Materials & Design 210 (2021) 110026

Contents lists available at ScienceDirect

Materials & Design

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

Design optimization of lattice structures with stress constraints

Rossana R. Fernandes, Ali Y. Tamijani ⇑
Embry-Riddle Aeronautical University, Daytona Beach, FL 32114, United States

h i g h l i g h t s g r a p h i c a l a b s t r a c t

Computational and experimental

evaluations of effective mechanical
properties.

Homogenization-based stress-
constrained optimization of lattice
structures.

Implementation of unit cell
orthotropic properties in the
optimization process.

Experimental investigation of
optimized lattice structure stiffnesses
and strengths.

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents an experimentally validated framework used to perform topology and orientation
Received 4 June 2021 (morphology) optimization of lattice structures subject to stress constraints. The effective stiffnesses
Revised 14 July 2021 and yield stresses of a unit cell are obtained using numerical homogenization and validated experimen-
Accepted 1 August 2021
tally. Due to the orthotropic behavior of the unit cell, the modified Hill’s yield criterion is used to describe
Available online 3 August 2021
the lattice strength. The effective orthotropic properties are implemented via macrostructure topology
optimization to further improve the lattice structure stiffness. Homogenization-based optimization is
Keywords:
performed using a coarse mesh and the optimized design is projected onto a fine mesh. This reduces
Topology optimization
Morphology optimization
the computational cost significantly. Finally, the projected design is post-processed to ensure the fabri-
Homogenization cation feasibility of the optimized lattice structure. The framework is tested for two cases: an L-shaped
Cellular solids bracket and a single-edge notched bend (SENB) problem. A comparison of the compliance-based and
Additive manufacturing stress-constrained designs used in the two cases demonstrates that the changes in the optimal material
distribution that occur upon implementing the stress constraint result in higher yield strength. The SENB
lattice structures are additively manufactured and the stiffnesses and yield strength of the optimized
designs are compared to those obtained numerically.
Ó 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction ogy and morphology optimization have received considerable

attention in recent years due to the emergence of new additive
A lattice structure comprises a network of cells with nodes and manufacturing techniques with the ability to fabricate microstruc-
struts that offers exceptional properties such as high stiffness, tures. Among the various multiscale optimization methods used to
energy absorption, and acoustic insulation. Lattice structure topol- design lattice structures, the homogenization-based approach has
been investigated the most. This is because the effective properties
⇑ Corresponding author.
of local microstructures can be obtained using homogenization
theory. A surrogate model is then obtained for various microstruc-
E-mail addresses: fernanr5@my.erau.edu (R.R. Fernandes), tamijana@erau.edu
(A.Y. Tamijani). ture densities, and this surrogate model is used within the opti-

https://doi.org/10.1016/j.matdes.2021.110026
0264-1275/Ó 2021 The Authors. Published by Elsevier Ltd.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

mization process. This significantly reduces the computational tions, two variables must be used to describe the cell. To this
complexity of microstructure analysis during the optimization pro- end, Donders [12] studied minimization of the L2 norms of lattice
cedure. Projection of homogenized designs is inspired by the work structure stresses in square cells with rectangular holes using two
of Pantz and Trabelsi [1], in which the homogenized properties of variables. The effective stress was obtained using homogenized
rank-two laminates were used to optimize microstructure topolo- stress–strain relations. Then an amplification factor, which took
gies and the principal direction was utilized to optimize cell mor- micro-stress fluctuations into account, was multiplied by the effec-
phologies. A mapping function based on the optimized orientation tive stress to mimic the microstructural stress. While both stress
was needed to project the homogenized design. However, since the amplification factor and effective allowable stress approaches were
optimized orientation was symmetric with respect to p, a regular- tested for microstructure topology optimization, cell orientations
ization approach was utilized to alleviate the discontinuities in the were kept unchanged in previous studies. However, various cells,
optimized orientation. Optimization was performed using a coarse such as square cells with rectangular holes that exhibit superior
mesh. The optimized design was projected onto a fine mesh using a orthotropic properties, are weak with regard to withstanding shear
mapping function and optimized density to generate the lattice stress. The orientation must be incorporated in the optimization
structure. Later, each of the three steps i.e., homogenization, regu- process in order to take advantage of the orthotropic properties
larization, and projection was enhanced in other research studies. of the cell.
Groen and Sigmund [2] interpolated the homogenized properties We previously studied compliance-based lattice structure
of a square cell with rectangular hole instead of using the explicit topology and morphology optimization [13] using load paths and
homogenized properties of rank-two laminates. They also intro- load flows as intermediate variables [14]. Later, we also studied
duced a connected-component labeling approach to simplify regu- optimization of material distribution and orientation and the sub-
larization. Allaire, Geoffroy-Donders, & Pantz [3] addressed the sequent projection for various types of cells and lattices using
conformality of square lattices. They also used an abstract manifold Fourier series representation [15]. The goal of the current study
for the computational domain to reconstruct lattice structures is to investigate the optimization of the topologies and morpholo-
from optimized homogenized designs. gies of lattice structures that are subject to stress constraints.
Despite the success of stiffness-based de-homogenization Square cells with rectangular holes with two characterizing
approaches in generating lightweight periodic lattice structures, parameters related to the void are used as the base cells. The effec-
higher stiffnesses cannot be achieved if the stress constraints are tive unit cell stiffness and yield stress are obtained using numerical
not implemented within the optimization process and the opti- homogenization and compared to those found experimentally.
mized lattice structure fails. In addition, utilizing stress constraints Effective properties are obtained for various cell parameters in
may change the optimized lattice structure topology and morphol- order to construct response surfaces. The response surfaces are
ogy. There are three sources of complexity associated with stress- implemented using a homogenization-based optimization algo-
constrained optimization using homogenized properties: (1) stress rithm. The modified Hill’s criterion is implemented to establish
singularities at zero density, (2) the large number of local stress the stress constraints and a material indicator variable is used to
constraints, and (3) the local nature of micro-stresses, which address the stress singularity issue. The stress constraints are rep-
depends heavily on the cell micro-architecture. In the context of resented using a p-mean function and a clustering approach is uti-
macrostructure topology optimization, stress singularities at zero lized to preserve some of the local nature of the stress. The
density are addressed by using a polynomial, Kreisselmeier- optimized homogenized designs are then projected onto a fine
Steinhauser, or reciprocal function to smooth the feasible design mesh to generate the lattice structure. Finally, the projected lattice
space [4,5]. The issue of large numbers of stress constraints has designs are post-processed to remove thin and floating members
been addressed by aggregating the local constraints into a global and enforce a minimum manufacturable feature size. The frame-
constraint using the Kresselmeier-Steinhauser, p-norm, or p- work is tested for two cases: an L-shaped bracket and a single-
mean functions [6,7]. However, aggregations of stresses may not edge notched bend (SENB) problem. Comparison of the
represent local stresses in structures. In order to preserve some compliance-based and stress-constrained designs for these cases
of the local nature of the stresses while avoiding excessive compu- clearly demonstrates that the optimized material distribution is
tational cost, regional stress approaches have been proposed in different when the stresses are considered as constraints. In addi-
which the design domain is split into several clusters and the stress tion, the projected SENB design is additively manufactured and
constraints of the elements within each cluster are aggregated into experimentally evaluated. The optimized design is shown to be
a single constraint [8,9]. manufacturable and to exhibit yield strength similar to those
Unlike the above two issues, i.e., stress singularities and large implemented via stress constraints. The four major contributions
numbers of local stress constraints, which exist in both macro of this study are: (1) comparison of the effective stiffnesses and
and microstructural optimization, macrostructure optimization yield stresses obtained using numerical homogenization and
with microstructural stress considerations is less investigated. experimental evaluation; (2) implementation of stress constraints
Microstructural stress constraints can be implemented within the and related sensitivity analyses in homogenization-based opti-
macrostructure optimization process using two approaches: (a) mization of lattice structures; (3) incorporation of unit cell ortho-
finding the effective allowable stresses and (b) amplifying the tropic properties in the optimization process to further improve
homogenized stress. Cheng, Bai, and To [10] investigated topology the stiffness; and (4) demonstrating the fabrication feasibility of
optimization of lattice structures with stress constraints. A cubic the lattice design and experimentally validating the optimized
lattice structure with a single-density design variable was consid- yield strength.
ered. The Hill’s yield criterion was used as the strength constraint.
The effective yield strength in Hill’s model was obtained by per-
forming finite element analyses (FEAs) of lattice structures for var- 2. Effective cell properties
ious densities. Yu et al. [11] studied topology optimization of shell-
lattice structures with stress constraints. This study also used the Square cells with rectangular holes are considered in the cur-
Hill yield criterion. The effective yield strength was found for var- rent study. As shown in Fig. 1, the cell is parameterized using
ious densities. In both of these studies, a single design variable was two parameters (h1 and h2 ). The member thicknesses can be found
employed to control the lattice structure topology. However, to use using the corresponding parameter tn ¼ 1  hn (n is the parameter
the full potential of the lattice structures in perpendicular direc- number) and the corresponding cell volume fraction is 1  h1 h2 .
2
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

