asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt.

Dwarka Prasad Mishra Indian Inst of Information Tec

Junjian Fu
State Key Lab of Digital
Manufacturing Equipment and Technology,
Huazhong University of Science and Technology,
Wuhan 430074, China
e-mail: [email protected]

Liang Xia Topology Optimization of Periodic

State Key Lab of Digital
Manufacturing Equipment and Technology,
Huazhong University of Science and Technology,
Structures With Substructuring
Wuhan 430074, China
Topology optimization of macroperiodic structures is traditionally realized by imposing
e-mail: [email protected]
periodic constraints on the global structure, which needs to solve a fully linear system.
Therefore, it usually requires a huge computational cost and massive storage requirements
Liang Gao1 with the mesh refinement. This paper presents an efficient topology optimization method for
Professor
periodic structures with substructuring such that a condensed linear system is to be solved.
Mem. ASME
The macrostructure is identically partitioned into a number of scale-related substructures
State Key Lab of Digital
represented by the zero contour of a level set function (LSF). Only a representative sub-
Manufacturing Equipment and Technology,
structure is optimized for the global periodic structures. To accelerate the finite element
Huazhong University of Science and Technology,
analysis (FEA) procedure of the periodic structures, static condensation is adopted for
Wuhan 430074, China
repeated common substructures. The macrostructure with reduced number of degree of
e-mail: [email protected]
freedoms (DOFs) is obtained by assembling all the condensed substructures together.
Solving a fully linear system is divided into solving a condensed linear system and parallel
Mi Xiao recovery of substructural displacement fields. The design efficiency is therefore significantly
State Key Lab of Digital improved. With this proposed method, people can design scale-related periodic structures
Manufacturing Equipment and Technology, with a sufficiently large number of unit cells. The structural performance at a specified scale
Huazhong University of Science and Technology, can also be calculated without any approximations. What’s more, perfect connectivity
Wuhan 430074, China between different optimized unit cells is guaranteed. Topology optimization of periodic,
e-mail: [email protected] layerwise periodic, and graded layerwise periodic structures are investigated to verify
the efficiency and effectiveness of the presented method. [DOI: 10.1115/1.4042616]
Hao Li
State Key Lab of Digital
Manufacturing Equipment and Technology,
Huazhong University of Science and Technology,
Wuhan 430074, China
e-mail: [email protected]

1 Introduction unit cell. Periodic microstructures are obtained by uniformly distrib-

uting the optimized unit cell on the global design domain. Based on
A periodic structure consists fundamentally of a number of repet-
this design paradigm, one can design the periodic microstructures
itive unit cells that are assembled to form a continuous structure. As
with extreme elastic stiffness [12], extreme thermal expansion coef-
an important branch of modern structures, periodic structures are
ficient [8], negative Poisson’s ratio [13–15], and maximum fluidic
widely used in structural systems of metamaterials, additive manu-
permeability [16], etc. It should be noted that the design paradigm
facturing, lattice structures, architectures, bridges, and wheels of
is a microscale design method for periodic structures. Because the
cars. Periodic structures are mainly designed for manufacturing,
homogenization method is assumed to be separated of scales, the
aesthetics, and functionality. Periodic structures can be manufac-
optimized microstructures are actually infinite small. To further con-
tured by repetitive unit cells and then assembled together. And in
sider the macrostructural influence, the multiscale topology optimi-
some designs of architectures and bridges, periodic structures are
zation based on the homogenization method is developed [17–28]. In
also applied to show aesthetic appeal visually [1]. Besides, periodic
the multiscale design method, the topology of the periodic micro-
structures can be extended to functional graded structures to achieve
structure is optimized for the macrostructural performance. The
multifunctions [2,3]. Currently, there are two methods available for
homogenization method is applied to evaluate the macroeffective
the topological optimization of periodic structures: method based on
material properties of microstructures. The macrostructural displace-
the homogenization [4,5] and method with macroperiodicity con-
ments, in turn, affect the optimization of microstructures. However,
straints [6,7].
periodic structures designed by the homogenization method do not
Topology optimization of periodic structures based on the homog-
consider the scale effects. The first problem may occur is that the
enization traces back to the inverse design of material microstruc-
connectivity between the multiple type of microstructures cannot
tures [4,8,9]. In the homogenization method [10,11], the stress
be guaranteed. As a result, methods achieving microstructural con-
and strain fields are assumed to be periodic at the macroscopic
nectivity in multiscale topology optimization are presented [29–
scale. Periodic displacement boundary conditions are imposed on
31]. Second, the design objective such as compliance would be over-
the boundary of a unit cell. The finite element analysis (FEA) and
estimated once a specific length scale is prescribed because the
topology optimization are conducted for the design of microscale
periodic structures are theoretically infinite small [32]. To this end,
a scale-related higher-order homogenization method [33] or the
1
multiscale finite element method (MsFEM) without scale separa-
Corresponding author.
Contributed by the Design Automation Committee of ASME for publication in the
tions [34] can be regarded as the viable alternatives of classical
JOURNAL OF MECHANICAL DESIGN. Manuscript received May 3, 2018; final manuscript homogenization method in the design of scale-related periodic
received January 2, 2019; published online March 13, 2019. Assoc. Editor: Xu Guo. microstructures.

