Beruflich Dokumente
Kultur Dokumente
net/publication/320127827
CITATION READS
1 315
3 authors, including:
Some of the authors of this publication are also working on these related projects:
HyperCool: Topology optimisation of passive coolers for LED applications View project
All content following this page was uploaded by Susana Rojas Labanda on 18 October 2017.
Abstract Structural topology optimization problems are com- Keywords Topology optimization · Density based meth-
monly defined using continuous design variables combined ods · Optimization methods
with material interpolation schemes. One of the challenges
for density based topology optimization observed in the re-
view article Sigmund and Maute (2013) is the slow conver- 1 Introduction
gence that is often encountered in practice, when an almost
solid-and-void design is found. The review article Sigmund and Maute (2013) studies, among
The purpose of this forum article is to present some pre- many other topics, the way optimization algorithms gener-
liminary observations on how designs evolves during the ate the designs in structural topology optimization, i.e. the
optimization process for different choices of optimization design iteration history. The design parametrization is based
methods. Additionally, the authors want to open a discussion on density based topology optimization combined with the
on how to properly define and identify the boundary transla- finite element method for structural analysis and the Solid
tion that is often observed in practice. The authors hope that Isotropic Material with Penalization (SIMP) interpolation
these preliminary observations can open for fruitful discus- scheme (see e.g. Zhou and Rozvany (1991) and Bendsøe and
sions and stimulate further investigations concerning slowly Sigmund (1999)). The computational behaviour reported in
moving boundaries. Although the discussion is centered on Sigmund and Maute (2013) is that the optimization algo-
density based methods it may be equally relevant to level-set rithms suggest the final topology in relatively few iterations.
and phase-field approaches. A substantial number of iterations are then required to find
the final geometry by slowly moving the boundaries.
Sigmund and Maute (2013) use a mechanism design ex-
This research is funded by the Villum Foundation through the research ample to show that “the optimization rapidly finds a fairly
project Topology Optimization – the Next Generation (NextTop). good design but requires a very large number of iterations
S. Rojas-Labanda for just slight improvements in objective function but rather
Technical University of Denmark, Department of Wind Energy large changes in geometry.” This effect is also illustrated
Frederiksborgvej 399, 4000 Roskilde, Denmark
by using two different starting points for a minimum com-
Tel.: (+45) 8191 0527
E-mail: srla@dtu.dk pliance problem resulting in a two-bar truss like structure.
From this example, Sigmund and Maute (2013) conclude as
O. Sigmund
Technical University of Denmark, Department of Mechanical Engi- follows: “This shows that the SIMP approach is very good
neering in locating a good design for a uniform grey starting guess
Nils Koppels Allé, Building 404, 2800 Kgs. Lyngby, Denmark but that it behaves similar to a phase field method, i.e. it
Tel.: (+45) 4525 4256
operates with a solid and void phase and changes shape by
E-mail: sigmund@mek.dtu.dk
moving boundaries, when the design has reached or is ini-
M. Stolpe
tialized in a solid-void state.”
Technical University of Denmark, Department of Wind Energy
Frederiksborgvej 399, 4000 Roskilde, Denmark The underlying reasons for the slowly moving bound-
Tel.: (+45) 2151 8240 aries could be as fundamental as the choice of design parame-
E-mail: matst@dtu.dk trization, which would be unfortunate since it is widespread
2 S. Rojas-Labanda et al.
use and difficult to remedy. The observed behaviour could The mechanism design problem in the nested approach
also be because of the choice of regularization. A sensitiv- used in this article is formulated as
ity based filtering technique is used in Sigmund and Maute
maximize uout = lT u(t)
(2013). It is claimed that the behaviour also occurs for den- n
t∈R
sity based filtering and this is supported by the numerical subject to vT t ≤ V (2)
experiments in this study. In fact, it could be a combination uout ≤ 0
of several of these issues or something completely different. 0 ≤ t ≤ 1,
Unfortunately, the numerical studies performed herein
The output displacement is defined using a unit length vec-
do not lead to any definitive conclusions regarding the above.
tor l with zeros in all the degrees of freedom expect at the
Nevertheless, the study provides insight that may spur fur-
output. The domain contains an input and an output spring
ther studies in this important challenge. Also, the study gives
stiffness (denoted kin and kout , respectively). An example of
valuable insight in the convergence behavior of different op-
a mechanism design problem is found in Figure 3.
timization methods.
The complete formulation of the problems can be found
in, for instance, Rojas-Labanda and Stolpe (2015).
