Finite Elements in Analysis and Design 139 (2018) 49–61

Contents lists available at ScienceDirect

Finite Elements in Analysis and Design

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

Topology optimization of binary structures using Integer Linear

Programming
R. Sivapuram,a,* , R. Picelli b
a
Structural Engineering, University of California, San Diego, La Jolla, CA, 92093, USA
b Cardiff School of Engineering, Cardiff University, Queen’ s Buildings, 14-17, The Parade, Cardiff, CF24 3AA, United Kingdom

A R T I C L E I N F O A B S T R A C T

Keywords: This work proposes an improved method for gradient-based topology optimization in a discrete setting of design
Topology optimization variables. The method combines the features of BESO developed by Huang and Xie [1] and the discrete topology
Integer Linear Programming
optimization method of Svanberg and Werme [2] to improve the effectiveness of binary variable optimization.
Relaxation
Herein the objective and constraint functions are sequentially linearized using Taylor’s first order approximation,
Filter
Truncation error
similarly as carried out in [2]. Integer Linear Programming (ILP) is used to compute globally optimal solutions
Compliance for these linear optimization problems, allowing the method to accommodate any type of constraints explicitly,
Displacement constraint without the need for any Lagrange multipliers or thresholds for sensitivities (like the modern BESO [1]), or
heuristics (like the early ESO/BESO methods [3]). In the linearized problems, the constraint targets are relaxed
so as to allow only small changes in topology during an update and to ensure the existence of feasible solutions
for the ILP. This process of relaxing the constraints and updating the design variables by using ILP is repeated
until convergence. The proposed method does not require any gradual refinement of mesh, unlike in [2] and
the sensitivities every iteration are smoothened by using the mesh-independent BESO filter. Few examples of
compliance minimization are shown to demonstrate that mathematical programming yields similar results as
that of BESO for volume-constrained problems. Some examples of volume minimization subject to a compliance
constraint are presented to demonstrate the effectiveness of the method in dealing with a non-volume constraint.
Volume minimization with a compliance constraint in the case of design-dependent fluid pressure loading is also
presented using the proposed method. An example is presented to show the effectiveness of the method in
dealing with displacement constraints. The results signify that the method can be used for topology optimization
problems involving non-volume constraints without the use of heuristics, Lagrange multipliers and hierarchical
mesh refinement.

1. Introduction lem into a density distribution problem, used in the homogenization-

based method and the SIMP (Solid Isotropic Material with Penalization)
The methods for structural topology optimization have been under method [6,7]. The SIMP model rapidly became a popular material inter-
intense research during the last couple of decades. The ultimate goal of polation based method with the work by Ref. [8] and his role in dissem-
topology optimization is to obtain binary solutions representing optimal ination of the method [9]. The method was successfully applied to solve
structural layouts. The topology optimization problem can be modeled a range of important problems, such as design for nonlinear responses
using binary variables, 0 and 1 representing the void and solid regions [10–12], stress-based design [13], design for fluid flow [14,15], dynam-
of the structure, respectively. This type of binary variable optimiza- ics design [16] and others. In such methods, gray transition material
tion presents an extremely challenging large-scale integer programming regions are intrinsically allowed between solid and void in the con-
problem and alternatives to this formulation were proposed [4]. tinuous variables definition. This leads to optimal solutions with non-
Towards obtaining 0∕1 structures, the first widely accepted ideas explicitly defined structural boundaries and, despite their popularity,
considered topology optimization of continuum structures with the this is challenging the development of these methods in problems where
relaxation of binary constraint {0,1} by using continuous density design explicit boundary description is important, e.g., in design-dependent
variables [5]. This transforms the binary variable optimization prob- multiphysics problems.

* Corresponding author.
E-mail address: [email protected] (R. Sivapuram).

https://doi.org/10.1016/j.finel.2017.10.006
Received 17 July 2017; Received in revised form 13 October 2017; Accepted 24 October 2017
Available online 10 November 2017
0168-874X/© 2017 Elsevier B.V. All rights reserved.
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 1. Filtering - Areas of averaging for nodal and filtered elemental sensitivities.

Some researchers proposed few techniques to reduce or eliminate of sensitivity numbers corresponding to the objective function, and the
intermediate density (gray scale) elements in the final solutions, such thresholds are set based on the evolutionary ratios. In this work, the use
as projection methods. These techniques consider Heaviside functions of mathematical programming enables to update the design variables
[17,18] or morphology-based operators [19,20] to project filtered den- without using any thresholds. Another discrete topology optimization
sities into 0∕1 solution space while aiming length scale control. method was proposed by Svanberg and Werme [40] where the authors
There are some gradient-based methods created for obtaining 0∕1 effectively proposed a sequential integer linear programming approach,
optimal solutions. One class consists of boundary-description based where one starts with a coarse mesh to solve an optimization prob-
methods, e.g. the level-set method (LSM), in which an implicit func- lem and uses the final solution of this problem as the initial solution
tion is used to describe the structure with clearly defined boundaries. for optimization on a refined mesh and so on. They also fix a region
In these methods, the design variables are boundary points and shape of the structure every iteration to speed up the optimization [2]. The
sensitivities are derived to predict design changes [21–23]. The opti- method proposed in this work uses a BESO filter to smoothen the sen-
mization is based on the structural shape movements and, thus, the sitivities, which makes it robust and removes the need for hierarchical
final solution is strongly affected by the initial design configuration. A mesh refinement.
novel method with explicit boundary representation was developed by One more discrete structural optimization method uses Genetic
Christiansen et al. [24] with the combination of shape and topology Algorithms. These are derivative-free techniques based on natural selec-
optimization. Other class of gradient-based methods that obtain 0∕1 tion. The design variables are genetically encoded and a pool of solu-
designs employs a discrete approach, which is the focus of this paper. tions are heuristically updated over generations [41,42]. There are
The most established discrete topology optimization method is the too many ways of designing these algorithms and convergence is not
Bi-directional Evolutionary Structural Optimization (BESO) [25]. The always guaranteed. The heuristics used for cross over and selection are
idea is to switch variables between void and solid using a design update problem-dependent and greatly affect convergence.
scheme where sensitivities serve as indicators of the design variable per- While density-based methods are very well developed but do not
formance. Although the idea was initially proposed by Xie and Steven present explicit boundaries during optimization, the BESO method
[26], the method as it is currently used was developed by Huang and showed effective potential as a binary approach such as in prob-
Xie [1], presenting convergent and mesh-independent solutions. Com- lems where boundary identification is important, e.g., in fluid-structure
prehensive reviews on the BESO methods are given in Refs. [27,28]. interaction problems [43]. The modern BESO method [1] updates the
The method has been applied to a wide range of problems like nonlin- design variables relying on a fixed change in volume fraction every
ear structures [29], natural frequency maximization [30,31], material iteration, and is not based on mathematical optimization. Thus, it is not
optimization and multiscale problems [32–37], multiphysics problems guaranteed that each iteration of BESO is an optimal step. The early
[38,39], etc. The design variables are updated based on the thresholds ESO/BESO methods solved volume minimization problems with stress