Journal of Mechanical Design Copyright © 2019 by ASME JULY 2019, Vol. 141 / 071403-1
 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
Another method to design periodic structures is the macroscale 2 Topology Optimization Formulation for Periodic
topology optimization with periodicity constraints [6,7]. In this Structures
method, the macrostructure is equally partitioned into a number
of unit cells with specific length scale size. The elemental densities 2.1 Structural Representation by Level Set Method. In the
within each cell are regarded as optimization variables. The period- framework of the LSM, a structural interface is represented implic-
icity is defined that the elemental densities on the same position of itly by the zero contour of higher dimensional LSF Φ(x), which is
the unit cells have the same value and so do the sensitivities. This described by the following definition:
method is later compared with the homogenization-based topology ⎧
⎨ Φ(x) > 0, x ∈ Ω(Solid)
optimization. Comparing with the homogenization-based topology
Φ(x) = 0, x ∈ ∂Ω(Interface) (1)
optimization, it is shown that with the refinement of cells, the opti- ⎩
mized periodic structure is gradually convergent to a similar struc- Φ(x) < 0, x ∈ D/(Ω ∪ ∂Ω)(Void)
ture designed by the homogenization method [32,35]. Although an where Ω describes the structure, ∂Ω is the interface of the structure
efficient bidirectional evolutionary structural optimization (BESO) represented implicitly as the zero level set of LSF Φ(x), and D is the
method [36,37] is utilized for the topology optimization, this design domain that contains the structure completely. Figure 1
design method with periodic constraints solves a fully linear shows an implicit representation by a LSF.
system. With the refinement of finite element discretization,
solving the fully linear system with direct solver would experience
a huge computational burden and massive storage requirements. 2.2 Minimization of Compliance of Periodic Structures.
Consequently, an efficient design method for periodic structures For simplicity, the topology optimization problem is limited to
considering scale effects is still in demand. the minimization of compliance with isotropic materials. As illus-
This paper presents an efficient topology optimization method for trated in Fig. 2, the 2D macrostructure is partitioned into m × n sub-
periodic structures with substructuring. The substructure method, structures, where m and n are the number of substructures in i and j
also known as static condensation [38,39] or domain decomposition directions, respectively. Φi,j is the LSF that represents the geometric
[40], is a common method used in the parallel solution of large FEA. boundary of a substructure. The periodicity is defined such that all
The basic idea of the substructure method is to partition the physical the substructures within the macrostructure have the identical LSF.
domain into a number of subdomains such that processors can deal Thus, only one LSF is involved for the topology optimization. By
with the computation and storage of the subdomains in parallel. considering all the identical LSFs, the compliance minimization
This idea is consistent with the definition of periodic structures, in problem of periodic structures is defined as follows
which the macrostructure is equally partitioned into a number of 
scale-related substructures. Besides, it would be of great advantage min: J(Φ) = εT (u)Dε(u)dΩ
if the finite element model contains many repetitions of the same Ω
n 
geometry. Then, the same reduced substructure stiffness matrix m 
= εT (ui, j )Dε(ui, j )dΩi, j
applies to all substructures. It happens that in the topology optimiza- i j Ωi, j (2)
tion of periodic structures, the periodicity is defined that all the sub- ⎧
⎨ a(u, v) = l(v) ∀v ∈ U
structures have the same geometric representation. 
Based on the consistency and the advantage, the substructure s.t. G(Φ ) =
⎩ i, j dΩi, j − f V
≤0
method is utilized to condense a substructure unit cell to a super Ωi, j
element with only the boundary nodes. A reduced global equilib-
rium equation of the macrostructure is obtained by assembling all where J is the structural compliance, u is the displacement field, ε is
the condensed substructures. The number of degree of freedoms the strain field, and D is the elasticity matrix of the base material.
(DOFs) is decreased dramatically to a lower order of magnitude, Subscript “i, j” indicates the quantities of a substructure, such as
contributing to a high computational efficiency. Besides, the
scale-related nature of the substructure method would guarantee
the connectivity between different optimized substructures.
In this research, the structural boundary of the substructure is rep-
resented by the zero contour of a level set function (LSF). Level set
method (LSM) is initially proposed by Osher and Sethian to track a
moving boundary [41]. Unlike the density method [42] and the
BESO method, level set method builds the structural boundary
into one higher dimensional surface as the zero contour. The level
set method has emerged as an alternative for the structural optimiza-
tion based on the implicit representation [43–45]. Many variations of
the level set-based topology optimization method have been devel- Fig. 1 Level set representation: (a) level set function and (b)
oped, including the semi-implicit method [46], the semi-Lagrange contour of zero level set
approach [47], the sequential linear programming level set method
[48], the level set method with a fictitious energy [49,50], and para-
metric level set method (PLSM) [51,52]. This paper simply applies
the PLSM to solve the optimization problem. It should be noted
that the substructure method for condensation is independent of
the topology optimization method. Density method, BESO
method, and moving morphable component (MMC)-based method
[53–55] can also be used in the design framework of this paper.
The remaining of this paper is organized as follows: Sec. 2
deals with the level set representation and the topology optimization
formulation of periodic structures. Section 3 explains the static con-
densation of both the substructure and the macrostructure in detail.
Section 4 derives the sensitivity formulation based on the PLSM.
Section 5 gives the optimization procedure. Section 6 investigates
three examples to demonstrate the efficiency and the effectiveness
of the proposed method. Section 7 concludes the paper. Fig. 2 Design domain with m × n periodic substructures

071403-2 / Vol. 141, JULY 2019 Transactions of the ASME

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
the displacement vector ui, j . Virtual displacement field v belong-
ing to the space U is spanned by the kinematically admissible set of
displacements. G(Φi, j ) is the volume constraint; f is the volume

denotes the volume of the substruc-
fraction of a substructure; and V
tural design domain. The displacement fields of all the substructures
are obtained by solving the equilibrium equation in Eq. (2). To
improve the computational efficiency of FEA, this equilibrium
equation is solved in a condensed manner, which is explained in
detail in Sec. 3.