2 Problem formulations
3 Optimization methods and implementation details
Two structural topology optimization problems are consid-
The aim of the article is to assess by example, whether the
ered throughout this article, the classical minimum compli-
boundary translation might be affected by the optimization
ance problem (Bendsøe and Sigmund, 2003) and the com-
solver choice. For this purpose, different optimization algo-
pliant mechanism design problem, see e.g. Sigmund (1997)
rithms are considered. The classical first-order structural op-
and Sigmund (2009). More details regarding topology opti-
timization Method of Moving Asymptotes (MMA), (Svan-
mization problems can be found in e.g. the text book Bendsøe
berg, 1987) is compared to the interior point method im-
and Sigmund (2003). The problems are formulated in their
plemented in IPOPT (Wächter and Biegler, 2006), and a se-
discretized version. Isotropic material and a regular mesh
quential quadratic programming method (SQP) implemented
with a constant density throughout each element are con-
for structural optimization problems see e.g. Boggs and Tolle
sidered. Additionally, only a single and design-independent
(1995) and Rojas-Labanda and Stolpe (2016).
load is assumed. The SIMP material interpolation scheme
All the optimization methods attempt to solve1 the nested
is used to produce almost solid-and-void designs. Thus, the
formulation of the topology optimization problem (1) and
stiffness matrix is defined as in Bendsøe (1989), Bendsøe
(2).
and Sigmund (2003), and Andreassen et al (2011). Finally,
Interior point methods, such as IPOPT, approximately
a density filter is used in order to regularize the problem
solve a sequence of barrier problems
(Bourdin, 2001).
The discretized minimum compliance problem in the nested minimize f (x) − µk φ (x, s)
x,s
approach is formulated as
subject to g(x) + s = 0.
occurs when a good topology is found relatively quickly, but Table 2: Objective function value and the number of itera-
numerous iterations are spent in moving the bars. To the best tions when the minimum compliance problem (see Figure 1)
of our knowledge, no formal definition has been established. is solved with MMA, IPOPT, and TopSQP. The table shows
the results using two different starting points illustrated in
Figure 2. .
4.1 Slowly moving boundaries using different starting
points in the minimum compliance problem
Solver Starting point Compliance Iterations
The design domain considered for the compliance minimiza- MMA Uniform 11.257256 24
tion problem is shown in Figure 1. The design domain is MMA Bar 11.536553 149
discretized as in Sigmund and Maute (2013) using 50 by 20 IPOPTa Uniform 12.034256 36
elements. The volume fraction V is set to 0.2. The radius fil- IPOPTa Bar 11.344896 290
IPOPTm Uniform 11.803645 58
ter is fixed to 1.5 = 0.03 × 50. Two different starting guesses IPOPTm Bar 11.682916 98
are studied; a uniform density (of 0.2) and a straight, solid TopSQP Uniform 11.247049 33
and vertical bar located in the middle of the design domain, TopSQP Bar 11.526452 144
and satisfying the volume constraints, see Figure 2.
This example is suggested in Sigmund and Maute (2013)
as a good benchmark problem for the purpose of studying
slowly moving boundaries. Table 2 contains the informa- The barrier parameter update strategy chosen (monotone
tion of the final designs found by the different optimiza- vs. adaptive) considerably affects the amount of intermedi-
tion methods using two different starting points. The table ate design variables (i.e. the amount of grey). The monotone
includes the objective function value (compliance) and the approach turns the solid and void design in an almost com-
number of outer iterations (gradient evaluations) needed to pletely grey design, while the adaptive strategy has more dif-
solve the problem. For this numerical experiment, IPOPT ficulties in changing the whole design to grey. The scaling
is run using both, the monotone (IPOPTm ) and the adaptive of the problem can also considerably affect the behaviour of
(IPOPTa ) barrier parameter update strategy. interior point methods.
The difference between the number of iterations when In order to understand this effect it is important to keep
different starting points are considered is remarkable. MMA in mind that interior point methods solve sequences of bar-
and TopSQP divide the solid bar into two members and then rier sub-problems (see Section 3). The bound constraints of
slowly move the boundaries outwards. At every iteration, the original problem are included in the objective function
the design is (visually) almost solid-and-void. In contrast, of the barrier sub-problem, and thus, the bounds are penal-
throughout the first iterations of the interior point methods ized by the barrier parameter. For large values of µ, interior
the design domain turns greyish, see Figures 4 and 5. In par- point methods try to reduce both the objective function and
ticular, the initial design of IPOPTm disappears and then the the barrier function. However, while µ goes to zero, the rel-
two bars are appearing in the correct position (see Figure 5). evance of the barrier function decreases.
