XXXVI Ibero-La n Ameri an

Congress on Computa onal

Methods in Engineering

Rio de Janeiro, 22-25 Nov

UNCERTAINTY BASED TOPOLOGY OPTIMIZATION USING THE

TOPOLOGICAL DERIVATIVE CONCEPT
Renatha Batista dos Santos
Antonio Andre Novotny
[email protected]
[email protected]
Laboratorio Nacional de Computaca o Cientfica LNCC/MCT
Coordenaca o de Matematica Aplicada e Computacional
Av. Getulio Vargas 333, 25651-075 Petropolis - RJ, Brasil
Andre Jacomel Torii
[email protected]
Universidade Federal da Paraba - UFPB
Departamento de Computaca o Cientfica
Cidade Universitaria, Joao Pessoa, CEP 58051-900, Brasil
Abstract. Optimization of structural topology has been subject of intense research for many
decades. In most cases, the parameters of the problem are taken as deterministic values. However, some parameters, such as loads and material properties, are frequently uncertain in practice. For this reason, several authors proposed approaches to address optimization problems
subject to uncertainties, establishing the field of uncertainty based optimization. In this paper
we present an approach for uncertainty based topology optimization. The topology optimization
is made using an efficient algorithm based on topological derivatives and uncertainties on load
magnitudes are considered.
Keywords: Topology Design; Structural Optimization; Robust Optimization; Uncertainties;
Topological Derivative

CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

INTRODUCTION

Optimization of structural topology has been subject of intense research for many decades.
On the other hand, in most cases, the parameters of the problem are taken as deterministic values. However, some parameters, such as loads and material properties, are frequently uncertain
in practice. For this reason, several authors proposed approaches to address optimization problems subject to uncertainties, establishing the field of Uncertainty Based Optimization, see for
example the reviews presented by Beyer and Sendhoff (2007); Schueller and Jensen (2008);
Lopez and Beck (2012) and references therein for details.
It has been observed that uncertainties consideration may lead to solutions conceptually different from deterministic topology optimization. This fact supports the application of optimization under uncertainties in several cases of practical interest. However, in general, uncertainty
based topology optimization requires much more computation effort than its deterministic counterpart. This fact limits the range of practical applications of uncertainty based optimization. In
particular, most works on the subject can be grouped into three main classes: Reliability Based
Design Optimization (RBDO), Robust Optimization and Risk Based Optimization. In the case
of RBDO the constraints of the optimization problem are stated as to impose a maximum failure
probability of the system (Maute and Frangopol, 2003; Luo et al., 2014). In the case of Robust
Optimization, the goal is to obtain an optimum design that is least sensitive to variations and
uncertainties of the variables (Dunning and Kim, 2013; Zhao and Wang, 2014). In the case of
Risk Based Optimization, the failure probability of the system is used to evaluate the total cost
of failures, that together with other costs compose the objective function of the problem (Beck
and Gomes, 2012).
In this context, the goal of this paper is to present an approach for robust compliance topology optimization that is both general and computationally efficient. The compliance is evaluated
considering a point-wise worst case scenario, found within an event set of possible outcomes
of the random parameters. Analogously to Sequential Optimization and Reliability Assessment
(SORA) (Du and Chen, 2004), the resulting robust optimization problem can be decoupled into
a deterministic topology optimization step and a reliability analysis step. Since the proposed
approach decouples the nested robust optimization problem, it is computationally efficient and
easy to implement provided that it requires only existing algorithms. In the reliability analysis step the point-wise worst case scenarios are found as in the Performance Measure Approach
(PMA). The topology optimization algorithm proposed by Amstutz and Andra (2006) is used to
solve the associated deterministic problem, which is based on the topological derivative concept
and a level-set domain representation method. In addition, from the mathematical point of view,
the topological derivative concept has been proved to be robust with respect to uncertainties on
the data (Hlavac ek et al., 2009). In this paper therefore, the theory developed by Hlavac ek
et al. (2009) is also confirmed from the numerical point of view. In particular, the numerical
examples presented at the end of this paper address the case of topology optimization with load
magnitudes as random variables. The general statement of the problem is also presented and
can be extended to other random variables as well, which however may require some additional
computational effort.
This paper is organized as follows. The robust design optimization problem is stated in
Section 2. In particular, we describe how the SORA concept can be used to decouple the nested
robust optimization problem into a deterministic optimization step and a reliability analysis step.
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