Fig. 2. Design domains and their loading configurations.

50
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 3. Cantilever beam (Compliance Minimization subject to volume constraint) - Evolution of topologies.

Fig. 4. Cantilever beam (Compliance Minimization subject to volume constraint) - Optimal solution and convergence.

51
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 5. Michell beam (Compliance Minimization subject to volume constraint) - Optimal solution and convergence.

Fig. 6. Michell beam (Volume Minimization subject to compliance constraint) - Evolution of topologies.

52
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 7. Michell beam (Volume Minimization subject to compliance constraint) - Optimal solution and convergence.

Fig. 8. Cantilever beam (Volume Minimization subject to compliance constraint) - Optimal solution and convergence.

[26], displacement and frequency constraints [44]. These methods are

based on heuristics [3] and are non-convergent [1]. The modern BESO
method uses Lagrange multipliers to deal with non-volume constraints.
The selection and updating of these multipliers is not trivial. This paper
aims to create an improved discrete topology optimization method by
using ILP for the binary design variable updating. The objective and
constraint functions are linearized using Taylor’s expansion and the
goal becomes to obtain the optimal change in design variables so as to
improve the structural performance in each optimization step, similar
to that of [2]. The maximum number of elements that can be deleted or Fig. 9. MBB beam - Design domain and loading configuration.
added every iteration may be constrained to keep the truncation error
low, but the optimizer is free to choose the optimal solution without
exactly satisfying the constraints of linear problems. The mathemati- and compliance constrained problems (subsection 4.2). Volume min-
cal programming allows the use of non-volume constraints explicitly, imization subject to compliance constraint in design-dependent fluid
without the use of multipliers. Herein, the sensitivity analysis is simi- pressure loading problems is also investigated, and the results are dis-
lar to that of BESO. Any method of sensitivity analysis can be used as cussed. Compliance minimization subject to volume and displacement
long as only the extreme values 0∕1 are obtained by design variable constraints is also presented (subsection 4.3) to demonstrate effective-
updating. In the proposed method, a soft-kill approach of elements and ness of the method for different displacement constraint values. Section
extrapolation of sensitivities to void regions by filtering are used. 5 presents the summary and key contributions of the work.
The remainder of the paper is organized as follows. Section 2
describes the formulation and implementation details of the proposed 2. Proposed methodology
method. The linearization of a generic binary variable optimization
problem and the relaxation of constraints is given in subsection 2.1. 2.1. Problem linearization
The details of ILP implementation are discussed in subsection 2.2. The
description of sensitivity analysis is presented in subsection 2.3 and sen- The proposed method uses binary design variables, one for each
sitivity filtering is given in subsection 2.4. Section 3 presents the vari- of the finite elements in the mesh of a structural domain, similar to
ous steps used for the proposed method of discrete topology optimiza- that of BESO. The design variable 1 implies the finite element is filled
tion using ILP. Section 4 shows the results obtained by using the pro- with material, 0 indicates void. Consider an objective function f (𝐱), con-
posed method for compliance minimization problems (subsection 4.1) strained by gi (𝐱) ≤ g i . i ∈ [1, Ng ], with 𝐱 being the design variables. The

53
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 10. MBB Beam (Volume minimization subject to compliance constraint - Oscillations during the evolution of topologies).

binary variable optimization problem can be formulated as follows: ation specified in (4).

Minimize f (𝐱) ⎧ −𝜖1 g k ∶ g < (1 − 𝜖1 )g k

𝐱 ⎪
Δg = ⎨ g − g k ∶ g ∈ [(1 − 𝜖1 )g k , (1 + 𝜖2 )g k ] (4)
Subject to gi (𝐱) ≤ g i i ∈ [1, Ng ] (1) ⎪
⎩ 𝜖2 g k ∶ g > (1 + 𝜖2 )g k

