K. Izui S. Nishiwaki

Department of Mechanical Engineering and Science,
Kyoto University, Kyotodaigaku-katsura,
Nishikyo-ku, Kyoto 6158540, Japan
e-mail: takayuki@me.kyoto-u.ac.jp

1 Introduction
This paper presents Matlab code for a level set-based
topology optimization method incorporating a reaction diffusion equation for a compliance minimization problem.
Concerning level set-based structural optimization methods, many have been proposed since the pioneering works
by Allaire et al. (2002, 2004) and Wang et al. (2003).
Programming code for a level set approach is freely available: Scilab code (Allaire et al. 2004) can be downloaded
via http://www.cmap.polytechnique.fr/allaire/levelset en.
html. Also, Challis (2010) presented Matlab code for
a discrete level set approach, which is available in the
manuscript. The Matlab code discussed in this paper can be
used as an aid to understand the similarities and differences
between this level set-based method and other topology
optimization methods.
The level set method was first proposed by Osher and
Sethian (1988) as a method to implicitly represent the
evolution of interfaces. In such analyses, the evolution
of interface boundaries is tracked by solving a so-called
Hamilton-Jacobi equation, using an appropriate velocity
normal to the interface of the moving boundary. In level
set-based approaches, structural boundaries are represented
by the zero iso-surface of the level set function and the
level set function is defined so that the structural domain
is represented wherever the level set function has a positive
value. With respect to boundary representation methods,
we classify the method presented here as a level set-based
approach, which was developed by several of the authors
(Yamada et al. 2010).
In this method, the level set function is updated by
solving a reaction-diffusion equation based on the topological derivative of the objective functional. Therefore,
this method allows topological changes that generate new

1160

M. Otomori et al.

boundaries during the optimization procedure and it does

not require re-initialization of the level set function, which
is typically required in Hamilton-Jacobi-based approaches
to ensure accuracy when solving the Hamilton-Jacobi partial differential equation. In addition, our method enables
qualitative control of the geometrical complexity of the
optimized configurations by using appropriate values when
setting a regularization parameter. In the following, we
provide a rigorous derivation of the topological derivative for a mean compliance minimization problem. We
note that although the term representing the effect of
boundary conditions when a hole is created in the design
domain was ignored in previous research (Yamada et al.
2010), this effect is now considered in the method
presented here.
The rest of this paper is as follows. Section 2 describes
the formulation of the level set-based topology optimization method using a reaction diffusion equation and the
formulation of an optimization problem for a compliance
minimization problem. Section 3 explains the Matlab code
we use, and Section 4 provides numerical examples and
an extension of the code. A conclusion is provided in
Section 5, the Matlab code is provided in Appendix A, and
the derivation of the topological derivative is provided in
Appendix B.

2 Formulation
2.1 Topology optimization
For structural optimization, optimization problems are formulated using a domain  filled with a material domain, a
void domain, and boundaries , as follows:


inf F [] =


fd (x, x, x)d +


fb (x, x, x)d,


(1)

where fd and fb are arbitrary real functions defined for

domain  and boundary , respectively. x represents a point
located in  and x represents the state variables.
In topology optimization, as Fig. 1 shows, the above optimization problem is replaced with a material distribution
problem within a fixed design domain, D, using the characteristic function  . The optimization problem is then
defined as


inf F [ ] =


fd (x,  ) d +
D

fb (x,  ) d,


(2)

Fig. 1 Fixed design domain, D

where the  is defined as


1 x 
 (x) =
0 x D \ .

(3)

In the above optimization problem, topological changes

such as an increase or decrease in the number of holes are
allowed during the optimization procedure, in addition to
changes in the shape represented by outer or inner boundaries. However, the continuity of the characteristic function
is not guaranteed and the value of the characteristic function
can be discontinuous at every point. That is, since the characteristic function  is defined as a subset of a bounded
Lebesgue space, L , which only assures integrability, the
obtained solutions can be discontinuous at every point in the
fixed design domain. The ill-posed nature of the optimization problem is due to the lack of regularity of the admissible
shapes.
In other words, because increasing the number of holes
in a given design without changing its volume, by concomitantly decreasing their size, results in improved values of
the objective functional, the solution of the optimization
problem is rendered nonexistent. This is caused by the lack
of closedness of the set of feasible designs. To overcome
this obstacle, the optimization problem must be regularized so that a solution can exist. Homogenization-based
approaches (Bendse and Kikuchi 1988; Allaire 2002) and
fictitious density approaches such as the SIMP method
(Bendse 1989) are two popular methods used to regularize
optimization problems.
Homogenization-based approaches use homogenization theory to relax the design domain, that is, enlarge the
set of possible designs. On the other hand, fictitious density
approaches replace the characteristic function with a fictitious material whose elasticity tensor is assumed to be a continuous function of the material density. The SIMP (Solid
Isotropic Material with Penalization) method is a type of
fictitious density approach that uses a continuous fictitious