3 Static Condensation of Substructures and the

Macrostructure
3.1 Static Condensation of a Substructure. In the finite
element method (FEM), a finite number of elements are used to dis- Fig. 4 Condensed m × n periodic substructures
cretize a solid model. Where there are elements, there are nodes.
These nodes can be partitioned into two parts: the boundary part
and the internal part. Based on this classification, the equilibrium This brings about great computational efficiency with the mesh
equation of a finite element model can be partitioned into a bound- refinement of both the macrostructure and the substructure.
ary set (boundary DOF) and an internal set (interior DOF) [38]. The
“b” and “i” subscript symbols over the matrix indicate the boundary 3.2 Assembly of the Macroperiodic Structures. Since the
partition and the internal partition, respectively. finite element model of periodic structures contains many repetitions
  of the same geometry, the same static condensation applies to all sub-
Kbb Kbi Ub Ub
= (3) structures. Each of the substructures can be condensed to a substruc-
Kib Kii Ui Fi ture or a super element. Only the external nodes connected with
where K, U, and F are the stiffness matrix, displacement vector, and neighboring elements are retained. The condensed substructure
force vector of the substructure, respectively. Collecting and rear- may lose its internal shape complexity. But the reduced stiffness
ranging the second row of Eq. (3), matrix of the substructure is constructed by implicitly considering
the level set representation of the uncondensed substructure.
Ui = K−1
ii (Fi − Kib Ub ) (4) Repeated assembly of the same reduced stiffness matrix using Eq.
(11) provides the condensed global stiffness matrix. An illustration
Substituting Eq. (4) into the first row of Eq. (3) of m × n condensed periodic structures is displayed in Fig. 4. Com-
paring with the original illustrative periodic structures in Fig. 2,
(Kbb − Kbi K−1 −1
ii Kib )Ub = Fb − Kbi Kii Fi (5) the condensed one has a lower number of nodes or DOFs, leading
The internal force partition Fi is generally a zero vector. The con- to a high computational efficiency improvement in the macro-FEA.
densed system of Eq. (5) is rearranged as
K* U* = F* (10)
Ksub Usub = Fsub (6)
with
with 
*
K = Ksub dV (11)
Ksub = Kbb − Kbi K−1
ii Kib (7) V

where K*, U*, and F* are the condensed global stiffness matrix, dis-
Usub = Ub (8) placement vector, and force vector, respectively.
After solving the condensed global equilibrium equation (10), the
Fsub = Fb (9) displacement vector ui,j of a substructure is recovered by substituting
the corresponding external displacements from U* back into Eq. (4).
where Ksub, Usub, and Fsub are the condensed stiffness matrix, dis- Because the condensed global displacement vector and the substruc-
placement vector, and force vector of the substructure, respectively. tural displacement vector are calculated in the same scale, separation
After solving equilibrium equation (6), internal nodal displacements of scales as assumed in the homogenization method [10] is not
Ui for the substructure can be recovered by Eq. (4). A typical illus- existed in the substructure method. As a result, multiple-type sub-
tration of the condensation of a substructure is shown in Fig. 3. No structures would be optimized with interconnections such that exter-
structural information is lost since all the active and inactive internal nal forces can pass throughout the macroperiodic structures.
nodes of the original structure contribute to the substructure by
reduced stiffness matrix Ksub. The dimensionality of Ksub is much
smaller when compared with the original full stiffness matrix K. 4 PLSM-Based Optimization and Sensitivity Analysis
4.1 Parametric Level Set Method. In the conventional LSM,
the evolution of the structural interface is realized by adding a
pseudo time t to the implicit LSF. Taking the derivative of
Φ(x, t) on both sides with respect to t and giving the outward
unit normal n = −∇Φ/|∇Φ|, we obtain the level set equation as a
Hamilton–Jacobi equation.
∂Φ(x, t)
− Vn |∇Φ| = 0 (12)
∂t
Equation (12) defines the motion of the interface on the Eulerian
grid under the normal velocity Vn. In the conventional LSM, explicit
Fig. 3 Illustration of the condensation of a substructure schemes such as the upwind differencing and the weighted

Journal of Mechanical Design JULY 2019, Vol. 141 / 071403-3

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
essentially nonoscillatory are generally used to implement the dis- The derivative of the Heaviside function is known as Dirac func-
crete level set evolution. In order to combine the evolution with tion, which is derived as
the mature gradient-based optimizer, this paper focuses on a ⎧  
further developed PLSM to realize the optimization procedure. ⎨ 3(1 − ξ) Φ2
1− 2 |Φ| ≤ Δ
The PLSM not only inherits the clear boundary from conventional δ(Φ) = 4Δ Δ (22)
⎩
LSM but also introduces the automatic generation of new holes. 0 |Φ| > Δ
Besides, gradient-based mathematical optimizers, such as optimal-
ity criteria (OC) method [56] and the method of moving asymptotes The sensitivity in Eqs. (19) and (20) should be modified slightly
(MMA) [57], can be used to update the design variables. to adapt to the design of periodic structures. As for the periodic
The scalar LSF is parametrized by a compactly supported radial structures, all the substructures share the same level set representa-
basis function (CSRBF)-based interpolation. tion and volume constraint. But the displacements within each sub-
structure are different. The sensitivities are constructed to update a

N
single representative substructure such that the assembled macro-
Φ(x, t) = φk (x)αk (t) = Ψ(x)α(t) (13) structure achieves the optimal performance. The modified sensitiv-
k
ities for the objective and the constraint are given as
with a vector of spatial shape functions and a vector of time- n 
dependent expansion coefficients ∂J  m 
=− εT (ui,j )Dε(ui,j )φk (xi,j )δ(Φi,j )dΩi,j (23)
∂αk Ωi,j
Ψ(x) = φ1 (x) φ2 (x) . . . φN (x) (14) i j

α(t) = [α1 (t) α2 (t) . . . αN (t)]T (15) ∂G
= φk (xi,j )δ(Φi,j )dΩi,j (24)
∂αk Ωi,j
with φk (x) is the 2D Wendland’s CSRBF with C2 continuity [58]