R.B. Santos, A.A. Novotny, A.J. Torii

The resulting deterministic topology optimization problem is presented in Section 3, which is

solved using the topological derivative concept together with a level-set domain representation
method. Some numerical examples are presented in Section 4, showing the efficiency and
simplicity of the proposed approach as well as the importance of considering uncertainties in
the topology optimization problem. Finally, some concluding remarks are presented in Section
5.

PROPOSED APPROACH

Let us introduce a hold-all structural domain D R2 and a subdomain D. We

consider the minimization of the structural compliance under volume constraint, which can be
written as:
Find d D , such that:

Minimize K =
,
D

Subject to

(1)

|| M ,

where || is the Lebesgue measure of (i.e. volume of the structure) and M is a prescribed
amount of material. The quantity is twice the strain energy density, with and used
to denote the stress and the strain tensors, respectively.
When some parameters of the problem (e.g. load magnitudes) are random variables, the
compliance becomes a random variable itself. In this case the compliance is not known in the
deterministic sense and the deterministic problem from Eq. (1) can result in inefficient designs
in practice. In order to take into account such uncertainties it is necessary to apply some Robust
Optimization strategy.
In this paper the Robust Optimization strategy adopted is a common non-probabilistic approach called Worst Case Design Optimization (WCDO) (Guo et al., 2009). In this approach,
the deterministic optimization problem is solved considering a Worst Case Scenario (WCS)
among the possible outcomes of the unknown parameters. The WCS is given by the combination of parameters that lead to the worst performance of the system being optimized.
An important aspect of WCDO is how the set of possible outcomes is defined. Taking the
set too large may result in too conservative designs. Taking the set too small, on the other hand,
can lead to non robust solutions. In this work we follow the WCDO approach, since it leads to
a computationally efficient decoupling of the Robust Optimization problem. However, we use
statistical information of the unknown parameters in order to build the bounded set of possible
outcomes.

2.1

The bounded set of possible outcomes

Here we assume that the unknown parameters are represented by a random vector x with
known statistical information. In order to write the bounded set of possible outcomes in a
concise manner we assume that the random vector x is composed by independent Gaussian
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

components xi . Then we can obtain the normalized random vector u by using the transformation
T : x u given by
ui =

x i i
,
si

(2)

where ui are independent Normal random variables with mean equal to zero and unitary standard deviation and i and si are the mean and the standard deviation of the random variable
xi .
Using the normalized random vector u = T (x) we can define the event set as
E = {x Rm : kT (x)k t },

(3)

where t is some prescribed parameter. Note that the event set is composed only by outcomes
as distant as t from the mean value T (x) = 0. In other words, only outcomes obtained within
t standard deviations from the mean value are considered.
From the definition of the cumulative distribution function (CDF)
P (u t ) = (t )

(4)

where is the cumulative distribution function (CDF) of a Normal random variable. Consequently, by imposing kT (x)k t we are actually imposing a target probability of occurrence
Pt = P (kT (x)k t ) = (t ).

(5)

From Eq. (5) we note that by prescribing a given value t we are actually prescribing that
the set E is comprised only by outcomes which CDF correspond to a target probability Pt .
Consequently, we can obtain the value of t for a given Pt from
t = 1 (Pt ).

(6)

By setting Pt = 0.95, for example, the event set will comprise all outcomes that correspond to
95% of the occurrences. The value of Pt can then be adjusted to consider a larger or smaller
portion of the possible outcomes as needed. Finally, we note that t is related to the HasoferLind reliability index (Haldar and Mahadevan, 2000; Melchers, 1999), in the sense that it defines
a distance from the expected value in the normalized space.