Matlab code for a level set-based topology optimization method

isotropic material whose elasticity tensor is assumed to be a

function of material density, penalized using an exponential
parameter. In both the Homogenization-based approaches
and fictitious density approaches, optimized configurations
are represented as density distributions expressed using
continuous values that range from 0 to 1. Therefore, the
obtained optimized configurations often include grayscale
areas where the density assumes an intermediate value
between 0 and 1. Furthermore, the application of a boundary condition at the structural boundaries is problematic in
these approaches because the structural boundaries are not
clearly expressed.
To fundamentally overcome these problems, approaches
based on a level set method were proposed (Allaire 2002,
2004; Wang et al. (Wang et al. 2003)), in which the boundaries of the optimal configuration are implicitly represented
using a level set function. However, because these particular
level set-based methods are based on a boundary advection
concept, topological changes that generate new boundaries
during the optimization procedure are not allowed, although
the number of holes in an existing configuration can be
decreased. We note that, for 3-dimensional problems, structural boundaries can easily merge to create holes in the
fixed design domain, but this is different from the generation of holes inside the material domain, which is still
disallowed. Therefore, although the above mechanism for
the generation of new holes makes the method much more
flexible, conventional level set-based approaches that use
the Hamilton-Jacobi equation for advecting the shape are
subject to limitations of topological changes even for 3dimensional problems. As a result, the initial configuration
settings greatly affect the obtained optimized configurations. To overcome this difficulty, Allaire et al. (2005)
applied the bubble method (Eschenauer et al. 1994) to
a level set-based structural optimization method in which
boundaries are updated using the shape derivative, and holes
are inserted based on the value of the topological derivative.
In Allaire et al. (2005), hole nucleation based on the topological derivative is performed in several steps during the
optimization procedure, as the level set function is updated
using the shape derivative.
2.2 Level set-based topology optimization using a reaction
diffusion equation
This subsection describes the level set-based topology optimization method applied in this paper (Yamada et al. 2010).
In this method, the level set function is updated by solving a
reaction-diffusion equation based on the topological derivative of the objective functional. Therefore, this method
allows topological changes that generate new boundaries
during the optimization procedure, and re-initialization of
the level set function is not required.

1161

As shown in Fig. 2, the structural boundaries in a

level set-based topology optimization method are implicitly
represented using the iso-surface of the level set function ,
as follows.

1  (x) > 0 x  \ 
(x) = 0
x 
(4)

0 > (x)  1 x D \ .
The above level set function is used to represent the
boundaries of the target structure; positive values represent
the material domain, negative values represent the void
domain, and zero represents the structural boundaries. The
level set function has upper and lower limits imposed for the
regularization term, used for regularizing the optimization
problem, which will be explained in the next subsection. To
avoid confusion concerning the sign of the level set function, we note that even though we define positive values of
the level set function as corresponding to points in the material domain, as in (Wang et al. 2003; Yamada et al. 2010),
an inverse definition where negative values correspond to
points in the material domain, as in (Allaire 2002; Allaire
et al. 2004; Challis 2010), is equally valid. We also note that
the level set function used in the method presented here is
not a signed distance function, such as is typically used in
many level set-based approaches. The optimization problem
that minimizes an objective functional F under an inequality
constraint G is then formulated as follows, using the level
set function defined above.


fd (x, )d + fb (x, )d
(5)
inf F [ ] =

D


subject to G[ ] =
g(x, ) d Gmax  0 ,
(6)
D

where g is the density function

and Gmax is the upper limit

value of the response D g(x, ) d. The characteristic
function () is now defined as

1 if  0
() =
(7)
0 if < 0 .
The above optimization problem is now replaced with
an unconstrained optimization problem, using Lagranges
method of undetermined multipliers, as follows.
inf F [ , ] = F + G ,

(8)

where is the Lagrange multiplier. Based on the above

formulation, the KKT (Karush-Kuhn-Tucker) conditions of
this optimization problem are described as
F  = 0,

G = 0,

 0,

G  0,

(9)

where the notation F  represents the derivative of the

Lagrangian F .
Level set functions that satisfy the above KKT conditions are candidate solutions of the level set function that
represent optimized configurations. However, it is nearly

1162

M. Otomori et al.

Fig. 2 Fixed design domain D

and level set function

impossible to find such optimized solutions directly, so

the optimization problem is replaced with a time evolution
equation by introducing a fictitious time t. The level set
function is then updated by solving this equation, and an
optimized configuration is ultimately obtained, as explained
below.
2.3 Time evolution equation
For the following formulation, which introduces fictitious
time t, it is assumed that the variation of the level set function is proportional to the gradient of Lagrangian F , as
follows.

= K F  ,
t

(10)

where K > 0 is a coefficient of proportionality. The optimization problem as formulated above is an ill-posed problem, so it is regularized by adding a regularization term, as
follows.