xj ∈ {0, 1} j ∈ [1, Nd ]
where g is any of the Ng constraints, and Δg is the right hand side of
where Nd is the number of design variables, i.e., the number of finite the constraint which is relaxed from g − g k at iteration k. The param-
elements in the mesh. Using first order Taylor’s approximation, (1) can eters 𝜖1 , 𝜖2 are chosen such that ILP has a feasible solution. They are
be converted into a linearized form as shown in (2). usually small numbers so as to maintain the linearization valid. Any
other definition for relaxation can be used as long as the constraints are
𝜕f ||
Minimize Δ𝐱k relaxed in a similar way as shown in (4).
Δ𝐱k 𝜕𝐱 ||𝐱k There are many methods in literature where an optimization prob-
𝜕gi || (2) lem is solved by sequentially solving simpler approximate problems.
Subject to Δ𝐱k ≤ g i − gik i ∈ [1, Ng ]
𝜕𝐱 ||𝐱k Griffith and Stewart [45] proposed sequential linear programming
where linearization is used to create subproblems. Svanberg [46] used
Δxj ∈ {−xj , 1 − xj } j ∈ [1, Nd ]
a nonlinear convex approximation for creating subproblems. Sequential
Quadratic Programming (SQP) uses quadratic approximate subprob-
where gik is the value of constraint gi at iteration k of optimization. After
lems which account for curvature of the functions due to which saddle
each iteration, the design variables are updated as shown in (3).
point solutions are avoided [47]. Other methods like [48] use the same
idea of linearization to create subproblems.
𝐱k+1 = 𝐱k + Δ𝐱k (3)

We can see from the optimization problem (2) that the formulation 2.2. Integer linear programming
is generic and can be used for problems involving any type of con-
straints. The binary linear optimization problem in (2) can be solved An integer programming problem is a linear programming problem
using ILP. with the additional restriction that the feasible space of solutions is the
The linearization approximation is valid only for small changes in integer space. Branch and bound algorithm is a famous technique used
the objective and constraint functions at every iteration. To achieve to solve an ILP. The technique initially solves the ILP as a Linear Pro-
this, we use a relaxation scheme on targets g i of the constraints such gram (LP), i.e., without any integer restraints. The solution obtained is
that the differences between the constraint targets and the current val- used as the initial solution and different LPs are solved adding extra
ues of constraints are small. This also ensures that the ILP has a feasible bounds on the design variables which force the optimizer to yield inte-
solution. Whenever gik is away from g i , we use a constraint target relax- ger solutions in the branches. In general, the computational complexity

54
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 11. MBB Beam (Volume Minimization subject to compliance and volume addition and removal constraints) - Evolution of topologies.

Fig. 12. MBB beam (Volume Minimization subject to compliance and volume addition and removal constraints) - Optimal solution and convergence.

of an ILP problem is higher than that of an LP problem [49]. 2.3. Sensitivity analysis
In this work, we used the intlinprog function of MATLAB for
solving ILP problems. We used some options in the optimizer to ensure In this paper, we focused on solving problems related to linear elas-
better performance. While solving an ILP problem, additional con- ticity with compliance and volume as either objective or constraint
straints called cuts are imposed by the optimizer to restrict the feasi- functions. If C is the compliance of a structure with design-independent
ble space to integer solutions. The CutGeneration option is set to loads, the sensitivity with respect to design variable xj is given by (5).
intermediate which uses the most cut types [50]. Primal-Simplex
algorithm is used for solving the LP problems formed in branches of 𝜕C 1
the optimizer. The number of nodes searched by the branch and bound ≈ − 𝐮Tj 𝐊ej 𝐮j (5)
𝜕xj 2
method of ILP HeuristicsMaxNodes is chosen to be 100. The deci-
sion on whether a design variable enters the basis is taken by checking
if the reduced cost corresponding to the design variable exceeds a tol- where 𝐊ej is the stiffness matrix corresponding to solid finite element j,
erance. This tolerance is set in the LPOptimalityTolerance option and 𝐮j is the displacement vector corresponding to finite element j. In
to 10−10 . These options yielded better performance of ILP over a wide this work, we have used the soft-kill approach, i.e., the void elements
range of problems. are assembled in finite element analysis using xj = 0.001, to prevent
singularity of the global stiffness matrix 𝐊. We also observed that the
oscillations in topology optimization are reduced by using the soft-kill
approach. The sensitivity of volume V of the structure with respect to

55
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

2.4. Sensitivity filter

The sensitivities need to be filtered to make sure the existence of

solutions. Filtering is also required to obtain non-checkerboard and
mesh-independent optimal solutions [28]. A nodal sensitivity field is
created from the design variable sensitivities, by averaging the sensi-
tivities corresponding to finite elements attached to the nodes. Fig. 1a
shows nodes 1 and 2 and the corresponding finite element regions E1
and E2 whose sensitivities are used for averaging at the nodes. The
nodal sensitivity (with respect to a virtual nodal design variable yn ) at
a node n is calculated as shown in (10).
𝜕C 1 ∑ 𝜕C
= (10)
𝜕yn |E| e∈E 𝜕xe
Fig. 13. Cantilever beam - Initial Solution.
where E is the set of finite elements having n as a node, and |E| is the
number of such elements. The filtered sensitivity field with respect to
design variable xj is shown in (6). the density design variables is reconstructed from the nodal sensitivity
field by using weighted-distance averaging in a neighborhood of each
𝜕V
≈ Vj (6) finite element. Fig. 1b shows the filtering regions N1 and N2 whose
𝜕xj
nodal sensitivities are used for averaging to compute the filtered sen-
where Vj is the volume of finite element j. In this work, unit-sized finite sitivities for elements 1 and 2. The filtered sensitivity with respect to
elements are used in topology optimization, i.e., Vj = 1, ∀j ∈ [1, Nd ]. density xj is computed as shown in (11).
We also presented topology optimization of a structure with design- ∑ 𝜕C
dependent fluid pressure loads [51]. The sensitivity of compliance with ̃
𝜕C i∈N wji 𝜕yi
= ∑ (11)
respect to design variable xj is given by (7). 𝜕xj i∈N wji