2.2

The robust optimization approach

In order to build a Robust Optimization approach, we search for the WCS point-wisely
inside the structural domain. At each point of the domain, the strain energy density WCS x for
events in the set E can be found by solving the following problem:
Find x Rm , such that:

Maximize (x) (x) ,

xE

Subject to kT (x)k ,
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

(7)

R.B. Santos, A.A. Novotny, A.J. Torii

where (x) (x) represents the fact that the strain energy depends on unknown parameters
x and the constraint kT (x)k states that we search for a solution in the set E as defined in
Eq. (3).
Minimization of the strain energy density considering a point-wise WCS can be achieved
by substitution of x into (1). The resulting optimization problem is:
Find p D , such that:

Minimize K =
(x ) (x ) ,
D

Subject to

(8)

|| M ,

where (x ) (x ) indicates that the strain energy density is evaluated considering the pointwise WCS x as defined in (7).
In the context of this work it is important to note the difference between d , solution to
(1), and p solution to (8). Note that p is the optimum solution of the Robust Optimization
problem while d is the solution of the deterministic optimization problem.

2.3

Sequential Optimization and Reliability Assessment

Unfortunately, it is not efficient to address the problem from (8) directly, since x is defined
implicitly by means of (7). This leads to a nested optimization problem, where it is necessary to
solve the maximization problem (7) at each point of the domain before evaluating the objective
function from (8).
However, it is possible to decouple these two optimization problems using concepts from
Sequential Optimization and Reliability Assessment (SORA). We first define the solution of the
problem from (8) as the pair {p , xp }. We note that, x is defined as the WCS obtained for an
arbitrary topology, while xp is defined as the special WCS obtained with the optimum topology
p .
If, for some reason, xp was known beforehand, it would not be necessary to solve the
problem from (7) before solving the problem from (8). In this case, the optimum topology p
can be found directly by taking the WCS xp in the problem from (8). This puts in evidence that,
once the WCS xp is known, the problem (8) becomes a deterministic topology optimization
problem. On the other hand, if the optimum topology p is known, the WCS xp can be found
directly by solving (7) point-wisely with the optimum topology p .
The problem from (7) is stated in the same form frequently encountered in RBDO using the
Performance Measure Approach (PMA) (Lopez and Beck, 2012; Tu et al., 1999). Consequently,
it can be solved by efficient schemes already developed in the literature. Here this problem is
solved using the Advanced Mean Value (AMV) algorithm, originally proposed by (Wu et al.,
1990).
Obviously, neither p nor xp are known beforehand, unless the problem is already solved.
However, it is possible to start from initial approximations and iterate until accurate solutions
are found. In fact, let us indicate some iteration number with (k), then an iterative algorithm for
solving the Robust Optimization problem can be summarized as presented in Algorithm 1. The
procedure is started with an arbitrary topology (0) . With this topology we find the WCS x(0)
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

at each point of interest by solving (7) with (0) . With the WCS x(0) we find a new topology
(1) and so on, until convergence is achieved.
Convergence of the solution can be checked on changes of or x between two successive
iterations. Besides, the point-wise WCS x is evaluated only at interest points of the domain. In
the case the problem is solved using the Finite Element Method, for example, the WCS can be
evaluated only at integration points.
Algorithm 1: Robust Optimization algorithm proposed
input : Initial topology (0)
output: {p , xp }
1
2
3
4
5
6
7
8

Set k 0;
Compute x(0) by solving (7) with = (0) ;
while convergence is not achieved do
Make k k + 1;
Compute (k) by solving (8) with x = x(k1) ;
Compute x(k) by solving (7) with = (k) ;
end while
Return {? , x? } as solution.