ity PA 12 in a Multi Jet Fusion (MJF) printer (HP Jet Fusion 5200
Series). Through the MJF, fusing and detailing agents were depos-
ited along with heat onto thin layers of powder. This process was
repeated until all layers were formed. The material properties were
determined by testing five specimens according to ASTM D638-14.
The Young’s modulus of the material was E ¼ 1288:30  45MPa;
the yield strength was rY ¼ 18:3  0:9MPa, and the Poisson’s ratio
was m ¼ 0:375  0:1.
The lattice specimens were subjected to tensile loads using an
MTS Criterion Model 43 test machine with a 50 kN load cell at a
constant crosshead displacement rate of 0.875 mm/min. The test
was terminated when the samples fractured fully. The strains were
measured using digital image correlation (DIC). The square cell was
5mm long (L) and 3 mm thick. In the first set of experiments,
h2 ¼ 0:5L was kept constant and five different h1 values
(h1 ¼ 0:1L; 0:3L; 0:5L; 0:7L and 0:9L) were considered. In the other
set of experiments, h1 was kept constant at 0:5L and four different
h2 values (h2 ¼ 0:1L; 0:3L; 0:7L and 0:9L) were considered. Table 1
Fig. 1. Parameterization of a square cell with a rectangular hole.
shows the measured dimensions of the 3D printed samples. There
is an average error of 2.42% between the CAD model and the
 printed sample. The results presented are the average of five sam-
The homogenized stiffness tensor (C ) is found via numerical ples that were 3D printed for each configuration. Fig. 3 (a) shows
homogenization [16–18] over a representative volume element the printed model for a cellular solid with
(RVE): h1 ¼ 0:7L and h2 ¼ 0:5L. The strain distribution for the middle of
R @ vkl  the sample was obtained via DIC and a single cell was extracted
 epq @@yvji dY ¼ 0
ð0Þkl
p
Y
C ijpq dyq (Fig. 3 (b) and (c)). The results are compared to the distribution
 R     
ð0Þkl ð0Þij
C ijkl ðrÞ ¼ jY1j Y C mspq ðr; yÞ epq  epq ðvkl Þ ems  ems ðvij Þ dY obtained via homogenization theory (e ¼ M S r) in Fig. 3 (d). The
numerical homogenization contour shows that high strains appear
ð1Þ around the hole with the highest strains near the four corners of
where Y is the cell domain, e ð0Þkl
are the three macroscopic unit the hole. In contrast, the experimental contour shows a high but
more distributed strain on the side of the hole with the highest
strains, vkl are the displacement fields, v i is the virtual displacement
strain at the corner of the hole and expanding to the corner of
field, and C is the stiffness tensor of the cell material. Using epq ðvkl Þ
the cell. Quantitatively, the experimental and numerical homoge-
allows the relation between the local micro-stress (r) and effective
   nization strains are within similar ranges. Fig. 3 (e)-(h) compares

macro-stress (r ¼ C e; e is the effective strain) to be established the experimental and computational strain contours for
[19]: ðh1 ¼ 0:5L and h2 ¼ 0:5L) and (h1 ¼ 0:5L and h2 ¼ 0:3LÞ, for which
observations similar to those from ðh1 ¼ 0:7L and h2 ¼ 0:5LÞ can
  1 
r ¼ CM S r; Mijkl ¼ dik djl þ dil djk  eij ðvkl Þ ð2Þ be drawn.
2
Stress–strain curves are plotted for various h1 values in Fig. 4
where dij is the Kronecker delta, Mijkl is a local structure tensor that (a). All of the samples deform linearly before yielding. This is fol-

relates the effective strain and micro-strain [19], and S is the inverse lowed by nonlinear deformation and fracture. It is also observed
of the homogenized stiffness tensor. Since it is assumed that the cell that the strength and elastic modulus decrease as h1 increases. A
is composed of an isotropic material, the von Mises criterion is cho- larger h1 implies that a smaller area carries the load; this results
sen to describe the yield at the microstructure level. The von Mises in higher stresses. The stress–strain curves produced when h2 var-
yield criterion is expressed as ies (Fig. 4 (b)) exhibit trends similar to those noted when h1 is
0 pffiffiffi !2 1
allowed to vary, although the change in the yield strength is not
 2  2 as significant. The experimental and computational yield strength
rs @ rrs22 rrs11 rkl rrs 3rrs A
SR þ  11 222 þ 12
1¼0 ð3Þ and elastic modulus of each sample are compiled in Table 2 and
rY rY rY rY
Table 3. The experimental yield stresses are determined from the
0.2% offset lines of the stress–strain curves (Fig. 4).
where SRrs in Eq. (3) represents the strength ratios and rY is the
The difference between the numerical homogenization and
material yield strength. The effective yield stress of the cell is
experimental results is less than 16%. This validates the results
obtained by multiplying the applied distributed load by the
obtained using the homogenization method. The largest difference
strength ratio. Four macroscopic unit distributed loads (rð0Þrs ;
of 15% is observed for h1 ¼ 0:9L (Table 2). This can be attributed to
rs ¼ 11; 22; 12; 44), including uniaxial in each direction, pure shear
geometric differences between the CAD model and the actual
and hydrostatic, as shown in Fig. 2, are applied in order to obtain
printed model. More specifically, the printer resolution forces the
SRrs . Since the unit distributed loads are applied, the macroscopic
 holes to be printed with some curvature instead of being com-
effective yield stresses (rY ) are obtained based on the strength pletely rectangular. Rounded holes have lower stress concentra-
ratios tions. To compensate for this, a filtering technique based on the
Y weighted average distance function [20] is applied to the homoge-
rrs ¼ SRrs ð4Þ nization results to remove concentrated stress from the corners of
In order to validate the properties obtained using the homoge- the holes. The effect of filter radius on maximum of micro-strain of
nization method and evaluate the effect of the characterizing unit cell is shown in Fig. 5. As can be seen, when there is no filter-
parameters on the effective tensile yield strength, 14  4 cell lat- ing, a high concentration in strain exists at the corners. However,
tice structures were manufactured using an HP 3D High Reusabil- this issue is alleviated when a filter radius above 0.05 is considered.

3
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 2. Unit distributed loads applied to the representative volume element.