A short numerical study on the optimization methods influence on topology optimization 5
Fig. 4: Intermediate and final designs at different optimization iterations for IPOPTa .
Fig. 5: Intermediate and final designs at different optimization iterations for IPOPTm .
For the minimum compliance problem, the barrier func- Thus, it cannot blur the solid bar completely as it does in the
tion increases when the design variables tend to the bounds. monotone strategy (IPOPTm ). This behaviour indicates that
Therefore, at the beginning of the optimization history, which the log-barrier term dominates the objective in the barrier
normally has a relatively large barrier parameter value, the problem despite the penalization term (SIMP) in the com-
obtained designs are mostly grey. This same reasoning can putation of compliance.
be used to explain the difference in the design history when
different barrier parameter update strategies are used. If the Therefore, it is possible to find a set of parameters such
adaptive strategy is chosen (IPOPTa ), the barrier penalty de- that interior point methods with a solid bar as initial point
creases fast at the beginning of the optimization process. behave similar to the same method applied to the uniform
starting guess. This explains why the number of iterations
6 S. Rojas-Labanda et al.
for IPOPTm reported in Table 2 is almost the same for the Table 3: The table contains the objective function value and
different starting points. the number of iterations when the inverter problem is solved
with MMA, SQP, and IPOPTa .
Remark 1 During the preliminary numerical experiments,
we experienced that the robustness of the solvers is sensi-
tive to the scaling of the problem. The scaling of the objec- Solver uout Number of iterations
tive function is highly important in the behaviour of interior MMA −1.9924 1, 000∗
point methods since they balance the minimization of the ob- SQP −1.9753 143
jective function and the penalization of the bounds. For some IPOPTa −1.9741 325
poorly scaled problems, the behaviour observed in this arti-
cle may not be visible. For the same reason IPOPTm starts
with a relatively large barrier parameter equal to µ0 = 10. iteration 100 and 1, 562 is 5.5%. Additionally, it includes a
This parameter is responsible of penalizing the bounds. density difference plot between those two designs. In this
On the other hand, when linearly constrained nonlinear example, OC slowly moves the bars of the design.
problems are considered, MMA and TopSQP belong to the We perform the same study using different optimiza-
same class of methods, i.e. sequential convex approximation tion methods with the aim of indicating the difference be-
methods and thus behave similarly. Both solvers produce in- tween optimization solvers with respect to the slowly mov-
termediate designs which are almost black and white. ing boundaries effect. More specifically MMA, SQP, and
IPOPTa are used to solve the mechanism design problem.
Remark 2 The considered problem is in general noncon- Table 3 shows the objective function value (output dis-
vex. The objective function values reported in Table 2 corre- placement, uout ) and the number of iterations needed for the
spond to different local minima. We would like to emphasize different solvers. MMA stopped due to the maximum num-
that the purpose of this study is not to judge which method ber of iterations (1, 000). In contrast, SQP and IPOPTa con-
finds the best design, although this is of course highly rele- verge within relatively few iterations.
vant, but instead to illustrate the qualitative the behaviour of
the algorithms. For a comparative study of the performance Remark 3 The examples in Sigmund and Maute (2013) are
of the solvers we refer to Rojas-Labanda and Stolpe (2015). solved by the MATLAB code described in Andreassen et al
(2011). This code stops if the norm of the difference between
Since interior point methods generally start, or can be two consecutive primal variable vectors is less than some
instructed to start, with a relatively large value of the barrier threshold. This criterion thus stops the algorithm if it is not
parameter, the behaviour of the solver is (relatively) not af- making primal progress. This criterion thus ignores the pos-
fected by the starting point. This observation does not mean sibility that the algorithm may be adjusting the working set
that using interior point methods can avoid the slowly mov- and the corresponding Lagrange multipliers (which is ex-
ing boundaries which are normally observed at the end of actly what is in progress when the boundaries are slowly
the optimization process. A disadvantage is its inability to moving). The differences in stopping criteria, the choice of
take advantage of qualified starting guesses. norms and the scale of the problem explain the difference in
the number of iterations reported in Tables 2 and 3, and in
Sigmund and Maute (2013).