STRUCTURAL TOPOLOGY OPTIMIZATION PROBLEM

The robust topology optimization we are dealing with consists in solving problem (8) evaluated at the worst case scenario, solution to the problem (7). Without loss of generality, we
restrict ourselves to uncertainties on the intensity of the applied loads (note that the problem
stated through (7) and (8) is general and can be applied when other parameters are random variables). In this case, we can use linear superposition of the effects in order to save computational
effort. Let us consider that the applied load can be written as a linear combination of qi independent loads, with i = 1, ..., m. Then, the stress and strain tensors at each point can be
written as
:=

m
X

xi (ui )

and

:=

i=1

m
X

xi (ui ) ,

(9)

i=1

where, in this context, xi can be interpreted as load scale factors. The canonical strain (ui )
and stress (ui ) tensors are obtained from each individual load qi . For this purpose, we must to
solve the following set of canonical variational problems related to the structural response when
each load qi is applied:
Find ui V , such that
Z
Z

(ui ) () =
D

qi

V ,

(10)

with (ui ) := C(ui ),

where
1
() = ( + ()> )
2
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

(11)

R.B. Santos, A.A. Novotny, A.J. Torii

is the linearized Green tensor,

C=

E
((1 )I + I I)
1 2

(12)

is the elasticity tensor, I and I are the second and fourth identity tensors, respectively, E is the
Young modulus and the Poisson ratio. The piecewise constant function

1, if x ,
(13)
(x) =
, if x D \ ,
0

with 0 < 0  1, is used to mimic voids. That is, the original structural problem, where the
structure itself consists of the domain of given elastic properties and the remaining part D \
of the hold-all is empty (has no material), is approximated by means of the two-phase material
distribution given by (13) over D where the empty region D \ is occupied by a material (the
soft phase) with Youngs modulus, 0 E, much lower than the given Youngs modulus, E, of the
structure material (the hard phase). The space V is defined as
V := { H 1 (; R2 ) : |D = 0} .

(14)

Here, D and N are Dirichlet and Neumann boundaries, respectively, such that D = D N
with D N = , and qi is the prescribed traction on N . See details in Fig. 1.

Figure 1: Sketch of the elasticity problem.

3.1

Deterministic Topology Optimization Method

The volume constraint in problem (8) is trivially imposed via the Augmented Lagrangian
Method (Campeao et al., 2014). In particular, it can be rewritten as an unconstrained optimization problem as follows:
Minimize J = K + 1 g+ +
D

2 + 2
(g ) ,
2

(15)

where g+ := max{g , 1 /2 }, with g = (|| M )/M , 2 > 0 is a fixed multiplier and 1

is updated according to the following recursive formula:
(0)

(n+1)

= 0
(n)

= max{0, 1 + 2 g } .

(16)

CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

The deterministic structural compliance optimization problem (8) is solved by using the
topological derivative concept, which has been shown to be robust with respect to uncertainties
on the data (Hlavac ek et al., 2009). The topological derivative measures the sensitivity of a
given shape functional with respect to an infinitesimal singular domain perturbation, such as
the insertion of holes, inclusions or source-terms (Novotny and Sokoowski, 2013). In particular, the topological derivative of the shape functional J with respect to the nucleation of a
small circular inclusion with different material property from the background, represented by a
contrast , is given by the sum
DT J = DT K max{0, 1 + 2 g } ,

(17)

where means that if we remove material the volume becomes smaller, on the other hand, if
we insert material the volume increases. The topological derivative for the compliance denoted
by DT K is know and can be found in the book (Novotny and Sokoowski, 2013, ch.5 pp.158),
for instance. It is given by:
DT K = P ,

(18)

where the Polya-Szego polarization tensor P is



1
1
1
II ,
(1 + a2 )I + (a1 a2 )
P=
1 + a2
2
1 + a1

(19)

with
a1 =

1+
1

and a2 =

3
.
1+

(20)

In addition, the contrast in the material property, is defined as follows

, if x ,
0
(x) =
1 , if x D \ .
0

(21)

In fact, the above result together with a level-set domain representation method is used for
solving the deterministic compliance topology optimization problem (8) with x = x(k) fixed,
necessary in step (3) of the algorithm proposed in Section 2.3.