𝜕C 1 where N is the set of nodes in a neighborhood of the element j. In

≈ − 𝐮Tj 𝐊ej 𝐮j + 𝐮Tj Δ𝐋𝐏f (7)
𝜕xj 2 this work, we use a circular neighborhood of radius r with the center
being the centroid of finite element j for filtering, and wji = max(0, r −
where 𝐏f is the fluid pressure in the fluid finite element which shares
dist(j, i)), with dist(j, i) being the distance between centroid of finite
boundary with the solid finite element j, and Δ𝐋 = 𝐋∗ − 𝐋 is a semi-
element j and node i.
analytical sensitivity of the fluid-solid coupling matrix, where 𝐋 and 𝐋∗
are the fluid-solid coupling matrices obtained before and after changing
the solid element j to fluid. 3. Steps in the proposed method
When an inequality constraint is applied on a displacement degree
of freedom in the structure (Equation (8)), we solve an adjoint problem This section presents the algorithmic steps for the proposed method.
for computing the sensitivities. 1. Consider the topology optimization of a structure. Discretize the
d≤d (8) design domain, apply loads and boundary conditions. The model
can be reduced if any symmetry is present.
The adjoint problem and sensitivity of the displacement constraint 2. Choose the parameters like 𝜖1 and 𝜖2 (based on the general range of
(8) with respect to xj are given by: constraint values), 𝛽a , 𝛽r , filter radius r and convergence criteria.
3. Perform Finite Element Analysis (FEA) for the structure and com-
𝐊𝝀 + 𝐇 = 0 pute sensitivities for the objective function and constraints with
𝜕d (9) respect to the design variables 𝐱 using (5)–(7) and (9), etc. The val-
≈ 𝝀Tj 𝐊ej 𝐮j ues of the objective and constraint functions are calculated.
𝜕xj
4. Filter the sensitivities for each function (volume sensitivities need
where 𝝀 is the vector of adjoint variables, and has the same size as that not be filtered since they are uniform for a uniform mesh) using the
of the global displacement vector 𝐮, 𝐇 is the vector having only one details presented in subsection 2.4.
non-zero entry 1 at the location corresponding to displacement d and 𝝀j 5. The constraints are relaxed so that the structure does not change
is the vector of adjoint variables corresponding to finite element j. very much using (4).

Fig. 14. Cantilever beam (Volume Minimization subject to compliance and volume addition and removal constraints) - Optimal solution and convergence.

56
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

6. Some extra constraints to restrain the maximum number of elements design variables is computed.
added and removed may be used, as described in subsection 4.2. 8. The design variables are updated using (3).
7. The filtered sensitivities are supplied to ILP and the optimization 9. The structure generated using the updated design variables is tested
problem (2) is solved after replacing the right hand side of con- for convergence. If convergence is achieved, continue to post-
straints with the relaxed constraint targets. The optimal change in processing, otherwise go to step 3.
10. Post-processing like plotting the optimized structure and conver-
gence of the functions is finally done.

4. Examples

This section aims to demonstrate the performance of the method

proposed in this work. First, it is shown that volume constraints used
in the standard BESO are well accommodated in the present methodol-
ogy. Different examples of volume minimization subject to compliance
constraints are then used to demonstrate the flexibility of the method in
dealing with non-volume constraints. An example is presented for mini-
mizing the volume subject to compliance constraints when the structure
is subjected to design-dependent loads. Finally, the effectiveness of the
method in dealing with displacement constraints is explored.

4.1. Compliance minimization subject to volume constraints

We present some examples of topology optimization for minimizing

compliance with volume being constrained. Fig. 2 shows the examples
considered for topology optimization. Fig. 2a shows a cantilever beam
(1.6 ∶ 1 aspect ratio) ABCD with a unit vertical load applied at Point E
Fig. 15. Design domain with fluid pressure loading. on the right edge BC. Due to symmetry, only the lower half of design

Fig. 16. Arch-like structure (Volume Minimization subject to compliance (5 Nm) and volume addition and removal constraints) - Optimal solution and convergence.

Fig. 17. Arch-like structure (Volume Minimization subject to compliance (6 Nm) and volume addition and removal constraints) - Optimal solution and convergence.

57
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 18. Arch-like structure (Volume Minimization subject to compliance (7 Nm) and volume addition and removal constraints) - Optimal solution and convergence.

domain is considered for optimization. Fig. 2b shows the symmetric

portion of a Michell beam (2 ∶ 1 aspect ratio) ABCD with a unit point
vertical load applied at A. The design domains are well-supported to
avoid rigid body motions. The Young’s Moduli of the solid and void
finite elements are 1.0 and 0.001 respectively, while the Poisson’s ratio
is 0.3 for both types of elements. The cantilever design domain in Fig. 2a
is discretized using 160 × 100 bilinear quadrilateral finite elements.
Topology optimization requires an initial solution to start iterating
so as to find a local optimal solution. A design domain completely filled
with solid finite elements is used as the initial guess for optimization.
The upper bound on volume fraction is 0.5 in both the examples. We
used 𝜖1 = 𝜖2 = 0.01, r = 6 for the cantilever beam and r = 3 for the
Michell structure in terms of number of finite elements. Few topolo-
gies during the evolution of cantilever beam are shown in Fig. 3. We
can observe that the regions near Points B and C become void at the
very early stages of evolution because these regions experience negligi-
ble strains. Members are formed inside the structure to keep the com-
pliance low while deleting some material to achieve the volume con-
straint. Fig. 4 shows the optimum structure obtained after 36 iterations
and the convergence history. We can observe that both the objective
and constraint functions converge smoothly.
The Michell beam design domain in Fig. 2b is discretized using
100 × 100 bilinear quadrilateral elements. Fig. 5 shows the optimum
structural design and convergence history of the Michell beam. The con-
vergence is achieved at 66 iterations and both compliance and volume Fig. 19. Design domain and boundary conditions [53].

fraction converge smoothly.