= K F  2 ,
t

(11)

The above equation is a reaction diffusion equation, and the

diffusive term 2 ensures the smoothness of the level set
function in the presented method. Furthermore, the value
of the regularization parameter affects the degree of this
diffusivity, with larger values of providing increased diffusivity for the level set function. Thus, an appropriately
set value of can prevent the generation of structures that

have excessive geometrical complexity, so that an optimized configuration that has a desired degree of geometrical
simplicity can be obtained.
Upper and lower limits are imposed on the level set function so that the smoothing effect only operates on points that
are close to the structural boundaries. Therefore, this regularization term functions implicitly as a kind of perimeter
control. That is, the complexity of the optimized configuration can be controlled by adjusting the value of the
regularization parameter . For relatively small values of ,
relatively complex optimized configurations are obtained,
and the converse is true. The details are provided in the
numerical examples. We note that adding the regularization term to (10) to obtain (11) makes it currently difficult
to guarantee that the objective functional monotonically
decreases, although the advection velocity is rigorously
guaranteed to decrease in a descent direction in Allaire et al.
(2004). We hope to address this issue in future research.
Setting appropriate boundary conditions for (11), the
following equations are obtained.

= K F  2 in D
t

=0
on D.

(12)

We note that although boundary D is defined for values

of the level set function that are equal to zero, this does not
mean that D represents structural boundaries. The use of
this definition allows domains that are close to the boundary of the design domain to remain unaffected by boundary

Matlab code for a level set-based topology optimization method

1163

settings. Other boundary conditions might work fine. However, we use the Dirichlet boundary condition = 0 on the
boundary of the design domain for simplicity on implementation. The optimized configuration can now be obtained by
solving the above time evolution problem. In this method,
topological derivative dt F is used for the F  term.

The second term on the right-hand side in the above equation can be replaced using Greens formula, as follows:
F =



ti ui d + u i ti d
u i,j Cij kl uk,l d





+
d Vmax .

Figure 3 shows the fixed design domain D and boundary

conditions for a mean compliance minimization problem.
The displacement is fixed at boundary u and traction ti is
applied at boundary t . The objective is to find the optimum
layout of the design space that minimizes the mean compliance under a given volume constraint. The optimization
problem is defined as

inf F =
ti ui d
(13)


s.t. G =
d Vmax  0
(14)



(15)
div Cij kl uk,l = 0 on 
ui = u i
ti = ti

(19)

2.4 Optimization problem

in u

(16)

in t ,

(17)

where Vmax is the upper limit of the volume constraint, Cij kl

is the elastic tensor, ui is the displacement and ti = ij nj =
Cij kl uk,l nj is the traction. u i and ti are constant values that
represent the given displacement and traction, respectively.
Using Lagranges method of undetermined multipliers,
the above optimization problem is replaced with an unconstrained optimization problem, using Lagrangian F , and
Lagrange multiplier u i and , as follows:




ti ui d +
u i div Cij kl uk,l d
F =




d Vmax .
+
(18)


2.5 Topological derivative

Figure 4 shows the concept of the topological derivative,
which is a measure of the influence when a hole  with
radius
is created at a certain point in the domain . 

represents the boundary of the created hole. The topological

derivative dt F of objective functional F is defined as
dt F := lim

(F + F ) F

0 meas( \  ) meas()

(20)

The topological derivative for Lagrangian (19) is given

as follows.
F
dt F = lim
= u 0i,j Aij kl u0k,l ,
3

0 4

(21)

where the superscript 0 indicates the value without creating

holes, and Aij kl is defined as follows:

Aij kl


(1 14 + 15 2 )E
3(1 )
=
ij kl
2(1 + )(7 5)
(1 2)2

+ 5E(ik j l + il j k ) . (22)

The details for the derivation of topological derivative are

provided in Appendix B.

3 Implementation
3.1 Reaction-diffusion equation

Fig. 3 Fixed design domain D and boundary conditions

We now present a scheme for implementing the reactiondiffusion equation (Yamada et al. 2010). As discussed
above, the level set function is updated using a reaction diffusion equation, and the diffusive term 2 ensures that
the optimization has sufficient smoothness and numerical
stability.
First, we introduce a parameter C to normalize the sensitivities so that the value of can be chosen regardless of the

1164

M. Otomori et al.

Fig. 4 Concept of the

topological derivative

particular problem being solved. We then replace (12) with

dimensionless equations, as follows.

= Cdt F 2 in D
t
(23)

=0
on D.
Here, K in (12) is set to 1. The effect of this term is explained
in Yamada et al. (2010). We note that the F  term in (12)
is replaced by Cdt F . The opposite in sign is due to the
relationship between the definitions of topological derivative and the sign of level set function. C is defined so that

D d
.
(24)
C=
|
d
t F | d
D
Using the finite difference approach, the equations in (23)
are discretized in the time direction as follows:

(t)
(t + t)
2 (t + t) = Cdt F +
(25)
t
t
= 0 on D,
where t is the step for fictitious time t. Using the finite element approach, the above equations are expressed in weak
form as follows:
 (t+t)



dD
+ D T (t + t) dD

D t


(26)
+ (t) dD

for 
Cd
=
F
t

D
t

= 0 on D,
is the functional space defined such that
where 


= (x)|(x) H 1 (D) with = 0 on D .


