Lattice Structure Design and Optimization With Additive Manufacturing Constraints

1546 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 15, NO.
4, OCTOBER 2018
Lattice Structure Design and Optimization With

Additive Manufacturing Constraints
Yunlong Tang, Guoying Dong, Qinxue Zhou, and Yaoyao Fiona Zhao
Abstract— Lattice structures with different desired physical I. I NTRODUCTION

properties are promising for a broad spectrum of applications.
The availability of additive manufacturing (AM) technology has
relaxed the fabricating limitation of lattice structures. However,
manufacturing constraints still exist for AM-fabricated lattice
T HE word “lattice” originally describes a framework or
structure of crossed woods or metal strips and can be
extended to regular geometrical arrangements of points or
structures, which have a significant influence on the printing objects over an area and space [1]. In this paper, the lattice
quality and mechanical properties of lattice struts. In this structure is defined as a mesoscale (0.1–10 mm) or
paper, a design and optimization strategy is proposed for lattice
structures with the consideration of manufacturability to ensure microscale (<0.1 mm) truss-like structure which consists
desired printing quality. The concept of manufacturable element of interconnected struts and nodes with a certain repeated
is used to link the design and manufacturing process. A meta- arrangement in 3-D space. In this sense, lattice structures can
model is constructed by experiments and the artificial neural net- be regarded as a type of cellular structure on a mesoscale.
work to obtain the manufacturing constraints. Sizes of struts are Compared to other types of cellular structures including foams
optimized by a bidirectional evolutionary structural optimization-
based algorithm with these manufacturing constraints. An arm of and honeycombs, lattice structures are more flexible to achieve
quadcopter is redesigned and optimized to validate the proposed a wide range of different desired physical properties, such
method. Its result shows that optimized heterogeneous lattice as high stiffness-weight ratio [2], low thermal expansion
structures can improve the stiffness of the model compared coefficient [3], negative Poisson ratio [4], and high heat
to the homogeneous lattice structure and the original design. dissipation rate through active cooling [5]. Moreover, it can
Both the Von-Mises stress and the maximum displacement are
reduced without increasing the weight of designed part. And also be designed as a bioimplant to enhance the ossesoin-
by considering the manufacturability constraints, the optimized tegration as well as alleviating stress-shielding effect. Due
design has been successfully fabricated by the selected additive to its outstanding performance, lattice structures have been
manufacturing process. used in a broad spectrum of applications, including bone and
dental implants [6]–[8], ultralight structures [9], [10], energy
Note to Practitioners—Lattice structures might fail to be fab-
ricated by the additive manufacturing technique if the designed absorbers [11], low thermal expansion structures [12], and
model exceeds the processability of the machine. Our approach conformal cooling [13].
has the capability of considering the manufacturing constraints in Traditionally, lattice structures can be fabricated via casting,
the design and optimization process. We conducted experiments sheet metal forming, or wire bonding processes [14]. However,
to investigate the manufacturability and proposed a method manufacturing constraints of these processes severely restrict
that can give the domain of the design variables for a selected
the complexity of designed lattice structures. These processes
manufacturing process. And we also designed an algorithm that
can optimize the lattice structure inside the domain of design can only be applied to fabricate lattice structures with few
variables. It ensures that the lattice model can be successfully simple unit cell topologies in a regular shape on a macroscale.
fabricated by the selected process and the performance is This manufacturing limitation has been relaxed by using
dramatically increased compared to the original design. Engi- additive manufacturing (AM) to fabricate lattice structures.
neers can use our approach to optimize the lattice structure
By fabricating a part layer by layer, AM enables the design
automatically without knowing the knowledge of optimization
and manufacturability. of complex structures without significantly increasing the
cost. Thus, the geometrical freedom provided by AM greatly
Index Terms— Additive manufacturing (AM), design, lattice
structure, manufacturing constraints.
enlarges the design space of lattice structures. Lattice struc-
tures with multiscale complexities can be easily fabricated for
Manuscript received October 4, 2016; revised December 8, 2016; accepted a better functional performance. For instance, on a macroscale,
January 23, 2017. Date of publication April 27, 2017; date of current version conformal lattice structures [15] can be produced to fit a
October 4, 2018. This paper was recommended for publication by Associate complex macroshape with a relatively smooth surface bound-
Editor C. C. L. Wang and Editor Y. Chen upon evaluation of the reviewers’
comments. This work was supported in part by the National Sciences ary. On a mesoscale or microscale, complex lattice unit cells
and Engineering Research Council of Canada Discovery under have been designed and fabricated to achieve a given gradient
Grant RGPIN 436055-2013, in part by the China Scholarship Council of elastic properties [16]. However, it should also be noted
under Grant 201306020032, and in part by the Rio Tinto-Richard Evans
Fellowship in Engineering. (Corresponding author: Yaoyao Fiona Zhao.) that every manufacturing process has limitations. AM is no
The authors are with the Mechanical Engineering Department, McGill Uni- exception. In particular, some recent research has observed and
versity, Montreal, QC H3A 0G4, Canada (e-mail: yaoyao.zhao@mcgill.ca). studied on the manufacturing constraints of lattice structures
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. fabricated by different AM processes. The manufacturability
Digital Object Identifier 10.1109/TASE.2017.2685643 of gyroid lattice structures made of 316L stainless steel
1545-5955 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
TANG et al.: LATTICE STRUCTURE DESIGN AND OPTIMIZATION WITH AM CONSTRAINTS 1547
this lattice structure can achieve a certain elastic properties

gradient along the x-axis.
For homogeneous lattice structure, most of research is
focusing on its unit cell. Topology optimization and homog-
enization techniques can be applied to generate the lattice
Fig. 1. Comparison between homogeneous lattice and heterogeneous unit cell which can achieve the desired properties. In order to
lattice [9]. (a) Homogeneous lattice. (b) Heterogeneous lattice. achieve the fabricatable unit cell, the manufacturing constraints
such as the minimum member size [21] can be considered in
the optimization procedure of microcell structure. To further
with selective laser melting (SLM) process has been studied improve the functional performance of homogeneous lattice
in [17]. In this research, the lattice structures with unit cell structures, some concurrent optimization methods [22]–[24]
sizes ranging from 2 to 8 mm with 15% volume fraction have been developed to update the cell topology as well
can be successfully manufactured. Based on the fabrication as the macroshape of lattice simultaneously. Compared to
experiments, it shows that the minimum unit cell size which other methods only focusing on single design scale, these
can be fabricated mainly depends on particle sizes of metal concurrent methods generally show a better performance on
powders, since the trapped powders with a big particle size in multifunctional applications [22].
a small unit cell are difficult to be removed. Moreover, it also Compared to homogeneous lattice structures, heterogeneous
concludes that the unit cell whose size is larger than 8 mm lattice structures generally provide more design freedom.
is problematic to be fabricated due to the longer distance Designers can control the distribution of some lattice parame-
of the overhang area. Similar research has been done by ters to achieve a better functional performance. To achieve this
Santorinaios et al. [18] on the lattice cells with vertical and goal, several design and optimization methods for this type of
cross struts fabricated by SLM process. This research also lattice structures are proposed. A general framework for lattice
shows that the lattice structures with large cell size (larger structures fabricated by AM processes is proposed in [10]. This
than 5 mm) are difficult to be fabricated. Since their struts tend general framework can be divided into two stages. In the first
to “slag” during the manufacturing process. To further eval- stage, size optimization is applied to optimize the size of lattice
uate the manufacturability of lattice structures with different struts for a better mechanical performance. In the optimization
geometrical configurations, a manufacturability table is used procedure, the size of lattice struts is constrained in a prede-
in [19]. In this research, two different topologies of lattice fined range based on the manufacturability of the machine.
cells, gyroid and diamond, are fabricated by SLM process Then in the second stage, another optimization process is
with Ti6Al4V powders. The building conditions of each type applied to obtain the optimal process parameters for each strut
of cells with different cell sizes and volumes are recorded. based on the proposed concept manufacturable element (ME).
The result shows the lattice structures with a low volume Besides the size optimization, several density-based topology
fraction and 4–5 mm cell sizes mostly failed. Instead of the optimization methods [25]–[30] have also been applied to
study on some lattice topologies, design rules are provided by design the heterogeneous lattice or cellular structures for a bet-
Wang et al. [20] to achieve a better manufacturability of 316L ter mechanical performance. Among them, different types of
lattice structures fabricated by SLM process. Several set of material interpolation functions are used to establish the rela-
benchmark parts have been designed to individually evaluate tionship between the properties and parameters of mesoscale
the effects of inclined angles and sizes of minimum struts structures. Based on these interpolation functions, a topology
on the manufacturability of lattice structures [20]. Based on optimization problem is formalized and solved to get the opti-
these benchmarks, a design guideline shows that the struts mal relative density distribution. Then, the mapping functions
with inclined angles smaller than 35° are hard to fabricate. are further used to convert the result of topology optimization
It also indicates that the minimum fabricatable size of struts to the distribution of the lattice relative density. In the mapping
is 0.15 mm. functions of some proposed methods, like Rosen’s method,
To help designers consider both freedom and constraints of predefined manufacturing constraints are considered. In alter-
AM processes for lattice structures, several design methods for native of density-based optimization methods, bidirectional
lattice structures have been proposed. Generally, these design evolutionary structural optimization (BESO) method [9] has
methods can be divided into two groups: the design of homo- also been modified and applied to design heterogeneous lattice
geneous lattice structures and the design of heterogeneous structures. This method tries to simulate the bone remolding
lattice structures. As it is shown in Fig. 1(b), a homogeneous process by moving materials from low stress struts to high
lattice structure consists of unit cells with exactly the same stress struts. Compared to density-based optimization methods,
shape periodically distributed in the design space. Thus, this BESO-based method can easily deal with those stress-related
type of lattice structures can be considered as homogeneous constraints during the optimization process. To further reduce
materials on a macroscale during the design process. Com- the size of design variables, a heuristic optimization method
pared to the shapes and sizes of cells in homogeneous lattice has been developed by Nguyen et al. [31]. This method gener-
structures, those of cells in heterogeneous lattice structures ally converts the original problem into the optimization process
may vary from point to point to achieve a certain type of with two design variables: maximum and minimum sizes of
functional gradient. A typical example of heterogeneous lattice struts. Then, size matching and scaling method have been
structures is shown in Fig. 1(a). By varying thickness of struts, applied to determine the size of each strut in the design domain
1548 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 15, NO. 4, OCTOBER 2018
with the consideration of local stress states. Like other methods