These examples demonstrate that the proposed method yields good
results for compliance minimization with volume constraint similar to in the structure over iterations to improve the objective (volume) while
that of BESO, by using ILP instead of sensitivity number thresholds. satisfying the inequality constraint (compliance). Fig. 7 shows the opti-
The only parameters in this algorithm are 𝜖1 , 𝜖2 which are used for mal solution obtained after 197 iterations and the convergence history.
relaxing the constraints thereby creating feasible points for ILP, and r We can see that convergence of the objective and constraint functions is
for sensitivity filtering. The method produces good solutions directly smooth for this compliance-constrained problem. We can also observe
using fine meshes, unlike [2] who used hierarchical mesh refinement. that since the compliance constraint chosen is the optimal compliance
value from the compliance minimization of Michell structure (5), the
4.2. Volume minimization subject to compliance constraints optimal volume fraction of the current volume minimization problem is
equal to the volume fraction constraint of the former problem, i.e, 0.5.
The design domains in Fig. 2a and b are considered for minimization This demonstrates that the method is capable of reproducing structures
of volume subject to compliance constraint. The design domains com- obtained through compliance minimization, even when compliance is
pletely filled with solid finite elements are used as the initial solutions. used as a constraint.
The parameters chosen are 𝜖1 = 𝜖2 = 0.0005, r = 5 and r = 3 finite ele- Fig. 8 shows the optimal structure and convergence history for the
ments for the cantilever and Michell problems respectively. volume minimization of cantilever beam, with an upper bound on
The upper bound of compliance for the Michell beam topology opti- compliance at 1.88 × 105 Nm. We can observe the optimal structure
mization is 10.7 × 104 Nm. Fig. 6 shows few topologies during the evo- obtained looks similar to that obtained in the compliance minimization
lution of Michell beam. At the beginning of evolution, the finite ele- example (Fig. 4). The optimal cantilever obtained in the compliance
ments near Point C are subject to negligible strains. These finite ele- minimization case has an objective of 1.88 × 105 Nm and a volume
ments get deleted at an early stage of optimization. Members are formed fraction of 0.5. We can observe that the volume fraction of optimal

58
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

structure for the volume minimization problem is 0.5. This means that
equivalent structures are obtained by changing the roles of compliance
and volume fraction between objective and constraint functions.
Fig. 9 shows the symmetric portion of design domain of a
Messerschmitt-Bölkow-Blohm (MBB) beam with a unit vertical load
applied at point D. The design domain is constrained to prevent any
rigid body motions. The Young’s modulus is 1.0 and Poisson’s ratio
is 0.3 for a solid finite element. The void elements are dealt using
the soft-kill approach. The design domain is discretized using 120 × 40
finite elements. The parameters used are 𝜖1 = 𝜖2 = 0.001, r = 4 finite
elements, and the upper bound on compliance used is 9 × 105 Nm.
Fig. 10 shows few topologies during the evolution of the compliance-
constrained MBB beam. It is observed that the evolution is noisy with
large volumes of material removed and added every iteration after 45th
iteration. The members in the structure also oscillate creating a noisy
evolution.
The parameters 𝜖1 , 𝜖2 relax the constraints, i.e., allow the constraint
(here compliance) to vary by a small amount each iteration. They do
not restrict the amount of volume removed and added each iteration.
This could make the truncation error of linearization larger, making the
linear approximations erroneous. The volume of the material removed
and added at every iteration needs to be constrained so as to keep the
truncation error (O(‖Δ𝐱‖22 )) low. To achieve this, we introduce extra
constraints on the removal and addition of material every iteration.
The constraints are given by (12).
∑
Δxi ≤ 𝛽a Nd V = {i|xik = 0}
i∈V
∑ (12)
Δxi ≥ −𝛽r Nd S = {i|xik = 1}
i∈S

where 𝛽a and 𝛽r are positive fractions and restrain the number of design
variables that are allowed to be removed and added from the structure
respectively. The sets V and S consist of the indices of finite elements
which are void and solid in the current iteration k respectively.
The MBB beam is re-optimized to minimize volume subject to com-
pliance constraint and the additional constraints (12) to restrict the
removal and addition of material every iteration. The fractions used
are 𝛽a = 𝛽r = 0.02.
Fig. 11 shows few topologies during evolution of the optimal struc-
ture. In the initial iterations, the regions with negligible strains like the
regions near Point C and midpoint of edge AD become void. Structural
members begin to form to minimize the volume while maintaining the
relaxed compliance constraint at every iteration. We can see that the
topologies evolve smoothly after adding the extra constraints. Fig. 12
shows the optimal structure of the MBB beam obtained after 99 itera-
tions and the convergence history. We can observe that the objective
and constraint functions converge smoothly. The addition of 𝛽a and
𝛽r has similar goals as that of ARmax and ER parameters of the BESO
method, with a small difference that the target volume fraction is pre-
decided every iteration in BESO, while the maximum number of ele-
ments getting added and deleted every iteration is constrained using 𝛽a
and 𝛽r in the current method [52]. developed a connectivity constraint
for discrete topology optimization methods based on an associativity
matrix which describes the connectivity between active elements. The
connectivity of the structure can be controlled by changing the target
connectivities of active elements in the structure.
The volume minimization of cantilever beam subject to compliance
constraint of 1.8 × 105 Nm is performed and the initial design used is
shown in Fig. 13. The white part signifies the void region and black part
corresponds to the material in the structure. The parameters used are
𝜖1 = 𝜖2 = 0.001, r = 6 finite elements, 𝛽a = 𝛽r = 0.2. The optimal struc-
ture obtained after 60 iterations and the convergence history is shown if
Fig. 14. We can observe that the functions have a smooth convergence.
This example also demonstrates the robustness of the method by using Fig. 20. Optimum solutions for compliance minimization subject to displacement and
an initial solution different from the fully filled design domain. volume constraints.
Fig. 15 shows the symmetric portion of a design domain (2 ∶ 1