3.2

Deterministic Topology Design Algorithm

The resulting deterministic topology optimization algorithm is now explained in details

for the reader convenience. It has been proposed by Amstutz and Andra (2006) and consists
basically in looking for a local optimality condition for the minimization problem (15), written
in terms of the topological derivative and a level-set function. Therefore, the elastic part as
well as the complacent material D \ are characterized by a level-set function L2 (D) of
the form:
= {(x) < 0 a.e. in D} ,
D \ = {(x) > 0 a.e. in D} ,
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

(22)
(23)

R.B. Santos, A.A. Novotny, A.J. Torii

where vanishes on the interface . A local sufficient optimality condition for Problem (15),
under the considered class of domain perturbation given by circular inclusions, can be stated as
Amstutz (2011)
DT J (x) > 0 x D .
Therefore, let us define the quantity

D J (x), if (x) < 0,

T
g(x) :=
D J (x), if (x) > 0,
T

(24)

(25)

allowing for rewrite the condition (24) in the following equivalent form

g(x) < 0, if (x) < 0,

g(x) > 0, if (x) > 0.

(26)

We observe that (26) is satisfied wether the quantity g coincides with the level-set function
up to a strictly positive number, namely > 0 : g = , or equivalently
#
"
hg, iL2 (D)
=0,
(27)
:= arccos
kgkL2 (D) kkL2 (D)
which shall be used as optimality condition in the topology design algorithm, where is the
angle between the functions g and in L2 (D). Let us now explain the algorithm. We start by
choosing an initial level-set function 0 L2 (D). In a generic iteration n, we compute function
gn associated with the level-set function n L2 (D). Thus, the new level-set function n+1 is
updated according to the following linear combination between the functions gn and n
0 L2 (D),


gn
1
sin((1 )n )n + sin(n )
n N ,
n+1 =
sin n
kgn kL2 (D)

(28)

where n is the angle between gn and n , and is a step size determined by a linear-search
performed in order to decrease the value of the objective function Jn , with n used to denote
the elastic part associated to n . The process ends when the condition n  and at the same
time the required volume |1 |n |/M | M are satisfied in some iteration, where  and M
are given small numerical tolerances. In particular, we can choose
0 S = {x L2 (D); kxkL2 (D) = 1} ,

(29)

and by construction n+1 S, n N. If at some iteration n the linear-search step size is

found to be smaller then a given numerical tolerance  > 0 and the optimality condition is not
satisfied, namely n >  , then a uniform mesh refinement of the hold all domain D is carried
out and the iterative process is continued.

NUMERICAL EXAMPLES

In the numerical examples we assume that the structures are under a plane stress state. The
Youngs modulus and the Poisson ratio are respectively given by E = 1.0 and = 0.3, while
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

the contrast 0 = 104 . The angle defined in Amstutz and Andra (2006), representing the
optimality condition, has converged to a value smaller than 1o in all cases. The mechanical
problem is discretized into linear triangular finite elements and two steps of uniform mesh
refinement were performed during the iterative process. Since only load magnitudes are taken
as random variables, linear superposition of effects is used wherever it is possible. Finally, the
mean compliance and the compliance standard deviation of the structures are computed using
Monte Carlo Simulation with 104 samples (Haldar and Mahadevan, 2000).

4.1

Example 1

Let us consider a square panel of size 1 1 clamped on the top and submitted to a pair
of loads, as shown in Fig. 2. The loading consists of two forces q1 = (2.0, 1.0) and q2 =
(2.0, 1.0) applied on the middle of the bottom edge. The hold-all domain is discretized into
a uniform mesh with 6400 elements and 3281 nodes. The required volume fraction is set as
M = 25%, while the parameter 2 = 1.0.

Figure 2: Example 1: initial guess and boundary conditions.

If topology optimization is made considering all parameters as deterministic, the optimal

topology obtained is that presented in Fig. 3, which is a benchmark solution to this problem.