mentioned here, predefined upper and lower bounds of struts’
size are considered as constraints from manufacturing.
In summary of those existing design and optimization meth-
ods of lattice structures for AM, it is notable that the functional
performance of parts can be further improved by using hetero-
geneous lattice structures. Moreover, as to the manufacturabil-
ity of AM-fabricated lattice structures, most design methods
only simply consider the constraints on the thickness of struts.
However, other geometrical parameters such as the inclined
angle [20] and the horizontal overhang [17] may also affect the
fabrication quality and manufacturability of lattice structures, Fig. 2. Example of FSs and FVs [35].
and the effects of those parameters are not totally independent.
Even though some design rules [20] have been provided, most of the information of the design boundary for the detailed
of them failed to consider the interaction between different design stage. It can be considered as the input for the proposed
geometrical parameters. Thus, it is difficult to directly integrate design method. This input can be obtained from the functional
those rules into a general design and optimization process of design step in the conceptual stage [33] or directly extracted
lattice structures. from the existing parts. Generally, a physical entity should
To solve the mentioned issues, a design and optimiza- include two types of information: geometrical information and
tion method of lattice structures is proposed in this paper material information. Material information includes a set of
with the consideration of manufacturability constraints of feasible materials which can achieve the defined functions of
AM processes. In this method, a concept called ME is used a physical entity. In most cases, the shape of physical entity
to link the design and manufacturing process. To predict is only defined on a macroscale. As to the information related
the quality and manufacturability of each ME, the artificial to mesostructure or microstructure of physical entity, they are
neural network (ANN) is established as a meta-model for the also considered as the material information in this paper.
selected AM process for a certain type of material. Based Besides the material information, concepts of functional
on the predicted value, the manufacturing constraint of each surfaces (FSs) and functional volumes (FVs) [34] are used
ME can be obtained. Then, obtained constraints will be used to represent the geometrical information of a physical entity.
in the BESO-based lattice optimization process to get the In this paper, an FS is referred to the key surface of a physical
optimal size distribution of lattice struts. In this paper, fused entity to realize the requested functions, such as assembly
deposition modeling (FDM) process, one of the most widely surfaces or loading surfaces. An FV is a geometrical volume
used AM processes for thermal plastic polymers, is used as an of a physical entity which is used to link the surfaces and assist
example to illustrate the proposed method. However, it should FSs to achieve the required functions. A typical example of
be noted that this methodology can be further applied to FSs and FVs is provided and shown in Fig. 2. In this figure,
other different types of AM processes with different materials. those green-tagged surfaces represent FSs, while the FV is
Moreover, compared to existing design methods, the proposed shown in the gray color to connect those FSs. The FSs of this
method can automatically consider the effects of different example are designed for three different purposes. The bottom
geometrical parameters on the manufacturability of lattice two surfaces are used to provide a vertical support. Another
structures in the optimization procedure. Thus, it can guarantee two cylindrical surfaces with connected surfaces on their top
the processablity of designed lattice structures. are designed for the assembly of bolts. The cylindrical hole on
In order to clearly describe the proposed method, this paper the top of the part is used to hold a bearing. It should be noted
is organized as follows. In Section II, several basic concepts that in some design cases, the FSs or FVs for a given physical
which are used in the proposed method are presented first. entity after the conceptual design stage is not totally fixed.
Based on these concepts, the proposed design method and its For these design cases, the parametric representation for FSs
related meta-model of the manufacturing process are carefully and FVs is needed. The detailed information of parametric
explained in Section III. In Section IV, a case study is used representation is beyond the scope of this paper. A detailed
to further illustrate and validate the proposed method, and discussion of parametric representation of FV can be found
a discussion on the design results under different types of in [33].
constraints is provided. At the end, this paper is wrapped up
with conclusions and future research directions. B. Lattice Unit Cell Model
II. BASIC C ONCEPTS To represent the topology of a repeating element inside the
lattice structures, lattice unit cell model is proposed. In this
A. Physical Entity model, the node of a lattice cell is defined in a unit cubic space
To represent the design space of lattice structures, the con- shown in Fig. 3, which can be represented by a 3-D tuple p
cept named as physical entity is used in this paper. According
p = (e1 , e2 , e3 ) ∈ P ⊆ [0, 1]3 (1)
to the previous research [32], physical entity is defined as a
concrete entity which is implemented for physical behaviors where ei is the i th component of the tuple and P is the set
required by given functions. Physical entity mainly consists which contains all the nodes of a lattice unit cell. To represent
Fig. 3. Unit cubic defined in a parametric space.
Fig. 5. General design flow of proposed design method.
to represent the midline of a strut, while the cross section

can be defined by the shape and its related parameters, such
as the diameter of a circle or the size of a square. Besides
geometrical data, the material data are another very important
factor. To clearly describe material compositions of fabricated
Fig. 4. Graphical view of data structure of ME for lattice structures.
lattice struts, both the material type and its grade should be
each strut of lattice unit cell, the node pair l is used and included in the data model of ME. The last key factor for
expressed as ME is the process data. It should be noted that different
AM machines may have different manufacturing capabilities,
l = { pi , p j }, pi p j ∈P. (2) even though they belong to the same type of AM technologies.
Based on the definition of lattice nodes and struts, the topology Thus, as to process data of ME, a certain AM machine should
of lattice unit cell can be expressed as a unidirectional graph G be determined first. Then, based on the selected machine, its
related fabrication strategy and fabrication parameters can be
G = {P, S} (3) decided. Among these three types of data, the geometrical
where S is the set that contains all the struts inside the lattice data are the primary concern of the proposed design method
unit cell. According to different design requirements, different in this paper. Thus, they have been considered as design
lattice unit cell topologies are selected by designers. Based variables during the optimization process. As to material
on the defined lattice unit cell model, lattice frame which data and process data, they can both be decided during the
represents the topological skeleton of lattice structures can be early conceptual design stage based on material selection and
built. This process is discussed in Section III-C. process selection strategy. Thus, these two types of information
are regarded as the input of the proposed design method.
C. Manufacturable Element of Lattice Structures
III. M ETHODOLOGY
To link the design and manufacturing process, a concept
called ME is used in this paper. This concept is originally A. General Design Flow
defined in [10] as a predefined, parametrized decomposition The general design flow of the proposed design method is
of a volumetric region of a part. Based on this original concept shown in Fig. 5. As it is shown in this figure, four types
and the characteristics of lattice structures focused in this of data are considered as the input of the proposed design
paper, an ME of lattice structure is defined as a lattice strut method. Among them, functional and design requirements are
with its related geometry, material, and process information. the most important factors, since it should be determined
To parametrically represent each ME of lattice structures in the at the beginning of the whole design process. This type
proposed design method, a data structure of ME is proposed of data describes the functional behaviors and the related
and its graphical view is shown in Fig. 4. requirements of a designed part or product. Based on the
It is clear that ME of lattice structures consists of three functional and design requirements, other three types of data
types of data. They are geometrical data, material data, and can be obtained from the conceptual stage. The general goal
process data. The geometrical data of ME can be further of this proposed method is to generate a manufacturable lattice
divided into two categories: the line segment and the cross- structure with a better performance under given functional
sectional shape. A line segment consists of a pair of nodes and design requirements. To achieve this purpose, the general
design flow of the proposed design method is given and