4.2 Slowly moving boundaries in the inverter mechanism
problem MMA relatively fast finds a “good” solid and void de-
sign. However, it needs a significant number of iterations to
The second example presented in Sigmund and Maute (2013) move the bars in order to find better designs. Similar be-
is the compliant mechanism problem described in e.g. Sig- haviour is observed in SQP, but since it uses a more efficient
mund (1997) and Sigmund (2009). Figure 3 shows the de- second-order approximation, the movement of the bars re-
scription of the design domain of the inverter problem. In quires fewer iterations. On the other hand, intermediate de-
particular, the design domain is discretized using 80 × 40 el- signs of IPOPTa contains more grey elements and the solver
ements, the volume fraction limit is set to V = 0.3, and the tries to find a good design without producing prematurely
input and output springs are set to kin = 1 and kout = 0.001, a solid and void design (see Figure 6). These grey designs
respectively. The radius filter is set to rmin = 0.03×80 = 2.4. do not disappear, as it occurs in MMA (and SQP), due to
The review article Sigmund and Maute (2013) includes the barrier penalty parameter. The barrier function is penal-
an example in which the Optimality Criteria method (OC, ized at the beginning of the optimization process, and with
(Rozvany and Zhou, 1991), (Zhou and Rozvany, 1991)) needs it, “black-and-white” designs.
1, 562 iterations before reaching the stopping criteria and in IPOPTa spends most of the iterations in finding an op-
which the difference in objective function values between timal geometry but also in improving the objective function
A short numerical study on the optimization methods influence on topology optimization 7
(e) Iteration 160 (f) Iteration 200 (g) Iteration 250 (h) Iteration 320
value. This observation is clearly illustrated in Figures 7, the percentage of improvement between objective function
8, and 9. The objective function value at each iteration, the values is less than or equal to 10% (rimprov = 10).
number of intermediate design variables (variables between The percentage of improvement in the objective function
0.01 and 0.95), and the norm of the KKT conditions3 are value throughout the optimization process is clearly differ-
presented for each solver. entiated between MMA, SQP, and IPOPTa . Both MMA and
Figure 7 and 8 show that MMA and SQP quickly im- SQP at the 40-th iteration, they obtain an objective function
prove the objective function and then level out. In contrast, value that is “only” 10% higher than the optimal found (at
the objective function values of IPOPTa (see Figure 9) de- most). However, IPOPTa needs 194 iterations to achieve this
creases slower and more gradually than in MMA and SQP. threshold. At this point, the visual difference in the designs
The same occurs for the number of grey elements. IPOPTa of interior point methods are much smaller than the differ-
spends more iterations to remove all the grey elements, while ence in the designs of MMA and SQP. In the two latter cases,
MMA and SQP remove the intermediate design variables the bars have been clearly moved (see Figure 10b and 11b).
at the very beginning of the optimization process, and then These numerical experiments indicate how interior point
they slowly move the bars. methods and convex approximation type methods behave.
From the density history of the solvers, it is possible to The former spends more iterations without suggesting a good
identify when the designs are “visually good”. For MMA, design. However when a “visually good” design is found, it
the design is identified in a very early stage, between the 38- is in fact much closer and accurate to the final design than
th and the 100-th iteration. SQP, in turn, needs about 60 to the “visually similar” designs suggested by MMA and SQP
70 iterations to define the design. In contrast, IPOPTa needs methods.
around 180 − 200 iterations, and these intermediate designs These preliminary and limited numerical experiments
still include grey areas. indicate that the choice of optimization method affects slowly
In addition, Figures 10a, 11a, and 12a, show the ratio moving boundaries, at least visually. The objective func-
of improvement between the intermediate objective function tion value in solvers such as MMA and SQP decreases too
value and the optimal one, i.e it is represented rapidly and the grey elements disappear in few iterations.
u∗out Thus, the final design is obtained by slowly moving the bound-
rimprov = ( − 1) × 100. aries. In contrast, IPOPT balances the improvement of ob-
uiout
jective function value with the reduction of grey elements in
Here, i indicates the outer optimization iteration. For those the early stage. This produces that the boundaries are prop-
positive objective function values, the value represented is erly located and therefore the solver does not need to move
the maximum possible, i.e. 100. A design difference be- them. These preliminary experiments suggests that the
tween a given intermediate iteration and the final iteration term “visually similar” for interior point methods can
is also presented in Figures 10b, 11b, and 12b. In particu- be related to “relatively close objective function value”.
lar the choice of this intermediate iteration is chosen such as This relationship is not found when using MMA or SQP.