Figure 3: Example 1: result for the deterministic case.

We now assume that the load scale factors are represented by Gaussian random variables.
The event set is defined with = 2.0 and the load scale factors have unitary means, namely
1 = 2 = 1.0. In the first case, the standard deviations are equal and given by s1 = s2 = 0.20.
The optimum topology is presented in Fig. 4. As expected, the optimum solution of the robust
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

R.B. Santos, A.A. Novotny, A.J. Torii

optimization problem leads to a V-bracket structure, able to handle the horizontal components
of the loads that result for cases where the two forces do not have the same magnitude.

Figure 4: Example 1: result for the robust case with s1 = s2 = 0.20.

The distributions of the load scale factors in the worst case scenario x1 and x2 are shown in
Fig. 5.

(a) index x1

(b) index x2

Figure 5: Example 1: load scale factors x1 and x2 for the robust case with s1 = s2 = 0.20.

Finally, we repeat the same experiment by setting s1 = 0.20 and s2 = 0.02. As expected,
in this case the final topology is not symmetric as can be seen in Fig. 6.

Figure 6: Example 1: result for the robust case with s1 = 0.20 and s2 = 0.02.

The distributions of the load scale factors in the worst case scenario x1 and x2 are shown in
Fig. 7.
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

(a) index x1

(b) index x2

Figure 7: Example 1: load scale factors x1 and x2 for the robust case with s1 = 0.20 and s2 = 0.02.

4.2

Example 2

Now, let us consider the design of a supported beam submitted to the loads as shown in Fig.
8. The hold-all domain is rectangular with dimensions 21. It is discretized into a uniform mesh
with 6400 elements and 3321 nodes. The loading consists of a pair of forces q1 = (0.0, 1.0)
and q2 = (0.0, 1.0) applied on the bottom edge as shown in Fig. 8. The required volume
fraction is set as M = 30%, while the parameter 2 = 1.0. The load scale factors have unitary
means 1 = 2 = 1.0 and identical standard deviations given by s1 = s2 = 0.20.

Figure 8: Example 2: initial guess and boundary conditions.

The optimal topology obtained considering all parameters as deterministic is presented in

Fig. 9. The mean compliance and the compliance standard deviation of this structure were
found to be 59.83 and 17.17, respectively. This results in a coefficient of variation of 0.29.

Figure 9: Example 2: result for the deterministic case.

Now we consider the robust case, with the load scale factors represented again by Gaussian
random variables. The event set is defined with = 2.0. The optimum topology is presented
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

R.B. Santos, A.A. Novotny, A.J. Torii

in Fig. 10. The mean compliance and the compliance standard deviation of this structure are
given respectively by 56.88 and 15.78, which results in a coefficient of variation of 0.28. This
result indicates that the proposed robust approach is able to obtain better solutions than its
deterministic counterpart.

Figure 10: Example 2: result for the robust case.

The distributions of the load scale factors in the worst case scenario x1 and x2 are shown
in Fig. 11. It is interesting to note that the point-wise worst case scenarios present significant
variation over the structural domain.

(a) index x1

(b) index x2

Figure 11: Example 2: load scale factors x1 and x2 for the robust case.

CONCLUDING REMARKS

In this paper an alternative approach for robust topology optimization has been presented.
Thanks to the SORA approach, the problem can be rewritten as a deterministic optimization
problem with modified parameters, that are obtained with inverse reliability analysis. This
allows standard deterministic algorithms already developed to be used in the context of robust
optimization. Here, the deterministic topology optimization problem has been solved using an
efficient method based on the topological derivative concept together with a level-set domain
representation method as proposed in Amstutz and Andra (2006).
The examples presented here considered only the magnitude of the applied load as random
variables. In this case, the stress and strain tensors can be written as a linear combination
of the load scale factors and the canonical stress and strain tensors, reducing drastically the
computational effort. The examples presented show that the approach is able to obtain optimum
solutions that take into account uncertainties of some parameters.