(27)

Discretizing (26) using the finite element method, we obtain

the following:

T (t + t) = Y
(28)
= 0 on D,
where (t) is a vector that expresses the nodal value of the
level set function at time t. T and Y are now described as

N  

1 T
(29)
T=
N N + T N N dVe
Ve t
e=1

Y=

N  


e=1 Ve

Cdt F +


(x, t)
NdVe ,
t

(30)

where N is the 
number of elements and Ve is the volume
of an element. N
e=1 represents the union set of the elements, where e is the element number. N is the interpolation
function of the level set function.
After updating the level set function using (28), the level
set function is replaced based on the following rule, so that
the upper and lower limit constraints of the level set function
in (4) are satisfied.
if > 1

then = sign().

(31)

3.2 Volume constraint

The volume constraint is dealt with using the augmented
Lagrangian method, in which the Lagrange multiplier , is
updated according to the following scheme.

=

D dt F d
D

d

exp



G
+d
p Gmax

(32)

where p and d are parameters that adjust the position of

the curve. This scheme can be considered as a modified
version of the augmented Lagrangian method, in which
Lagrange multiplier is updated as i+1 = i +cG, where c
is a penalty parameter. We note that the Maclaurin expansion of an exponential function is expressed as ex = 1 + x +
1 2
2 x + , so the standard updating scheme can be considered as a first-order expansion of the above updating scheme
when the volume constraint is active, with the value of pi
used as the penalty parameter c.
When the constraint is sufficiently satisfied, that is,
when the value of G is very small, the value of in (32)
approaches 0. The sensitivity of the constraint functional
then becomes relatively small compared with that of the
objective functional, in the sensitivity expression F  + G .
In this way, the optimization is primarily affected by the
value of the objective functional. On the other hand, when
the constraint is far from being satisfied, the value of
becomes very large, causing the sensitivity of the constraint
functional to become relatively large compared with that of
the objective functional. When this occurs, the optimization
is primarily affected by the need to satisfy the constraint

Matlab code for a level set-based topology optimization method

1165

functional. Although other methods can be applied to satisfy

the volume fraction, we use this scheme for its simplicity.
When the volume fraction of an initial guess is greater
than the maximum allowable volume fraction Vmax , the
volume constraint is relaxed according to the following
formula, to stabilize the convergence.

where Vmax is the upper limit of the volume fraction and tau
is the regularization parameter .



i
 0,
dVmax (V0 Vmax ) max 0, 1
nvol
D
(33)

E0, Emin, and nu are parameters used for analysis, Youngs


G=

where i is the current iteration number and V0 is the volume fraction of the initial guess. The first term represents
the volume of the configuration at the current iteration. The
third term in the right-hand side of (33) is added to the primal volume constraint to relax the upper limit of the volume
constraint so that the constraint is gradually tightened during
nvol iterations. After nvol iterations, the constraint functional
(33) represents the original volume constraint.

4.1 Parameter definition: lines 313

Explanations for the various parameters are given in Table 1.
modulus in the material and void domains, and Poissons
ratio, respectively. nvol corresponds to the iteration number
for the volume constraint, nvol in (33). When the initial volume is greater than the upper limit of the volume
constraint Vmax , the volume constraint is relaxed, and then
it is gradually tightened during the iterations prescribed by
nvol. dt, d, and p are optimization parameters. dt corresponds
to the fictitious time step t in (29) and (30). d and p are
parameters used in the updating scheme applied to Lagrange
multiplier for the volume constraint in the augmented
Lagrangian method formulated in (32). phi and str represent the level set function , and the material distribution,
respectively.

4 Matlab code

4.2 Preparation of finite element analysis: lines 1440

The simple Matlab code for the mean compliance minimization problem is provided in Appendix A. The program
consists of four parts: parameter definitions (lines 313),
finite element analysis preparation (lines 1440), loads and
boundary settings (lines 4153), and the main loop (lines
5488). Figure 5 shows the design domain and boundary
conditions of the problem. Displacement is fixed at the left
boundary and a downward force is applied at the center of
the right boundary. The design domain is discretized using
a rectangular mesh. The numbers of elements are nelx in the
horizontal direction and nely in the vertical direction.
The Matlab code is evoked with the following call:

These lines of code define the preparation for solving the

displacement field and reaction diffusion equation, and
computing the topological derivative. This code follows the
same procedure for computing the global stiffness matrix
as that of Andreassen et al. (2011), where assembly of the
global stiffness matrix is efficiently performed using the
sparse function in Matlab. This procedure is advantageous
from the standpoint of computation time because it avoids
the use of for loops. Lines 1619 define the components of
the element stiffness matrix KE for the displacement field
and these are assembled to create the global stiffness matrix
K at each iteration.