shown in Fig. 5. Its entire process can be generally divided
into four steps which are discussed in the following sections,
respectively. Besides these four major steps, an ANN-based
meta-model of selected AM process is built and applied to
obtain the manufacturing constraints for each ME during the
design and optimization process. Thus, this model and its
related experiments are also going to be explained in detail.
B. Initial Design Fig. 6. Comparison between: (a) uniform lattice and (b) conformal lattice [9].
In the initial design step, some macroscale design parame-
ters such as the weight or volume target for each FV should be design optimization process. Generally speaking, the volume
determined. These parameters will be regarded as the design constraints generated in the initial design stage provide a
constraints in the following optimization step. To achieve this foundation for the following optimization process.
purpose, the density-based topology optimization is used in
the initial design stage. Specifically, for a design problem of C. Lattice Frame Generation
a minimum compliance with a constrained volume, the math- In the second step of the proposed design method, lattice
ematical formulation of density-based topology optimization frame which represents the topological skeleton of lattice
can be expressed as structures is generated for the given FVs. To achieve this
Find : x ∈R n , x = (x 1 , x 2 , . . . ,x n ) objective, several parameters including type of lattice frame
Minimize: c = f T u and size of lattice cell need to be determined. Generally,
lattice structures can be divided into two different types
s.t. : K (x)u = f
according to its frame configuration. They are uniform lattice
xi vi ≤ V structures (also known as periodic lattice structures) and
0 <ε<x i ≤ 1 (4) conformal lattice structures. A comparison between these two
types of lattice frames is given in Fig. 6. As it is shown
where x represents an n-D vector which contains the relative in Fig. 6, the lattice frame of uniform lattice structures
density of each element inside the FVs of physical entity; c is consists a periodically distributed lattice cell. Thus, each cell
the structural compliance which can be calculated from the in the uniform lattice has the same size, shape, and topology.
global external force vector f and a global displacement vec- As to the conformal lattice, the size and shape of lattice
tor u; K (x) is the global stiffness matrix; v i is the volume for cell’s frame may vary to adapt to the macroshape of design
i th element, and V is the upper limit of the total volume; ε is boundary. Compared to uniform lattice structures, conformal
a small positive value which is slightly larger than 0 to avoid lattice structures can guarantee the integrity of cells on the
the singularity of the stiffness matrix during the optimization boundary of an FV. Thus, it can avoid some poorly connected
process. In this paper, ε is selected as 0.001. To solve the struts located on the boundary of FV which are shown
topology optimization problem defined in (1), several methods in Fig. 6(a). However, it should be noted that to generate
including solid isotropic material with penalization (SIMP) conformal lattices for an FV with a complex geometry is
[36] can be applied. Based on these methods, the optimal not an easy task. Even though there is some research on
relative density distribution can be obtained. This distribution conformal lattice generation [32], [37], most of them are
can be used as a reference to determine the volume target for limited to simple or regular shape. Moreover, the property
each FV. Particularly, for FV j , its volume target VFV j can be of uniform lattices is comparatively easy to be controlled by
calculated based on the following equation: carefully choosing the right cell size and topology. To consider
both advantages and disadvantages of those two types of
VFV = xi vi (5)
lattice structures, a general guideline is provided here to assist
where x i is the volume of the i th element inside F V j , and designers in decision-making. Generally, if an FV is bounded
v i is its optimal relative density obtained from topology by the FSs which play roles as aesthetic or assembly purposes,
optimization. For those FVs with solid material, this volume then this FV is suggested to be filled with conformal lattice
constraint can be achieved by removing materials in the low structures. Otherwise, uniform lattice structures can be used.
relative density area like traditional topology optimization After the determination of lattice types, the topology of the
routine. As to those FVs filled with lattice structures, the cal- lattice unit cell also needs to be determined. For both uniform
culated volume constraint from (5) can be used to calculate and conformal lattice structures, the topology of a unit cell can
the struts’ thickness of homogeneous lattice structures which be selected from the cell library which is a database for unit
will be regarded as the initial design in the BESO-based cell models. In this database, different lattice cell topologies
lattice optimization method. Since BESO-based optimization are linked to different types of properties. For example, based
method used in this paper only moves materials from the low on the loading condition of lattice struts, lattice cells can be
stress region to the high stress region, it can guarantee that divided into two types: bending dominant lattice and stretching
the total volume of lattice structures unchanged during the dominant lattice [38]. In the structural design for a better
stiffness, stretching dominant lattice cell is preferred, since the is fixed. The decreasing of the cell size will significantly
strut of this type of lattice is only under an axial load, which increase the minimum achievable relative density. It limits
may exhibit better stiffness than its counterpart. However, the freedom of designers in the sebsequent optimization
in the design of compliant structures or energy absorption process. Moreover, it will also make the lattice structures
structures, bending dominant lattice structures are preferred. lose the porosity which leads some issues on the simulations;
To further assist designers in selecting a suitable lattice unit therefore, the value of ρl∗ defined in (6) is suggested to be
cell from the unit cell library, material selection strategy [39] equal or smaller than 0.3 for efficient optimization as well as
can also be used based on the material chart with effective simulation. It should also be noted that the change of cell size
properties of different cell topologies. inside the bounded region defined by the above two guidelines
The size of lattice structure is another parameter needs will lead slight change of the optimal functional performance.
to be considered in the current design stage. The size of In order to achieve the optimal performance, designers can
the uniform lattice structure can be precisely controlled by do the optimization with several different cell sizes and then
three independent parameters. They are sizes on X, Y, Z -axes select the cell size which can achieve the best performance.
of the cell coordinate system. As to conformal lattice struc- Based on those design parameters determined at the begin-
tures, their size is defined as a diameter of a circumscribed ning of this step, the lattice frame can be generated inside
sphere of bounding box of lattice cell. Unlike uniform lattice the design space. The general lattice frame generation method
structure, the size of conformal lattice structures cannot be is given in this paper. Its process can be further divided
accurately controlled, since both size or shape of lattice cells into two substeps. In the first substep, hexahedron primitives
may vary inside the FV. Thus, in this design step, an average can be generated inside the given FVs. For a uniform lattice
size of conformal lattice is used. To help designers select an structure, its hexahedron primitives should be exactly periodic.
appropriate cell size, two design guidelines are provided. Thus, the kernel generation method discussed in the previous
Guideline 1: The lattice cell size is suggested to be smaller research [9] can be used. As to conformal lattice, the vol-
than the minimum size of functional features. ume meshing technique can be used to generate hexahedron
Guideline 2: The lattice cell size l is suggested to satisfy primitives in FVs with regular geometry. For an FV volume
the following condition: with complex geometry, auxiliary volume [32] can be built
to convert a complex geometry into a simple and mapped
p(l, tmin ) < ρl∗ (6)
shape. Since both ways have been exhaustively discussed
where the ρl∗ is the upper bound of the minimum relative in [9] and [32], this paper will not repeat them again.
density that lattice structures can achieve; tmin is a vector In the second substep, the selected lattice cell topology can
which contains the minimum fabricatable thickness of each be populated into the generated primitives from the last sub-
strut in the lattice unit cell. Once the cell topology and process step. The generated hexahedron primitive can be represented
data are determined, the value of vector tmin is fixed. p(l, t) is by the positions of eight corners which are denoted as Ci j k ,
the function to calculate the relative density of a given lattice i, j, k = 0, 1. Once the positions of eight corners of hexahe-
cell. The form of this function varies between cells with dron primitive are determined, the positions of all the lattice
different topologies and struts’ cross-sectional shape. However, node inside this hexahedron primitive can be calculated by a
they all depend on two independent variables: cell size l and tri-linear interpolation function. For example, the position c p
a vector t which contains the thickness of each strut inside of node p inside the given hexahedron primitive Ci j k can be
the cell. For example, the function pcubic(l, t) which is used calculated by
to calculate the relative density of cubic cell with struts in
c j k = c1 j k e1 +c0 j k (1−e1 ) (8)
square cross-sectional shape can be expressed as
ck = c1k e2 +c0k (1−e2 ) (9)
t 2 t 3
pcubic(l, t) = 1 − 3 1 − +2 1− (7) c p = c1 e3 +c0 (1−e3 ) (10)
l l
In (7), all the struts are assumed to have the same thickness. where (e1 , e2 , e3 ) is the 3-D tuple represents the node p
Thus, in this equation, t is a scalar which represents the inside the lattice unit cell model discussed in Section II. After
uniform thickness of struts in a lattice unit cell. calculating the position of each node, the frame of lattice
These two guidelines generally provide two limitations on struts can be built based on unidirectional graph G by linking
the size of a lattice cell. Based on the first guideline, it can the related nodes. Finally, to trim those connected struts with
be concluded that the cell size cannot be too large. If it is the boundary of FV, the final frame of lattice structures can
larger than the minimum size of functional features, the lattice be obtained. To illustrate these two substeps of lattice frame
structures cannot be guaranteed to fully fill every corner of generation, an example of arc shape FV is given and shown
the design space. In the other word, there might be some void in Fig. 7.
regions in the design space when the first guideline is violated.
Designers will lose controllability on those regions in the sub- D. Manufacturable Element Construction
sequent optimization process. The second guideline provides To link the design and manufacturing process, a concept of
a lower bound of cell size based on the manufacturability ME is used in this paper. As mentioned in Section II-C, each
of selected AM process. Once the AM process and material ME mainly consists of three types of information. They are
are determined, the minimum fabricatable strut’s thickness geometrical data, material data and process data. Both material
all the possible sizes of a lattice strut is set as the size