59
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

Fig. 21. Convergence of Compliances for the three cases of displacement constraints. Fig. 22. Convergence of displacement constraints for the three cases.

aspect ratio) surrounded by a fluid with constant pressure P0 . The its updating is not trivial. The Lagrange multiplier is non-zero only
design domain is discretized using a mesh of 100 × 100 bilinear quadri- when the inequality displacement constraint is not satisfied. When the
lateral elements. The design domain is subjected to volume minimiza- constraint is just satisfied near convergence, the Lagrange multiplier
tion and a compliance constraint. The parameters used are 𝜖1 = 𝜖2 = switches to zero and the problem becomes a compliance minimization
0.001, r = 4 finite elements, 𝛽a = 𝛽r = 0.1. The fully solid structure is problem subject to volume constraint. This allows the optimizer to min-
used as the initial design for topology optimization. imize compliance without any regard to dB at that iteration, thereby
Figs. 16–18 show the optimal structures and convergence histories potentially failing to satisfy the displacement constraint. This may also
of the volume minimization when the upper bounds used on compli- lead to an oscillatory convergence of the displacement constraint. In
ance are 5 Nm, 6 Nm and 7 Nm respectively. In all three cases, arch-like this work, the constraints are handled explicitly, i.e., no Lagrange mul-
structures are created which are optimal for bearing design-dependent tipliers or penalty-based methods are used. Fig. 22 shows the conver-
fluid pressure loads [51]. As the upper bound on compliance increases, gence of displacement constraint for the three cases. The convergence of
the thickness of the arch decreases due to the removal of more mate- the constraints is smooth, because the constraint does not get removed
rial from the structure. We can clearly observe that the compliance and from optimization even when it is satisfied. Thus the optimizer min-
volume fraction converge smoothly in the three cases. This example imizes compliance without violating the displacement constraint. This
demonstrates that the proposed method works well for compliance con- example demonstrates that the method proposed in this work is capable
straints in problems with design-dependent loads. of handling the non-volume constraints without the use of any Lagrange
multipliers.
4.3. Compliance minimization subject to a displacement constraint
5. Conclusions
This example [53] is a compliance minimization problem subject to
volume and displacement constraints, when a unit point vertical load A discrete method is proposed for topology optimization of struc-
is acting on the structure. Due to symmetry, only half of the model tures to obtain binary (void (0) and solid (1)) designs. The proposed
(50 mm × 50 mm) is considered for optimization, as shown in Fig. 19. method uses sequential linear approximations of the objective and con-
The Young’s Modulus of the material is 1.0 GPa and the Poisson’s ratio straint functions to obtain integer linear problems, solved using Integer
is 0.3. A 100 × 100 mesh of bilinear quadrilaterals is used to discretize Linear Programming (ILP). The proposed method uses mathematical
the geometry. The upper bound on volume fraction of the structure is programming for design variable updating, unlike the BESO method.
0.3. We consider three problem cases where horizontal displacement This helps the method to accommodate non-volume constraints explic-
at Point B is constrained to be 1.4 mm, 1.2 mm and 1.0 mm respec- itly, unlike the modern BESO method which uses Lagrange multipli-
tively. The parameters used are 𝜖1 = 𝜖2 = 0.01 for the volume con- ers. The present method is also mesh-independent because of the use
straint, 𝜖1 = 𝜖2 = 0.25 for the displacement constraint, 𝛽a = 𝛽r = 0.05 of mesh-independency filter taken from the modern BESO method [1].
and r = 3 finite elements. The method does not require any hierarchical mesh refinement dur-
Fig. 20 shows the optimal topologies obtained when the horizontal ing optimization, unlike Svanberg and Werme [2]. A soft-kill approach
displacement at point B dB is constrained to be below 1.4 mm, 1.2 mm is used for dealing with the void elements so as to obtain a positive
and 1.0 mm. The compliance of the optimal solutions is 192.63 Nmm, definite global stiffness matrix for the model. A constraint relaxation
194.95 Nmm and 205.83 Nmm respectively. Fig. 21 shows the conver- method is used to allow only small changes in the constraint functions
gence of compliances. We can observe that the convergence is smooth in each iteration thereby assuring the existence of feasible solutions for
all the three cases. The optimal topologies have features similar to that ILP. Few examples are demonstrated for compliance minimization sub-
of [53], and the compliance values of the optimal solutions are com- ject to volume constraint. The results demonstrate that the proposed
parable. This example demonstrates the effectiveness of the method for method works well in the same way as BESO method for compliance
problems with a displacement constraint. minimization problems. Some examples are solved for volume mini-
[53] use a Lagrange multiplier approach for dealing with the dis- mization subject to compliance constraint, and the results show smooth
placement constraint, so as to modify the topology optimization prob- convergence to optimal structures. We observed oscillations in evolving
lem to optimization of an objective function subject to volume con- topologies because of the unrestrained removal and addition of mate-
straint. The choice of Lagrange multiplier used for the constraint and rial every iteration. This is overcome by constraining the removal and