levelset88(nelx,nely,Vmax ,tau),
Table 1 Parameter definition
parameter

meaning

nelx

number of elements in x direction

number of elements in y direction
maximum allowable volume
regularization parameter
Youngs modulus in material domain
Youngs modulus in void domain
Poissons ratio
iteration number for volume constraint
step size for fictitious time t
parameter for augmented Lagrangian method

nely
Vmax
tau
E0
Emin
nu
nvol
dt
d, p

Fig. 5 Fixed design domain and boundary conditions of design

model 1

: set according to the users function call

1166

M. Otomori et al.

Due to the similarity in the formulation of the topological derivative u 0i,j Aij kl u0k,l defined in (22), and the mutual
strain energy density u i,j Kij kl uk,l (Nishiwaki et al. 1998;
Howell 2001), the topological derivative can be computed
using the same procedure as that for computing the mutual
strain energy density, by replacing Youngs modulus E and
Poissons ratio as follows:
E

4A22
,
A1 + 2A2

(34)

A1
,
A1 + 2A2

(35)

where


3(1 ) 1 14 + 15 2
A1 =
E,
2(1 + )(7 5)(1 2)2
A2 =

15E(1 )
.
2(1 + )(7 5)

(36)

node is different for the displacement field (two variables:

u, v) and the level set function (one variable: ), the procedure for computing NN and NNdif is same as that for
computing global stiffness matrix K.
4.3 Loads and boundary settings: lines 4153
Lines 4248 define the boundary conditions for the displacement field. Line 45 defines the load vector that is
applied at the center of right-hand boundary. The length
of the line where the load is applied is 1/16th that of the
entire length of the right-hand boundary. Line 49 defines the
stiffness matrix T shown in (29) for the reaction-diffusion
equation. The load vector Y shown in (30) for the reactiondiffusion equation is defined in the main loop, since it
includes the level set function that varies during optimization. Lines 4953 define the boundary conditions for the
reaction-diffusion equation.

(37)

The detailed derivation of this relationship is provided in

Appendix C. a1 and a2 in the code correspond to A1 and
A2 in the above equation, respectively, and A corresponds
to Aij kl . We note that when the given displacement at
boundary u is set to u = 0 as in Fig. 5, the adjoint problem (46)(48) is equivalent to the state problem (15)(17).
Therefore the optimization problem becomes self-adjoint
problem and u i,j = ui,j .
nodenrs is a matrix that indicates the node number for
all elements and edofVec is a vector representing the first
DOF index for all elements. The i-th column of edofMat
consists of the node number of the i-th element. iK and jK
are index vectors used when assembling the global stiffness matrix using the element stiffness matricies. Details
of this procedure are given in Andreassen et al. (2011).
Next, lines 30 through 40 compute the global matrices NT N,
NT N for the reaction-diffusion equation. NN and NNdif
respectively correspond to NT N and NT N defined in
(29). Although the number of variables defined for each

4.4 Optimization loop: lines 5488

The optimization loop consists of four parts: finite element
analysis and sensitivity computation, convergence check,
computation of Lagrange multiplier for the volume constraint, and the level set function update.
4.4.1 Finite element analysis and sensitivity computation
sK is a vector that is created by reshaping the element
stiffness matrix, and global stiffness matrix K is constructed
using the sparse function. SED and TD are the strain
energy density and topological derivative, respectively. TD

describes the value of the topological derivative in each

element and these values are mapped onto nodes. TDN is the
value of the topological derivative used on the nodes. In line
65 and 66, the objective functional, objective, and the volume of the current configuration, vol, are calculated. Lines
6770 print the optimization results and plot the material
distribution str.

Fig. 6 Optimized configurations for design model 1 obtained using different meshes: a 8064; b 160128; c 320256

Matlab code for a level set-based topology optimization method

1167

Fig. 7 Optimized configurations for design model 1 obtained using different regularization parameter settings: a 2 104 ; b 6 104 ;
c 1 102

4.4.2 Convergence check

Lines 7275 are for the convergence check of the algorithm.
The convergence check is performed after nvol iterations.
The optimization terminates if both of the following conditions are satisfied: the volume is within 0.005 of the required
value, Vmax , and the five previous values of the objective
functional differ by less than 1 %.
4.4.3 Updating the Lagrange multiplier for the volume
constraint
ex corresponds the second and third terms in (33), and
indicates the relaxed upper limit of the volume constraint
so that the constraint will be gradually tightened during nvol iterations. lambda corresponds to the Lagrange
multiplier defined in (32). Line 79 computes the normalization parameter C defined in (24). And g2 in

line 80 expresses the nodal topological derivative in vector

notation.
4.4.4 Updating the level set function
Matrix Y is computed based on the Lagrange multiplier
lambda (line 82) and the level set function phi obtained during the previous iteration. The level set function phi is then
updated by solving the reaction-diffusion equation (line 83).
The updated level set function may not satisfy the upper and
lower limit constraints of the level set function. To ensure
that these constraints are satisfied, the level set function is
replaced based on the rule described in (31) (line 84 in the
code). Nodal level set function phin is mapped onto the elements. An updated structure is obtained using the elemental
level set function phie. The process then returns to the first
step of the optimization loop.