parameter in this step. The manufacturing quality of this set
of size parameters will be evaluated based on the proposed
meta-model discussing in the following section. Based on the
predicted results, a subset of them which can satisfy the quality
requirement is chosen as the design space of BESO-based
lattice optimization algorithm discussed in Section III-F.
E. Meta-Model for AM Constraints
The objective of using meta-model in this research is to
obtain the manufacturing constraints for each ME during the
design and optimization process. To simplify the process of
getting meta-model, only the influence of the geometrical
parameters on the manufacturability is considered in this paper.
Therefore, the material and process parameters are set to be
constant in the whole process. Based on the minimum recog-
nition of the machine and the properties of lattice structure,
like porosity, a design domain of geometrical data of MEs is
obtained. This predefined domain is imported to the meta-
model, and by setting printing quality in the meta-model,
the design domain is adjusted to satisfy this printing quality.
The meta-model is expressed as
W = (U, q); W ⊂ U (11)
where U is the predefined design domain, q is the required
manufacturing quality, represents the function of the meta-
Fig. 7. Lattice frame generation. model, and W is the design domain with the consideration of
manufacturing constraints.
To obtain the meta-model of a selected AM process, several
and process data can be obtained as the input of the proposed
experiments are conducted to get the raw data of the printing
design method. Thus, in this section, the focus is on how to
quality at first. To obtain the relation between design parame-
build geometrical data based on the input lattice frame and
ters of lattice struts and its geometrical deviation, ANN model
physical entity.
trained by the obtained raw data is utilized in this paper. For
As shown in Fig. 4, the geometrical data of lattice ME can
different types of MEs, the geometrical data could be different.
be further divided into two types of data. They are line seg-
Multiple ANNs would be trained by different categories of
ments to represent the midline of the strut and its cross section.
raw data. Then, the meta-model of a selected AM process
It should be noted that the line segment of the lattice ME is
is constructed by the trained ANNs. In this paper, FDM
defined in the processing coordinate system rather than the
process is selected to fabricate samples for the experiment, and
design coordinate system where the lattice frame is generated.
the obtained meta-model is only valid for this manufacturing
Thus, to construct the line segment of ME, the transformation
process. However, the method of constructing the meta-model
matrix between the design coordinate system and the process-
can be implemented on other AM process.
ing coordinate system is needed. This transformation matrix
1) Design of Experiment:
can be easily calculated once the fabrication orientation is
a) Geometrical parameters of the lattice structure: In the
determined. To simplify the process, the FVs with lattice struc-
experiment, three types of geometrical data, the diameter D,
tures will be considered as a solid volume temporarily. Based
the length L, and the inclined angle θ are defined to build
on this simplification and summarization of existing research
the ME. D is the size of the cross section of the ME, L is
related to orientation determination, two general guidelines are
the length of the line segment of the ME, and θ is the angle
provided in this paper. These two general guidelines should be
between the line segment of the ME and the printing platform.
considered sequentially.
The formulation of L and θ are defined as
Guideline 1: To select the orientation which can guarantee
the geometrical quality of FSs. L = |− p s −−
→ →
pe | (12)
− −
Guideline 2: To select the orientation with minimal volume → →
− →
ps − pe · z

θ = 90° − cos−1 (13)
|−
→
p s −−
→
of support structures. pe |
Based on the processing orientation, a lattice strut can be
easily converted to a line segment for the geometrical data where ps and pe are the position of the start point and endpoint
of ME. As for cross-sectional data, the shape of the cross of the strut, z is the unit vector perpendicular to the printing
section can be predefined by users. But, it is impossible to platform.
finally decide the size of a lattice strut in current step. Thus, According to the different tool paths shown in Fig. 8,
instead of using one value as the size parameter, a set of lattice struts can be divided into three types, the horizontal
TABLE I
P ROCESS PARAMETERS
Fig. 10. Deflection of horizontal struts. (a) Deformed area. (b) Definition
of Df.
Fig. 8. Tool paths for different types of struts. (a) Horizontal strut.
(b) Inclined strut. (c) Vertical strut.
Fig. 11. Cantilever area of slanted struts.
2) Measurement and Result:

a) Criteria of manufacturability: To find the manufactur-
Fig. 9. Models of test samples in the experimental and geometrical data.
(a) Horizontal group. (b) Nonhorizontal group. ing constraints for lattice structures under a certain fabricating
process, proper criteria of manufacturability should be pre-
defined. There are several critical issues for horizontal struts
struts (θ = 0°), vertical struts (θ = 90°), and slanted
and slanted struts. For horizontal struts, while printing the first
struts (0° <θ < 90°). Because the tool path of vertical struts
layer of overhang struts, the filament is melt and extruded
and slanted struts are similar, vertical struts can be seen
by the nozzle at a high temperature, so it cannot be cooled
as a special case of slanted struts with 90° inclined angle.
immediately. Because of no support structure, the first layer
Therefore, the experiment contains two groups of samples:
tends to deform due to its gravity and the deposition force of
one is the horizontal group, and the other is the nonhorizontal
the second layer. Therefore, the thickness in the middle of the
group, as shown in Fig. 9.
strut is larger than its two ends, which is shown in Fig. 10(a).
After doing the preliminary test of the selected FDM
If the deflection of the bottom is too large, the geometry
machine, proper values of defined geometrical data are cho-
will be inconsistent with the design model. Consequently,
sen for the experiment to find the relationship between the
the deflection can be a criterion for horizontal struts. The
manufacturability and these data. For the horizontal group,
deflection, which is shown in Fig. 10(b), is defined as
L has six levels from 10 to 60 mm. For the nonhorizontal
group, θ has nine levels from 10° to 90°. And for both groups, D f = Tmax − Tmin (14)
D has five levels from 1 to 5 mm. It should be noted the
number of experimental groups is not always fixed. If the where D f is the deflection of the strut, and Tmax and Tmin are
trained ANN model is not accurate enough, more groups of the maximum and minimum thickness of the strut.
experiments can be added in those critical regions following For slanted samples, the difficulty is concentrated on the
the procedure as the initial set of experiments. These extra cantilever area of each new layer, which is shown in Fig. 11.
experiments can further refine the ANN model and make it With the inclined angle decreasing, the length of the cantilever
more accurate especially on those highly nonlinear regions. area L c is getting longer. It is encountering the same problem
b) Manufacturing process: This paper concentrates on as the first layer of the horizontal strut. L c is calculated by
the influence of manufacturing constraints on topology opti- L c = t/ sin θ (15)
mization under a certain fabricating condition. Therefore, man-
ufacturing parameters are constants during the whole printing where t is the thickness of each layer, θ is the inclined angle
process. FDM process is used to fabricate the MEs. The of the slanted strut.
STereoLithography (STL) file is sliced by Z-Suite software. L c will affect the thickness of the strut in certain directions.
And the FDM printer is Zortrax M200. The printing material ta and tb are defined as the thickness of the strut in two
is Z-ABS [40]. All the process parameters are summarized directions to illustrate the different influence of L c , as shown
in Table I. The printing speed of M200 only has two levels in Fig. 12. The thickness in certain direction is significantly
without a specific value, and normal level is selected in this influenced by L c , as shown in Fig. 12(b). But in some direc-
paper. To get a more reliable result, each sample is printed tions, the phenomenon is not obvious, as shown in Fig. 12(a).
three times, and the result is obtained from the average value. Therefore, tb , as defined in Fig. 12(b), is measured to
Fig. 12. Thickness of slanted struts in different directions. (a) Thickness in

the direction that is not affected by L c . (b) Thickness in the direction that is
greatly affected by L c .
Fig. 15. Experimental result of nonhorizontal group.
Fig. 13. Image analysis of the horizontal strut. (a) Original image. c) Result of nonhorizontal group: The thickness of
(b) Sharpened image for edge detection. inclined struts tb is measured by a caliper with 0.01 mm
accuracy. The ratio of the thickness deviation of slanted
struts Rs is defined in (17) to measure the discrepancy ratio
compared to the designed thickness