ACKNOWLEDGEMENTS
This research was partly supported by CNPq (Brazilian Research Council), CAPES (Brazilian Higher Education Staff Training Agency) and FAPERJ (Research Foundation of the State
of Rio de Janeiro). These supports are gratefully acknowledged.
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

Uncertainty Based Topology Optimization Using Topological Derivative

REFERENCES
Amstutz, S. Analysis of a level set method for topology optimization. Optimization Methods
and Software, 26(4-5):555573, 2011.
Amstutz, S. and Andra, H. A new algorithm for topology optimization using a level-set method.
Journal of Computational Physics, 216(2):573588, 2006.
Beck, A. T. and Gomes, W. J. S. A comparison of deterministic, reliability-based and risk-based
structural optimization under uncertainty. Probabilistic Engineering Mechanics, 28(0):18
29, 2012.
Beyer, H. and Sendhoff, B. Robust optimization a comprehensive survey. Computer Methods
in Applied Mechanics and Engineering, 196(3334):31903218, 2007.
Campeao, D. E., Giusti, S. M., and Novotny, A. A. Topology design of plates consedering
different volume control methods. Engineering Computations, 31(5):826842, 2014.
Du, X. and Chen, W. Sequential optimization and reliability assessment method for efficient
probabilistic design. ASME Journal of Mechanical Design, 126(2), 2004.
Dunning, P. D. and Kim, H. A. Robust topology optimization: Minimization of expected and
variance of compliance. American Institute of Aeronautics and Astronautics Journal, 51(11):
26562664, 2013.
Guo, X., Bai, W., Zhang, W., and Gao, X. Confidence structural robust design and optimization under stiffness and load uncertainties. Computer Methods in Applied Mechanics and
Engineering, 198(4144):3378 3399, 2009.
Haldar, A. and Mahadevan, S. Reliability assessment using stochastic finite element analysis.
John Wiley & Sons, New York, 2000.

Hlavac ek, I., Novotny, A. A., Sokoowski, J., and Zochowski,

A. On topological derivatives for
elastic solids with uncertain input data. Journal of Optimization Theory and Applications,
141(3):569595, 2009.
Lopez, R. H. and Beck, A. T. Reliability-based design optimization strategies based on form:
a review. Journal of the Brazilian Society of Mechanical Sciences and Engineering, 34(4),
2012.
Luo, Y., Zhou, M., Yang, M. W., and Deng, Z. Reliability based topology optimization for
continuum structures with local failure constraints. Computers & Structures, 143:7384,
2014.
Maute, K. and Frangopol, D. M. Reliability-based design of {MEMS} mechanisms by topology
optimization. Computers & Structures, 81(811):813 824, 2003.
Melchers, R. Structural reliability analysis and prediction. John Willey & Sons, Chichester,
1999.
Novotny, A. A. and Sokoowski, J. Topological derivatives in shape optimization. Interaction
of Mechanics and Mathematics. Springer, 2013.
CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015

R.B. Santos, A.A. Novotny, A.J. Torii

Schueller, G. and Jensen, H. Computational methods in optimization considering uncertainties:

an overview. Computer Methods in Applied Mechanics and Engineering, 198(1):213, 2008.
Tu, J., Choi, K. K., and Park, Y. H. A new study on reliability-based design optimization. ASME
Journal of Mechanical Design, 121(4), 1999.
Wu, Y. T., Cruse, T. A., and Millwater, H. R. Advanced probabilistic structural analysis method
for implicit performance functions. American Institute of Aeronautics and Astronautics Journal, 28(9):16631669, 1990.
Zhao, J. and Wang, C. Robust structural topology optimization under random field loading
uncertainty. Structural and Multidisciplinary Optimization, 50(3):517522, 2014.

CILAMCE 2015
Proceedings of the XXXVI Iberian Latin-American Congress on Computational Methods in Engineering
Ney Augusto Dumont (Editor), ABMEC, Rio de Janeiro, RJ, Brazil, November 22-25, 2015