Two-Scale Topology Optimization with Microstructures

BO ZHU, MÉLINA SKOURAS, DESAI CHEN, WOJCIECH MATUSIK, MIT CSAIL

Fig. 1. Our two-scale topology optimization framework allows to optimize continuous material properties mapping to printable microstructures (left) to
fabricate high-resolution functional objects (middle) and minimum compliant structures (right).

In this paper we present a novel two-scale framework to optimize the struc- object of interest into small elements and optimizing the material dis-
ture and the material distribution of an object given its functional specifi- tribution over these elements in such a way that the functional goals
cations. Our approach utilizes multi-material microstructures as low-level are satisfied [Bendsøe and Sigmund 2004]. Traditionally, topology
building blocks of the object. We start by precomputing the material prop- optimization focused on designs made of homogeneous materials
erty gamut – the set of bulk material properties that can be achieved with and was concerned with macroscopic changes in the object geom-
all material microstructures of a given size. We represent the boundary
etry. With the advent of multi-material 3D printing techniques,
of this material property gamut using a level set field. Next, we propose
an efficient and general topology optimization algorithm that simultane-
it is now possible to play with materials at a much higher resolu-
ously computes an optimal object topology and spatially-varying material tion, allowing to obtain much finer designs and, thus, improved
properties constrained by the precomputed gamut. Finally, we map the functional performances. Unfortunately, standard techniques for
optimal spatially-varying material properties onto the microstructures with topology optimization do not scale well and they cannot be run
the corresponding properties in order to generate a high-resolution print- on objects with billions of voxels. This is because the number of
able structure. We demonstrate the efficacy of our framework by designing, variables to optimize increases linearly with the number of cells
optimizing, and fabricating objects in different material property spaces on in the object. Since many current 3D printers have a resolution of
the level of a trillion voxels, i.e several orders of magnitude higher than what 600DPI or more, a one billion voxel design occupies only a 1.67 inch
can be achieved with current systems. cube.
CCS Concepts: •Computing methodologies → Physical simulation; One direction to handle this issue is to work with microstruc-
Additional Key Words and Phrases: microstructures, metamaterials, 3D tures corresponding to blocks of voxels instead of individual voxels
printing, topology optimization directly. Some recent works followed this direction and proposed to
decouple macro structural design and micro material design [Coelho
ACM Reference format:
et al. 2008; Nakshatrala et al. 2013; Rodrigues et al. 2002]. However,
Bo Zhu, Mélina Skouras, Desai Chen, Wojciech Matusik. 2017. Two-Scale
Topology Optimization with Microstructures. ACM Trans. Graph. 34, 4, these approaches remain computationally expensive and, in most
Article 84 (July 2017), 17 pages. cases, limited to the well-known minimal compliance problem. The
DOI: 10.1145/nnnnnnn.nnnnnnn second direction to reduce the problem complexity is to temporar-
ily ignore the geometry of the microstructures and consider only
1 INTRODUCTION their macroscopic physical behaviour. However, this introduces
new difficulties as the space of material properties covered by all
Many engineering problems focus on the design of complex struc- printable microstructures is much wider than the properties of the
tures that needs to meet high level objectives such as the capability base materials. For example, microstructures made of alternating
to support localized stresses, optimal tradeoffs between compliance layers of soft and stiff isotropic materials exhibit an anisotropic
and mass, minimal deformation under thermal changes, etc. One behaviour as they are able to stretch more easily in one direction
very popular approach to design such structures is topology opti- that in the others. This implies that not only the ranges but also
mization. Topology optimization generally refers to discretizing the the number of physical parameters needed to describe the physical
© 2017 ACM. This is the author’s version of the work. It is posted here for your behaviour of these microstructures increases. Therefore, in order to
personal use. Not for redistribution. The definitive Version of Record was published in work with the material properties of microstructures, one needs to
ACM Transactions on Graphics, http://dx.doi.org/10.1145/nnnnnnn.nnnnnnn.
solve two challenging problems: (i) computing the gamut – i.e the

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:2 • B. Zhu et al.

set – of the material properties achievable by all microstructures, et al. 2013; Sigmund and Torquato 1996], multifunctional structure
(ii) efficiently optimizing the distribution of these high-dimensional design [Yan et al. 2015], coupled structure-appearance optimization
material properties inside the layout of the object. [Martı́nez et al. 2015]. Many algorithms have been proposed to
Most previous algorithms working in the material space focused numerically solve the optimization problem itself. We refer to the
on optimizing a single material property such as density or mate- survey by Sigmund and Maute [2013] for a complete review. In
rial stiffness, for which analytical formulas describing the property the very popular SIMP (Solid Isotropic Materials with Penalization)
bounds exist [Allaire and Kohn 1993]. On the contrary, optimiz- method, the presence of material in a given cell is controlled by
ing the structure and material distribution of an object in a high locally varying its density. A binary design is eventually achieved
dimensional material property space remains an open problem. In by penalizing intermediate values for these densities. In practice,
this work, we propose a new computational framework for topol- this method works well for two-material designs (e.g., a material
ogy optimization with microstructures that supports design spaces and a void), but generalizing this method to robustly handle higher
of multiple dimensions. We start by computing the gamut of the dimensional material spaces remains challenging. Instead of con-
material properties of the microstructures by alternating stochastic sidering only discrete structures, free material optimization [Haber
sampling and continuous optimization. This gives us a discrete rep- et al. 1994; Ringertz 1993] optimizes structures made of continuous
resentation of the set of achievable material properties, from which material distributions constrained by analytical bounds. Another
we can construct a continuous gamut representation using a level class of methods rely on homogenization. They replace the material
set field. We then reformulate the topology optimization problem in in each voxel of the object by a mixture of the base materials whose
the continuous space of material properties and propose an efficient material properties can be analytically derived. While optimal mi-
optimization scheme that finds the optimized distributions of mul- crostructures are known for certain classes of problems (laminated
tiple material properties simultaneously inside the gamut. Finally, composites in the case of the minimum compliance problem), this
in order to obtain fabricable designs, we map the optimal material is not the case in the general setting, for which using a specific
properties back to discrete microstructures from our database. subclass of microstructures can lead to suboptimal results. In a
Our general formulation can be applied to a large variety of sense, our work is a generalization of these approaches and aims to
problems. We demonstrate its efficacy by designing and optimiz- handle a wider range of materials for which theoretical bounds on
ing objects in different material spaces using isotropic, cubic and the material properties are not known a priori.
orthotropic materials. We apply our algorithm to various design Although they are largely used in engineering, standard meth-
problems dealing with diverse functional objectives such as min- ods for topology optimization suffer from a major drawback : the
imal compliance and target strain distribution. Furthermore, our parametrization of the problem at the voxel level makes them ex-
approach utilizes the high-resolution of current 3D printers by sup- tremely expensive and largely impedes their use on high resolutions
porting designs with trillions of voxels. We fabricate several of our models such as the ones generated by modern 3D printing hardware.
designs, thus, demonstrating the practicality of our approach. High-performance GPU implementations with careful memory han-
The main contributions of our work can be summarized as follows: dling can be used to push the limits of what can be done (a couple
• We present a fully automatic method for computing the of million variables in the implementation by Wu et al. [2016]), but
space of material properties achievable by microstructures such approaches rely on specificities of the minimum compliance
made of a given set of base materials. problem and are difficult to generalize. To counteract the effects of
• We propose a generic and efficient topology optimization the explosion of variables in finely discretized layouts, Rodrigues et
algorithm capable of handling objects with a trillion voxels. al. [2002] alternatively proposed an interesting formulation where
The key of our approach is a reformulation of the prob- microstructure designs and macroscopic layouts using the effective
lem to work directly on continuous variables representing properties of the underlying microstructures were hierarchically
the material properties of microstructures. This allows us coupled and treated simultaneously. This initial work has been
to cast topology optimization as a reasonably sized con- extended in multiple ways [Coelho et al. 2008; Nakshatrala et al.
strained optimization problem that can be efficiently solved 2013; Xia and Breitkopf 2014; Yan et al. 2014]. Alexandersen and
with state of the art solvers. Lazarov [2015] proposed a fast simulation algorithm for optimiz-
• We validate our method on a set of test cases and demon- ing complete macroscopic structures made of layered or periodic
strate its versatility by applying it to various design prob- microstructures. However, these methods still need to handle vari-
lems of practical interest. ables defined at the microstructure level and therefore they remain
relatively costly. The most related work is the method proposed
2 RELATED WORK by Xia et al.[2015b], which also relies on a database to speed up
Topology Optimization. Topology optimization is concerned with computations. However, their work specifically targets minimum
the search of the optimal distribution of one or more materials within compliance problems in the structural design which allows them to
a design domain in order to minimize some input objective function approximate the macroscale behaviour of the microstructures with
while satisfying given constraints [Bendsøe and Sigmund 2004]. Ini- a particular strain-based interpolating function.
tially applied to the structural design in engineering [Bendsøe 1989],
topology optimization has been extended since then to a variety of
problems including micromechanism design [Sigmund 1997], mass Fabrication-oriented Optimization. The last decade has witnessed
transfer [Challis and Guest 2009], metamaterial design [Cadman an increasing interest by the computer graphics community in the