|tb −D|
Rs = ×100% (17)
D
where tb is the thickness defined in Fig. 12(b) and D is the
design diameter.
The results of Rs are shown in Fig. 15. As the plot shows,
the value gets a huge increase when θ decreased to 10°.
Rs of 1- and 2-mm struts is higher than that of thicker struts.
And 1- and 2-mm struts tend to be less stable when the
θ is less than 60°. It can be concluded that the thickness of
Fig. 14. Experimental result of horizontal group. the discrepancy is influenced by the diameter as well as the
inclined angle of the strut.
investigate the influence of an inclined angle on the deviation 3) Artificial Neural Network: In this paper, the ANN is used
of the strut thickness. to predict the manufacturability of the lattice structure and
b) Result of horizontal group: Because it is hard to find build the meta-model. ANN is a massively parallel distributed
the largest deflection of the horizontal strut directly measured processor consists of simple processing units. Due to large
by a caliper. The image analysis method is used to find scale of parallel distributed structure as well as the ability
the maximum deflection point and measure it by pixels. The to learn and generalize, ANN has computing power to solve
image of one strut is shown in Fig. 13. Fig. 13(a) shows the complex problems that are currently intractable [42].
photograph taken by a camera. Then the image is sharpened Because the data obtained from the experiment are not
to find the edge of the strut by MathWorks Image Processing sufficient to determine the manufacturability of the lattice
Toolbox [41], which is shown in Fig. 13(b). A gauge block is struts with specific geometrical data. The relationship between
measured by this approach to calibrate the size of the pixel. the manufacturability and geometrical data should be more
The resolution of this measurement method is 0.03 mm. The comprehensively established. Since ANN is a nonlinear model.
deflection ratio Rh is defined in (16) to illustrate the deflection And it is often used when the relationship between the input
of horizontal struts compared to the strut thickness and output variables is not completely understood or even
unknown. Therefore, it is suitable for this research to build
Df
Rh = ×100% (16) the nonlinear relationship between the manufacturability and
D geometrical parameters. And this model can analyze more
where D f is the deflection and D is the design diameter of input parameters without much effort so that if designers want
the horizontal strut. to consider the influence of process parameters as well as
The result of Rh is shown in Fig. 14. The result indicates design parameters the ANN model is also valid to predict the
that the Rh is increasing with the increase of the L. When result. Another advantage of ANN is that when new data are
D gets larger, the percentage of deflection will decrease. So it obtained in the later fabrication, it can be imported to the
can be concluded that the ratio of the deflection of horizontal network to improve the performance.
struts is not only related to the length of the strut, but also Multilayer perceptrons are used for this research. Typically,
related to the design diameter. Therefore, when considering the the network consists of three types of layers, the input layer
manufacturability of horizontal struts, both the diameter and that collects the input data, one or more hidden layers that
the length of horizontal line segments should be considered. makes connections between the input and output layer and
Fig. 16. Regression plot of horizontal group.

Fig. 17. Regression plot of nonhorizontal group.
analyze data, and an output layer that returns the learning

results. The algorithm used for this problem is a back-
propagation. It contains a forward pass which fixes the synap-
tic weights, and a backward pass which adjusts the synaptic
weights in accordance with an error-correction rule [42].
A sigmoidal nonlinearity defined by the logistic function, as
shown in (18), is used to model the nonlinear relationship
1
yj = (18)
1 + exp(−v j )
where v j is the weighted sum of all synaptic inputs plus bias
of neuron j , and y j is the output of the neuron. The weights
will be tuned in the training process until the error reduces to
an acceptable level.
The data obtained from the experiment are trained in Fig. 18. Interpolation of horizontal group.
the MATLAB ANN tool. The input data of the horizon-
tal group contain 30 samples. Each sample consists of the inclined angle. The output sample only has one elements
2 elements, L and D. The output data only have one element, that is Rs . The training group, validation group, and testing
which is the Rh . Thirty samples are divided into the training group have 35 samples, 5 samples, and 5 samples, respectively.
group (22 samples), validation group (4 samples), and testing Other settings are the same with those of horizontal struts. The
group (4 samples). The number of hidden neurons is set regression plot of the nonhorizontal group is shown in Fig. 17.
to be 5. Mean square error is used as the error evaluation After establishing the neural network of the input data and
function. And the regression R values measure the correlation output data, it can be used as an interpolation approach to
between outputs and targets. The regression plots across all predict the result that was not obtained by the experiment. For
samples for the percentage of deflection are shown in Fig. 16. the horizontal group, D is interpolated from 0.5 to 5 mm with
It shows the relationship between the actual network outputs 0.1 mm as the interval, and L is from 1 to 60 mm with 1 mm
and the associated target values. If the linear regression fits as the interval. The interpolation result by the neural network
to this output–target relationship closely alongside the line is shown in Fig. 18. The z-axis is the ANN predicted Rh . For
connecting the bottom-left and top-right corners of the plot, the nonhorizontal group, θ is from 1° to 90° with 1° as the
it means the network has been successfully trained to fit the interval. The range of the D is the same as that of horizontal
data [43]. struts. The interpolation result is shown in Fig. 19. The
The data of the nonhorizontal group contain 45 samples. z-axis is the ANN predicted Rs . These two interpolation plots
The input sample consists of two elements, the diameter and visualize the ANNs for horizontal and nonhorizontal struts.
Fig. 22. Failure at the joint of the strut with Rs equal 20%.
the acceptable design domain can be determined by the quality

tolerance and manufacturing constraints are obtained.
F. BESO-Based Lattice Optimization

The last step of the proposed method is to optimize the
Fig. 19. Interpolation of nonhorizontal group.
size of each strut for a given lattice frame with respect to the
volume constraints from initial design stage. In this process,
the manufacturing constraints from the meta-model described
in Section III-E. should also be satisfied. To achieve this
purpose, a BESO-based optimization algorithm is used. This
algorithm is originally proposed by Querin et al [44] to update
the shape and topology of structures for a better functional
Fig. 20. (a) Inclined strut with stringing. (b) Horizontal strut with stringing. performance. It simulates the remodeling process of human’s
bone which is known as Wolff’s law [45]. Recently, this
algorithm has been modified and successfully applied to the
design of heterogeneous lattice structures [9]. Compared to
other optimization methods for lattice structures, there are
three advantages of BESO method for lattice structures. First,
like other heuristic optimization method, it does not require the
calculation of gradients. Thus, it is easier for implementation.
Fig. 21. Quality of inclined strut with Rs (a) over and (b) less than 10%. Second, compared to other heuristic optimization methods
such as pattern search, the convergence rate can be increased
If more types of geometrical data are defined, the ANNs could by considering the stress inside the heuristic function dur-
be more complex which may not be visualized. When using the ing the design optimization process. Apart from these two
meta-model, the first step is to choose the ANN according to advantages, the manufacturing constraints are also easier to
the type of the strut. A user-defined quality tolerance is set for be considered during the optimization. Due to these reasons,
the meta-model. Particularly for this research, it is found that the BESO-based optimization method is chosen in this paper.
the horizontal struts whose deflection ratio is larger than 25% Its mathematical representation can be expressed as
and that the vertical or slanted struts whose discrepancy ratio
To find : t =(t1 , t2 , . . . , tn )
is larger than 10% may have the printing quality issues like i=n
stringing phenomenon as shown in Fig. 20. The filament will
Minimize: P I = σVMi Vi (ti ) /(F · L)
be detached and the printed shape is no longer the cylindrical
i=1
shape as what it supposed to be. For the horizontal strut, s.t. : K (t)u − P = 0
the maximum deflection only happens at the middle region

i=n
of the strut. It is found that the acceptable ratio is larger Vi (ti ) ≤ Vconst
than the inclined strut. For the inclined struts, as shown in i=1
Fig. 21, it is found if the discrepancy ratio is larger than i
tmin ≤ti ≤tmax
i
(19)
10%, the stringing phenomenon happens, and the strut gets
unstable at the tip which will influence the connection between where t is the n-dimensional vector contains the size of each
each strut. This has been verified by printing the unit cell strut ti in the generated lattice frame. PI is the performance
of the strut (shown in Fig. 22). This may not significantly indicator of a structure to measure how well the overall
influence the strength of a single strut, but it may affect the structure is performing against an idealized fully stressed
quality of the following printing and the strength of the whole design [44]. σVMi and Vi represent the maximum Von-Mises
structure. Due to this reason, those specific values are selected stress of the i th strut and its related volume, respectively.
as the quality tolerance in this paper. It should be noted F and L are two parameters used to describe an idealized
that different designers can select different quality tolerance load case, where F is a representational force and L is a
based on their requirements. In some special cases, more than reference length. Their values can be predefined by designers
one type of quality tolerance is needed. Then, the predefined based on the size of FV and its related loading condition.
design domain will be imported into the meta-model. Finally, During the optimization process, the values of these two
Fig. 24. Arm of quadcopter and its loading condition, (a) arm of quadcopter,
original design, (b) loading condition.
Fig. 23. General work flow of BESO-based optimization algorithm for

heterogeneous lattice structures.
parameters will keep unchanged. K is the global stiffness

matrix of the lattice structure which can be regarded as a
function of design variable t in this problem. u is the vector
of nodal displacement in lattice structures and P is the nodal
load. Vconst is the value of volume constraint which can be
calculated from the initial design stage by (5). Vi (ti ) is a
function to calculate the volume of i th strut inside the lattice
structure. The form of this function may vary depending on
i
the cross-sectional shape of the strut and its length. tmin , tmax
i
represent the lower and upper boundary of the size of lattice