φk (x) = (max(0, 1 − r))4 · (4r + 1) (16) 5 Numerical Implementation

with The LSF is generally defined on the Eulerian grid. The optimized
 designs would be plotted as the zero contour of the LSF in this
(x − xk )2 + (y − yk )2 paper. In order to realize the repeated FEA in topology optimization,
r(x − xk ) = (17)
dmI the physical model is approximated by elements with different den-
where r(‖x − xk‖) is the compactly supported radius defined on sities using the Heaviside function (21).
Euclidean grid and dmI is the radius of influence domain. The flowchart of the topology optimization of periodic structures
Substituting Eq. (13) into Eq. (12), the parametrized level set is shown in Fig. 5. The optimization process begins with the initial-
equation is written as ization of the LSF, structural mesh, the number of substructures,
and some other optimization parameters related to the PLSM.
Ψ(x)α̇α̇(t) − Vn |∇Ψ(x)α(t)| = 0 (18) After that, the substructures are condensed by the substructure
method and then assembled to a reduced macrostructure. With the
With this parametrization, the original space and time coupled reduced macrostructure, the macrodisplacement field is obtained
Hamilton–Jacobi partial differential equation is transferred to a by an efficient macro-FEA. The displacement fields of substructures
system of ordinary differential equations. are recovered by substituting the corresponding boundary nodal

4.2 Sensitivity Analysis. The shape derivative [43,44] is uti-

lized to derive the sensitivity of boundary perturbations with
respect to the pseudotime variable t. The Lagrange multiplier
method, the adjoint variable method, and the chain rule are used
to obtain the sensitivities of the objective and the constraint with
respect to expansion coefficients. The topology optimization formu-
lation of periodic structures in this paper is a popular compliance
minimization problem which is frequently applied in many works
of literature [45]. We would not go through with the derivation
process in this paper. Readers can refer to Refs. [51,27] for more
details. The derivatives of the objective and constraint with
respect to the expansion coefficient αk are given by

∂J
= − εT (u)Dε(u)φk (x)δ(Φ)dΩ (19)
∂αk Ω


∂G
= φk (x)δ(Φ)dΩ (20)
∂αk Ω

The Heaviside function H(Φ) is implicitly embedded in Ω. It is

defined as
⎧
⎪ ξ Φ < −Δ
⎪
⎨  
3(1 − ξ) Φ Φ3 1+ξ
H(Φ) = − + −Δ ≤ Φ ≤ Δ (21)
⎪
⎪ 4 Δ 3Δ3 2
⎩
1 Φ>Δ

where ξ = 0.001 is a small positive number and Δ is half of the

bandwidth for the numerical approximation. Fig. 5 Flowchart of the optimization procedure

071403-4 / Vol. 141, JULY 2019 Transactions of the ASME

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
displacements in the macrodisplacement field into Eq. (4). With all are optimized in the following three numerical examples. All the
the substructural displacement fields, the sensitivity information is examples are limited to compliance minimization problems as
calculated. The OC algorithm is finally used to update the design defined previously. The mathematical optimizer of the structural
variables. optimization model is the OC. The termination criterion is con-
In the PLSM, both OC and MMA can be used to update the structed that the absolute difference of successive objective is less
design variables. The problem considered in this paper is compli- than 1e − 5 or the program achieves the maximum loop number
ance minimization with single volume constraint. The OC algo- 200. The material properties of all the structures are specified as a
rithm is effective enough to handle the optimization problem with solid isotropic material with elasticity modulus E = 1 and Poisson’s
a single constraint. ratio v = 0.3. For unity, all the substructures are discretized by 40 ×
40 plane stress quadrilateral elements.

6 Numerical Examples
6.1 Periodic Structures. The first example is presented for the
This section investigates some examples to demonstrate the effi- periodic design of a double-clamped beam. The design domain of
ciency and effectiveness of the proposed method. Periodic, layer- this beam is a rectangle with an aspect ratio of L : H = 2 : 1.
wise periodic, and graded layerwise periodic substructures Figure 6 shows that both of the left and the right boundaries are
clamped. A concentrated load F = −1 is applied at the center. In
the periodic design, the design domain is partitioned into six
cases: 2 × 1, 8 × 4, 16 × 8, 32 × 16, 64 × 32, and 128 × 64 substruc-
tures, respectively. The volume fraction of each substructure is
set to f = 0.5.
The optimized substructures and the assembled macroperiodic
structures are listed in Table 1. The optimized substructures of the
first several cases are topologically different. After the partition is
refined larger, the optimized substructures are gradually convergent
to a structure with similar topology. It is shown that as the number
of partitions increase, the compliance of the macrostructure also
increases from 5.735 to 12.759. Since the stiffest multiscale struc-
Fig. 6 The design domain of a periodic structure ture in compliance minimization design is a macroscale structure