5 Numerical examples
5.1 Mesh independency
First, we examine three cases whose degree of discretization is subject to the following mesh parameters:
80 64, 160 128, and 320 256. The regularization
parameter is set to 2 104 and a volume fraction upper
limit of 0.5 is used in all cases. The Matlab code calls for
these three cases are respectively:
levelset88(80,64,0.5,2e-4),
levelset88(160,128,0.5,2e-4), and
levelset88(320,256,0.5,2e-4).

Fig. 8 Fixed design domain and boundary conditions of design

model 2

Figure 6 shows the optimized configurations for each

case. The results indicate that the optimized configuration
has minimal dependency on the mesh discretization.

1168

M. Otomori et al.

Fig. 9 Optimized configurations for design model 2 obtained using different values for the regularization parameter : a 2 104 ; b 6 104

5.2 Effect of the regularization parameter

Next, we examine three cases in which different degrees of
diffusivity are imposed by setting parameter to the following values: = 2 104 , 6 104 , and 1 102 . A mesh
discretization of 320 256 and a volume fraction upper
limit of 0.5 is used in all cases. Figure 7 shows the optimized configurations for each case. The results show that
relatively complex optimized configurations are obtained
when is set to smaller values, and vice versa. We note that
in stiffness maximization problems for cantilever structures,
a structure with an infinite number of infinitely thin members is known to be optimal. Therefore, as the regularization
parameter is set to smaller values, the optimized configuration becomes more complex and correspondingly better
values of the objective functional are obtained. The results
indicate that essentially any desired degree of geometrical
complexity in the optimized configuration can be obtained,
by setting the regularization parameter to an appropriate
value.
5.3 Other boundary conditions
Last, we offer an additional example in which different
boundary conditions are used. Figure 8 shows the analysis
domain and boundary conditions. A load is applied at the
bottom center of the analysis domain. Vertical and horizontal displacements at the bottom left corner of the analysis
domain are fixed, as is the horizontal displacement at the
bottom right corner of the analysis domain. The design
domain is symmetric along the y-axis and a roller constraint
is imposed on the boundary at the bottom right corner, so
only the right half of the design domain is considered in the
optimization problem.
To enable different load and boundary conditions for this
optimization problem, lines 44, 45, 49 and 50 of the code
are changed, as follows.
Line 44:
F(2*(nely+1):2*(nely+1):2*(nely+1)*
(round(nelx/32)+1),1) = 1;

Line 45:
fixeddofs = [1:2:2*(nely+1) 2*(nely+1)*
(nelx+1-round(nelx/32)):2*(nely+1):2*
(nely+1)*(nelx+1)];

line 49, 50:

fixeddofs phi = [];

Figure 9 shows the optimization results obtained using two

different settings for the regularization parameter . A mesh
discretization of 120 120 (half model) and a volume
fraction upper limit of 0.5 is used in both cases. We can confirm that the optimization obtained appropriate optimized
configurations.

6 Conclusions
This paper presented Matlab code for a level set-based
topology optimization method that uses a reaction diffusion equation to update the level set function. A rigorous
derivation of the topological derivative for a compliance
minimization problem was provided. Although the term representing the effect of boundary conditions when a hole
is created in the design domain was ignored in previous
research (Yamada et al. 2010), this effect was considered in
the method presented here.
We hope that dissemination of this code will enable
readers to more easily understand the operation of this
topology optimization method and also allow them to
compare results obtained when using different topology
optimization methods. We believe that this code is very
compact, comprehensible, and computationally efficient. It
employs the loop vectorization and memory preallocation
detailed in Andreassen et al. (2011).

Acknowledgments The authors would like to express their gratitude

to Prof. Toshiro Matsumoto and Prof. Hiroshi Isakari for their helpful
discussions concerning topological derivatives and encouragement of
this work.

Matlab code for a level set-based topology optimization method

Appendix A: Matlab code

1169

1170

M. Otomori et al.

Appendix B: Topological derivative derivation

The boundary value problem for a created hole can be
described as



div Cij kl uk,l + uk,l = 0

 \ 

in

(38)

ui + ui = u i
ti + ti = ti

on u

(39)

on t

(40)

ti + ti = 0

on

(41)


.