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:3

Base materials Microstructures Discrete point cloud Continuous level set

Material
𝐸 Space 𝐸
Sampling
Microstructure Rigid
Database Generation Soft
𝜈 𝜈
𝜌 𝜌
Mapping Optimization
Push
Target
deformation

Multi-scale Topology
Optimization

Fabrication Map to discrete microstructures Topology optimization in Design goal

continuous material space

Fig. 2. Algorithm overview. We start by precomputing the gamut of material properties that can be achieved with all material microstructures of a given size.
Next, we run our topology optimization algorithm that optimizes the material properties of the object within this gamut such as to minimize some functional
objective. Finally, we map the optimal continuous material properties back to microstructures from our database to generate a printable object.

design of tools and algorithms targeting digital fabrication of phys- et al. [2013] introduced a fabrication-specific language and a pro-
ical artifacts. The range of media and applications addressed in gramming pipeline for a procedural material synthesis that lift this
previous literature is very diverse and we focus our discussion on limitation.
systems targeting 3D printing. The problem of optimizing the ma-
terial assignment for the individual voxels of an object in order to
control its large scale behaviour has been studied in different con-
texts. Starting with optical properties, Hašan et al. [2010] and Dong Microstructures and Metamaterials. Microstructures can be de-
et al. [2010] provided methods for printing objects with desired fined as small scale assemblies made of one or several base materi-
subsurface scattering properties. Stava et al. [2012] later considered als, whose macroscale properties can be very different from those
stability of 3D printed objects, Zhou et al. [2013] explored structural of the original materials. Many materials found in the nature are
strength while Chen et al. [2014] focused on rest shape optimization. microstructures when observed at a sufficiently small scale. Mi-
Closer to our present work, frameworks for the design of objects crostructures can also be engineered so as to define composites
with desired mechanical behaviours have been proposed by Bickel with improved capabilities or even metamaterials with exceptional
et al. [2010] and Skouras et al. [2013]. Like these works, our system properties. For example, Lakes [1987] presented in 1987 the first
allows to match given input deformations. However, while these man-made structure with negative Poisson’s ratio, i.e., a structure
previous systems assume a small set of available base materials and which transversally expands when it is axially stretched. The design
use these base materials in relatively coarse discretizations, our sys- of composites and metamaterials is an active research field inspiring
tem combines the base materials into microstructures to expand the myriads of works [Andreassen et al. 2014; Babaee et al. 2013; Cadman
design possibilities. Also relevant is the tool presented by Xu et al. et al. 2013; Sigmund 1997; Sigmund and Torquato 1996; Wang et al.
[2015] that allows to interactively design heterogeneous materials 2014]. While many of these works are concerned with the inverse
for elastic objects subject to prescribed displacements and forces, modeling of specific microstructures or families of microstructures,
and the material optimization approach proposed by Panetta et al. the study of the space of properties that these microstructures can
[2015]. However, these methods may output materials that are not achieve as a whole has been investigated much less. Theoretical
available in the real world for non-convex manifolds of material bounds have been derived without experimental validation [Lip-
properties. By contrast, we guarantee that all the microstructures ton 1994; Milton and Cherkaev 1995; Ting and Chen 2005]. Taking
used are always realizable in such cases, which is one of the key into account additive manufacturing constraints, Schumacher et al.
contributions of our work. Lastly, in an effort to unify individ- [2015] and Panetta et al. [2015] recently investigated the design
ual contributions when dealing with inverse modeling problems, of tileable and printable microstructures. In the first part of this
Chen et al. [2013] proposed an abstraction mechanism to facilitate paper, we further explore this line of research and focus on the gen-
the development of goal-based methods. The output of most of eration and characterization of databases of microstructures with
these systems is a per-voxel material composition, which cannot maximal material property coverage. In particular, we present a
be efficiently represented using simple surface meshes. Vidimče novel approach combining a probabilistic search and a continuous
optimization that allows us to fully automatically explore the gamut
of material properties that can be achieved by assembling given
base materials.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:4 • B. Zhu et al.

3 OVERVIEW the mechanical behaviour of the entire object and the microstruc-
Given as input a set of base materials, an object layout, and func- tures. In such a setting, the relation between the linear strain ϵ and
tional objectives, the goal of our system is to compute the material Cauchy stress σ at every material point is given by
distribution inside the object in order to optimize these functional σ = Cϵ , (1)
objectives. In our approach, we do not solve the problem directly,
instead we work with microstructures made of the base materi- where C, the so-called elasticity tensor, can be described by 21
als and the space of physical material properties spanned by them. parameters [Bonet and Wood 1997].
The complete pipeline of our system, illustrated in Figure 2, can be Working in such a high dimensional space is prohibitive and
decomposed into three stages. therefore we focus on materials having a certain numbers of sym-
metries, such as orthotropic materials for which the elasticity tensor
Material Space Precomputation. In the first stage, we estimate the is defined by 12 parameters (4 parameters in 2D), cubic materials
gamut of material properties covered by all possible microstructures defined by 3 parameters, and isotropic materials defined by 2 pa-
made by spatial arrangement of base materials. Since exhaustively rameters. For example, the tensor for a 3D cubic structure can be
computing the properties of all these microstructures is, in practice, written (using the Voigt notation) as
intractable, we progressively increase the material space by alter-
(1−ν )Ê ν Ê ν Ê
nating a stochastic search and a continuous optimization. The first © ν Ê (1−ν )Ê ν Ê ª
step introduces discrete changes in the materials of the microstruc- C = ­ ν Ê
­ ®
ν Ê (1−ν )Ê ® (2)
tures and allows emergence of new types of microstructures. The ­ µ ®
µ
second step allows to locally push the material space boundaries by « µ¬
refining the microstructure shapes. After completing this stage, we
with Ê = E/((1 − 2ν )(1 + ν )) and where E is the Young’s modulus of
obtain a discrete representation of the space of material properties
the material, ν its Poisson’s ratio, and µ its shear modulus.
and the mapping between these properties and the corresponding
Alternatively, one can also use Lamé’s parameters to define the
microstructures.
tensor C, which simplifies the derivation of the tensor with respect
Gamut-based Continuous Topology Optimization. In the second to the elastic parameters. In this case, the tensor has the form
stage, we construct a smooth continuous gamut representation of 2µ+λ λ λ
the material property space by using a level set field. We define our © λ 2µ+λ λ ª
C=­ λ λ 2µ+λ ® .
­ ®
topology optimization problem directly in this space. Our approach (3)
­ µ ®
minimizes the objective function over possible material parameters µ
« µ¬
while asking for strict satisfaction of the physics constraints – typ-
ically, the static equilibrium – as well as the strict satisfaction of The tensor for a 2D orthotropic structure can be written as
 E x νyx E x 
the physical parameter bounds. Taking advantage of our gamut
representation as a level set, we formulate this last constraint as C = c xy y
ν E E y (4)
µ/c
limiting the material properties to stay on the negative side of the
level set. This guarantees that the material properties that we use with c = 1/(1 − ν xy νyx ), and where E x and Ey are the Young’s
in the optimization are always physically realizable. moduli along the two principal axes, µ is the shear modulus, ν xy is
the Poisson’s ratio corresponding to a contraction in the direction
Fabrication-oriented Microstructure Mapping. In the last stage, we y when an extension is applied along the x axis, and νyx verifies
generate a printable result by replacing each cell in the object layout νyx E x = ν xy Ey .
with a microstructure whose material properties are the closest to Letting R n denote the space of n material properties, we then
the continuous material assignment resulting from the optimization. write each point p ∈ R n as p = [ρ, e], where ρ is the density of
We also take into account the boundary similarity across adjacent the material and e are the other material parameters. Our gamut,
cell interfaces to improve the connectivity between microstructures. i.e. the set M ⊂ R n of material properties corresponding to mi-
This results in a complex, high-resolution, multi-material model crostructures of a given resolution, is made of a finite number of
with optimized functional specifications. points. However, by increasing the resolution of the microstructures
this gamut gets denser and denser so that we assume that it can be
4 MECHANICS approximated by the union of continuous n-dimensional manifolds
In this section, we briefly introduce the background material for and can be represented using a distance field.
simulating deformable objects. We will use these concepts when
computing the material properties of the microstructures (Section 5) 4.2 Discretization
and in the topology optimization algorithm (Section 6). We refer to Following standard finite element methodology, we discretize the
the course by Sifakis and Barbic [2012] for a more comprehensive object in regular voxels and compute its deformed state when subject
exposition. to external forces fext using the well-known relation

4.1 Material Model Ku = fext , (5)

Most available materials for 3D printers are elastic materials. As- where K is the stiffness matrix of the system, and u are the displace-
suming small deformations, we use linear elasticity to compute both ments at the nodes of the voxels.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:5

Fig. 3. Level set gamuts for two dimensional cubic microstructures (top row) and three dimensional cubic microstructures (bottom row). The first column
shows the projection of the sample points in the space parametrized by the density ρ, the normalized Young’s modulus Ê and the Poisson’s ratio ν . The
second to the fourth columns show three slices of the four dimensional level sets corresponding to different values for the shear modulus G.