Table 1 Optimized results of a double-clamped beam

Partitions Substructure Macrostructure LSF Compliance Time per iteration

2×1 5.735 0.74

8×4 9.89 1.44

16 × 8 10.457 3.79

32 × 16 11.203 13.60

64 × 32 11.978 57.05

128 × 64 12.759 326.19

Journal of Mechanical Design JULY 2019, Vol. 141 / 071403-5

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
with solid microstructures, the optimal design is changed from the The recovery of substructure internal displacement field is carried
macroscale design to the microscale design as the number of unit out substructure-by-substructure sequentially. It is suggested to
cell increases, the objective function value increases. Some opti- use the result of K−1
ii Kib from Eq. (7) and substitute it into Eq. (4)
mized substructures are slightly dissimilar to the results obtained directly during optimization. In this case, K−1
ii Kib needs to be calcu-
by Huang and Xie [7]. The main differences are observed that lated only once. One may, otherwise, calculate K−1 ii Kib for m × n + 1
there are more holes and features in our optimized substructures. times. Considering the similar example in literature [34], the com-
To the knowledge of authors, these dissimilarities may be caused putational efficiency of the proposed method is very impressive.
by the difference of the level set topology optimization method It should be noted that the displacement field of the periodic struc-
and the ESO method. On the one hand, the initialization of the ture is obtained by the substructure method without any approxima-
LSF can affect the optimized topology. On the other hand, the tions. Furthermore, parallel computing can be applied for the three
level set topology optimization method has no sensitivity filtering steps of the FEA to achieve a higher computational efficiency.
within the numerical implementation.
The main advantage of the presented method in this paper is the 6.2 Layerwise Periodic Structures. Not only does the period-
increased computational efficiency with regard to the FEA. As dis- icity can be defined over the whole macrodomain, but it can also be
played in Fig. 7, the numbers of the condensed DOFs of the six defined in a layerwise manner. With this example, we investigate a
cases are given as 558, 6018, 22, 146, 84, 738, 3, 31, 266, and 1, layerwise periodic cantilever beam design. The layers in this work
309, 698. Comparing with the numbers of the full DOFs without are defined along the Y direction. Substructures in the same row are
a static condensation, which are given by 6642, 103, 362, 411, grouped into the same layer. To do so, the original objective func-
522, 1, 642, 242, 6, 561, 282, and 26, 229, 762, the condensed tion is modified to the following form. For layer j, the objective
number of DOFs are decreased dramatically even to a lower order function is defined as
of magnitude. The decreased DOFs significantly reduce the compu-
m 

tational cost of FEA in each step of the iteration in topology optimi-

J = εT (ui,j )Dε(ui,j )dΩi,j j = 1, . . . , n (25)
zation. This efficiency would be much more rewarding with the
i Ωi,j
mesh refinement of the partitions and the substructures.
The computational environment is a personal desktop computer The design domain is a n-layer cantilever beam with an aspect
with an Intel i7-6900K CPU and 64 GB memory, installing ratio of L : H = 2 : 1. As illustrated in Fig. 9, a concentrated load
Windows 10 operating system and MAtlab 2017a. The program F = −1 is applied on the middle point of the right boundary. The
of this example is run with a single computational thread. left boundary of the design domain is fixed. Although the proposed
Figure 8 shows the average time per iteration during the optimiza- method is able to efficiently handle a large number of substructures,
tion process. This time contains the computational cost of FEA, the number of substructures is limited for better visualization. Five
updating of design variables, and plotting of figures in each itera- cases of i × j substructures are given as 4 × 2, 8 × 4, 12 × 6, 16 × 8,
tion. The FEA is the most time-consuming procedure. This proce- and 20 × 10. Each of the substructure is discretized by 40 × 40
dure is divided into three steps: the assembly of global stiffness plane stress quadrilateral elements. The volume fraction of each
matrix K* (Eq. (11)), solving of condensed global equilibrium substructure is prescribed as f = 0.5.
equation (Eq. (10)), and the recovery of substructure internal displa- The optimized layerwise substructures for this five partitions are
cement field (Eq. (4)). The assembly of global stiffness matrix is given in Table 2. All the layers of the macrostructures are optimized
performed with the sparse function in MAtlab to avoid the use of independently to obtain different topologies. It can be observed that
for loops by following the style of 88-line MAtlab code [59]. all optimized substructures are well connected with each other
Global equilibrium equation (10) is solved by the direct solver. without the artificial enforcement of passive connectors as done
in Ref. [26]. The reason is that scale-related FEA is performed
that all layers are included in one integral analysis. The optimization
objective drives the LSF to one that forces can be transferred
throughout the whole design domain effectively. The boundary
conditions of this cantilever beam lead to a symmetric distribution
of structural strain energies. Even though the layers in the upper half
part optimized independently with the layers in the lower half part,
they have symmetric topologies.
Comparison of the designs of layerwise with periodic substruc-
tures is listed as in Table 2. In the periodic design, each substructure
is optimized adapted to the forces averaged by the whole domain.
Unlike the periodic design, layerwise design is optimized adapted
to the forces by layer. And there are more design freedoms in the
layerwise design. Consequently, the structural performance in the
layerwise design outperforms in stiffness than that of the periodic
Fig. 7 Comparison of the number of DOFs design. As shown in Fig. 10, the compliance of layerwise design
is increased from 73.6743 to 97.9765, while the compliance of

Fig. 8 The average time per iteration Fig. 9 Design domain of a cantilever beam

071403-6 / Vol. 141, JULY 2019 Transactions of the ASME

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
Table 2 Layerwise design versus periodic design

Partitions Layerwise design Periodic design Compliance (layer-wise/periodic)

4×2 73.6743 / 101.9019

8×4 83.6019 / 125.1113

12 × 6 92.7152 / 129.7955

16 × 8 94.2075 / 129.7955

20 × 10 97.9765 / 130.6799

the periodic design is increased from 101.9019 to 130.6799. The have the same volume fractions. The volume fractions are then
compliance of the layerwise design is much lower than that of the rounded to the nearest two decimal digits. The volume fractions
periodic design. Still, the compliance of both designs is increased of the upper and the lower half part are symmetric. Each half part
as the mesh refinement of the partitions just as the same in Sec. 6.1. has a group of graded distributed volume fractions. The mean
average volume fraction of all the layers equals 0.5. Each
6.3 Graded Layerwise Periodic Structures. This example is
further extended to a graded layerwise periodic design. The macro-
design domain and the boundary conditions are the same as in Sec.
6.2. But the volume fractions of each layer of substructures are not
uniform. As illustrated in Fig. 11, the initial volume fractions of the
substructures are calculated as elemental densities by the 99-line
MAtlab code [56] with the penalty factor equals 1 and additional
layerwise constraints. The number of substructures is restricted to
20 × 10 for better visualization. Substructures in the same layer

Fig. 11 Graded density distributions

Fig. 10 Comparison of compliance of layerwise design and peri-

odic design Fig. 12 Graded layerwise design of periodic substructures

Journal of Mechanical Design JULY 2019, Vol. 141 / 071403-7

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
51705165), National Basic Scientific Research Program of China
(JCKY2016110C012), and Program for HUST Academic Frontier
Youth Team.