60
R. Sivapuram and R. Picelli Finite Elements in Analysis and Design 139 (2018) 49–61

addition of material during optimization. The results have shown clear [23] R. Sivapuram, P.D. Dunning, H.A. Kim, Simultaneous material and structural
topologies with nice convergence. Volume minimization subject to com- optimization by multiscale topology optimization, Struct. Multidiscip. Optim. 54
(2016) 1267–1281.
pliance constraint is also solved in problems with design-dependent [24] A.N. Christiansen, M.N. gensen Jr., , N. Aage, O. Sigmund, J.A. Baerentzen,
fluid pressure loading for different upper bounds on compliance. In Topology optimization using an explicit interface representation, Struct.
all cases, nice arch-like structures are created in the optimal solution Multidiscip. Optim. 49 (2014) 387–399.
[25] Y.M. Xie, X. Huang, Evolutionary Topology Optimization of Continuum Structures:
with smooth convergence. We also demonstrated effectiveness of the Methods and Applications, first ed., John Wiley & Sons, Ltd, West Sussex, 2010.
method for a compliance minimization problem subject to volume and [26] Y.M. Xie, G.P. Steven, A simple evolutionary procedure for structural optimization,
displacement constraints. The proposed method obtains smooth conver- Comput. Struct. 49 (1993) 885–896.
[27] D.J. Munk, G.A. Vio, G.P. Steven, Topology and shape optimization methods using
gence of the displacement constraint because the constraint is not dealt evolutionary algorithms: a review, Struct. Multidiscip. Optim. 52 (2015) 613–631.
using Lagrange multiplier. Parameters are required in the proposed [28] L. Xia, Q. Xia, X. Huang, Y.M. Xie, Bi-directional evolutionary structural
method for constraint relaxation, filtering and for restraining addition optimization on advanced structures and materials: a comprehensive review,
Archiv. Comput. Methods Eng. (2016) 1–42.
and removal of material every iteration. The proposed method has the
[29] X. Huang, Y.M. Xie, Topology optimization of nonlinear structures under
potential to deal with non-volume constraints without using heuristics, displacement loading, Eng. Struct. 30 (2008) 2057–2068.
Lagrange multipliers or hierarchical mesh refinement, which is a sig- [30] X. Huang, Z.H. Zuo, Y.M. Xie, Evolutionary topological optimization of vibrating
nificant improvement over the existing discrete topology optimization continuum structures for natural frequencies, Comput. Struct. 88 (2010) 357–364.
[31] R. Picelli, W.M. Vicente, R. Pavanello, Y.M. Xie, Evolutionary topology
methods (like the early ESO/BESO, the modern BESO, discrete topol- optimization for natural frequency maximization problems considering
ogy optimization by Ref. [2]). The current work demonstrates topology acoustic-structure interaction, Finite Elem. Anal. Des. 106 (2015) 56–64.
optimization with compliance and displacement constraints. The appli- [32] Y.M. Xie, X. Yang, J. Shen, X. Yan, A. Ghaedizadeh, J. Rong, X. Huang, S. Zhou,
Designing orthotropic materials for negative or zero compressibility, Int. J. Solids
cation of the method to other types of constraints and to microstructural Struct. 51 (2014) 4038–4051.
optimization is a work for the future. [33] L. Xia, P. Breitkopf, Concurrent topology optimization design of material and
structure within nonlinear multiscale analysis framework, Comput. Methods Appl.
References Mech. Eng. 278 (2014) 524–542.
[34] L. Xia, P. Breitkopf, Multiscale structural topology optimization with an
approximate constitutive model for local material microstructure, Comput.
[1] X. Huang, Y.M. Xie, Convergent and mesh-independent solutions for the
Methods Appl. Mech. Eng. 286 (2015) 147–167.
bi-directional evolutionary structural optimization method, Finite Elem. Anal. Des.
[35] B. Xu, X. Huang, S. Zhou, Y. Xie, Concurrent topological design of composite
43 (2007) 1039–1049.
thermoelastic macrostructure and microstructure with multi-phase material for
[2] K. Svanberg, M. Werme, Topology Optimization by Sequential Integer Linear
maximum stiffness, Compos. Struct. 150 (2016) 84–102.
Programming, Springer Netherlands, Dordrecht, 2006, ISBN: 978-1-4020-4752-7.,
[36] W.M. Vicente, Z.H. Zuo, R. Pavanello, T.K.L. Calixto, R. Picelli, Y.M. Xie,
pp. 425–436, https://doi.org/10.1007/1-4020-4752-5_42.
Concurrent topology optimization for minimizing frequency responses of two-level
[3] G.I.N. Rozvany, A critical review of established methods of structural topology
hierarchical structures, Comput. Methods Appl. Mech. Eng. 301 (2016) 116–136.
optimization, Struct. Multidiscip. Optim. 37 (2009) 217–237.
[37] X. Wang, S. Xu, S. Zhou, W. Xu, M. Leary, P. Choong, M. Qian, M. Brandt, Y.M.
[4] J.D. Deaton, R.V. Grandhi, A survey of structural and multidisciplinary continuum
Xie, Topological design and additive manufacturing of porous metals for bone
topology optimization: post 2000, Struct. Multidiscip. Optim. 49 (2014) 1–38.
scaffolds and orthopaedic implants: a review, Biomaterials 83 (2016) 127–141.
[5] M.P. Bendsøe, N. Kikuchi, Generating optimal topologies in structural design using
[38] W.M. Vicente, R. Picelli, R. Pavanello, Y.M. Xie, Topology optimization of
a homogenization method, Comput. Methods Appl. Mech. Eng. 71 (1988)
frequency responses of fluid-structure interaction systems, Finite Elem. Anal. Des.
197–224.
98 (2015) 1–13.