3 The KKT condition of IPOPT cannot be obtained with the inter-
face used in these numerical examples, and thus cannot be presented in Remark 4 Enforcing grey designs at the beginning of the
Figure 9. optimization process is not suitable for all multiphysics prob-
8 S. Rojas-Labanda et al.
-0.4
2
-0.6
3
1
-0.8
-1 2.5 0
-1.2 -1
2
-1.4
-2
-1.6
1.5
-3
-1.8
-2 1 -4
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Iterations Iterations Iterations
(a) Objective function value (b) Number of intermediate design variables (c) Norm of the optimality conditions
Fig. 7: Objective function value at each optimization iteration, number of intermediate design variables (logarithmic scale),
and norm of the KKT conditions (logarithmic scale) using MMA as solver.
3.5 2
Objective function value
3 0
-0.5
2.5 -2
-1
2 -4
-1.5
1.5 -6
-2 1 -8
0 50 100 150 0 50 100 150 0 50 100 150
Iterations Iterations Iterations
(a) Objective function value (b) Number of intermediate design variables (c) Norm of the optimality conditions
Fig. 8: Objective function value at each optimization iteration, number of intermediate design variables (logarithmic scale),
and norm of the KKT conditions (logarithmic scale) using SQP as solver.
3.5
Objective function value
3
-0.5
2.5
-1
2
-1.5
1.5
-2 1
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Iterations Iterations
(a) Objective function value (b) Number of intermediate design variables
Fig. 9: Objective function value at each optimization iteration and number of intermediate design variables (logarithmic
scale) using IPOPTa as solver.
A short numerical study on the optimization methods influence on topology optimization 9
lems. Some physics situation require clearly defined bound- and guide the community to answer some research questions
aries, such as fluid-structure interaction problems. such as how to properly define the slowly moving bound-
aries, how to identify them, and finally how to avoid them.
5 Closing remarks and future directions Acknowledgements We would like to thank Professor Krister Svan-
berg at KTH in Stockholm for providing the implementation of MMA.
We also express our sincere thanks to the two reviewers and the edi-
Several observations, interpretations, and open questions can tor for their honest comments and suggestions which lead to improve-
be made based on the presented numerical results. Since the ments of the article.
observations are supported by only a handful of numerical
experiments they do, of course, not provide conclusive ev-
idence. They do, however, give indications in which areas
future theoretical and numerical investigations should be di- References
rected.
Amestoy PR, Duff IS, L’Excellent JY (2000) Multifrontal
– It is still unclear how the slowly moving boundaries can
parallel distributed symmetric and unsymmetric solvers.
be identified during the optimization process and how to
Computer Methods in Applied Mechanics and Engineer-
properly and theoretically define them.
ing 184(24):501–520
– The TopSQP method behaves similar to MMA for this
Andreassen E, Clausen A, Schevenels M, Lazarov BS, Sig-
type of problems. However, since TopSQP and SQP use
mund O (2011) Efficient topology optimization in MAT-
efficient second-order approximations, they generally con-
LAB using 88 lines of code. Structural and Multidisci-
verge using fewer iterations.
plinary Optimization 43(1):1–16
– Interior point methods behave (if they are appropriately
Bendsøe MP (1989) Optimal shape design as a material dis-
scaled and with appropriate parameters) different com-
tribution problem. Structural Optimization 1(4):192–202
pared to sequential convex approximation methods. At
Bendsøe MP, Sigmund O (1999) Material interpolation
the beginning of the optimization process, relatively large
schemes in topology optimization. Archive of Applied
barrier parameter values in interior point methods can
Mechanics 69(9–10):635–654
turn the initial point to a greyish design. Thus, they per-
Bendsøe MP, Sigmund O (2003) Topology optimization:
form similarly independently of the starting point.
Theory, methods and applications. Springer
– Interior point methods (with proper scale and parame-
Boggs PT, Tolle JW (1995) Sequential Quadratic Program-
ter selection) can keep intermediate design values for a
ming. Acta Numerica 4:1–51
longer period than MMA. Thus, they are able to properly
Bourdin B (2001) Filters in topology optimization. Inter-
identify the location of the structural boundaries, avoid-
national Journal for Numerical Methods in Engineering
ing the expensive operation of moving members once
50(9):2143–2158
the design is solid and void.