References
[1] Zuo, Z. H., Xie, Y. M., and Huang, X., 2011, “Reinventing the Wheel,” ASME
J. Mech. Des., 133(2), 024502.
[2] Liu, C., Du, Z., Zhang, W., Zhu, Y., and Guo, X., 2017, “Additive
Manufacturing-Oriented Design of Graded Lattice Structures Through Explicit
Topology Optimization,” ASME J. Appl. Mech., 84(8), 081008.
[3] Aremu, A., Brennan-Craddock, J., Panesar, A., Ashcroft, I., Hague, R. J.,
Wildman, R. D., and Tuck, C., 2017, “A Voxel-Based Method of Constructing
and Skinning Conformal and Functionally Graded Lattice Structures Suitable
for Additive Manufacturing,” Addit. Manuf., 13, pp. 1–13.
[4] Sigmund, O., 1994, “Materials with Prescribed Constitutive Parameters:
An Inverse Homogenization Problem,” Int. J. Solids. Struct., 31(17), pp. 2313–
2329.
Fig. 13 Iteration history of compliance and volume fraction [5] Cadman, J. E., Zhou, S., Chen, Y., and Li, Q., 2013, “On Design of
Multi-Functional Microstructural Materials,” J. Mater. Sci., 48(1), pp. 51–66.
[6] Zhang, W., and Sun, S., 2006, “Scale-Related Topology Optimization of Cellular
substructure is discretized by 40 × 40 plane stress quadrilateral ele- Materials and Structures,” Int. J. Numer. Methods Eng., 68(9), pp. 993–1011.
[7] Huang, X., and Xie, Y., 2008, “Optimal Design of Periodic Structures Using
ments as stated previously. Evolutionary Topology Optimization,” Struct. Multidiscipl. Optim., 36(6),
The optimized macrostructure is displayed in Fig. 12. The opti- pp. 597–606.
mized macrostructure can be regarded as a sandwich structure. [8] Sigmund, O., and Torquato, S., 1997, “Design of Materials With Extreme
The upper and lower layers are initialized as solid to provide suffi- Thermal Expansion Using a Three-Phase Topology Optimization Method,”
J. Mech. Phys. Solids., 45(6), pp. 1037–1067.
cient stiffness to prevent bending deformation from the external [9] Bendsøe, M. P., and Sigmund, O., 1999, “Material Interpolation Schemes in
load. And the middle layers are optimized as porous substructures Topology Optimization,” Arch. Appl. Mech., 69(9–10), pp. 635–654.
to resist the concentrated force. [10] Hassani, B., and Hinton, E., 1998, “A Review of Homogenization and Topology
The substructures in each layer are characterized by specially Optimization i-Homogenization Theory for Media With Periodic Structure,”
Comput. Struct., 69(6), pp. 707–717.
designed structures adapted to the force of the layer. The substruc- [11] Andreassen, E., and Andreasen, C. S., 2014, “How to Determine Composite
tures in different layers are well connected even though they may Material Properties Using Numerical Homogenization,” Comput. Mater. Sci.,
have different volumes. The topologies of these substructures are 83, pp. 488–495.
different from that of the layerwise design since the volume frac- [12] Xia, L., and Breitkopf, P., 2015, “Design of Materials Using Topology
Optimization and Energy-Based Homogenization Approach in Matlab,” Struct.
tions in this example are distributed in a graded manner. It should Multidiscipl. Optim., 52(6), pp. 1229–1241.
be noted that the graded distribution of volume fractions is achieved [13] Wang, Y., Luo, Z., Zhang, N., and Kang, Z., 2014, “Topological Shape
by a density-based topology optimization. The graded distribution Optimization of Microstructural Metamaterials Using a Level Set Method,”
is superior to the uniform volume fraction distribution with Comput. Mater. Sci., 87, pp. 178–186.
[14] Clausen, A., Wang, F., Jensen, J. S., Sigmund, O., and Lewis, J. A., 2015,
respect to the structural compliance. As a result, the compliance “Topology Optimized Architectures With Programmable Poisson’s Ratio Over
of the graded layerwise design, 86.1830, is lower than that of the Large Deformations,” Adv. Mater., 27(37), pp. 5523–5527.
layerwise design in the case of 20 × 10 partitions. [15] Vogiatzis, P., Chen, S., Wang, X., Li, T., and Wang, L., 2017, “Topology
The iteration history, shown in Fig. 13, is very stable during the Optimization of Multi-Material Negative Poisson’s Ratio Metamaterials using a
Reconciled Level Set Method,” Comput. Aided Des., 83, pp. 15–32.