Note that we use the same approach to simulate both the mechan- generate new microstructures with desired properties. This second
ical behaviour of the microstructures and the object macroscopic option is effective in reaching locally optimal values in the material
behaviour. However, we work at two different scales. To simulate property space. However, the function that maps the material assign-
the microstructures, we assume that each of its voxels is made of an ment to material properties is nonlinear. In particular, very different
homogeneous base material, whereas for determining the large scale microstructures can correspond to the same point in the material
behaviour of the object, we assume that each of its cells corresponds property space. Additionally, since the ratio of materials in each
to a microstructure. The properties of the individual microstructures cell is bounded between zero and one, the continuous optimization
are determined from 6 harmonic displacements (or 2 displacements converges slowly or stops moving when material distributions in
in 2D) using numerical coarsening as described by Kharevych et al. many cells are at the lower or upper bound. Being able to jump out
[2009]. of a local optimum and discovering different variants is important
We solve the static equilibrium Equation 5 using a fast multigrid in order to provide new exploration regions. We leverage these two
solver based on the implementation by Dick et al. [2011]. approaches by combining them in a scheme that alternates between
a stochastic search and a continuous optimization. We provide the
5 MATERIAL SPACE EXPLORATION technical details in the rest of this section.
The first step in our pipeline is to determine the space of physical
properties that can be achieved when combining the base materials 5.1 Discrete Sampling of Microstructures
into microstructures of a predefined size. We aim at sampling the space of material assignments, i.e. mi-
Computing the mechanical properties of microstructures, when crostructures, in such a way that we maximize the number of sam-
arranged in periodic tilings, can be performed by probing the struc- ples corresponding to microstructures whose material properties
ture using a physical simulation. This approach, based on the ho- lie in the vicinity of the material gamut boundaries. We do not
mogenization theory, is a common practice and has been widely draw all samples at once but progressively enrich the database of
used in the past [Allaire 2012; Panetta et al. 2015; Schumacher et al. microstructures as we refine our estimation of the material gamut
2015]. However, while inferring the homogenized properties of in- boundaries. This sampling strategy is motivated by the observation
dividual microstructures is not particularly challenging, analyzing that a small change in the material assignment of a microstructure
the space covered by all combinations of base materials is much generally – but not always – translates to a small change of its
more difficult due to the combinatorial explosion in the number of material properties. By modifying microstructures located near the
possible material arrangements. As an example, 16 × 16 × 16 lattices current boundaries of the material property gamut, we are likely to
made of only two materials corresponds to 24096 microstructures: generate more structures in this area, some of which will lie outside
exhaustively probing all microstructures is clearly an impossible of the current gamut.
task. To address this issue, two possible avenues can be pursued:(i) Given a population of microstructures to evolve, we generate
we can try to sample the space of the microstructures, (ii) we can new samples from each microstructure by changing its material at
rely on the continuity between material parameters of the individual random voxel locations. To rationalize computational resources,
voxels and macroscopic properties of the microstructures in order to we want to avoid revisiting the same voxel twice. But we do not

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:6 • B. Zhu et al.

Discrete Sampling Continuous Optimization

Rigid Rigid
material material

Soft Soft
material material

𝛻𝜑(𝐩)
𝐩

Poisson’s ratio

Poisson’s ratio
Poisson’s ratio
Poisson’s ratio

Relative Young’s Relative Young’s Relative Young’s Relative Young’s

modulus modulus modulus modulus

Fig. 4. One cycle of computing the microstructure gamut. Given a set of samples, we compute a signed distance function approximating the material gamut
(left) and randomly perturb microstructures lying near the boundary to provide new seeds to the continuous algorithm (middle left). We then update the
distance field and use the gradient of the signed distance function at the the boundary to define new target material points (middle right). These target
material points are used in a continuous optimization that generates new samples (right).

want to privilege any particular order either. Ritchie et al. [2015] Algorithm 1 Procedure for generating new microstructures
recently presented a Stochastically-Ordered Sequential Monte Carlo procedure genMicrostructure(input: microstructure Mi , out-
(SOSMC) method that provides a suitable approach. In SOSMC, a put: microstructure Mo )
population of particles (here, our microstructures) corresponding to Mo ← Mi
instances of a procedural program (here, the sequential assignment while some voxels of Mo have not been visited do
of materials to the voxels of the microstructures) are evolved so as to while microstructure Mo is unchanged do
represent a desired distribution. During this process, the programs pick a random voxel v of Mo that has not been visited
are executed in a random order and particles are regularly scored assign a randomly chosen material to v
and reallocated in regions of high probability. In our particular if Mo is manifold and Mo  Mi then
settings, we use the scoring function accept the change
Φ(pi ) 1 end if
s(pi ) = × , (6) end while
D(pi ) D(pi )
end while
where Φ(pi ) is the signed distance of the material properties of end procedure
particle i to the gamut boundary (see Section 5.3) and D(pi ) is the
local sampling density at the location pi . We define the sample

 5.2 Continuous Optimization of Microstructures

density as
D(pi ) = ϕ k (pi ) , (7) The goal of the continuous optimization is to refine the geometry of
k the microstructures located at the boundary of the gamut in order
 4 to further expand the gamut along the normal directions (Figure 5).
| |p−pk | |22
where ϕ k (p) = 1 − are locally-supported kernel func-
h2 We start continuous sampling by selecting a subset of microstruc-
tions that vanish beyond their support radius h, set to a tenth of tures lying on the boundary of the gamut as starting points for
the size of the lattice used for the continuous representation of the the continuous optimization. The discrete structures are mapped
material gamut (see Section 5.3). to continuous values close to 0.5. We used 0.5 ± 0.3 in our exper-
The first term in Equation 6 favors microstructures located near iments. Doing so allows the topology optimization algorithms to
the gamut boundary. The normalization by D allows us to be less move freely in the first steps and discover new structures. We show
sensitive to the local microstructures density and to hit any location an example of reducing the Poisson’s ratio of an initial structure in
corresponding to the same level-set value with a more uniform Figure 5. In the plot, the initial Poisson’s ratio is close to 0.4 since
probability. The second product is used to additionally privilege the starting point is similar to a homogeneous block.
under-sampled areas. For each starting structure, we identify target material parameters
Particles are resampled using systematic resampling scheme [Douc using the gradient of the level set Φ at the initial discrete sample
2005] that is also used to initiate the population of particles. These point p (see Section 5.3) defined by q = p + ∇Φ(p). We translate
particles are then evolved according to Algorithm 1. Additional this target material parameters into an elasticity tensor C0 and
details regarding the implementation of our algorithm are provided density ρ 0 . Here ρ is the ratio of the two base materials in the
in the supplementary material. microstructure.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:7

Either interpolation allows us to threshold the final continuous

distribution and obtain a similar discrete sample. We tolerate small
deviations introduced by the thresholding since our goal is to obtain
a microstructure lying outside of the gamut rather than reaching a
particular target. In practice, we observed that the material proper-
ties of the final discrete structures often did not change significantly
after the thresholding step.

5.3 Continuous Representation of the Material Gamut

We represent the gamut of material properties using a signed dis-
tance field that is computed from the material points associated to
Fig. 5. The continuous sampling step uses topology optimization to expand the sampled microstructures. First, we normalize each coordinate
the gamuts by refining existing structures. A good starting point is necessary pi of p to constrain the scope of the level set to an n-dimensional
for the optimization to find better solutions. We use discrete microstructures
unit cube. Then we compute the level set values on the cell cen-
near the boundary to initialize topology optimization. At convergence, we
threshold the values to obtain a new discrete sample.
ters of an n-dimensional Cartesian grid that encloses this unit cube.
We draw inspiration from the methods for surface reconstruction
used in particle fluid rendering [Ando et al. 2013; Bhatacharya et al.
2011; Zhu and Bridson 2005] and extend it to n dimensions. In this
Note that our problem formulation does not restrict us to a par- case, a signed distance field is generated from a set of points by
ticular topology optimization algorithm or material distribution evaluating an implicit distance function Φ at each point p ∈ M.
parametrization. We have experimented with two objective func- We initialize the signed distance field using the implicit function
tions that worked equally well for our purposes. The first objective Φ(p) = ||p − p̄|| − r from [Zhu and Bridson 2005] where || · || is the
uses an energy based formulation [Xia and Breitkopf 2015a] to com- Euclidean distance between two points in M, and p̄ is the average
pute and optimize the elasticity tensor directly. At a high level, the position of the neighboring points of p within a range of 2r . Note
optimization problem is