i
struts. The minimum size of struts tmin may vary strut by Fig. 25. Physical entity of the arm of quadcopter.
strut, since it is mainly determined by the manufacturability
of struts. Its value can be obtained from the proposed meta-
model described in Section III-E. As to the upper bound of where tri is the target size of lattice after material removal
i , it is mainly controlled by the maximum
struts thickness tmax process; A(ti ) is the function to calculate the area of cross
allowable porosity of designed lattice structures. This value section with respect to the size of cross section ti . li is the
can be predefined by designers. total length of i th strut.
To solve the problem defined in the (19), the BESO-based Between two iterations, the incremental value ri is used to
optimization algorithm is used in this paper, and its general update the current RR based on the following equation:
working flow is shown in Fig. 23. The key of this algorithm
R R n+1 = R R n + ri (22)
is the relocation of materials from low stress struts to high
stress struts. To control this process, three parameters can be where R R n and R R n+1 are the R R values used in the nth and
used. They are rejection ratio (RR), size remove ratio (RT), n + 1th iteration, respectively.
and RR incremental value ri. Among them, RR is the ratio to To judge the convergence of the optimization process,
control the threshold of stress for the struts whose material the difference of PI values between two iteration steps is used.
needs to be removed. Suppose σMAXVon is the maximum If the difference is smaller than a given value, the optimization
Von-Mises stress of all the struts in the lattice frame, process can be stopped. For the detailed steps of this algorithm,
if the Von-Mises stress σVMi of strut si satisfy readers can refer the previous research paper [9].
σVMi < R R × σMAXVon, then a certain amount of volume After the optimization process, size of each lattice strut can
should be removed from this strut until it reaches the lower be obtained. Based on the data and lattice frame, geometrical
i
bound of strut’s size tmin . model of lattice structures for a given FV can be generated.
RT is used to control the material removal rate for lattice To combine this model with other FVs with solid material,
struts. The following equations can be used to calculate the the final design can be obtained.
removed volume Vrtotal of each lattice strut:

RT × ti , if ti > tmin
i
/(1 − RT) IV. C ASE S TUDY AND D ISCUSSION
tri = i (20)
tmin , if ti ≤ tmin /(1 − RT)
i
To further illustrate the proposed design and optimization

i=n method, a design case of the quadcopter arm is given in this
Vrtotal = (A(ti ) − A(ti − tri ))li (21) section. Its original design and loading condition are shown
i=1 in Fig. 24.
TABLE II
M ATERIAL P ROPERTIES OF ABS
The major function of this part is to transfer the lifting

force from the propeller to the frame of quadcopter. To achieve
this function, four design requirements which are summarized
based on general design and fabrication guidelines of quad-
copter [46], [47] are listed in the following.
1) The maximum displacement at point c should be smaller
than 7 mm under given load condition.
2) The maximum stress should be smaller than the yield
stress of selected materials.
3) The total weight of part should be smaller or equal
to 33 g.
4) The main body of frame should be porous to minimize
the drag force when the air passes through it vertically.
Based on the functional description and design require- Fig. 26. Topology optimization of initial design.
ments mentioned above, a physical entity with 16 FSs and 3
FVs (shown in Fig. 25) is built on the conceptual design stage. TABLE III
Acrylonitrile-butadiene-styrene (ABS) is selected as material C ONSTRAINTS OF V OLUME
for this physical entity. Its material properties used in this
paper are summarized in Table II. This generated physical
entity is regarded as the input of the proposed design method.
Among those sixteen FSs, FS1 to FS9 are assembly surfaces
for the motor, while FS10 to FS16 are assembly surfaces
for the main frame of a quadcopter. To connect those FSs,
three FVs are used. Among them, FV1 and FV2 are filled
with solid material, while lattice structure is used in FV3 to
reduce drag force when air passes through. Based on those
FSs and FVs, the design space and nondesign space can be
built for the topology optimization. They are shown in Fig. 26.
The topology optimization formulation shown on (4) has been
applied to this design and solved by an SIMP-based topol-
ogy optimization solver called OptiStruct [48]. In this initial
Fig. 27. Updated FVs after initial design.
optimization process, the volume constraint has been added
to control the total volume of optimized shape. The result is
TABLE IV
shown in Fig. 26. Based on this result, volume constraints of
D IMENSION OF THE U NIT C ELL
each FV are calculated by (5) and listed in Table III. Since
the FV1 and FV2 are filled with the solid material, its shape
can be directly obtained from the initial topology optimization
process by removing the material in the low relative density
region. The optimized shape of FV1 and FV2 is shown
in Fig. 27. As to FV3, the volume constraint from the initial
design stage is regarded as the volume constraint for lattice boundary of FV3 smooth. In other words, the FV3 can be
optimization which will be discussed in the following portion divided by lattice cell size without remainders on the boundary.
of this section. Based on these rules, the cell size is selected and summarized
After the initial design, the lattice frame is built to fill in Table IV, and the uniform lattice frame has been built which
the FV3. To build the lattice frame, two types of parameters can full fill the FV3.
need to be determined. The first parameter is the lattice In order to provide the manufacturing constraints for the
topology. In this case study, a cubic-center lattice cell is following lattice optimization process, the MEs of designed
selected due to its stretching-dominant properties. As to cell lattice frames need to be built. During the construction process
size, general guidelines provided in Section II have been of MEs, the printing orientation of lattice structures needs
considered. Moreover, since FV3 is in a regular shape, with to be decided at first. To determine this parameter, three
careful selection of cell size, it is possible to make the criteria are considered. First, to avoid the support structure
TABLE VI
VALUE OF PARAMETER U SED IN THE O PTIMIZATION A LGORITHM
Fig. 28. Printing orientation.
TABLE V
G EOMETRICAL PARAMETERS AND C ONSTRAINTS
IN O PTIMIZATION A LGORITHM
Fig. 29. PI index with respect to number of iterations.
in the cylindrical surface, those functional surfaces FS2–FS9