Byrd RH, Hribar ME, Nocedal J (1999) An interior point
– The performance of interior point methods is little af-
algorithm for large-scale nonlinear programming. SIAM
fected by the initial point. A direct consequence is that
Journal on Optimization 9(4):877–900
the solver cannot take advantage of qualified starting
Dennis JE, Moré JJ (1977) Quasi-Newton Methods, Moti-
guesses.
vation and Theory. SIAM Review 19(1):46–89
– In the numerical experiments presented, visually similar
Forsgren A, Gill PE (1998) Primal-dual interior methods
designs relate to similar objective function values when
for nonconvex nonlinear programming. SIAM Journal on
IPOPT is used.
Optimization 8(4):1132–1152
– In many practical applications, the optimization solvers
Morales JL, Nocedal J, Wu Y (2010) A sequential quadratic
are stopped after a relatively small number of iterations.
programming algorithm with an additional equality con-
Interior point methods might not produce solid-and-void
strained phase. Journal of Numerical Analysis 32(2):553–
designs since the barrier penalization parameter values
579
could be relatively large for these intermediate iterations.
Nocedal J (1980) Updating Quasi-Newton matrices
– The use of continuation methods to delay the solid-and-
with limited storage. Mathematics of Computation
void designs in MMA and SQP methods might help to
35(151):773–782
reduce the slowly moving boundaries but it would in-
Nocedal J, Wright SJ (1999) Numerical Optimization.
crease the number of iterations.
Springer
The authors of this article hope that the observations pre- Nocedal J, Wächter R, Waltz RA (2009) Adaptive barrier
sented here can lead the field to further investigate the slowly update strategies for nonlinear interior methods. SIAM
moving boundaries issue. We hope these results will help Journal on Optimization 19(4):1674–1693
10 S. Rojas-Labanda et al.
2
(f ∗ /fi − 1) × 100 (log scale)
-2
-4
-6
-8
-10
0 200 400 600 800 1000
Iterations
(a) Percentage between f ∗ and fi (logarithmic scale) (b) Density difference between an intermediate and the final design
Fig. 10: The left figure presents the percentage between the objective function value at each optimization iteration and
minimum objective function value found by the same solver (MMA). The right figure shows the design difference between
the final design (at iteration 1000) and an intermediate design (at iteration 40). The difference in the objective function value
for these two designs is less than or equal to 10%.
0
(f ∗ /fi − 1) × 100 (log scale)
-2
-4
-6
-8
-10
-12
-14
0 50 100 150
Iterations
(a) Percentage between f ∗ and fi (logarithmic scale) (b) Density difference between an intermediate and the final design
Fig. 11: The left figure presents the percentage between the objective function value at each optimization iteration and
minimum objective function value found by the same solver (SQP). Figure 11b shows the design difference between the
final design (at iteration 143) and an intermediate design (at iteration 40). The difference in the objective function value for
these two designs is less than or equal to 10%.
Rojas-Labanda S, Stolpe M (2015) Benchmarking optimiza- Sigmund O (1997) On the design of compliant mechanisms
tion solvers for structural topology optimization. Struc- using topology optimization. Journal of Structural Me-
tural and Multidisciplinary Optimization 52(3):527–547 chanics 25(4):492–526
Rojas-Labanda S, Stolpe M (2016) An efficient second- Sigmund O (2009) Manufacturing tolerant topology opti-
order SQP method for structural topology optimiza- mization. Acta Mechanica Sinica 25(2):227–239
tion. Structural and Multidisciplinary Optimization Sigmund O, Maute K (2013) Topology optimization ap-
53(6):1315–1333 proaches. Structural and Multidisciplinary Optimization
Rozvany GIN, Zhou M (1991) The COC algorithm, part I: 48(6):1031–1055
Cross-section optimization or sizing. Computer Methods Svanberg K (1987) The method of moving asymptotes
in Applied Mechanics and Engineering 89(1–3):281–308 - A new method for structural optimization. Interna-
A short numerical study on the optimization methods influence on topology optimization 11
1
(f ∗ /fi − 1) × 100 (log scale)
-1
-2
-3
-4
-5
-6
-7
0 50 100 150 200 250 300 350
Iterations
(a) Percentage between f ∗ and fi (logarithmic scale) (b) Density difference between an intermediate and the final design
Fig. 12: The left figure presents the percentage between the objective function value at each optimization iteration and
minimum objective function value found by the same solver (IPOPTa ). The right figure shows the design difference between
the final design (at iteration 325) and an intermediate design (at iteration 194). The difference in the objective function value
for these two designs is less than or equal to 10%.