that the signed distance is initially defined only near the boundary
arg min f (x) = (C(x) − C0 )2 + w ρ (ρ − ρ 0 )2 , ρ = x i , (8) of the gamut. In order to sample the distance on the entire domain,
x we propagate the 0-level set surface using the fast marching algo-
i
rithm and solve an explicit mean curvature flow problem defined as
where x is the ratio of materials in each cell, and w ρ controls the ∂Φ/∂t = ∆Φ [Osher and Fedkiw 2006] .
weighting between the displacement term and the density term. The Having a continuous representation of the gamut of materials
authors of the method developed parameter heuristics to optimize achievable by the microstructures, we can now reformulate the
for difficult cases such as negative Poisson’s ratio structures. We topology optimization problem directly in the material space.
naturally arrive at structures with negative Poisson’s ratio without
the parameter varying step in [Xia and Breitkopf 2015a] since our 6 TOPOLOGY OPTIMIZATION
discrete samples allow us to explore a wide variety of initial designs. A classic topology optimization problem consists of optimizing the
The second objective is formulated using harmonic displace- shape and structure of a given object defined by a prescribed domain
ments [Kharevych et al. 2009; Schumacher et al. 2015] G instead of in order to minimize some cost function. For example, the standard
the elasticity tensor directly. G is a 6 × 6 symmetric matrix where topology optimization minimizes the compliance of the object while
each row corresponds to a strain in vector form. We use the target satisfying the static equilibrium and the total weight constraint.
elasticity tensor C0 to compute the target harmonic displacements Since the topology of the object is unknown a priori, a method
matrix G0 and minimize the objective function: of choice is to define the shape of the object through its material
f (x) = (G(x) − G0 )2 + w ρ (ρ − ρ 0 )2 . (9) distribution and to locally work with material densities. To this end,
the design layout is voxelized and a density variable is assigned to
This objective matches soft structures more accurately since entries every cell of the discretized domain. By penalizing intermediate
of G are inversely proportional to material stiffness. values for these densities, a binary distribution corresponding to
Following the work by Andreassen et al. [2014], we use the the object’s final layout can be eventually obtained.
method of moving asymptotes (MMA) [Svanberg 1987]) to optimize In this work, we extend the traditional topology optimization
the objectives using an implementation provided in the NLOPT algorithm in multiple ways. First, we do not compute a binary ma-
package [Johnson 2014]. We run at most 50 iterations since it usually terial distribution at the cell level as commonly done. Instead, we
converges to a solution within 20-30 steps (Figure 5). MMA makes leverage our database of microstructures and ask for each cell to be
large jumps during the optimization while keeping track of the filled with one of the microstructures. By doing so we change the
current best solution, thus causing the oscillation of the objective topology of the object at a finer scale, i.e. within each cell. This is
value. To force continuous material ratios towards discrete values, done by working with the macro-scale material properties of the
we experimented with the SIMP model with the exponent set to 3 microstructures instead of their geometry directly. The second dif-
and the Hashin-Shtrikman bound for isotropic materials described ference is that our algorithm can be used with parametrizations of
by Bendsøe and Sigmund [1999]. the material property space that are more complex than the single

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:8 • B. Zhu et al.

density parameter per cell that is commonly used in the standard the adjoint method. The optimizer only needs to take the function
topology optimization algorithm. Indeed, in our generalized topol- values of S and Φ along with their gradients as input.
ogy optimization problem, each cell c i contains an n-dimensional
material parameter pi ∈ R n . We use p to denote the stacked vector 6.1 Elasticity Objectives
of material parameters in all cells. Given a signed distance func- We used two different types of objective functions for the elasticity
tion Φ(pi ) that defines the gamut, our new topology optimization term in our topology optimization algorithm. These two types of
problem is then written as objectives allowed us to design a wide range of objects.
min : S(p, u) Target Deformation. Our algorithm takes a vector of nodal tar-
p
s.t . : F (p, u) = 0 (10) get displacements and boundary conditions (external forces, fixed
points, etc.) as input. Then, it automatically optimizes the material
: Φ(pi ) ≤ 0, 1 ≤ i ≤ Nc distribution over the object domain to achieve the desired linear
where S is a real-valued objective function that depends on the deformation assuming a linear elastic behavior.
material parameters and the displacement vector u of the entire We define the deformation objective as
object at the elasticity equilibrium. The equality constraint F = 0 Cd (e, u) = (u − û)T D(u − û), (13)
requires u to satisfy the elasticity equilibrium and the inequality
constraint Φ ≤ 0 guarantees that the material properties of each where û is the vector of the target displacements, D is a diagonal
cell stay inside the precomputed gamut. matrix that determines the importance of each nodal displacement.
In our examples, the material parameter p consists of the density We use D to define the subset of nodes that we are interested in.
ρ and the elasticity parameters e. We split our objective function For example, we can set most entries of D to zero and focus on a
into an elasticity term C(e, u) that controls the deformation behavior portion of the domain (see Figure 13).
(see Section 6.1) and an optional density term V(ρ) that controls Minimum Compliance. We have experimented with the same
the overall mass of the object.The density term can be written as objective as the one used in the standard topology optimization

Nc algorithm where the compliance Cc is defined as
V(ρ) = ( ρ i Vi − M̂)2 , (11) Cc (e, u) = uT K(e)u. (14)
i=1
In the commonly used SIMP algorithm, the stiffness matrix Ki of
where Vi is the cell volume and M̂ is the target overall mass. When each cell i depends on the artificial density value ρ i through an
one of the base material is void, the use of the density term allows analytical formula such as Ki = ρ i3 K0 where K0 corresponds to the
to modify the topology of the object at a larger scale than the one stiffness matrix of the base material. In contrast, the stiffness matrix
of the microstructures, and thus to change the external shape of in our objective function is directly computed from the material
the object. In fact, even for multi-material designs involving base parameters of the material space and forced to correspond to a
materials with similar mass densities, we noted that we could use realizable material thanks to our gamut constraints.
the density term to encourage the presence of soft material in the Like in standard algorithms, we regularized the problem to avoid
structure. By removing the external cells entirely made of the soft checkerboard solutions by applying a smoothing kernel on the
material, we could then decrease the mass of the structure without material properties that favors smooth variations of the material
significantly changing its mechanical behaviour. Alternatively, the parameters over the object layout. Our optimizer supports multiple
density term can also be used to control other quantities related to objectives by linearly combining weighted objective functions.
the ratios of the different materials such as the cost of the object. For
specific problems, we can also add spatially-varying weight control 7 MAPPING MATERIAL PROPERTIES TO
terms to Equation 11. For example, we can control the target weight MICROSTRUCTURES
of each individual cell by adding a local term (ρ i − ρ̂ i )2Vi .
After running the topology optimization
Assuming static equilibrium, the elasticity constraint is written
algorithm, we generate a printable result
as
by replacing each cell in the object lattice
F (e, u) = K(e)u − fext = 0, (12) by a microstructure whose material prop-
where fext are the external loads applied to the object. erties match the optimal ones.
The gamut constraint for a point pi in the material property space Material properties of the microstruc-
is described by an n-dimensional level set function Φ(p). We have tures are computed using the homogeniza-
Φ(pi ) < 0 for a point inside the gamut, Φ(pi ) > 0 for a point outside tion theory which is more accurate with
the gamut, and Φ(pi ) = 0 for a point on the boundary of the gamut. a smooth transition between the geome-
The value of Φ represents the n-dimension Euclidean distance to tries of neighboring cells. While smooth-
the level set boundary. The gradient of Φ are evaluated by a finite ness in the material parameters can be eas-
difference operation on the signed distance field. ily enforced, it does not imply topological
We used a standard gradient-based numerical optimizer (Ipopt similarity of nearby microstructures. For
[Wächter and Biegler 2006] in our implementation) to solve Equa- example, any translation of a given microstructure in a periodic
tion 10. We enforced the elasticity equilibrium constraint using tiling will result in a microstructure geometrically different but with

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:9

Fig. 6. Gamuts computed with our discrete-continuous sampling scheme for 2D cubic structures (left) , 2D orthotropic structures (second from left), 3D cubic
structures (second from right) and 3D cubic structures with 0.35 as Poisson’s ratio (right). The plots show the results for the projection of the gamuts on
the plane defined by the macroscale Young’s modulus along the x axis (normalized by the Young’s modulus of the stiffest base material) and the Poisson’s
ratio corresponding to a contraction along the y-direction when the material is stretched along the x-direction. The blue dots correspond to the generated
samples,the orange dots correspond to the microstructures from Schumacher et al. [2015] and the yellow dots correspond to the microstructures from Panetta
et al. [2015].