[6] G.I.N. Rozvany, M. Zhou, T. Birker, Generalized shape optimization without
[39] R. Picelli, R. van Dijk, W. Vicente, R. Pavanello, M. Langelaar, F. van Keulen,
homogenization, Struct. Optim. 4 (1992) 250–254.
Topology optimization for submerged buoyant structures, Eng. Optim. 49 (1)
[7] M.P. Bendsøe, O. Sigmund, Topology Optimization - Theory, Methods and
(2017) 1–21.
Applications, Springer Verlag, Berlin Heidelberg, 2003.
[40] K. Svanberg, M. Werme, Sequential integer programming methods for stress
[8] O. Sigmund, Design of Material Structures Using Topology Optimization, Ph.D.
constrained topology optimization, Struct. Multidiscip. Optim. 34 (2007) 277–299.
thesis, Department of Solid Mechanics, Technical University of Denmark, 1994.
[9] O. Sigmund, Design of multiphysics actuators using topology optimization - Part I:
[41] J.W. Duda, M.J. Jakiela, Generation and classification of structural topologies with
one material structures, Comput. Methods Appl. Mech. Eng. 190 (2001)
genetic algorithm speciation, J. Mech. Des. 119 (1997) 127–131.
6577–6604.
[42] D. Guirguis, M.F. Aly, A derivative-free level-set method for topology optimization,
[10] T. Buhl, C.B. Pedersen, O. Sigmund, Stiffness design of geometrically non-linear
Finite Elem. Anal. Des. 120 (2016) 41–56.
structures using topology optimization, Struct. Multidiscip. Optim. 19 (2) (2000)
[43] R. Picelli, W.M. Vicente, R. Pavanello, Evolutionary topology optimization for
93–104.
structural compliance minimization considering design-dependent FSI loads, Finite
[11] T.E. Bruns, D.A. Tortorelli, Topology optimization of non-linear elastic structures
Elem. Anal. Des. 135 (Supplement C) (2017) 44–55, https://doi.org/10.1016/j.
and compliant mechanisms, Comput. Methods Appl. Mech. Eng. 190 (26–27)
finel.2017.07.005 ISSN 0168-874X.
(2001) 3443–3459.
[44] H. Guan, Y.J. Chen, Y.C. Loo, Y.M. Xie, G.P. Steven, Bridge topology optimisation
[12] G.H. Yoon, Y.Y. Kim, Element connectivity parameterization for topology
with stress, displacement and frequency constraints, Comput. Struct. 81 (3) (2003)
optimization of geometrically nonlinear structures, Int. J. Solids Struct. 42 (7)
131–145, https://doi.org/10.1016/S0045-7949(02)\ignorespaces00440&ndash;
(2005) 1983–2009.
00446 ISSN 0045-7949.
[13] C. Le, J. Norato, T. Bruns, C. Ha, D. Tortorelli, Stress-based topology optimization
[45] R.E. Griffith, R.A. Stewart, A nonlinear programming technique for the
for continua, Struct. Multidiscip. Optim. 41 (2010) 605–620.
optimization of continuous processing system, Manag. Sci. 7 (1961) 379–392.
[14] T. Borrvall, J. Petersson, Topology optimization of fluids in Stokes flow, Int. J.
[46] K. Svanberg, The method of moving asymptotes - a new method for structural
Numer. Methods Fluids 41 (2003) 77–107.
optimization, Int. J. Numer. Methods Eng. 24 (1987) 359–373.
[15] G.H. Yoon, Topology optimization for stationary fluid-structure interaction
[47] J. Nocedal, S.J. Wright, Numerical Optimization, second ed. edn, 2006.
problems using a new monolithic formulation, Int. J. Numer. Methods Eng. 82
[48] P. Dunning, H.A. Kim, Introducing the sequential linear programming level-set
(2010) 591–616.
method for topology optimization, Struct. Multidiscip. Optim. 51 (2015) 631–643.
[16] J. Du, N. Olhoff, Topological design of freely vibrating continuum structures for
[49] H.P. Williams, Integer Programming, Logic and Integer Programming, Springer US,
maximum values of simple and multiple eigenfrequencies and frequency gaps,
Boston, MA, 2009.
Struct. Multidiscip. Optim. 34 (2007) 91–110.
[50] MathWorks, Intlinprog Documentation, 2017. https://www.mathworks.com/help/
[17] J.K. Guest, J.H. Prvost, T. Belytschko, Achieving minimum length scale in topology
optim/ug/intlinprog.html[Accessed 30 June 2017].
optimization using nodal design variables, Int J. Numer. Methods Eng. 61 (2)
[51] R. Picelli, W.M. Vicente, R. Pavanello, Bi-directional evolutionary structural
(2004) 238–254.
optimization for design-dependent fluid pressure loading problems, Eng. Optim. 47
[18] A. Kawamoto, T. Matsumori, S. Yamasaki, T. Nomura, T. Kondoh, Heaviside
(10) (2015) 1324–1342.
projection based topology optimization by a PDE-filtered scalar function, Struct.
[52] D.J. Munk, G.A. Vio, G.P. Steven, A Bi-directional Evolutionary Structural
Multidiscip. Optim. 44 (1) (2011) 19–24.
Optimisation algorithm with an added connectivity constraint, Finite Elem. Anal.
[19] O. Sigmund, Morphology-based black and white filters for topology optimization,
Des. 131 (Supplement C) (2017) 25–42, https://doi.org/10.1016/j.finel.2017.03.
Struct. Multidiscip. Optim. 33 (4–5) (2007) 401–424.
005 ISSN 0168-874X.
[20] F. Wang, B.S. Lazarov, O. Sigmund, On projection methods, convergence and
[53] X. Huang, Y.M. Xie, Evolutionary topology optimization of continuum structures
robust formulations in topology optimization, Struct. Multidiscip. Optim. 43 (6)
with an additional displacement constraint, Struct. Multidiscip. Optim. 40 (2010)
(2011) 767–784.
409–416.
[21] G. Allaire, F. Jouve, A.M. Toader, Structural optimization using sensitivity analysis
and a level-set method, J. Comput. Phys. 194 (2004) 363–393.
[22] N. Jenkins, K. Maute, An immersed boundary approach for shape and topology
optimization of stationary fluid-structure interaction problems, Struct. Multidiscip.
Optim. 54 (2016) 1191–1208.

61