Table 1
Average measured dimensions of lattice specimens (width ¼ 20mm; thickness ¼ 3mm and L ¼ 5mm).

h1 ¼ 0:1L h1 ¼ 0:3L h1 ¼ 0:5L h1 ¼ 0:7L h1 ¼ 0:9L

Actual value % error Actual value % error Actual value % error Actual value % error Actual value % error
Width 19.85 ± 0.12 0.75 20.00 ± 0.02 0.02 20.03 ± 0.02 0.13 20.08 ± 0.03 0.38 20.13 ± 0.06 0.67
Thickn 3.06 ± 0.02 2.00 3.05 ± 0.03 1.56 3.03 ± 0.05 1.11 3.05 ± 0.02 1.56 3.06 ± 0.02 2.00
h1 0.51 ± 0.04 2.00 1.54 ± 0.03 2.67 2.53 ± 0.02 1.20 3.56 ± 0.04 1.71 4.56 ± 0.04 1.41
h2 2.52 ± 0.02 0.93 2.57 ± 0.04 2.57 2.57 ± 0.04 2.80 2.58 ± 0.03 3.07 2.49 ± 0.02 0.27
h2 ¼ 0:1L h2 ¼ 0:3L h2 ¼ 0:7L h2 ¼ 0:9L
Actual value % error Actual value % error Actual value % error Actual value % error
Width 20.06 ± 0.03 0.30 20.05 ± 0.07 0.27 20.06 ± 0.02 0.30 20.04 ± 0.06 0.20
Thickn 3.01 ± 0.01 0.33 3.05 ± 0.03 1.67 3.08 ± 0.07 2.67 3.06 ± 0.08 2.11
h1 2.59 ± 0.14 3.47 2.61 ± 0.01 4.13 2.51 ± 0.04 0.40 2.52 ± 0.03 0.93
h2 0.53 ± 0.01 5.33 1.53 ± 0.03 2.22 3.62 ± 0.10 3.43 4.68 ± 0.16 3.93

After experimental validation of the effective properties, prop- principal axes of anisotropy. The stresses and strains in the refer-
 
erties were determined for various cell parameters and used to

ence axes (rh and eh ) are obtained using the transformation

construct the response surfaces, C ðhn Þ and rY ðhn Þ, as shown in matrix (R).
Fig. 6.

The homogenized stiffness matrix C is derived with respect to 3. Material models and stress constraints
the principal axes of anisotropy. When the cell is rotated by angle
h, the stiffness matrix is updated using the transformation matrix As mentioned earlier, one issue related to stress constraints in
(R): topology optimization is that of stress singularities at low densi-
ties. This issue was first observed during a three-bar truss opti-
  mization [22]. It was shown that removing members is necessary
C h ðh; h1 ; h2 Þ ¼ RðhÞ C ðh1 ; h2 ÞRT ðhÞ ð5Þ to obtain a global minimum. However, the stress in a member
It should be noted that the stress and strain tensors are increases as the member’s thickness approaches zero. This pre-
expressed as vectors using Voigt notation. Due to the orthotropic vents removal of the member. Several approaches to addressing
behavior of the square cell with a rectangular hole, Hill’s yield cri- this issue are discussed in the introduction section. The stress
terion is chosen to describe the lattice yield strength. It is assumed interpolation scheme [8] is adopted in this work. In order to estab-
in Hill’s yield criterion that the hydrostatic pressure does not affect lish the stress interpolation scheme, a material indicator variable
the yield strength. However, this assumption is not accurate for (/) is implemented. The material indicator has previously been
lattice cells. Thus, a modified Hill’s criterion [21] is utilized that used to obtain stiffness-optimized, coated structures with ortho-
includes the hydrostatic yield strength. The effective yield stresses tropic infill [23]. In our work, the material indicator is used mainly
 to address the singularity issue during stress-constrained opti-
(rY ) are used in the macroscopic modified Hill’s yield criterion for 
each element: mization. The stiffness matrix (C h ), density (q), and failure index
(F re ) are updated using the material indicator variable:
pffiffiffiffiffiffiffiffiffiffiffiffiffi
 
ffi  
F e  1 ¼ 0; F e ¼ rT V r C h ðh; h1 ; h2 ; /Þ ¼ /q1 RðhÞ C ðh1 ; h2 ÞRT ðhÞ
 2  2  2  2
q ¼ /ð1  h1 h2 Þ ð7Þ
V11 ¼ Y1 þ 19 Y1 ; V22 ¼ Y1 þ 19 Y1 pffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
r11 r44 r22 r44  
   
2  
2   ! 2 2  2 F re ¼/ rT V r
q2

V12 ¼  12 1
Y þ 1
Y  Y
1
 29 1
Y ; V33 ¼ 12 1
Y where q1 and q2 are parameters that penalize the intermediate den-
r11 r22 r33 r44 r 12
sity. Based on previous research [8,23], q1 ¼ 3:0 and q2 ¼ 0:5 are
ð6Þ used in this work. The second challenge in stress-constrained topol-
The effective uniaxial, shear, and hydrostatic yield strengths ogy optimization is related to the large number of stress con-
Y Y Y Y Y straints. To address this issue, the element failure index can be
(r11 , r22 , and r12 are, r44 ) were found using Eq. (4). r33 is obtained
aggregated to a single constraint using the p-mean function. In
Y Y
from minðr11 ; r22 Þ. The yield criterion in Eq. (6) is developed in the order to preserve some of the local nature of the stress while avoid-
4
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 3. Lattice structure strain contours: (a) printed specimen, (b) DIC strain contour, (c) DIC image of one cell and (d) numerical homogenization strain contour for h1 ¼ 0:7L
subjected to the distributed load of 2:98 N=mm2 ; (e) DIC and (f) numerical homogenization strain contours of one cell with h1 ¼ 0:5L subjected to the distributed load of
7:44 N=mm2 ; (g) DIC and (h) numerical homogenization strain contours of one cell with h2 ¼ 0:3L subjected to the distributed load of 3:98 N=mm2 .

 Z 1p
ing excessive computational cost, a clustering approach is utilized 1  r p
[8,24]. In the clustering approach, elements in the design domain F Pm ¼ F e dX ð8Þ
Xm Xm
are sorted based on their failure indexes. Then, the sorted elements
are placed in N groups. The failure indexes of the elements in each
group are aggregated into a single value using a p-mean function: where m is the group number, Xm is the total volume of the ele-
ments in each group, and p is a tuning coefficient. Since the p-
5
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

40 35
h = 0.1L h = 0.1L
1 2
h = 0.3L h = 0.3L
1 30 2
h = 0.5L h = 0.5L
1 2
30
h = 0.7L 25 h = 0.7L
1 2
h = 0.9L
Effective Stress (MPa)

Effective Stress (MPa)

h = 0.9L
1 2
20
20
15

10
10

0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.05 0.1 0.15 0.2

Effective Strain (mm/mm) Effective Strain (mm/mm)

(a) (b)
Fig. 4. Stress vs. strain curves: (a) variation with h1 and (b) variation with h2 .

Table 2
Variation in effective properties with h1 .

h1 0.1 L 0.3 L 0.5 L 0.7 L 0.9 L

Young’s modulus variation (MPa)

Experimental 1105 ± 106.59 881 ± 30.85 728 ± 29.51 453 ± 11.09 178 ± 6.27
Numerical 1188 945 702 448 172
% Difference 7.24 6.97 3.63 1.00 3.70
Yield strength variation (MPa)
Experimental 14.93 ± 2.80 12.17 ± 1.18 9.3 ± 0.66 5.77 ± 0.81 1.55 ± 0.37
Numerical 14.5 11.05 8.27 5.33 1.8
% Difference 2.94 9.62 11.72 7.87 15.14

Table 3
Variation in effective properties with h2 .

h2 0.1 L 0.3 L 0.5 L 0.7 L 0.9 L

Young’s modulus variation (MPa)

Experimental 886 ± 33.16 700 ± 37.34 728 ± 29.51 642 ± 45.47 632 ± 51.25
Numerical 840 757 702 668 652
% Difference 5.34 7.79 3.63 4.00 3.16
Yield strength variation (MPa)
Experimental 11 ± 0.4 8.97 ± 0.15 9.30 ± 0.66 8.33 ± 0.29 8.43 ± 0.38
Numerical 9.71 8.35 8.27 8.15 8.1
% Difference 12.46 7.12 11.72 2.22 4.03

 k1
max F r
g m ¼ skm F Pm  1; skm ¼ ak  ek1 þ ð1  ak Þsm
k1
ð9Þ
P
Fm

where g m is the stress constraint for each group. The parameter ak is

selected based on the skm in two consecutive iterations; ak ¼ 0:5 if s
is oscillating, otherwise ak ¼ 1:0.