200 loops. The compliance reaches a relatively high value in the [16] Guest, J. K., and Prévost, J. H., 2007, “Design of Maximum Permeability Material
first five iterations and then decreases with the iteration number. Structures,” Comput. Methods Appl. Mech. Eng., 196(4–6), pp. 1006–1017.
This phenomenon is caused by the LSF initialization and OC [17] Rodrigues, H., Guedes, J. M., and Bendsoe, M., 2002, “Hierarchical Optimization
optimizer in the PLSM. Because the volume fraction of the LSF-rep- of Material and Structure,” Struct. Multidiscipl. Optim., 24(1), pp. 1–10.
[18] Liu, L., Yan, J., and Cheng, G., 2008, “Optimum Structure With Homogeneous
resented substructure is initialized larger than the prescribed value Optimum Truss-Like Material,” Comput. Struct., 86(13–14), pp. 1417–1425.
0.5, the LSF shrinks to meet the volume fraction at the first steps. [19] Huang, X., Zhou, S., Xie, Y., and Li, Q., 2013, “Topology Optimization of
The LSF shrinks to meet the volume fraction at the first steps. And Microstructures of Cellular Materials and Composites for Macrostructures,”
void area occurs on the path from the constraint location to the Comput. Mater. Sci., 67, pp. 397–407.
[20] Xia, L., and Breitkopf, P., 2014, “Concurrent Topology Optimization Design of
load point, leading to an amplified displacement field. After the Material and Structure Within Fe2 Nonlinear Multiscale Analysis Framework,”
volume fraction is satisfied, the LSF evolves to connect the locations Comput. Methods Appl. Mech. Eng., 278, pp. 524–542.
of the constraint and load such that the structure has better stiffness. [21] Xia, L., and Breitkopf, P., 2015, “Multiscale Structural Topology Optimization
with an Approximate Constitutive Model for Local Material Microstructure,”
Comput. Methods Appl. Mech. Eng., 286, pp. 147–167.
[22] Sivapuram, R., Dunning, P. D., and Kim, H. A., 2016, “Simultaneous Material
7 Conclusions and Structural Optimization by Multiscale Topology Optimization,” Struct.
Multidiscipl. Optim., 54(5), pp. 1267–1281.
An efficient level set topology optimization method for periodic [23] Wang, Y., Wang, M. Y., and Chen, F., 2016, “Structure-Material Integrated
structures with substructuring is presented in this paper. The static Design by Level Sets,” Struct. Multidiscipl. Optim., 54(5), pp. 1145–1156.
condensation is adopted for repeated common substructures to [24] Yan, J., Guo, X., and Cheng, G., 2016, “Multi-Scale Concurrent Material and
improve the computational efficiency. Scale-related periodic struc- Structural Design under Mechanical and Thermal Loads,” Comput. Mech.,
57(3), pp. 437–446.
tures with millions of DOFs can be solved efficiently. For the design [25] Da, D., Cui, X., Long, K., and Li, G., 2017, “Concurrent Topological Design of
of periodic structures, higher structural performance can be Composite Structures and the Underlying Multi-Phase Materials,” Comput.
achieved by the layerwise and graded layerwise design. The pre- Struct., 179, pp. 1–14.
sented method provides the potential to design well-connected [26] Li, H., Luo, Z., Gao, L., and Walker, P., 2018, “Topology Optimization for
Functionally Graded Cellular Composites With Metamaterials by Level Sets,”
element-wise mesostructures. Future research can be extended to Comput. Methods Appl. Mech. Eng., 328, pp. 340–364.
the design of scale-related hierarchical structures [60] and multiphy- [27] Li, H., Luo, Z., Gao, L., and Qin, Q., 2018, “Topology Optimization for
sics problems. Concurrent Design of Structures With Multi-Patch Microstructures by Level
Sets,” Comput. Methods Appl. Mech. Eng., 331, pp. 536–561.
[28] Fu, J., Li, H., Gao, L., and Xiao, M., 2019, “Design of Shell-Infill Structures by a
Multiscale Level Set Topology Optimization Method,” Comput. Struct., 212,
Acknowledgment pp. 162–172.
[29] Zhu, Y., Li, S., Du, Z., Liu, C., Guo, X., and Zhang, W., 2019, “A Novel
This research is partially supported by National Natural Science Asymptotic-Analysis-Based Homogenisation Approach towards Fast Design of
Foundation of China (51705166, 51675196, 51721092, and Infill Graded Microstructures,” J. Mech. Phys. Solids., 124, pp. 612–633.