exactly the same mechanical properties. Fortunately, our database

is very dense and multiple microstructures generally map to similar
points in the material property space, offering several variants. To
further increase the number of possibilities, we also incorporate an
additional exemplar for each microstructure by translating it by half
its size, which preserves its cubic or orthotropic symmetry without
changing its properties (see inset Figure). We then run a simple
but effective algorithm that picks the microstructure exemplars that
Nc
minimize the boundary material mismatch across adjacent cells. We
quantify this mismatch by I = i=1 Ii , where Ii is the contribution
associated to the cell i and corresponds to the number of boundary
voxels filled with materials that are different from the ones of the
voxels’ immediate neighbours across the interfaces.
Our algorithm proceeds as follows:
• For each cell, we define a list of possible candidates by
picking all the microstructures mapping to material points
lying in the vicinity of the optimal material point and we
randomly initialize the cell with one of the candidates.
• We compute the mismatch energy Ii associated to each cell Fig. 7. Gamut corresponding to 2D cubic microstructures made of two
i and sort the cells according to their energy. materials and void. The Young’s modulus of the microstructures is plotted
• We pick the first cell in the sorted list, i.e. the one with the using a logarithmic scale. We show above some examples of microstructures
highest energy and assign to it the microstructure candidate lying near the estimated boundary of the gamut, i.e. with extreme material
that decreases the energy the most. If we cannot decrease properties. The dark color corresponds to the softer material, while the light
the cell energy, we move to the next cell in the list. grey color is used for the stiffer material.
• We update the mismatch energies of all the impacted cells
and we update the priority list.
can be described with 4 parameters (3 elasticity parameters and
• We repeat the last two steps until the mismatch energy I
density) and 5 parameters (4 elasticity parameters and density) re-
cannot be decreased anymore.
spectively. In 3D we computed the gamut corresponding to cubic
structures with 4 parameters. In all cases, the size of the lattice
8 RESULTS for the microstructures was set to 16 in every dimension. We used
We first analyzed our microstructure sampling algorithm for 2D isotropic base materials whose Young’s modulus differed by a factor
and 3D microstructure gamuts. Then we used these precomputed of 1000 and having 0.48 as Poisson’s ratio. We initially computed
gamuts and we designed and optimized a wide variety of objects the databases for two-material microstructures, but also adapted
with our topology optimization algorithm. these databases for microstructures made of a void and a stiff ma-
terial. In the later case, we replaced the softer material by void,
8.1 Microstructure Sampling filter out all the microstructures with disconnected components
We evaluated our method on two- and three-dimensional microstruc- and, in the 3D case, filled the enclosed voids and recomputed the
tures made of one or two materials. For the 2D case we consid- homogenized properties. We provide a comparison between the
ered patterns with cubic and orthotropic mechanical behaviors that initial and postprocessed databases in the supplementary material.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:10 • B. Zhu et al.

×106 30
analytic analytic
15

Objective energy
25 cubic

Elastic energy
cubic
orthotropic orthotropic
20 SIMP-3
10
SIMP-1
15
5
10

0 5
0 20 40 60 0 20 40 60
# iterations # iterations
×108 20
3 32
32 18

Objective energy
64

Elastic energy
64
128 16 128
2
256 256
14

1 12
10
0 8
0 20 40 60 80 0 20 40 60 80
# iterations # iterations
7
×10 40
8 point 1 point 1

Objective energy
point 2

Elastic energy
point 2
6 30 point 3
point 3
Fig. 8. Resulting material property distributions when running our topol- point 4 point 4
4
ogy optimization algorithm in the orthotropic (top left), cubic (top right), 20
isotropic (middle left) and an analytically defined gamut E ≥ ρ 3 E 0 (middle 2
right), with the material property space dimensions ranging from five to 0
10
two. We compare our algorithm with the standard SIMP method with 0 20 40 60 0 20 40 60
power index p = 1 (bottom left) and p = 3 (bottom right). For these figures, # iterations # iterations

we computed the color of each cell by mapping every base vector of the
normalized parameter space to a color range and linearly interpolating the Fig. 9. Convergence tests. Variation of the objective energy (left) and the
colors associated to each of the parameters. In this example, the left side of elastic energy right of a beam being optimized for minimum compliance as
the cantilever is fixed while a discretized, linearly varying, distributed force the optimization progresses. The convergence plots correspond to the beam
is applied to the bottom side (see red arrows in the top left picture). of Figure 8 when optimized using different material spaces (top), different
resolutions for the beam lattice (middle) when using cubic microstructures,
and different initial material properties for the cubic microstructures (bot-
tom).
The resulting postprocessed gamuts are also depicted in Figure 6.
Our databases contain 274k, 388k and 88k 2D cubic, 2D orthotropic
and 3D cubic microstructures respectively and took from 15 hours
to 93 hours to compute, which correspond to 68, 19 and 5 sampling
cycles, respectively.
We first compared our results to the ones obtained by Schumacher Poisson’s ratio as Panetta’s, i.e 0.35, which affects the extremal val-
et al. [2015] and observed a significant increase in the coverage of ues of the obtained gamut. For the 643 microstructures, we used
the material space, even for 2D microstructures where we used a morphological operations in the discrete step and sensitivity filter-
coarser discretization. This comforts us with the idea that topology ing with a radius of 3 voxels in the continuous step to limit the
optimization only, while helpful to locally improve the microstruc- minimum feature size to 1/32 of the lattice size [Sigmund 2007].
ture geometries, is suboptimal when one aims to discover the entire Note that this comparison is provided for reference only since our
gamut of physical properties. The diversity of the microstructures microstructures are cubic while Panetta’s are isotropic (a subset of
that we obtained is also much richer, thus providing a larger set cubic). Furthermore, they target a different 3D printing technology
of options for the practical use of microstructures. Note that they with self-supporting constraints not imposed here. Finally, we also
employed some regularization to avoid thin features. For 163 mi- obtained a dense sampling in the interior of the space, as a result of
crostructures, we found regularization unnecessary since they are the randomness inherent to our approach. This reduces the need of
manifold and have a minimal feature size of 1/16 of the lattice size, running costly optimization in these areas and occurs even if we do
which is the same order of magnitude as the thinnest parts of Schu- not explicitly enforce any sampling there.
macher’s microstructures. For completeness, we also compared our We also experimented with three-material 2D cubic microstruc-
database of 3D microstructures to the one of Panetta et al. [2015] tures (two solid materials with Young’s moduli differing by a factor
at 163 and 643 grid resolutions (Figure 6, right). Our initial data- of 1000 and with 0.48 as Poisson’s ratio, plus a void material). The
base was computed with 0.48 as Poisson’s ratio and is shown in resulting database contains about 800k microstructures that can po-
the supplementary material. For this comparison, we then recom- tentially be printed. The corresponding gamut and some examples
puted the material properties of the microstructures using the same of the generated microstructures are shown in Figure 7.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:11

Fig. 11. Designing a soft ray. The wings of the ray are asked to flap up
and down when vertices on its spine contract and expand. Constrained
Fig. 10. Optimizing a beam to make it bend when it is squeezed. A beam vertices are colored in green. The deformations achieved with the optimized
with optimized material properties can take the desired ffSff shape (right) materials are displayed on the bottom row.
whereas a beam with homogeneous material properties can only axially
deform under small deformation (left). In this example, the vertices on the
vertical sides are fixed in their target positions, while the other vertices are forces (Figure 10). In order to avoid overfitting, we applied target
free to move. Target displacements are set on the nodes of the cells of the
displacements on the vertices of the boundary cells only. As de-
two horizontal boundary layers. The color plot for the bottom beams shows
picted in the figure, the use of microstructures largely improves
the deformation error of each cell as defined by Equation 13.
the global shape of the beam, which closely matches the target de-
formed shape. This becomes even more striking when compared to
8.2 Topology Optimization the behavior of a beam made of a homogeneous material.
We also validated our algorithm by designing a soft ray whose
We tested our topology optimization algorithm on a number of
wings can flap using a compliant mechanism (see Figure 11 and
simple test cases and large scale examples. Detailed analysis and
accompanying video). Boundary conditions are applied on two cir-
discussion of the results is provided below.
cular areas located along the spine of the ray. Each disk has one
Impact of the Material Space. We evaluated the impact of the degree of freedom for deformation, namely contracting or expand-
chosen material space on a 2D cantilever beam with optimized min- ing along the disk normals. This mechanism resembles the one of
imum compliance. We tested our topology optimization algorithm many hydraulics-driven soft robots. We define two target deforma-
on isotropic, cubic and orthotropic gamuts as well as with the virtual tion objectives corresponding to the flapping of the wings up and
materials used in the traditional SIMP approach and for which the down, when alternatively contracting and expanding the two disks’
stiffness of the material E = ρ p E 0 , p ≥ 1 is a function of the density boundaries. By running our multi-objective topology optimization
ρ of the cell and the stiffness of the base material E 0 . We also tested framework, we can compute an optimized material design that can
our algorithm on an analytical gamut with allowed stiffnesses E achieve both deformation modes when the corresponding boundary
defined by E ≤ ρ 3 E 0 . The results are shown in Figure 8. It can be conditions are exerted.
noted that, as the dimension of the material space increases, the final
Convergence and Robustness. We evaluated the convergence rate
energy of the system decreases. This is to be expected since higher
of our topology optimization both on the minimum compliance
dimensional space means larger gamuts. Thus, when using cubic
problem and with the target deformation objective. For the mini-
materials, the minimum compliance objective function reaches 3%
mum compliance problem, we used the same loading as the one of
lower energy than the standard SIMP method with power index 3.
Figure 8. The corresponding results are shown in Figure 9 where
This difference reaches 11% when we use orthotropic materials. It
we plot both the deformation energy of the structure as defined in
is worth noting that the lowest elastic energy is achieved when we
Equation 14 and the original objective of the problem 10 that also
use the traditional SIMP method with p = 1 (as shown in Figure 9).
includes the volume term defined by Equation 11. For all these ex-
However, this solution does not correspond to a realizable structure
amples, the algorithm converged after a couple of dozen iterations,
since some of the optimized materials do not correspond to any
irrespectively of the lattice resolution, i.e. the number of variables
microstructure.
and the number of non-linear constraints. This demonstrates the
Matching Quality. We evaluated for different examples the match- scalability of the our algorithm. We also tested the robustness of our
ing quality of the target deformation optimization. For the first test, algorithm by starting with different initial conditions. In this case,
we forced a beam to take an ‘S’ shape when undergoing tensile we initialized the material parameters of each cell with a random

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:12 • B. Zhu et al.