4. Topology and morphology design optimization algorithms

In each optimization iteration, the topology and morphology

optimizations are performed in two steps. The characteristic
parameters and material indicator variable are optimized in the
Fig. 5. The effect of filtering radius on maximum of micro-strain. first step and the optimized cell orientation is obtained in the sec-
ond step. Two topology optimization problems are considered in
this research: (1) minimizing the compliance (maximizing stiff-
mean function in Eq. (8) converges to a lower value of maxðF e Þ, a ness) subjected to the equilibrium equation and volume constraint
scaling factor (s) is implemented during each iteration to decrease (V g ), and (2) minimizing the volume subjected to the equilibrium
the difference between the p-mean function and the maxðF e Þ [8]: equation and stress constraints. For the compliance optimization
k k
problem, we start with initial variables (h1 ; h2 ; /k , and hk for
6
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 6. Response surfaces for the homogenized stiffness (MPa) and yield strength (MPa).

k ¼ 0) and perform the following steps during each optimization min SE=SE0 þ Ph =2P h0
ð11Þ
iteration: design variables  2p  h  2p

k k    

(a) Find C ðh1 ; h2 Þ. where SE is compliance SE ¼ eT C e , and SE0 is the compliance in
(b) Solve the elasticity problem, KU ¼ F, where K, U, and F are
the first iteration. V0 is the volume of the design domain, and V g is
the stiffness matrix, the displacement vector, and the force
the volume fraction constraint. Ph is a control function that penal-
vector, respectively.
k k izes sudden changes in orientation among neighboring elements
(c) Filter the design variables h1 ; h2 , and /k and Ph0 is the penalty function in the first iteration [25]:
(d) Obtain the sensitivity of the objective function and con-
f  
k k
straint with respect to the design variables h1 ; h2 , /k , and hk . Xne X
1 1
(e) Perform topology optimization (10) and update Ph ¼  cosð4he  4hi Þ ð12Þ
kþ1 kþ1 e¼1 i¼1
2 2
hn and / .
(f) Perform morphology optimization (11) and update hkþ1 .
where f is the number of elements connected to the element e. The
(g) The process continues (k = k þ 1) until convergence.
topology optimization problem presented in (10) is solved using the
min SE method of moving asymptotes [26]. The morphology optimization
R problem (11) is solved using the Broyden–Fletcher–Goldfarb–Shan
subjected to ðV10 Þ X qdX  V g  0 ð10Þ no (BFGS) algorithm from the open-source library NLopt [27]. The
design variables 0 < hn and /  1: n ¼ 1; 2 stress-constrained optimization steps are similar to those from
compliance optimization, but incorporate the following changes:
7
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

    
k k
V h1 ; h2 is also calculated in step (a) and topology optimization  tanh ðbgÞ þ tanh b / g
/¼ ; ð19Þ
with stress constraints (13) is performed in step (e). tanh ðbgÞ þ tanh ðbð1  gÞÞ
R
min V f ¼ ðV10 Þ X qdX
where b is the projection parameter used to control the intensity of
subjected to F pm 10 ð13Þ the projection. The parameter g specifies the inflection point and is
design variables 0 < hn and /  1: n ¼ 1; 2 set to g ¼ 0:5. It was previously shown that use of homogeneous
Neumann boundary conditions in the filtering formulation may
The sensitivity analyses for the compliance objective function,
cause several issues in the optimized design, such as forcing struc-
stress constraints, and volume fraction constraint are required in
tural members to be perpendicular to the boundary [31]. A bound-
the optimization process. The sensitivity of the volume fraction is:
ary padding approach is suggested to address this issue [31,32]. In
R
d
dh1e X v dX ¼ /h2e Xe the padding technique, the boundary of the design domain is
R extended by a width equal to the filter radius, except at the support
X v dX ¼ /h1e Xe ð14Þ
d
dh2e and load. This approach is adopted in the current research to
R
d
d/e X v dX ¼ ð1  h1e h2e ÞXe address the boundary effects of Helmholtz-type filtering.
The material properties, state variables, and objective and con-
where Xe is the area of each element. The sensitivities of the com- straint functions are obtained using the projected material indica-
pliance and p-mean stress measure are obtained using the adjoint 


method. The compliance objective function is self-adjoint and its tor variable / and filtered characteristic parameters hn . In order to
sensitivity is: retrieve the sensitivities of the objective and constraint functions
   (f ) with respect to the design variables / and hn , the following
dSE
dhne
¼ /q1 eT dh
dC
ne
e Xe ; chain rules are implemented [29,30]:
 
ðq1 1Þ T
dSE
d/e
¼ q1 / e C e Xe ; ð15Þ  

 @f h
q1  T  n
@f
e C eh Xe @hn
T
dSE
dhe
¼ 2/ dR
dhi @hn
¼ 
@hn

@hn ð20Þ
 
The sensitivity analysis of the p-mean stress measure is @f / @f @ / @ /
@/
¼  
obtained using Eq. (8): @ / @ / @/
0  1p

F Pm   Compliance-based and stress-constrained topology and mor-
dF Pm B   d C   r p1 /q2
¼ @/q1 e kTm R eþ Fe phology optimization are applied to an L-bracket problem. The
dhne dhne 2Xe Fe dimensions and boundary conditions for the L-shaped bracket
 !! problem, including the clamped top edge and distributed load
 dC   dV  applied to the right corner (F ¼ 1000 N=cm) are shown in Fig. 7.
 2rT V e þrT r Xe ð16Þ
dhne dhne The load is distributed over 0:1cm and a region with an area of
0:2  0:2 cm2 near the applied load is excluded from the design
0  1p 1 domain. The domain is discretized using 23,544 triangular ele-
P
dF Pm B   F m  r p1  
C ments. The filter radius is 0:2cm and the maximum value of the
¼ @q1 /q1 1 e kTm R C e þ Fe q2 /q2 1 F e AXe pffiffiffi
d/e 2Xe projection parameter is b ¼ rf = 3le [33], where le is the minimum
elemental edge length (le ¼ 0:033 cm) in this study. As suggested
where the adjoint variables are: in [33], b is updated every 100 iterations until it reaches the max-
imum value. The response surfaces (Fig. 6) are used to obtain
kTm K ¼ F am  
   ð17Þ homogenized properties C and rY during optimization. First, the
P R ð Þ  r p1 /q2  T  T
1p
F Pm
F am ¼ e2Xm Xe Xm Fe Fe
r V C R B dXe stress-constrained optimization is performed. Ten regional stress
constraints (m ¼ 10) and a p-mean parameter of p ¼ 10 are
The topology and morphology optimization framework are selected. The optimized volume fraction, compliance, and maxi-
developed using the open-source PDE solver FreeFem++ [28]. mum of the failure index are reported in Table 4. The convergence
P 1 -functions are utilized to discretize the displacements and histories for the volume fraction and maximum failure index are
adjoint variables and P0 -functions are implemented to discretize shown in Fig. 8. The optimized material distribution
all other variables, such as stress, strain, characteristic parameters,
the material indicator variable, and orientation. In order to
regularize the characteristic parameters and material indicator
variable, the Helmholtz-type filtering approach is adopted in this
research [29]:
 2   
r
 2pf ffiffi3 r2 hn þ hn ¼ hn on X @hn
¼ 0 on @ C ð18Þ
@C

where r f is the filter radius, X is the design domain, @ C is the


boundary of the design domain, and h n is the intermediate filtered

variable. P 1 -functions are used to discretize hn , which is then trans-
formed into a P0 -function to obtain the filtered characteristic
parameters. The same procedure is applied to obtain the filtered
material indicator variable (/). In addition, a smoothed Heaviside
projection [30] is applied to the material indicator variable to pro-
duce the 0/1 design variable: Fig. 7. Dimensions and boundary conditions for the L-shaped bracket test case.