071403-8 / Vol. 141, JULY 2019 Transactions of the ASME

 asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/141/7/071403/6402337/md_141_7_071403.pdf?casa_token=gSn-1Oj4C4oAAAAA:dthF37YJIzq9qE2HTpneqERvDK-_ILfaNlQSTwaNTx771wEM1S834_X2D0NV6rW6AO4IC7gv-g by Pt. Dwarka Prasad Mishra Indian Inst of Information Tec
[30] Du, Z., Zhou, X.-Y., Picelli, R., and Kim, H. A., 2018, “Connecting [46] Luo, J., Luo, Z., Chen, L., Tong, L., and Wang, M. Y., 2008, “A Semi-Implicit
Microstructures for Multiscale Topology Optimization With Connectivity Index Level Set Method for Structural Shape and Topology Optimization,”
Constraints,” ASME J. Mech. Des., 140(11), 111417. J. Comput. Phys., 227(11), pp. 5561–5581.
[31] Groen, J. P., and Sigmund, O., 2018, “Homogenization-Based Topology [47] Xia, Q., Wang, M. Y., Wang, S., and Chen, S., 2006, “Semi-Lagrange Method for
Optimization for High-Resolution Manufacturable Microstructures,” Level-Set-Based Structural Topology and Shape Optimization,” Struct.
Int. J. Numer. Methods Eng., 113(8), pp. 1148–1163. Multidiscipl. Optim., 31(6), pp. 419–429.
[32] Zuo, Z. H., Huang, X., Yang, X., Rong, J. H., and Xie, Y. M., 2013, “Comparing [48] Dunning, P. D., and Kim, H. A., 2015, “Introducing the Sequential Linear
Optimal Material Microstructures With Optimal Periodic Structures,” Comput. Programming Level-Set Method for Topology Optimization,” Struct.
Mater. Sci., 69, pp. 137–147. Multidiscipl. Optim., 51(3), pp. 631–643.
[33] Yvonnet, J., and Bonnet, G., 2014, “A Consistent Nonlocal Scheme Based on [49] Yamada, T., Izui, K., and Nishiwaki, S., 2011, “A Level Set-Based
Filters for the Homogenization of Heterogeneous Linear Materials With Topology Optimization Method for Maximizing Thermal Diffusivity in
Non-Separated Scales,” Int. J. Solids. Struct., 51(1), pp. 196–209. Problems Including Design-Dependent Effects,” ASME J. Mech. Des., 133(3),
[34] Alexandersen, J., and Lazarov, B. S., 2015, “Topology Optimisation of 031011.
Manufacturable Microstructural Details Without Length Scale Separation Using [50] Zhu, B., Wang, R., Li, H., and Zhang, X., 2018, “A Level Set Method With a
a Spectral Coarse Basis Preconditioner,” Comput. Methods Appl. Mech. Eng., Bounded Diffusion for Structural Topology Optimization,” ASME J. Mech.
290, pp. 156–182. Des., 140(7), 071402.
[35] Xie, Y. M., Zuo, Z. H., Huang, X., and Rong, J. H., 2012, “Convergence of [51] Luo, Z., Tong, L., and Kang, Z., 2009, “A Level Set Method for Structural Shape
Topological Patterns of Optimal Periodic Structures Under Multiple Scales,” and Topology Optimization Using Radial Basis Functions,” Comput. Struct.,
Struct. Multidiscipl. Optim., 46(1), pp. 41–50. 87(7–8), pp. 425–434.
[36] Huang, X., and Xie, M., 2010, Evolutionary Topology Optimization of Continuum [52] Luo, Z., Tong, L., and Ma, H., 2009, “Shape and Topology Optimization for
Structures: Methods and Applications, John Wiley & Sons, New York. Electrothermomechanical Microactuators Using Level Set Methods,”
[37] Xia, L., Xia, Q., Huang, X., and Xie, Y. M., 2018, “Bi-Directional Evolutionary J. Comput. Phys., 228(9), pp. 3173–3181.
Structural Optimization on Advanced Structures and Materials: A Comprehensive [53] Guo, X., Zhang, W., and Zhong, W., 2014, “Doing Topology Optimization
Review,” Arch. Comput. Methods Eng., 25(2), pp. 437–478. Explicitly and Geometrically a New Moving Morphable Components Based
[38] Guyan, R. J., 1965, “Reduction of Stiffness and Mass Matrices,” AIAA J., 3(2), Framework,” ASME J. Appl. Mech., 81(8), 081009.
pp. 380–380. [54] Zhang, W., Chen, J., Zhu, X., Zhou, J., Xue, D., Lei, X., and Guo, X.,
[39] Groen, J. P., Langelaar, M., Sigmund, O., and Ruess, M., 2017, “Higher-Order 2017, “Explicit Three Dimensional Topology Optimization via Moving
Multi-Resolution Topology Optimization Using the Finite Cell Method,” Morphable Void (MMV) Approach,” Comput. Methods Appl. Mech. Eng., 322,
Int. J. Numer. Methods Eng., 110(10), pp. 903–920. pp. 590–614.
[40] Borrvall, T., and Petersson, J., 2001, “Large-Scale Topology Optimization in 3d [55] Liu, C., Zhu, Y., Sun, Z., Li, D., Du, Z., Zhang, W., and Guo, X., 2018, “An
Using Parallel Computing,” Comput. Methods Appl. Mech. Eng., 190(46–47), Efficient Moving Morphable Component (MMC)-Based Approach for
pp. 6201–6229. Multi-Resolution Topology Optimization,” Struct. Multidiscipl. Optim., 58(6),
[41] Osher, S., and Sethian, J. A., 1988, “Fronts Propagating with Curvature- pp. 2455–2479.
Dependent Speed: Algorithms Based on Hamilton-Jacobi Formulations,” [56] Bendsoe, M. P., and Sigmund, O., 2013, Topology Optimization: Theory,
J. Comput. Phys., 79(1), pp. 12–49. Methods, and Applications, Springer Science & Business Media, Berlin/
[42] Sigmund, O., and Maute, K., 2013, “Topology Optimization Approaches,” Struct. Vienna/New York.
Multidiscipl. Optim., 48(6), pp. 1031–1055. [57] Svanberg, K., 1987, “The Method of Moving Asymptotes a New Method for
[43] Wang, M. Y., Wang, X., and Guo, D., 2003, “A Level Set Method for Structural Structural Optimization,” Int. J. Numer. Methods Eng., 24(2), pp. 359–373.
Topology Optimization,” Comput. Methods Appl. Mech. Eng., 192(1–2), [58] Wendland, H., 2004, Scattered Data Approximation, Vol. 17. Cambridge
pp. 227–246. University Press, Cambridge, England/London/New York.
[44] Allaire, G., Jouve, F., and Toader, A.-M., 2004, “Structural Optimization Using [59] Andreassen, E., Clausen, A., Schevenels, M., Lazarov, B. S., and Sigmund, O.,
Sensitivity Analysis and a Level-Set Method,” J. Comput. Phys., 194(1), 2011, “Efficient Topology Optimization in Matlab using 88 Lines of Code,”
pp. 363–393. Struct. Multidiscipl. Optim., 43(1), pp. 1–16.
[45] van Dijk, N. P., Maute, K., Langelaar, M., and Van Keulen, F., 2013, “Level-Set [60] Wu, Z., Xia, L., Wang, S., and Shi, T., 2019, “Topology Optimization of
Methods for Structural Topology Optimization: A Review,” Struct. Multidiscipl. Hierarchical Lattice Structures With Substructuring,” Comput. Methods Appl.
Optim., 48(3), pp. 437–472. Mech. Eng., 345(1), pp. 602–617.

Journal of Mechanical Design JULY 2019, Vol. 141 / 071403-9