15
arm
face
objective energy (log)

0.08
10 wave
6
push back
push back
pull back
pull back 0.07 push side
5 push side pull side
pull side
deformation energy (L2 norm)

0.06

deformation energy (Linf)

4
0.05

3 0.04

0 0.03
2
0.02

1
-5 0.01
0 50 100 150 200
0 0
#iteration 0 50 100 150 200 250 300 350 0 50 100 150 200
#iteration
250 300 350
#iteration

Fig. 12. Target deformation examples. The pictures on the left (a) show
the deformed optimized result when the boundary conditions illustrated
Fig. 13. Designing functional grippers. The top row shows the initial shape
by the pictures (b) are applied. The orange meshes (c) correspond to the
of the gripper (left), and the target deformation for the tip (right). The green
simulated deformations when a homogeneous material is used. The blue-
dots correspond to the fixed vertices while the blue arrows correspond to
to-red meshes indicate the relative deformation error of the unoptimized
the target displacements. The two middle rows correspond to the optimized
(d) and optimized (e) structures. Convergence rates corresponding to the
results obtained for the specified boundary conditions. The inset pictures
optimization of these three examples are reported on the bottom figure.
color-code the initial and final deformation error for the different exam-
ples.The convergence plots in the bottom row depict the change in the sum
of the deformation errors corresponding to all the cells (left) and the value
material point projected onto the boundary of the gamut. Similar of the maximal cell error contribution (right) as the optimization progresses.
to other topology optimization schemes, we have no guarantee that
we reach the global minimum of the function, and indeed, our algo-
rithm sometimes converges to different solutions. However we note optimization algorithm successfully converges to the specified de-
that these different solutions have a similar final objective value and formation behavior in less than 100 iterations for all the examples.
are therefore equally good. We also tested the convergence rate when optimizing for functional
For the evaluation of the target deformation optimization, we mechanisms. To this end, we designed several grippers that can
ran our topology optimization algorithm on scenarii similar to the grasp objects by moving their tips when external forces are applied
ones from Panetta et al. [2015] where different extruded structures to their extremities. We experimented with four sets of boundary
are asked to deform into prescribed shapes when being compressed conditions, namely, pulling and pushing the back of the gripper
between two plates (see Figure 12). As shown in the figure, our horizontally, and compressing and stretching the extremities of the

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:13

Fig. 16. Comparison of a simulated gripper with homogenized material

properties (inset picture, left) to the one using full microstructures (main
picture, left). The figures on the right show the vector field of vertex displace-
ment of the two models. The blue-to-red colors represent the magnitudes
of the displacements.

we show that the ratio between the magnitudes of the average ver-
tex displacement differences is between 4% and 7%, and the elastic
Fig. 14. Comparison of simulated beams with homogenized material prop- energy difference is between 10% and 19%. Finally, we also com-
erties (inset pictures) to the ones using full microstructures (large pictures). pared the behaviors of one of the grippers (Figure 16). The original
optimized gripper is made of 3k elements while the high resolution
version is made of 4M voxels. Overall, the two models exhibit similar
global deformation behaviors, in particular in the tip area. Some dif-
gripper vertically. As shown in Figure 13, these different settings ferences can be observed on the left side of the gripper for which the
lead to different material structures. The deformation errors of all high-resolution model exhibits a lower effective material stiffness
the four designs converge to a low level after a couple of hundreds than its homogenized counterpart. With the same displacement
of iterations. boundary conditions applied, the high-resolution model deforms
about 25% more than the homogenized model.

Fig. 15. Comparison of simulated cubes with different material patterns

modeled by homogenized cells (inset pictures) and full resolution microstruc- Fig. 17. Simulation of a cube made of a periodic arrangement of a single mi-
tures (large pictures). crostructure at different resolutions. Inset pictures correspond to the model
with homogenized material properties, while main pictures correspond to
the full resolution simulations.
Accuracy. We evaluated the accuracy of our algorithm on several
optimized structures by comparing the deformation obtained when
The differences observed between the homogenized model be-
using the optimized homogenized material properties for each cell
haviour and the full resolution simulation can be explained by two
to the one obtained by a high resolution simulation in which every
major factors: (i) numeral stiffness when using larger elements
cell is replaced by its mapped microstructure. We first evaluated
which tends to make the homogenized mesh slightly stiffer in par-
the accuracy on a deformable bar, one side of which was rigidly
ticular when bending deformation arises, (ii) violation of the peri-
attached while the other was subject to different sets of external
odicity assumption when replacing each cell by a single microstruc-
conditions (see Figure 14). We used a 8 × 2 × 2 lattice to represent
ture. This issue can be reduced by replacing each cell by a tiling of
the bar with homogenized cells, which translates into a 128 × 32 × 32
microstructures. This was verified on a cube made of a periodic ar-
grid for the full resolution mesh. Similar stretching, bending and
rangement of a single microstructure (see Figure 17). And indeed, as
shearing behaviors were obtained for both sets of models. From
we increase the resolution of the simulation grid, the error between
a quantitative point of view, the differences amount to 5-10% in
the homogenized model and the full resolution version decreases
terms of average vertex displacement and 9%-33% in terms of elastic
and converges to similar values.
deformation energy (see Table 1). We further evaluated the effects
of material patterns by running a similar comparison on a cube Orthotropic materials. We tested the behavior of our algorithm
made of periodic layers of similar microstructures and with random in a 5-dimensional space by using the gamut of 2D orthotropic
assignments of microstructures (Figure 15). As reported in Table 1, microstructures depicted in Figure 6 (middle). To this end, we used

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:14 • B. Zhu et al.

Table 1. Error statistics (SI units). The size of one microstructure is set to Table 2. Statistics on the 3D-printed models. The last row uses the database
1 × 1 × 1. of 643 microstructures.

Example Mean dis- Mean dis- Elastic energy Example Grid Size # Voxels Time Time per
placement placement homogenized/full resolution per FEM Step [s]
difference Solve [s]
Beam 1 6.47×10−3 6.04×10−4 6.85×10−5 6.17×10−5 Beam 96×24×4 38M 0.7 5
Beam 2 6.47×10−3 6.04×10−4 1.63×10−5 1.08×10−5 Flexure 32×32×16 67M 1 12
Beam 3 5.07×10−3 4.97×10−4 2.38×10−4 2.07×10−4 Gripper 64×32×8 67M 1.7 10
Beam 4 8.78×10−3 4.45×10−4 3.33×10−4 2.30×10−4 Bunny 32×32×32 134M 0.6 4
Cube 1 3.62×10−3 2.86×10−4 3.64×10−3 3.20×10−3 Bridge 128×64×32 1074M 27 81
Cube 2 4.35×10−3 1.94×10−4 6.82×10−3 5.94×10−3 Bridge 2 320×160×80 1074G 1.3k -
Cube 3 5.42×10−3 4.22×10−4 7.81×10−3 6.32×10−3
Gripper 1.32×10−2 6.90×10−3 8.67×10−3 5.70×10−3
Cube 4 5.22×10−3 4.89×10−4 2.08×10−2 1.63×10−2 Beams with controlled deformation behaviour. We started by de-
Cube 5 5.21×10−3 2.17×10−4 1.89×10−2 1.63×10−2
Cube 6 5.21×10−3 1.32×10−4 1.82×10−2 1.63×10−2
signing a 3D hollowed beam with a desired deformed shape. The
beam was stretched by moving vertices on two opposite sides. Our
topology optimization algorithm was run using a target deforma-
tion objective. The resulting optimized material properties and the
a regular lattice whose vertices on the left side where fixed and 3D-printed structure are depicted in Figure 19.
we applied parallel forces on the vertices of the opposite side. The
goal in the test was to minimize the compliance of the structure.
As can be seen in Figure 18 and in the accompanying video, we
experimented with different force directions. Unsurprisingly, when
a single cell is considered, the microstructure that we obtain has a
structure that is aligned with the direction of the forces (see Figure
18, left). For a higher resolution lattice this is no longer true and
the resulting overall structure becomes less intuitive (see Figure 18,
right). Note that the resulting material distribution varies smoothly.
By considering various alternative for each material point, our tiling Fig. 19. An optimized hollow beam with target deformation. The left figure
algorithm is able to map the material properties to microstructures shows the target deformation and optimized material distribution. The
right figure shows the 3D-printed structure and the achieved deformation.
which are well connected.