8
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Table 4 ever, the mapping functions fail to produce the projected lattices
L-bracket optimized design volume fractions, compliances, and maximum of failure due to irregularities in the optimized orientation (see Fig. 9 (b)
index.
and (e)). In order to create a locally continuous vector field, we
Optimization Problem Vf SE (Ncm) maxðF re Þ adapted the vector field combing suggested in [35]. In this proce-
Stress-constrained optimization 33.6% 28 1.0 dure, we start with an element in the lower left corner, the element
Compliance-based optimization 33.6% 20 3.1 is added to the visited element vector (i), the adjacent element (e)
with minimum x2 (see Fig. 7) is selected, and the following function
is calculated for four possible frame orientations (j), including

  ð0; p; p=2; p=2Þ:
(q ¼ /ð1  h1 h 2 Þ), orientation, and failure index are shown in
Fig. 9 (a), (b), and (c), respectively. As shown in the figure, the X
Rje ¼ ð1  v i :v e Þ ð24Þ
material at the re-entrant corner is removed during the topology
i
optimization process, the failure index is uniformly distributed,
and the stress constraint is satisfied. Next, the compliance-based The best frame orientation occurs at minimum of Rje . The ele-
optimization problem with volume fraction constraints is consid- ment (e) is then added to the visited vector and the next adjacent
ered. The upper bound of the volume fraction is taken as the opti- element is considered. While following this procedure does not
mized volume fraction of the stress-constrained optimized design ensure that singularities are prevented, they occur only in the void
(V g ¼ 33:6%). Upon comparing the material distributions and fail- or solid region and not in the area with intermediate density for
ure indexes of the two optimization problems in Fig. 9, it is evident the test cases considered in this research. After combing procedure,
that the stress constraint is not satisfied in the compliance-based the mapping functions are obtained using Eq. (22) and the homog-
design because of the stress concentration at the right-angle corner enized design is projected using Eq. (21). The projected design for
of the design. The results presented in Table 4 show that lower the periodicity parameter K ¼ 0:065 is shown in Fig. 10 (a). Post-
compliance is obtained by using compliance optimization for the projection treatment must be implemented to ensure the manufac-
same volume fraction. turability of the optimized designs. Details of the post-processing
framework are discussed in our previous work [15]. A summary
5. Projection and post-treatment of lattice structures of the process is outlined here. The process consists of five steps:
1) implementing the minimum feature size; 2) enforcing the den-
The homogenized design presented in Fig. 9 is obtained based sity threshold; 3) imposing the density boundary; 4) removing
on the assumption of an infinitesimal length scale. The projection small holes; and 5) eliminating floating members and smoothing
of the homogenized design to a finite length scale is discussed in boundaries. Within the first step, the cell size is defined based on
this section. As shown in [15,34], a cellular solid can be repre- the modified periodicity (tf ¼ Kec [3]), which is plotted in
sented by Fourier series expansions. For the square cell with a rect- Fig. 10 (b). Cells smaller than 2hmin (hmin ¼ 0:3K is selected for
angular hole considered in this research, implementing two cosine the L-bracket projected designs) are identified and two conditions
terms is sufficient to represent the lattice structure [1]: are used to ensure that all features are greater than the minimum
  threshold (hmin ). Small cells (tf < 2hmin ) in the region q  0:5 are
2pw1 ðx1 ; x2 Þ removed (1  hn ¼ 0). The small cells that fail that condition are
XM ¼ fðx1 ; x2 Þ 2 Dj cos
K made solid (1  hn ¼ 1). The Heaviside filter is then applied to
 
2pw2 ðx1 ; x2 Þ the material indicator value / eliminating some of the floating
> cosðp/ð1  h1 ÞÞ [ cos
K members at low density regions. Fig. 11 (a) shows the resulting
> cosðp/ð1  h2 ÞÞg ð21Þ design. The second step modifies the thicknesses of the features
in the tf 2hmin region based on the condition that if q > qth (for
where K is the periodicity parameter and wn is the mapping func- a given density threshold qth ) and tf hn < hmin , then tf hn is set equal
tion used to project the homogenized design. The mapping func- to hmin . Otherwise, the member is removed. The thicker members
tions are obtained using the optimized orientation among the small cells at the center left of the design in Fig. 11
(v ¼ ½cosðhÞ; sinðhÞ) [3]: (b) compared to Fig. 11 (a) are visible. The boundaries of the pro-
rw1 ¼ ec v ð22Þ jected design might exhibit gaps for some periodicities. In the third
step, this issue is addressed, and the boundaries are preserved. The
where c is a dilatation factor used to adjust the lattice spacing [3]: density distribution is modified by setting regions with q > 0:7 to

 
 1 and superimposing them on the design in step 2. Consequently, a
rc ¼ ðr  v Þ v þ r  v v ð23Þ thick boundary is seen in Fig. 11 (c). In the fourth step, the small

regions are filled to guarantee manufacturability and improve
and v is a vector that is perpendicular to v . The mapping function structural performance. For this effect, the analog lattice is thresh-

w2 can also be obtained by using Eq. (22) and changing v to v . How- olded with hmin and superimposed on the design from the previous

Fig. 8. (a) Volume fraction and (b) maximum failure index convergence histories for the design shown in Fig. 9 (a).

9
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 9. Optimized stress-constrained design: (a) material distribution; (b) orientation; and (c) failure index (F re ). Optimized compliance-based design: (d) material
distribution; (e) orientation; and (f) failure index (F re ).