and FS11–FS14 should be perpendicular to the printing plat-
form. Therefore, only two orientations can satisfy the first
criterion. The second criterion is that for the functional vol-
ume FV1, the printing orientation should minimize the support
structure. Obviously when the functional surfaces FS1 and
FS10 are placed at the bottom of the platform, FV1 can
be printed without supporting structures. The last criterion is
that the printing layout of the model should fit the printing
volume of the selected machine. Consequently, the printing
orientation is determined, as shown in Fig. 28. Besides the
printing orientation, the predefined design domain of struts’
thickness is also determined in this stage. This predefined
domain ranges from 0.6 to 5 mm. The upper and lower
bounds of this domain are determined based on the minimal
resolution of the machine as well as the porosity constraints.
Based on those parameters mentioned above, the MEs of
designed lattice structures can be built and input into the Fig. 30. Optimized quadcopter arm with manufacturing constraints.
established metal-model. the manufacturing constraints can be (a) CAD model. (b) FDM-fabricated part.
obtained with the given quality tolerance (Rh is set to be
less than 25% for horizontal struts and Rs is set to be less
than 10% for slanted and vertical struts in this case study). design and the design of homogeneous lattice structures, it is
Then the lower and upper bounds of the design domain with obvious that optimized heterogeneous lattice structures can
manufacturing constraints are found. All the geometrical para- significantly improve the structural stiffness without increasing
meters and constraints are summarized in Table V. The arm its weight. To further verify this conclusion, physical tests
of the quadcopter is optimized under those manufacturability on both homogeneous lattice and optimized heterogeneous
constraints and given parameters for optimization algorithm lattice have been done. In the physical tests, a tensile test
summarized in Table VI. After 52 iterations, the optimization machine (ADMET MicroEP series with 45N load sensor) is
converged (the change of the PI index is less than a given used to apply the given load and measure the displacement
value 0.001), which is shown in Fig. 29. The CAD model of on the end of arm. A clamp is used to fix the other end
the optimized quadcopter arm shown in Fig. 30(a) is success- of arm. The experimental setup used in this paper is shown
fully fabricated [Fig. 30(b)] by the selected FDM machine. The in Fig. 31. The results of physical tests are also summarized
arm of the quadcopter is also optimized without considering in Table VII. It shows a small deviation (less than 0.1 mm)
the manufacturability constraints of lattice structures. This part between experimental results and physical testing results. This
is failed during the manufacturing process. deviation is mainly caused by the anisotropic material proper-
The simulation results of different types of design configu- ties of printed ABS material which has not been considered in
rations are summarized in Table VII. Compared to the original the current paper. However, both simulation and experimental
TABLE VII
C OMPARISON OF S IMULATION R ESULTS
Fig. 31. Setup of physical test.
data show a clear trend that the designed lattice structure can
significantly improve the structural stiffness compared to its
original design (shown in Fig. 24) and homogeneous lattice.
Generally, the efficiency of the proposed design method has
been validated by both simulation and physical results.
To further evaluate the efficiency of the proposed lattice
design and optimization method, the part [shown in Fig. 32(b)]
directly obtained from traditional topology optimization rou-
tine (SIMP) has also been considered as a benchmark part
in this paper. The displacement contours of optimized lattice Fig. 32. Comparison between optimized lattice structures and topology
optimization result. (a) Displacement contour of optimized lattice structures.
structures and topology optimization result are compared and (b) Displacement contour of topology optimization result.
shown in Fig. 32. Even though the topology optimization
results can achieve slightly better stiffness, the large area of
the overhang region in the topology optimized part makes it summarized in Table V. Based on this fact, it can be inferred
difficult to be directly fabricated without support structure. that if the process related parameters, such as printing strategy
Moreover, as it is shown in Table VII, optimized lattice and process parameters, can be optimized to alleviate the
structures may achieve a smaller maximum Von-Mises stress existing manufacturing constraints, the performance of parts
compared to topology optimization result. can be further improved by the proposed design method.
Another interesting fact which can be observed from
Table VII is that the optimized lattice with uniform constraint V. C ONCLUSION
can achieve a better stiffness than that of nonuniform con- In this paper, a design method of lattice structures under
straints obtained from the proposed method as well as the the manufacturability constraints of AM process has been
result of topology optimization. In this paper, the uniform proposed. The meta-model for the selected AM process is
constraint refers smallest cylinders that can be printed by the obtained from experiments and ANN. A BESO-based opti-
selected machine. Its value is 0.6 mm for the selected machine. mization process is used to find the optimum struts’ thickness
This constraint is uniformly applied to all the struts during the distribution.
optimization process. Compared to the optimized lattice with Several conclusions can be drawn from this paper are as
nonuniform constraints, the uniform constraints can provide a follows.
larger design freedom for designers. Thus, the structure can 1) The lattice structures generated by the proposed design
be further optimized to achieve better performance. However, method can improve the stiffness of the model.
this structure cannot be fabricated with the selected machine, Von-Mises stress and displacement can be reduced with-
since the dimension of optimized struts violates the constraints out increasing the volume.
2) The heterogeneous lattice structure optimized by the [17] C. Yan, L. Hao, A. Hussein, and D. Raymont, “Evaluations of cellu-
proposed method has a better performance compared to lar lattice structures manufactured using selective laser melting,” Int.
J. Mach. Tools Manuf., vol. 62, pp. 32–38, Nov. 2012.
the homogenous lattice structure. [18] M. Santorinaios, W. Brooks, C. J. Sutcliffe, and R. A. Mines, “Crush
3) The meta-model obtained from the experiment and behaviour of open cellular lattice structures manufactured using selective
ANN has ensured the manufacturability of the lattice laser melting,” WIT Trans. Built Environ., vol. 85, pp. 481–489, 2006.
[19] A. Hussein, L. Hao, C. Yan, R. Everson, and P. Young, “Advanced lattice
structure by certain AM process. support structures for metal additive manufacturing,” J. Mater. Process.
4) If the design domain can be enlarged, which in this paper Technol., vol. 213, no. 7, pp. 1019–1026, Jul. 2013.
means that if the lower bound of the struts’ thickness can [20] D. Wang, Y. Yang, R. Liu, D. Xiao, and J. Sun, “Study on the
designing rules and processability of porous structure based on selective
be smaller, the performance can be further improved. laser melting (SLM),” J. Mater. Process. Technol., vol. 213, no. 10,
Future research will be focused on the following prospects. pp. 1734–1742, Oct. 2013.
First, relations between struts’ geometrical dimensions and [21] Z. Ming, F. Raphael, S. Yaw-Kang, T. Harold, and B. Jeffrey, “Progress
mechanical properties need to be further investigated. This in topology optimization with manufacturing constraints,” in Proc. 9th
AIAA/ISSMO Symp. Multidisciplinary Anal. Optim. Amer. Inst. Aeronau-
relation can provide a feedback to lattice simulation and tics Astron., 2002, p. 5614.
optimization model with more accurate material properties. [22] J. Yan, G.-D. Cheng, and L. Liu, “A uniform optimum material based
Second, other two aspects of MEs including the material data model for concurrent optimization of thermoelastic structures and mate-
rials,” Int. J. Simul. Multidisciplinary Design Optim., vol. 2, no. 4,
and the process data, can be further investigated in the future pp. 259–266, 2008.
research to build a more comprehensive meta-model for lattice [23] J. Deng, J. Yan, and G. Cheng, “Multi-objective concurrent topology
structures. Finally, the process parameter can be optimized to optimization of thermoelastic structures composed of homogeneous
porous material,” Struct. Multidisciplinary Optim., vol. 47, no. 4,
obtain a larger feasible area for design and optimization. pp. 583–597, Apr. 2013.
[24] X. Yan, X. Huang, Y. Zha, and Y. M. Xie, “Concurrent topology
R EFERENCES optimization of structures and their composite microstructures,” Comput.
Struct., vol. 133, pp. 103–110, Mar. 2014.
[1] Merriam-Webster. (2016). Lattice. [Online]. Available: http://www. [25] Y. Tang and Y. F. Zhao, “Design method for lattice-skin structure
merriam-webster.com/dictionary/lattice fabricated by additive manufacturing,” presented at the ASME Int. Mech.
[2] V. S. Deshpande, N. A. Fleck, and M. F. Ashby, “Effective properties Eng. Congr. Expo., Montreal, QC, Canada, 2014, p. V02BT02A030.
of the octet-truss lattice material,” J. Mech. Phys. Solids, vol. 49, no. 8, [26] V. G. Sundararajan, “Topology optimization for additive manufacturing
pp. 1747–1769, 2001. of customized meso-structures using homogenization and parametric
[3] J. Lehman and R. Lakes, “Stiff lattices with zero thermal expansion smoothing functions,” M.S. thesis, Dept. Mech. Eng., Univ. Texas
and enhanced stiffness via rib cross section optimization,” Int. J. Mech. Austin, Austin, TX, USA, 2011.
Mater. Design, vol. 9, no. 3, pp. 213–225, 2013.
[27] D. Brackett, I. Ashcroft, and R. Hague, “A dithering based method to
[4] L. Yang, D. Cormier, H. West, O. Harrysson, and K. Knowlson, “Non-
generate variable volume lattice cells for additive manufacturing,” in
stochastic Ti–6Al–4V foam structures with negative Poisson’s ratio,”
Proc. 22nd Annu. Int. Solid Freeform Fabrication Symp., Austin, TX,
Mater. Sci. Eng. A, vol. 558, pp. 579–585, Dec. 2012.
USA, Aug. 2011, pp. 8–10.
[5] Q. Zhang, Y. Han, C. Chen, and T. Lu, “Ultralight X-type lattice
[28] R. Rezaie, M. Badrossamay, A. Ghaie, and H. Moosavi, “Topology
sandwich structure (I): Concept, fabrication and experimental charac-
optimization for fused deposition modeling process,” Procedia CIRP,
terization,” Sci. China, E, Technol. Sci., vol. 52, no. 8, pp. 2147–2154,
vol. 6, pp. 521–526, 2013.
2009.
[6] P. Heinl, L. Müller, C. Körner, R. F. Singer, and F. A. Müller, [29] D. Brackett, I. Ashcroft, and R. Hague, “Topology optimization for addi-
“Cellular Ti–6Al–4V structures with interconnected macro porosity for tive manufacturing,” in Proc. 22nd Annu. Int. Solid Freeform Fabrication
bone implants fabricated by selective electron beam melting,” Acta Symp., 2011, pp. 348–362.
Biomaterialia, vol. 4, no. 5, pp. 1536–1544, Sep. 2008. [30] J. Wu, C. C. L. Wang, X. Zhang, and R. Westermann, “Self-supporting
[7] S. Bose, S. Vahabzadeh, and A. Bandyopadhyay, “Bone tissue engi- rhombic infill structures for additive manufacturing,” Comput.-Aided
neering using 3D printing,” Mater. Today, vol. 16, no. 12, pp. 496–504, Design, vol. 80, pp. 32–42, Nov. 2016.
Dec. 2013. [31] J. Nguyen, S.-I. Park, and D. Rosen, “Heuristic optimization method
[8] C. Emmelmann, P. Scheinemann, M. Munsch, and V. Seyda, “Laser for cellular structure design of light weight components,” Int. J. Precis.
additive manufacturing of modified implant surfaces with osseointegra- Eng. Manuf., vol. 14, no. 6, pp. 1071–1078, Jun. 2013.
tive characteristics,” Phys. Procedia, vol. 12, Part A, pp. 375–384, 2011. [32] Y. Tang, Y. Sheng, and Y. F. Zhao, “Design method for conformal lattice-
[9] Y. Tang, A. Kurtz, and Y. F. Zhao, “Bidirectional evolutionary structural skin structure fabricated by AM technologies,” presented at the ASME
optimization (BESO) based design method for lattice structure to be Int. Design Eng. Tech. Conf. Comput. Inf. Eng. Conf., Charlotte, NC,
fabricated by additive manufacturing,” Comput.-Aided Design, vol. 69, USA, 2016, p. V01AT02A037.
pp. 91–101, Dec. 2015. [33] Y. Tang, S. Yang, and Y. F. Zhao, “Sustainable design for additive
[10] D. W. Rosen, “Computer-aided design for additive manufacturing manufacturing through functionality integration and part consolida-
of cellular structures,” Comput.-Aided Design Appl., vol. 4, no. 5, tion,” in Handbook of Sustainability in Additive Manufacturing, vol. 1,
pp. 585–594, 2007. S. S. Muthu and M. M. Savalani, Eds. Singapore: Springer, 2016,
[11] Z. Ozdemir et al., “Energy absorption in lattice structures in dynamics: pp. 101–144.
Experiments,” Int. J. Impact Eng., vol. 89, pp. 49–61, Mar. 2016. [34] Y. Tang, J.-Y. Hascoet, and Y. F. Zhao, “Integration of topological
[12] J. Lehman and R. S. Lakes, “Stiff, strong, zero thermal expansion and functional optimization in design for additive manufacturing,”
lattices via material hierarchy,” Compos. Struct., vol. 107, pp. 654–663, in Proc. ASME 12th Biennial Conf. Eng. Syst. Design Anal., 2014,
Jan. 2014. p. V001T06A006.
[13] H. Brooks and K. Brigden, “Design of conformal cooling layers with [35] Y. Tang, K. Mak, and Y. F. Zhao, “A framework to reduce product
self-supporting lattices for additively manufactured tooling,” Additive environmental impact through design optimization for additive manu-
Manuf., vol. 11, pp. 16–22, Jul. 2016. facturing,” J. Cleaner Prod., vol. 137, pp. 1560–1572, Nov. 2016.
[14] H. N. G. Wadley, “Multifunctional periodic cellular metals,” Philos. [36] M. P. Bendsøe, A. Díaz, and N. Kikuchi, “Topology and generalized lay-
Trans. Roy. Soc. London A, Math. Phys. Sci., vol. 364, no. 1838, out optimization of elastic structures,” in Topology Design of Structures.
pp. 31–68, 2006. Dordrecht, The Netherlands: Springer, 1993, pp. 159–205.
[15] H. Wang, Y. Chen, and D. W. Rosen, “A hybrid geometric modeling [37] H. V. Wang, “A unit cell approach for lightweight structure and com-
method for large scale conformal cellular structures,” in Proc. ASME pliant mechanism,” Ph.D. dissertation, Dept. Mech. Eng., Georgia Inst.
Comput. Inf. Eng. Conf., Long Beach, CA, USA, 2005, pp. 421–427. Technol., Atlanta, GA, USA, 2005.
[16] J. Panetta, Q. Zhou, L. Malomo, N. Pietroni, P. Cignoni, and D. Zorin, [38] V. S. Deshpande, M. F. Ashby, and N. A. Fleck, “Foam topology:
“Elastic textures for additive fabrication,” ACM Trans. Graph., vol. 34, Bending versus stretching dominated architectures,” Acta Mater., vol. 49,
no. 4, p. 135, 2015. no. 6, pp. 1035–1040, Apr. 2001.
[39] M. F. Ashby, Y. J. M. Bréchet, D. Cebon, and L. Salvo, “Selection Guoying Dong was born in Luoyang, Henan, China,
strategies for materials and processes,” Mater. Design, vol. 25, no. 1, in 1993. He received the B.S. degree in aerospace
pp. 51–67, Feb. 2004. engineering from Northwestern Polytechnical Uni-
[40] Z-ABS Material Data Sheet, Olsztyn, Poland, Zortrax Inc, 2015. versity, Xi’an, China, in 2015. He is currently pur-
[41] (2016). MathWorks Image Processing Toolbox Documentation. [Online]. suing the Ph.D. degree in mechanical engineering
Available: http://www.mathworks.com/help/images/index.html with McGill University, Montreal, QC, Canada.
[42] S. Haykin, Neural Networks: A Comprehensive Foundation. Englewood His research interests include the finite-element
Cliffs, NJ, USA: Prentice-Hall, 1998. analysis of lattice structures, manufacturability of
[43] (2016). MathWork MATLAB R2016b Documentation. [Online]. lattice structures, and the characteristics of additive
Available: http://www.mathworks.com/help/nnet/examples/house-price- manufacturing on the mechanical performance.
estimation.html
[44] O. M. Querin, G. P. Steven, and Y. M. Xie, “Evolutionary structural
optimisation (ESO) using a bidirectional algorithm,” Eng. Comput.,
vol. 15, no. 8, pp. 1031–1048, 1998.
[45] H. M. Frost, “Wolff’s Law and bone’s structural adaptations to mechan- Qinxue Zhou was born in Wuhan, China, in, 1997.
ical usage: An overview for clinicians,” Angle Orthodontist, vol. 64, She has been pursuing the B.Eng. (Hons.) degree
no. 3, pp. 175–188, 1994. in mechanical engineering with McGill University,
[46] A. V. Javir, K. Pawar, S. Dhudum, N. Patale, and S. Patil, “Design, Montreal, QC, Canada, since 2014.
analysis and fabrication of quadcopter,” J. Int. Assoc. Adv. Technol. Sci., Her research interests is mostly focused on addi-
vol. 16, no. 11, 2015. tive manufacturing and design optimization.
[47] I. Gaponov and A. Razinkova, “Quadcopter design and implemen-
tation as a multidisciplinary engineering course,” in Proc. IEEE
Int. Conf. Teaching, Assessment Learn. Eng. (TALE), Aug. 2012,
pp. H2B-16–H2B-19.
[48] A. E. Inc, Optistruct User’s Guide V12.0, Altair Engineering Inc, Troy,
MI, USA, 2013.
Yaoyao Fiona Zhao received the B.E. from the
Beijing Institute of Technology, Beijing, China, in
2003, and the M.E. (First Class Hons.) and Ph.D.
degrees from the University of Auckland, Auckland,
Yunlong Tang was born in Chongqing, China. He New Zealand, in 2006 and 2010, respectively.
received the B.Eng. degree in mechanical engi- She was a Researcher at the National Institute
neering from the Harbin Institute of Technology, of Standards and Technology, Gaithersburg, MD,
Harbin, China, in 2010, and the M.Eng. degree from USA, and a Post-Doctoral Research Fellow at Ecole
Beihang University, Beijing, China, in 2013. He is Centrale de Nantes, France, from 2010 to 2012. She
currently pursuing the Ph.D. degree with the Depart- is currently an Assistant Professor with the Depart-
ment of Mechanical Engineering, McGill University, ment of Mechanical Engineering, McGill University,
Montreal, QC, Canada. Montreal, QC, Canada, where she is also the Head of the Additive Design
His research interests mainly focus on computer- and Manufacturing Laboratory. Her research expertise lies in design and
aided design and manufacturing, additive manufac- manufacturing systems integration, quality control and optimization. More
turing and its related design methods and tools. His specifically, her research focuses on manufacturing informatics, dimensional
current research topic is the design methodology for additive manufacturing metrology, additive manufacturing design and process modeling, as well as
process. manufacturing sustainability analysis.