8.3 3D-Printed Designs Multi-Objective Flexure Design. We tested our algorithm on a

Leveraging our two-scale approach, we used our topology optimiza- multi-target deformation setting by optimizing the structure of
tion algorithm to generate a wide variety of high resolution models a flexure mount with two different target shapes (see Figure 20).
that we 3D-printed. We used a Stratasys Objet Connex 500 and the Here, our goal is to design a flexure that resists vertical loads while
two base materials Vero Clear and Tango Black Plus and used the remaining compliant to horizontal loads. We assume that the object
database containing the three-dimensional cubic microstructures. mounted on the flexure is connected to the flexure using a cylindrical
The sizes and computation times of the resulting models are outlined connector that transmits the forces to the flexure via the connecting
in Table 2. area. In the first scenario, vertical forces are applied to the points
Since Ipopt performs a line search at each gradient step, one single of the cylindrical area and we ask the flexure to stay as close as
step may correspond to multiple simulations. We show the average possible to its rest configuration. In the second scenario, horizontal
time required for taking a step in the last column of Table 2. For these forces are applied to the points of the cylinder and we ask the flexure
large scale examples, Ipopt takes two hundred iterations in average shape to match the shape shown in the Figure 20.
to find a local minimum. Since our problem is formulated as a very
Gripper. We verified the functionality of our grippers by fabricat-
general constrained continuous optimization, it is independent of
ing two of them. For these results we ran the optimization on high
the optimization package that is used and its speed could potentially
resolution meshes of the version that grasps the object when the
be further improved by using alternative minimizers. We found
extremities of the gripper are pressed (see Figure 21). By changing
Ipopt to be a good choice for its capability to efficiently handle a
the parameter controlling the ratio of the soft material, different
large number of inequality constraints, which is not the case of
designs based on different mechanisms can be achieved. When
other popular minimizers used in topology optimization such as the
more soft material is used the gripper achieves its target deforma-
method of moving asymptotes (MMA).
tion thanks to out-of-plane bending, while for stiffer designs, the
Our algorithm is mainly directed towards engineering applica-
grasping motion is achieved via in-plane deformation.
tions and targets the design of objects undergoing small deforma-
tions. In the following examples, we sometimes intentionally ex- Minimal Compliance Examples. To demonstrate the scalability of
aggerated the target displacements (and scaled the external forces our algorithm, we ran our topology optimization algorithm on a
accordingly) for better visualization, which does not change the Stanford bunny (Figure 22) made of more than 100 million voxels
output of the algorithm with a linear material model. and subject to two load case scenarios. We also designed two bridges

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:15

Fig. 18. Optimizing the orthotropic material parameters of a single cell (left) and a 32 × 32 lattice of cells (right) subject to directional forces. The vertices on
the left side of the layout are fixed while forces are applied on the right vertices as depicted by the red arrows. Our simple but effective tiling algorithm allows
to nicely transition between microstructures of smoothly material properties (right, top).

Fig. 20. Optimizing a flexure mount. The flexure is connected to an object

thanks to a cylindrical connector. We leave space for this connector by
keeping a cylindrical area of the design layout empty of material. The
material distribution of the flexure is optimized for two sets of external
forces applied to the cylindrical area. Under vertical load, the flexure should
stay close to the rest shape while under horizontal load, the flexure should
deform according to the inset figure.

of increasing resolutions. The first bridge was optimized using a

lattice of half a million cells which corresponds to 1 billion voxels
(Figure 23). For the second bridge, we used the database of 643 Fig. 21. 3D-printed functional grippers. By setting different target ratios
microstructures and a layout made of 4 million cells, which amounts of the rigid material, different designs can be obtained. When more soft
to 1 trillion voxels. We initialized the topology optimization by material is used the grasping behaviour of the gripper is obtained via out-
of-plane bending (top), whereas more rigid material is used, the gripper
running the algorithm on a lower resolution grid with 1.4 million
deformation remains planar (bottom).
elements and used the resulting parameters as initial material pa-
rameter values for the higher resolution optimization.
gamut. Our second insight is to use this gamut as a constraint in a
9 CONCLUSION generalized topology optimization framework to assign spatially-
We have presented a computational framework for two-scale topol- varying material properties throughout the optimized object. Finally,
ogy optimization. Our approach can efficiently optimize high resolu- the volume with assigned material properties can be converted to
tion models that can be fabricated using multi-material 3D printing. a 3D model with corresponding spatially-varying microstructures.
Our first insight is to use a precomputation process to efficiently We demonstrated the effectiveness of our approach on multiple ex-
sample the space of microstructures and their corresponding ma- amples and showed improvements over traditional binary topology
terial properties in order to define a continuous material property optimization schemes.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
84:16 • B. Zhu et al.

REFERENCES
Joe Alexandersen and Boyan S. Lazarov. 2015. Topology optimisation of manufacturable
microstructural details without length scale separation using a spectral coarse basis
preconditioner. Computer Methods in Applied Mechanics & Engineering 290 (2015),
156fi??182.
Grégoire Allaire. 2012. Shape optimization by the homogenization method. Vol. 146.
Springer Science & Business Media.
G. Allaire and R.V. Kohn. 1993. Optimal bounds on the effective behavior of a mixture
of two well-ordered elastic materials. (1993).
Ryoichi Ando, Nils Thürey, and Chris Wojtan. 2013. Highly adaptive liquid simulations
on tetrahedral meshes. ACM Trans. Graph. (TOG) 32, 4 (2013).
Erik Andreassen, Boyan S. Lazarov, and Ole Sigmund. 2014. Design of manufacturable
Fig. 22. Stanford bunny optimized for two loading cases. Two sets of external 3D extremal elastic microstructure. Mechanics of Materials 69, 1 (2014).
forces are applied to the back and chest of the bunny as indicated by the Sahab Babaee, Jongmin Shim, James C. Weaver, Elizabeth R. Chen, Nikita Patel, and
Katia Bertoldi. 2013. 3D Soft Metamaterials with Negative Poisson’s Ratio. Advanced
arrows, and the color indicates the material distribution (left). Material Materials 25, 36 (2013).
parameters are mapped to microstructures to obtain an object that can be Martin P Bendsøe. 1989. Optimal shape design as a material distribution problem.
actually printed (right). Structural optimization 1, 4 (1989), 193–202.
Martin P Bendsøe and Ole Sigmund. 1999. Material interpolation schemes in topology
optimization. Archive of applied mechanics 69, 9-10 (1999).
Martin Philip Bendsøe and Ole Sigmund. 2004. Topology Optimization: Theory, Methods,
and Applications. Springer Science & Business Media.
Haimasree Bhatacharya, Yue Gao, and Adam Bargteil. 2011. A level-set method for skin-
ning animated particle data. In Proceedings of the 2011 ACM SIGGRAPH/Eurographics
Symposium on Computer Animation. ACM, 17–24.
Bernd Bickel, Moritz Bächer, Miguel A. Otaduy, Hyunho Richard Lee, Hanspeter Pfister,
Markus Gross, and Wojciech Matusik. 2010. Design and Fabrication of Materials
with Desired Deformation Behavior. ACM Trans. Graph. (Proc. SIGGRAPH) 29, 4
(2010).
J. Bonet and R. D. Wood. 1997. Nonlinear Continuum Mechanics for Finite Element
Analysis. Cambridge University Press.
Joseph E Cadman, Shiwei Zhou, Yuhang Chen, and Qing Li. 2013. On design of multi-
functional microstructural materials. Journal of Materials Science 48, 1 (2013), 51–66.
Vivien J Challis and James K Guest. 2009. Level set topology optimization of fluids in
Stokes flow. International journal for numerical methods in engineering 79, 10 (2009),
1284–1308.
Desai Chen, David I. W. Levin, Piotr Didyk, Pitchaya Sitthi-Amorn, and Wojciech
Matusik. 2013. Spec2Fab: A reducer-tuner model for translating specifications to
3D prints. ACM Trans. Graph. (Proc. SIGGRAPH) 32, 4 (2013).
Xiang Chen, Changxi Zheng, Weiwei Xu, and Kun Zhou. 2014. An Asymptotic Numer-
ical Method for Inverse Elastic Shape Design. ACM Trans. Graph. (Proc. SIGGRAPH)
Fig. 23. Optimizing a bridge. The initial layout corresponds to a 128x64x32 33, 4 (Aug. 2014).
regular grid. We apply uniform loads on the upper plane deck. We compute PG Coelho, PR Fernandes, JM Guedes, and HC Rodrigues. 2008. A hierarchical model
the material parameters and set cells with extremely low stiffness to void
for concurrent material and topology optimisation of three-dimensional structures.
Structural and Multidisciplinary Optimization 35, 2 (2008), 107–115.
(top left). We look up the microstructures and 3D print the bridge (top Christian Dick, Joachim Georgii, and Rüdiger Westermann. 2011. A real-time multigrid
right). We scale the problem to 1 trillion voxels by using a lattice of 4 million finite hexahedra method for elasticity simulation using CUDA. Simulation Modelling
elements where each element corresponds to a 643 microstructure (bottom Practice and Theory 19, 2 (2011).
left). We show a 20 × 20 × 1 patch on the bridge with filled microstructures
Yue Dong, Jiaping Wang, Fabio Pellacini, Xin Tong, and Baining Guo. 2010. Fabricating
Spatially-Varying Subsurface Scattering. ACM Trans. Graph. (Proc. SIGGRAPH) 29, 4
and a single microstructure with 643 voxels on the patch (bottom right). (2010).
Randal Douc. 2005. Comparison of resampling schemes for particle filtering. In 4th
International Symposium on Image and Signal Processing and Analysis (ISPA. 64–69.
RB Haber, P Pedersen, and JE Taylor. 1994. An analytical model to predict optimal
Limitations and Future Work. First, while our sampling method material properties in the context of optimal structural design. Urbana 51 (1994),
outperforms current approaches in terms of the material space cover- 61801.
age and the approach converges to stable gamuts, we do not provide Miloš Hašan, Martin Fuchs, Wojciech Matusik, Hanspeter Pfister, and Szymon
Rusinkiewicz. 2010. Physical Reproduction of Materials with Specified Subsur-
any theoretical guarantees that the gamut space cannot be further face Scattering. ACM Trans. Graph. (Proc. SIGGRAPH) 29, 4 (2010).
expanded. Second, we would like to investigate microstructures Steven G Johnson. 2014. The NLopt nonlinear-optimization package. (2014). http:
with additional properties, e.g., electrical or magnetic properties, //ab-initio.mit.edu/nlopt
Lily Kharevych, Patrick Mullen, Houman Owhadi, and Mathieu Desbrun. 2009. Nu-
their combined property gamuts, and the different applications merical Coarsening of Inhomogeneous Elastic Materials. ACM Trans. Graph. (Proc.
they enable. Finally, our framework builds upon linear elasticity SIGGRAPH) 28, 3 (2009).
Roderic Lakes. 1987. Foam structures with a negative Poisson’s ratio. Science 235, 4792
and optimizes the material distribution of objects subject to small (1987), 1038–1040.
deformations only. While this is enough for many engineering ap- Robert Lipton. 1994. Optimal Bounds on Effective Elastic Tensors for Orthotropic
plications, extending our algorithm to the nonlinear regime would Composites. Proceedings of the Royal Society A 444, 1921 (1994), 399–410.
Jonàs Martı́nez, Jérémie Dumas, Sylvain Lefebvre, and Li-Yi Wei. 2015. Structure and
be an interesting direction for future work. Appearance Optimization for Controllable Shape Design. ACM Trans. Graph. 34, 6,
Article 229 (Oct. 2015), 11 pages. DOI:http://dx.doi.org/10.1145/2816795.2818101
ACKNOWLEDGEMENTS Graeme W. Milton and Andrej V. Cherkaev. 1995. Which elasticity tensors are realizable?
Journal of Engineering Materials and Technology 117 (1995), 483.
This research is supported in part by the Defense Advanced Research PB Nakshatrala, DA Tortorelli, and KB Nakshatrala. 2013. Nonlinear structural design
Projects Agency (DARPA) and Space and Naval Warfare Systems using multiscale topology optimization. Computer Methods in Applied Mechanics
and Engineering 261 (2013).
Center Pacific (SSC Pacific) under Contract No 66001-15-C-4030.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.
 Two-Scale Topology Optimization with Microstructures • 84:17