We note that the boundary conditions expressed in (39)

and (40) indicate that the displacement and traction after
creating a hole must satisfy the original displacement and
traction constraints, respectively. (41) represents the free
surface on the boundary of the created hole.
The Lagrangian that includes the created hole is given as
F + F =

The third term on the right-hand side in the above equation

can be replaced using Greens formula, as follows:


ti ui d
u i,j Cij kl uk,l d =
(44)
\
u t




+
Cij kl u i,j nl uk d
div Cij kl u k,l ui d,


\

where ti is the derivative of Lagrange multiplier u i in the

normal direction and nl is the normal vector at boundary

. Substituting the above equation into (43) and considering the conditions, ui = 0 on u and ti = 0 on t , the
variation of the Lagrangian is then given as





F =
div Cij kl u k,l ui d +
(ti ti )ui d
\
t


+
(ui + u i )ti d +
u i,j Cij kl uk,l d



 u
Cij kl u i,j nl uk d
d.
(45)

u t

(ti ui + ti ui + ti ui )d

u i (ti + ti )d +
u i (ti + ti )d

u t

u i,j Cij kl (uk,l + uk,l )d

\


+
d Vmax .

The adjoint equation is now defined so that the integrals of

the term that includes ui or ti are canceled out, as follows:


(46)
div Cij kl u k,l = 0 in 
u i = ui
ti = ti

(42)

\

When the boundary condition (41) is inserted into (42), the

third term on the right-hand side becomes zero. Subtracting
the Lagrangian (19) from the Lagrangian (42), the variation
F of the Lagrangian then becomes

on u

(47)

on t .

(48)

The variation of the Lagrangian can be obtained as follows,

using the adjoint variable obtained by solving the above
adjoint field.


F =
u i,j Cij kl uk,l d
Cij kl u i,j nl uk d



d.
(49)

F =


u t

(ti ui + ti ui )d +

u t

u i,j Cij kl uk,l d +

\

d.


u i ti d

u i,j Cij kl uk,l d



(43)

The first term on the right-hand side of the above equation

represents state variable ui and adjoint variable u i , which
are known values. On the other hand, the value of ui in the
second term is unknown. We note that this second term was
ignored in previous research (Yamada et al. 2010), so that
the method did not consider the effect of boundary condition

that arises when a hole is created in the design domain.

Matlab code for a level set-based topology optimization method

1171

By subtracting the boundary value problem (15)(17)

from the boundary value problem (38)-(41), we obtain the
following boundary value problem by which the unknown
value u is governed.


div Cij kl uk,l = 0 in  \ 
(50)
ui = 0

on u

ti = 0

on t

(52)

ti = ti

on 
.

(53)

(51)

In the above problem, since the radius of the hole

is sufficiently small, the effect of boundaries u and t can be
ignored for solving ui on 
. ui is then a solution of the
following boundary value problem.


div Cij kl uk,l = 0
in  \ 
(54)
ti = ij0 nj + O(
)

on 
,

(55)

where the superscript 0 indicates the value without creating

holes, and ij = Cij kl uk,l . The solution of this boundary
value problem in sphere 
is given (Guzina and Bonnet
2004; Lurie and Belyaev 2005) as:


4 5 0
3 5 0
ui =
nj
nj + O() .
7 5 ij
4(7 5) jj
(56)
The variation of Lagrangian F is then given by:


3
3(1)
(114+15 2 )E
ij kl
F = 4
3
2(1+)(75)
(12)2


1
0 u0 ,
+
5E(

)
u

ik
j
l
il
j
k
i,j k,l
1

The first term on the right-hand side in (21) can then be

given as follows:


u i,j Aij kl uk,l = u i,j A1 ij kl + A2 (ik j l + il j k ) uk.l
= A1 (u 1,1 + u 2,2 + u 3,3 )(u1,1 + u2,2 + u3,3 )
+A2 (u 1,1 u1,1 + u 1,2 u1,2 + u 1,3 u1,3
+u 2,1 u2,1 + u 2,2 u2,2 + u 2,3 u2,3 + u 3,1 u3,1
+u 3,2 u3,2 + u 3,3 u3,3 )
+A2 (u 1,1 u1,1 + u 1,2 u2,1 + u 1,3 u3,1
+u 2,1 u1,2 + u 2,2 u2,2 + u 2,3 u3,2 + u 3,1 u1,3
+u 3,2 u2,3 + u 3,3 u3,3 )
= (A1 + 2A2 )u 1,1 u1,1 + (A1 + 2A2 )u 2,2 u2,2
+(A1 + 2A2 )u 3,3 u3,3
+A1 (u 1,1 u2,2 + u 2,2 u1,1
+u 2,2 u3,3 + u 3,3 u2,2 + u 3,3 u1,1
+u 1,1 u3,3 )
+A2 [(u 1,2 + u 2,1 )(u1,2 + u2,1 )
+(u 2,3 + u 3,2 )(u2,3 + u3,2 )
+(u 3,1 + u 1,3 )(u3,1 + u1,3 )].