with a larger periodicity (K ¼ 0:13Þ is obtained and shown in

Fig. 12 (a). The post-processed projected compliance-based design
with K ¼ 0:065 is also shown in Fig. 12 (b). The post-processed
projected designs are analyzed using ANSYS for the same loading
and constraint conditions as the optimization model. The failure
index for the three designs is shown in Fig. 13 (a)-(c). The small
cells in the stress-constrained design for K ¼ 0:065 produce a more
uniform stress distribution than those in the design for K ¼ 0:13.
The maximum stresses of the projected stress-constrained designs
are distributed over the central boundary of the L-shape. In con-
trast, the compliance-based design has a high stress concentration
at the sharp middle corner that leads to premature yield. The
results demonstrate the effectiveness of stress-constrained
Fig. 10. The projected (a) stress-constrained design and (b) feature size for
K ¼ 0:065. optimization.
The compliance and failure indexes are reported in Table 5.
Upon comparing Table 5 to Table 4, it is observed that post-
step (Fig. 11 (c)). The complement of the superposition is parti- treatment decreases the compliance by 29% and 25% for the
tioned into regions. Regions for q > qth with inscribed circles smal- stress-constrained and compliance-based designs, respectively.
ler than hmin of the diameter are filled. This produces the shape This difference is explained by the 17% and 26% increases in the
shown in Fig. 11 (d). In the last step, the closed holes inside the volume fractions of the stress-constrained and compliance-based
design are filled temporarily and their complement is divided into designs, respectively.
two regions: the solid and void. The shapes of these regions are
created using Delaunay triangulation. Adjusting the triangulation
factor produces smoother boundaries and exposes vertices inside 6. Experimental evaluation of lattice structures and solid
the boundaries of the regions. These vertices are added to the isotropic materials with penalization designs
regions, eliminating floating members. The complement of the
shape is taken and the interior holes remain unfilled (Fig. 11 (e)). The second test case considered in this research is the single-
Several iterations might be required to remove all of the floating edge notched bend problem. Figure 14 shows the dimensions
members. To produce a smoother shape, the boundary points cre- and boundary conditions of the SENB test case. The thickness of
ated using a non-uniform rational basis spline (NURBS) are added the structure is 20 mm. The notch causes stress concentration at
to the projected design. Finally, the projected design is imported its tip. The distributed load (F ¼ 240 N=mm) is applied over
into MeshLab and is further smoothed using Taubin Filtering with 6 mm in the middle of the top surface. A region with an area of
45 smoothing steps (Fig. 11 (f)). As will be shown later, the out- 12  6 mm2 near the applied load is excluded from the design
lined post-processing may increase the volume of the optimized domain. The domain is discretized using 62,580 elements. A length
structure. parameter of r f ¼ 3 mm is selected. As in the previous case, the
pffiffiffi
In addition to the stress-constrained design with K ¼ 0:065 maximum of the projection parameter is b ¼ rf = 3le with
(Fig. 11 (f)), the post-processed projected stress-constrained design le ¼ 0:04, and ten regional stress constraints and a p-mean param-
10
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 11. Projected design: (a) after implementing the minimum feature size; (b) after increasing the thicknesses of members where q > qth ; (c) after superposing the density
boundary; (d) after small regions have been filled; (e) without floating members; and (f) with smoothed boundaries.

Table 6. The reported effective specific stiffness (K eff ) is obtained

from K eff ¼ VF d, where d is the deflection at the top region where
f

the load (F) is applied. The post-processed lattice structures are

shown in Fig. 16 (a) and (c). The minimum manufacturable feature
size is hmin ¼ 0:8 mm. Some of the small holes in the post-
processed designs are filled and the walls thickened to comply
with the minimum manufacturable feature size. This leads to a
considerable increase in the volume fraction (Vg Þ as reported in
the projected design section of Table 6. The effective stiffnesses
and yield loads for the projected designs are also reported in
Table 6. As shown in the table, in the case of the compliance-
based design, the specific stiffnesses of the homogenized and pro-
Fig. 12. Post-processed projected designs produced via (a) stress-constrained jected designs differ by 15.6% because the projected model has a
optimization with K ¼ 0:13 and (b) compliance-based optimization with K ¼ 0:065. larger volume fraction that reduces the specific stiffness consider-
ably. The increased volume has a less adverse effect on the specific
stiffness of the stress-constrained design.
eter of p ¼ 10 are selected. The optimized material distributions for The test case is also optimized using a solid isotropic material
the compliance-based and stress-constrained designs are shown in with penalization (SIMP) method. In SIMP, the elastic modulus is
Fig. 15, and the corresponding volume fraction, specific stiffness, related to the density using a power law in order to penalize the
and yield load are reported in the homogenized design section of intermediate density and restrict the design space to obtain a

Fig. 13. Failure indexes for the three optimized designs: (a) a stress-constrained optimized design with K ¼ 0:065; (b) a stress-constrained optimized design with K ¼ 0:13;
and (c) a compliance-optimized design with K ¼ 0:065.

11
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Table 5 high stresses mostly at the bottom bulky region and the inner sides
The volume fraction, compliance, and failure index for post-treated L-bracket designs. of its legs. The compliance-based designs yield considerably sooner
Designs Vf SE (Ncm) maxðF e Þ than the stress-constrained designs due to high localized stresses
Stress-constrained, K ¼ 0:065 39.2 20 0.9
in the notch and at the inner sides of the legs (Fig. 16 (d) and
Stress-constrained, K ¼ 0:13 39.4 20 0.9 (h)). These results are expected since the stress constraint is not
Compliance-based, K ¼ 0:065 42.3 15 2.1 implemented in the compliance optimization approach. The
homogenization-based, compliance-based design also exhibits sig-
nificantly high stresses on its thin, outer members (Fig. 16 (d))
since they are quite thin and play an important role in transferring
the load to the supports.
Three specimens for each optimized configuration were addi-
tively manufactured using a PA12 printer with multi jet fusion
(MJF) technology. The models were tested experimentally in an
MTS system with a 50 kN load cell. Three modifications are intro-
duced to the designs to make them experimentally reliable (see
Fig. 17 (a), (c), (e), and (g). To allow more surface area for the roller
constraints and prevent the supports from sliding off the rollers,
the lower supports are extended. The first set of tests reveals that
the top of the model tends to slide off to the side (Fig. 18). To
address this rigid translation, a semi-circular crown is added such
Fig. 14. The SENB test case with dimensions in mm.
that the top roller can fit inside while remaining as centered as
possible throughout the test. The described modifications are
responsible for the increased volume fractions among the printed
solid-void design. Details of the SIMP method with stress con- designs in Table 6 and Table 7. Due to these alterations, Table 6
straint implementation are discussed in Ref. [8]. A compliance- and Table 7 distinguish between optimized and printed models.
based design and a stress-constrained design were generated using It is also observed that the support legs, especially those of the
the SIMP topology optimization. The SIMP designs are shown in stress-constrained designs, rotate too easily as the load increases.
Fig. 16 (e) and (g). The volume fractions, effective stiffnesses, and Thus, the models fail to reproduce the desired boundary condi-
yield loads of the SIMP designs are reported in Table 7. tions. A thick aluminum plate is placed between the parts and
Figure 16 shows the failure index distribution for a load of the support rollers, as shown in Fig. 17. However, the
1440 N. As in the optimization model, the region near the load homogenization-based, compliance-based design is tested without
application (blue rectangle) is neglected in the stress analysis. Con- the plate since the portion of its bottom region near the constraints
centrated stress is eliminated in the stress-constrained models, is low and would hit the plate at high loads. For these reasons, a
which exhibit more uniform stress distributions throughout the rotation was verified for this sample, as will be discussed later.
structure. The stresses on the SIMP stress-constrained design The displacements were measured using DIC. Symmetry was
(Fig. 16 (f)) are uniformly high throughout the structure, while assumed and the two DIC cameras were focused on one side of
the lattices in its homogenized counterpart (Fig. 16 (b)) seem to the structure to magnify the field of view and increase the resolu-
provide low stress and good load transfer from load application tion. This was especially important for lattice structures with smal-
to support through the solid regions. The latter design exhibits ler members. Figure 19 (a) shows the applied force per volume

Fig. 15. Optimized (a) stress-constrained and (b) compliance-based designs.

Table 6
Yield loads (F) and stiffnesses (K eff ) for homogenization-based designs.

Designs Optimized Design Printed Design

Homogenized design FEA for projected design % diff FEA Experim. % diff
Homogenization-based, stress-constrained design
Vg 22.9% 28.0% 30.4% 30.4%
FðNÞ 1440 1440 0.00 1350.00 1451.05 7.22
dðmmÞ 3.30 2.72 19.27 2.43 2.39 1.29
Keff ðN=mmÞ 1905.52 1890.76 0.78 1831.37 1993.95 8.50
Homogenization-based, compliance-based design
Vf 22.9% 28.5% 29.7% 29.7%
FðNÞ 360.00 370.18 2.79 420.00 415.00 1.20
dðmmÞ 0.63 0.60 3.40 0.72 0.78 6.95
Keff ðN=mmÞ 2515.28 2151.57 15.59 1949.77 1797.12 8.15

12
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 16. SENB-optimized designs and failure index distributions at 1440 N: (a) and (b), the homogenization-based stress-constrained design; (c) and (d), the homogenization-
based compliance-based design; (e) and (f), the SIMP stress-constrained design; and (g) and (h), the SIMP compliance-based design.