Stanley Osher and Ronald Fedkiw. 2006. Level set methods and dynamic implicit surfaces.
Vol. 153. Springer Science & Business Media.
Julian Panetta, Qingnan Zhou, Luigi Malomo, Nico Pietroni, Paolo Cignoni, and Denis
Zorin. 2015. Elastic Textures for Additive Fabrication. ACM Trans. Graph. 34, 4,
Article 135 (July 2015), 12 pages. DOI:http://dx.doi.org/10.1145/2766937
U. T. Ringertz. 1993. On finding the optimal distribution of material properties. Structural
& Multidisciplinary Optimization 5, 4 (1993), 265–267.
Daniel Ritchie, Ben Mildenhall, Noah D. Goodman, and Pat Hanrahan. 2015. Controlling
Procedural Modeling Programs with Stochastically-ordered Sequential Monte Carlo.
ACM Trans. Graph. (Proc. SIGGRAPH) 34, 4 (2015).
H Rodrigues, Jose M Guedes, and MP Bendsøe. 2002. Hierarchical optimization of
material and structure. Structural and Multidisciplinary Optimization 24, 1 (2002),
1–10.
Christian Schumacher, Bernd Bickel, Jan Rys, Steve Marschner, Chiara Daraio, and
Markus Gross. 2015. Microstructures to Control Elasticity in 3D Printing. ACM
Trans. Graph. 34, 4 (2015).
Eftychios Sifakis and Jernej Barbic. 2012. FEM Simulation of 3D Deformable Solids:
A Practitioner’s Guide to Theory, Discretization and Model Reduction. In ACM
SIGGRAPH 2012 Courses.
Ole Sigmund. 1997. On the Design of Compliant Mechanisms Using Topology Opti-
mization. Mechanics of Structures and Machines 25, 4 (1997).
Ole Sigmund. 2007. Morphology-based black and white filters for topology optimization.
Structural and Multidisciplinary Optimization 33 (2007), 401–424.
Ole Sigmund and Kurt Maute. 2013. Topology optimization approaches. Structural and
Multidisciplinary Optimization 48, 6 (2013), 1031–1055.
Ole Sigmund and Salvatore Torquato. 1996. Composites with extremal thermal expan-
sion coefficients. Applied Physics Letters 69, 21 (1996), 3203–3205.
Mélina Skouras, Bernhard Thomaszewski, Stelian Coros, Bernd Bickel, and Markus
Gross. 2013. Computational Design of Actuated Deformable Characters. ACM Trans.
Graph. (Proc. SIGGRAPH) 32, 4 (2013).
Ondrej Stava, Juraj Vanek, Bedrich Benes, Nathan Carr, and Radomı́r Měch. 2012. Stress
relief: improving structural strength of 3D printable objects. ACM Trans. Graph.
(Proc. SIGGRAPH) 31, 4 (2012).
Krister Svanberg. 1987. The method of moving asymptotes —a new method for structural
optimization. International journal for numerical methods in engineering 24, 2 (1987),
359–373.
T. C. T. Ting and Tungyang Chen. 2005. Poisson’s ratio for anisotropic elastic materials
can have no bounds. Quarterly Journal of Mechanics & Applied Mathematics 58, 1
(2005), 73–82.
Kiril Vidimče, Szu-Po Wang, Jonathan Ragan-Kelley, and Wojciech Matusik. 2013.
OpenFab: A Programmable Pipeline for Multi-material Fabrication. ACM Trans.
Graph. (Proc. SIGGRAPH) 32, 4 (2013).
A. Wächter and L. T. Biegler. 2006. On the Implementation of a Primal-Dual Inte-
rior Point Filter Line Search Algorithm for Large-Scale Nonlinear Programming.
Mathematical Programming 106, 1 (2006).
F. Wang, O. Sigmund, and J.S. Jensen. 2014. Design of materials with prescribed
nonlinear properties. Journal of the Mechanics and Physics of Solids 69 (2014).
Jun Wu, Christian Dick, and Rüdiger Westermann. 2016. A System for High-Resolution
Topology Optimization. IEEE Trans. on Visualization and Computer Graphics 22, 3
(2016).
Liang Xia and Piotr Breitkopf. 2014. Concurrent topology optimization design of
material and structure within nonlinear multiscale analysis framework. Computer
Methods in Applied Mechanics and Engineering 278 (2014).
Liang Xia and Piotr Breitkopf. 2015a. Design of materials using topology optimization
and energy-based homogenization approach in Matlab. Structural and Multidisci-
plinary Optimization (2015), 1–13.
Liang Xia and Piotr Breitkopf. 2015b. Multiscale structural topology optimization with
an approximate constitutive model for local material microstructure. Computer
Methods in Applied Mechanics and Engineering 286 (2015), 147–167.
Hongyi Xu, Yijing Li, Yong Chen, and Jernej Barbivč. 2015. Interactive material design
using model reduction. ACM Trans. Graph. 34, 2 (2015).
Xiaolei Yan, Xiaodong Huang, Guangyong Sun, and Yi Min Xie. 2015. Two-scale optimal
design of structures with thermal insulation materials. Composite Structures 120
(2015), 358–365.
X Yan, X Huang, Y Zha, and YM Xie. 2014. Concurrent topology optimization of
structures and their composite microstructures. Computers & Structures 133 (2014),
103–110.
Qingnan Zhou, Julian Panetta, and Denis Zorin. 2013. Worst-case Structural Analysis.
ACM Trans. Graph. (Proc. SIGGRAPH) 32, 4 (2013).
Yongning Zhu and Robert Bridson. 2005. Animating sand as a fluid. ACM Trans. Graph.
24, 3 (2005), 965–972.

ACM Transactions on Graphics, Vol. 34, No. 4, Article 84. Publication date: July 2017.