Lattice Structure Design and Optimization With Additive Manufacturing Constraints

Uploaded by

Copyright:

Available Formats

Lattice Structure Design and Optimization With Additive Manufacturing Constraints

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lattice Structure Design and Optimization With Additive Manufacturing Constraints

Uploaded by

Copyright:

Available Formats

1546 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 15, NO.

Lattice Structure Design and Optimization With

Abstract— Lattice structures with different desired physical I. I NTRODUCTION

this lattice structure can achieve a certain elastic properties

with the consideration of local stress states. Like other methods

Fig. 3. Unit cubic defined in a parametric space.

Fig. 5. General design flow of proposed design method.

to represent the midline of a strut, while the cross section

design flow of the proposed design method is given and

all the possible sizes of a lattice strut is set as the size

Fig. 11. Cantilever area of slanted struts.

2) Measurement and Result:

Fig. 12. Thickness of slanted struts in different directions. (a) Thickness in

Fig. 15. Experimental result of nonhorizontal group.

Fig. 16. Regression plot of horizontal group.

analyze data, and an output layer that returns the learning

the acceptable design domain can be determined by the quality

F. BESO-Based Lattice Optimization

Fig. 23. General work flow of BESO-based optimization algorithm for

parameters will keep unchanged. K is the global stiffness

represent the lower and upper boundary of the size of lattice

The major function of this part is to transfer the lifting

Fig. 28. Printing orientation.

Fig. 29. PI index with respect to number of iterations.

in the cylindrical surface, those functional surfaces FS2–FS9

Fig. 31. Setup of physical test.

You might also like