Table 7
Yield loads (F) and stiffnesses (K eff ) for SIMP designs.

Designs Optimized Design Printed Design

FEA FEA Experim. % diff
SIMP stress-constrained design
Vf 19.9% 22.1% 22.1%
FðNÞ 1440.00 1425.00 1360.00 4.67
dðmmÞ 3.52 3.35 3.21 4.26
Keff ðN=mmÞ 2056.77 1924.17 1916.31 0.41
SIMP compliance-based design
Vf 21.8% 24.6% 24.6%
FðNÞ 458.60 590.00 645.00 8.91
dðmmÞ 0.87 0.86 0.83 3.62
Keff ðN=mmÞ 2432.96 2786.93 3159.03 12.52

fraction versus the vertical displacement per unit of structure strained and rotates more than is predicted via FEA. Another rea-
height (H) for the four samples. The yield loads (F) for the experi- son for this discrepancy might be that bending of the real sample
mental evaluations reported in Table 6 and Table 7 are determined is not perfectly symmetric.
from the 0.2% offset line of the curve presented in Fig. 19 (a). As shown in the printed design computational and experimen-
The experimental and computational displacement contours at tal results in Table 6, the homogenization-based, stress-
yield are shown in Fig. 19 (b)-(i). The two sets of contours for each constrained design exhibits an enhanced yield load compared to
design are in good agreement and the computational model pre- the compliance-based design. Both the yield load and specific stiff-
dicts a slightly higher maximum displacement than is observed ness for the stress-constrained design are similar to those of the
experimentally. However, the homogenization-based, homogenized and projected designs. Overall, this design provides
compliance-based design seems to be an exception. For this model, a higher ultimate strength than the other three designs (Fig. 19
the experimental displacements are higher than the computational (a)). As the load increases, some of its thin walls deflect, buckle,
displacements. This might indicate that the real model is less con- and ultimately break (Fig. 17 (b)). The homogenization-based,

13
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

Fig. 17. Designs produced via additive manufacturing: the homogenization-based, stress-constrained design (a) before and (b) after testing; the homogenization-based,
compliance-based design (c) before and (d) after testing; the SIMP stress-constrained design (e) before and (f) after testing; and the SIMP compliance-based design (g) before
and (h) after testing.

no plate is used. FEA of the homogenization-based, compliance-

based design shows that the stiffness would be closer to that of
the SIMP compliance-based design (2678:5 N=mm instead of the
1949:8 N=mm shown in Table 6) if no rotation was allowed. This
demonstrates that similar specific stiffnesses can be achieved using
the SIMP and homogenization methods for the same boundary
conditions.
Upon comparing the printed design experimental and computa-
tional results reported in Table 6 and Table 7, it can be seen that
the stress-constrained lattice and SIMP designs exhibit similar
Fig. 18. A three-point bending test showing a sample sliding off the rollers. specific stiffnesses. Both the experimental (Fig. 17 (f)) and compu-
tational models of the SIMP stress-constrained design demonstrate
yield due to bending. The specific stiffnesses of the two models are
compliance-based design has the lowest stiffness and strength of in good agreement and differ by less than 1%. Experimentally, the
the four designs. It exhibits high plasticity and deforms consider- model deflects more on one side (Fig. 17 (f)) despite the effort
ably before some of its thin members buckle (Fig. 17 (d)). Although made to center it during test preparation. Comparison of the
the computational model exhibits some deformation among the printed and optimized models in Table 7 indicates similar specific
thin walls, it predicts yield due to a high stress concentration at stiffnesses and yield loads. This occurs because the supports are
the point where the bottom supports touch the rollers. This design thin and compliant and extending them for testing did not produce
has low stiffness because its supports rotate (see Fig. 17 (d)) since significant effects.

14
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

12000

10000

8000

(N)
f
6000

Force/V
4000
Homog Stress

2000 Homog Compliance

SIMP Stress
SIMP Compliance
0
0 0.05 0.1 0.15 0.2
/H

(a)

(b) (c)

(d) (e)

(f) (g)

(h) (i)

Fig. 19. Three-point bending results: (a) the force (per volume fraction) versus DIC vertical displacement (per unit of structure height) plot. Displacement contour plots [mm]
for the: (b) experimental and (c) computational homogenization-based, stress-constrained design; (d) experimental and (e) computational homogenization-based,
compliance-based design; (f) experimental and (g) computational SIMP stress-constrained design; and (h) experimental and (i) computational SIMP compliance-based
design.