(64)

u i,j Aij kl uk,l = (A1 + 2A2 )

11
11 + (A1 + 2A2 )
22
22 + (A1 + 2A2 )
33
33
+A1 (
11
22 +
22
11 +
22
33 +
33
22 +
33
11 +
11
33 )
+A2 (12 12 + 23 23 + 31 31 )


=
11
22
33 12 23 31

A1
A1 + 2A2 A1
A1
A1 + 2A2 A1

A1
A1
A1 + 2A2

0
0
0

0
0
0
0
0
0

(57)

0 4

(65)

0
0
0
A2
0
0

0
0
0
0
A2
0

0
0
0
0
0
A2

11

22

33

12

23
31

(66)

Similarly, for the plane stress problem, we have

(59)

u i,j Aij kl uk,l

Appendix C: Topological derivative implementation

(60)

A1
where c = A1 +2A
. Compare this with the following
2
formulation of the mutual strain energy density:

Namely,


3(1 ) 1 14 + 15 2
15E(1 )
.
A1 =
E, A2 =
2
2(1 + )(7 5)
2(1 + )(7 5)(1 2)
(61)

11
0
A1 + 2A2 A1

22
A1 + 2A2 0
=
11
22 12 A1
12
0
0
A2 0

1
c
0

11


= (A1 + 2A2 )
11
22 12 c 1 0
22 ,
12
0 0 1c
2


(67)

First, for simplicity, we reformulate the tensor Aij kl defined

in (22) as follows:
Aij kl := A1 ij kl + A2 (ik j l + il j k ) .

(63)

In the above formulation, we use u0i,j instead of ui,j for

simplicity. Now, using the following definition of strains,

ii = ui,i , and ij = ui,j + uj,i , the above equation is

transformed as follows:

where ij is Kroneckers delta function. Finally, the topological derivative of the Lagrangian is given as
F
dt F = lim
= u 0i,j Aij kl u0k,l ,
(58)
3
where Aij kl is defined as follows:

3(1)
(114+15 2 )E
ij kl
Aij kl = 2(1+)(75)
(12)2

+5E(ik j l + il j k ) .

(62)

u i,j Eij kl uk,l

"
#

11

 1 0
E

11
22 12 1 0
22 .
=
(1 )2
12
0 0 1
2
(68)

1172

M. Otomori et al.

The topological derivative can be therefore computed by

substituting A1 , A2 into E, in the procedure for computing
the mutual strain energy density, as follows:
E (A1 + 2A2 )(1 c)2 , c.

(69)

References
Allaire G (2002) Shape optimization by the homogenization method.
Vol. 146 of applied mathematical sciences. Springer, New York
Allaire G, Pantz O (2006) Structural optimization with FreeFem++.
Struct Multidiscipl Optim 32:173181
Allaire G, Jouve F, Toader AM (2002) A level-set method for shape
optimization. C R Acad Sci Paris Ser I 334:11251130
Allaire G, Jouve F, Toader AM (2004) Structural optimization using
sensitivity analysis and a level-set method. J Comp Phys 194:363
393
Allaire G, de Gournay F, Jouve F, Toader AM (2005) Structural optimization using topological and shape sensitivity via a level set
method. Control Cybern 34:5980
Andreassen E, Clausen A, Schevenels M, Lazarov BS, Sigmund O
(2011) Efficient topology optimization in MATLAB using 88 lines
of code. Struct Multidiscipl Optim 43:116

Bendse MP, Kikuchi N (1988) Generating optimal topologies in

structural design using a homogenization method. Comput Methods Appl Mech Eng 71:197224
Bendse MP (1989) Optimal shape design as a material distribution
problem. Struct Multidiscipl Optim 1:193202
Challis VJ (2010) A discrete level-set topology optimization code
written in Matlab. Struct Multidisc Optim 41:453464
Eschenauer HA, Kobelev VV, Schumacher A (1994) Bubble method
for topology and shape optimization of structures. Struct Multidisc
Optim 8:4251
Howell HH (2001) Compliant mechanisms. Wiley
Guzina BB, Bonnet M (2004) Topological derivative for the inverse
scattering of elastic waves. Q J Mech Appl Math 57(2):161179
Lurie AI, Belyaev A (2005) Theory of elasticity. Springer, New York
Nishiwaki S, Frecker MI, Min S, Kikuchi N (1998) Topology optimization of compliant mechanisms using the homogenization
method. Int J Numer Methods Eng 535559
Osher SJ, Sethian JA (1988) Fronts propagating with curvature dependent speed: algorithms based on Hamilton-Jacobi formulations. J
Comput Phys 79:1249
Sigmund O (2001) A 99 line topology optimization code written in
Matlab. Struct Multidisc Optim 21:120127
Wang MY, Wang X, Guo D (2003) A level set method for structural topology optimization. Comput Methods Appl Mech Eng
192:227246
Yamada T, Izui K, Nishiwaki S, Takezawa A (2010) A topology
optimization method based on the level set method incorporating
a fictitious interface energy. Comput Methods Appl Mech Eng
199:28762891