15
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

The experimental and computational SIMP compliance-based structures and comparison of their performances to those reported
designs yield expectedly due to high stress concentrations at the in this article is the subject of future research. Future research also
notch as the load increases (Fig. 17 (h)). The bulky bottom supports includes extension of this framework to 3D lattice structure
provide more stability and better load bearing, resulting in better designs.
stiffness than other models, especially relative to the stress-
constrained designs in which the long, thin supports tend to Declaration of Competing Interest
buckle. The FEAs of the printed and optimized designs indicate that
the modifications made to the SIMP compliance-based design to The authors declare that they have no known competing finan-
make it testable affect its yield load and stiffness. The experimental cial interests or personal relationships that could have appeared
model is 13% stiffer than computationally predicted. This might to influence the work reported in this paper.
indicate that the boundary conditions applied in the FEA model
do not fully resemble the real boundary conditions. That is, the alu-
Acknowledgements
minum plate provides a much stiffer contact than predicted com-
putationally. The printed designs also have higher yield loads
The authors gratefully acknowledge the support of the National
since their bulky supports become even stronger when additional
Science Foundation (NSF) under award number 1847133 with pro-
material is added.
gram manager Dr. Kathryn Jablokow. The authors also thank Pro-
Overall, the experimental and computational models are in
fessor Alberto Mello for the help he provided for the
good agreement with the maximum 12.5% difference in the stiff-
experimental evaluations of the optimized design.
ness of the SIMP compliance-based design. This difference can be
attributed to the difficulty of applying perfectly symmetric con-
straints and loading conditions experimentally and the difficulty Data availability
of reproducing the experimental boundary conditions via FEA.
The data that support the findings of this study are available on
request from the corresponding author, AYT.
7. Conclusions
References
The implementation of stress constraints in topological and ori-
entation optimization of lattice structures was discussed in this [1] O. Pantz, K. Trabelsi, A post-treatment of the homogenization method for
shape optimization, SIAM J. Control Optim. 47 (3) (2008) 1380–1398.
article. The effective properties were calculated using the numeri-
[2] J.P. Groen, O. Sigmund, Homogenization-based topology optimization for high-
cal homogenization method. The results were validated experi- resolution manufacturable micro-structures, Int. J. Numer. Meth. Eng. (2017).
mentally and then utilized in the optimization algorithm. [3] G. Allaire, P. Geoffroy-Donders, O. Pantz, Topology optimization of modulated
Compliance-based and stress-constrained optimization algorithms and oriented periodic microstructures by the homogenization method,
Comput. Math. Appl. 78 (7) (2019) 2197–2229.
were developed and homogenized designs were obtained. These [4] P. Duysinx, M.P. Bendsøe, Topology optimization of continuum structures with
designs were projected to generate optimized lattice structures. local stress constraints, Int. J. Numer. Meth. Eng. 43 (8) (1998) 1453–1478.
An advantage of the de-homogenization methodology is that opti- [5] G. Rozvany, J. Sobieszczanski-Sobieski, New optimality criteria methods:
forcing uniqueness of the adjoint strains by corner-rounding at constraint
mization can be performed efficiently using a coarse mesh and intersections, Struct. Optimiz. 4 (3–4) (1992) 244–246.
high-resolution lattice structures can be obtained via projection [6] R. Yang, C. Chen, Stress-based topology optimization, Struct. Optimiz. 12 (2–3)
onto a fine mesh. Two case studies were considered to demon- (1996) 98–105.
[7] P. Duysinx, O. Sigmund, New developments in handling stress constraints in
strate the feasibility of the proposed method. For the first case, optimal material distribution, in: 7th AIAA/USAF/NASA/ISSMO symposium on
the L-shaped bracket, it was observed that the implementation of multidisciplinary analysis and optimization, 1998, p. 4906.
stress constraints during the optimization process resulted in [8] C. Le, J. Norato, T. Bruns, C. Ha, D. Tortorelli, Stress-based topology optimization
for continua, Struct. Multidiscip. Optim. 41 (4) (2010) 605–620.
removal of material at the re-entrant corner and a more uniform [9] E. Holmberg, B. Torstenfelt, A. Klarbring, Stress constrained topology
failure index distribution. For the SENB test case, the optimization, Struct. Multidiscip. Optim. 48 (1) (2013) 33–47, https://doi.
homogenization-based stress-constrained and compliance-based org/10.1007/s00158-012-0880-7.
[10] L. Cheng, J. Bai, A.C. To, Functionally graded lattice structure topology
designs were compared to those obtained using SIMP. The two
optimization for the design of additive manufactured components with
stress-constrained designs exhibited similar structural perfor- stress constraints, Comput. Meth. Appl. Mech. Eng. 344 (2019) 334–359,
mance; the two designs had similar stiffnesses and yield strength. https://doi.org/10.1016/j.cma.2018.10.010.
Comparison to the compliance-optimized design indicated that [11] H. Yu, J. Huang, B. Zou, W. Shao, J. Liu, Stress-constrained shell-lattice infill
structural optimisation for additive manufacturing, Virt. Phys. Prototyp. 15 (1)
stress-constrained optimization is effective in inhibiting yield but (2020) 35–48.
compromises stiffness. Experimental evaluation of the optimized [12] P.G. Donders, Homogenization method for topology optmization of struc-tures
designs showed that the additively manufactured models could built with lattice materials, 2018.
[13] K. Gharibi, A.Y. Tamijani, Load-Path-Based Topology Optimization of Two-
reach the yield strength determined via the optimization algo- Dimensional Continuum Structures, AIAA J. (2021) 1–10.
rithm. However, the experimental results also demonstrated the [14] A.Y. Tamijani, K. Gharibi, M.H. Kobayashi, R.M. Kolonay, Load paths
adverse effects of enforcing a minimum manufacturable feature visualization in plane elasticity using load function method, Int. J. Solids
Struct. 135 (2018) 99–109, https://doi.org/10.1016/j.ijsolstr.2017.11.013.
size. This increases the volume fraction and thus decreases the [15] A.Y. Tamijani, S.P. Velasco, L. Alacoque, Topological and morphological Design
specific stiffnesses of the lattice structures. of Additively-Manufacturable Spatially-Varying Periodic Cellular Solids, Mater.
Both the computational and experimental results indicate the Des. (2020) 109155.
[16] J. Guedes, N. Kikuchi, Preprocessing and postprocessing for materials based on
importance of including buckling constraints in future studies. As the homogenization method with adaptive finite element methods, Comput.
shown in the SENB compliance-based design, thin members are Meth. Appl. Mech. Eng. 83 (2) (1990) 143–198.
susceptible to buckling. Depending on the minimum member [17] B. Hassani, E. Hinton, A review of homogenization and topology optimization
I—homogenization theory for media with periodic structure, Comput. Struct.
thickness, this may occur before the yield stress is reached.
69 (6) (1998) 707–717.
Another important observation is that while we did not intend to [18] E. Andreassen, C.S. Andreasen, How to determine composite material
obtain coated structures with infill lattices, the optimized designs properties using numerical homogenization, Comput. Mater. Sci. 83 (2014)
resemble them due to the use of a minimum manufacturable fea- 488–495.
[19] S.J. Hollister, N. Kikuchi, A comparison of homogenization and standard
ture size and a material indicator variable to address stress singu- mechanics analyses for periodic porous composites, Comput. Mech. 10 (2)
larities. The implementation of stress constraints to obtain coated (1992) 73–95.

16
R.R. Fernandes and A.Y. Tamijani Materials & Design 210 (2021) 110026

[20] J.K. Guest, J.H. Prévost, T. Belytschko, Achieving minimum length scale in [27] S.G. Johnson, The NLopt nonlinear-optimization package. http://github.com/
topology optimization using nodal design variables and projection functions, stevengj/nlopt.
Int. J. Numer. Meth. Eng. 61 (2) (2004) 238–254. [28] F. Hecht, O. Pironneau, freefem+, 2013.
[21] V.S. Deshpande, N.A. Fleck, M.F. Ashby, Effective properties of the octet-truss [29] B.S. Lazarov, O. Sigmund, Filters in topology optimization based on Helmholtz-
lattice material, J. Mech. Phys. Solids 49 (8) (2001) 1747–1769. type differential equations, Int. J. Numer. Meth. Eng. 86 (6) (2011) 765–781.
[22] G. Sved, Z. Ginos, Structural optimization under multiple loading, Int. J. Mech. [30] F. Wang, B.S. Lazarov, O. Sigmund, On projection methods, convergence and
Sci. 10 (10) (1968) 803–805, https://doi.org/10.1016/0020-7403(68)90021-0. robust formulations in topology optimization, Struct. Multidiscip. Optim. 43
[23] J.P. Groen, J. Wu, O. Sigmund, Homogenization-based stiffness optimization (6) (2011) 767–784.
and projection of 2D coated structures with orthotropic infill, Comput. Meth. [31] A. Clausen, E. Andreassen, On filter boundary conditions in topology
Appl. Mech. Eng. 349 (2019) 722–742, https://doi.org/10.1016/j. optimization, Struct. Multidiscip. Optim. 56 (5) (2017) 1147–1155, https://
cma.2019.02.031. doi.org/10.1007/s00158-017-1709-1.
[24] L. Alacoque, R.T. Watkins, A.Y. Tamijani, Stress-based and robust topology [32] B.S. Lazarov, F. Wang, O. Sigmund, Length scale and manufacturability in
optimization for thermoelastic multi-material periodic microstructures, density-based topology optimization, Arch. Appl. Mech. 86 (1–2) (2016) 189–
Comput. Meth. Appl. Mech. Eng. 379 (2021), https://doi.org/10.1016/j. 218, https://doi.org/10.1007/s00419-015-1106-4.
cma.2021.113749 113749. [33] G.A. Da Silva, N. Aage, A.T. Beck, O. Sigmund, Large scale three-dimensional
[25] F.C. Stutz, J.P. Groen, O. Sigmund, J.A. Bærentzen, Singularity aware de- manufacturing tolerant stress-constrained topology optimization, Int. J.
homogenization for high-resolution topology optimized structures, Struct. Numer. Meth. Eng. 122 (2) (2021) 548–578, https://doi.org/10.1002/
Multidiscip. Optim. 62 (5) (2020) 2279–2295, https://doi.org/10.1007/s00158- nme.6548.
020-02681-6. [34] R.C. Rumpf, J. Pazos, Synthesis of spatially variant lattices, Opt. Express 20 (14)
[26] K. Svanberg, The method of moving asymptotes—a new method for structural (2012) 15263–15274.
optimization, Int. J. Numer. Meth. Eng. 24 (2) (1987) 359–373. [35] J. Groen, F. Stutz, N. Aage, J.A. Bærentzen, O. Sigmund, De-homogenization of
optimal multi-scale 3D topologies, arXiv preprint arXiv:1910.13002, 2019.

17