You are on page 1of 10

Finite Elements in Analysis and Design 214 (2023) 103855

Contents lists available at ScienceDirect

Finite Elements in Analysis & Design


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

A reduced order model approach for finite element analysis of cellular


structures✩
Daniel A. White a ,∗, Jun Kudo a , Kenneth Swartz a , Daniel A. Tortorelli b , Seth Watts a
a
Lawrence Livermore National Laboratory, Livermore CA, USA
b
Department of Mechanical Science and Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, USA

ARTICLE INFO ABSTRACT

Keywords: Due to advances in modern manufacturing, there is an increased need for accurate and efficient simulation
Reduced order model capability for microarchitected cellular structures. It is quite expensive to simulate a large cellular structure
Finite element using a fully resolved finite element method, even on a modern supercomputer. A new method is proposed
Cellular structure
that is of intermediate complexity, it is more efficient than a fully resolved finite element simulation, but more
Higher order
general than other approximations such as beam theory or homogenization. The proposed method is based on
reduced order modeling and is compatible with any finite element simulation code that supports higher-order
basis functions.

1. Introduction and public-domain Finite Element Analysis (FEA) tools support simu-
lation of lattices via homogenization. Unfortunately, homogenization
Due to advances in modern manufacturing, in particular additive has the potential danger of being employed beyond its validity. Tradi-
manufacturing, there is an increased need for accurate and efficient tional homogenization analysis assumes an infinite periodic lattice of
simulation capability for microarchitected cellular structures. A recent identical unit cells and a uniform macroscale strain, this is a physical
review article [1] proposes a taxonomy of cellular structures as follows: approximation that must be used with care. Recent advances extend
a cellular structure is any structure that consists of a mix of solid and homogenization to quasi-periodic lattices in the context of lattice op-
void regions; the class of cellular structures consists of foams, honeycombs, timization [4] in which the density and/or orientation of the lattice
and lattices; foams are random structures without well-defined unit cells; can be optimized for a specific application. Homogenization has been
honeycombs are 2D structures extruded into 3D; lattices are 3D structures compared to fully resolved finite element simulation of finite lattices
consisting of unit cells with well-defined faces and edges; the class of with the conclusion that there exists scale effects, meaning the stiffness
lattices includes periodic lattices, quasi-periodic (or graded) lattices, and of the macroscale structure depends upon the size of the lattice unit
non-uniform lattices; quasi-periodic (or graded) lattices consist of a single cell [5–7]. The accuracy of homogenization can be improved by using
topology but allow for perturbations in density or orientation; non-uniform higher-order strain fields and relaxing periodicity constraints [8].
lattices allow for jump discontinuities between adjacent cells. Our focus
For large macroscale structures that consist solely of long, straight,
in this article is on general non-uniform lattices, which may include
slender members truss theory or beam theory [9] can be used to
combinations of truss-like structures, sheet-like like structures, and
simulate the entire structure. An advantage of this is that the structure
even solid material. For a very small collection of cells it is possible
need not be composed of similar unit cells. A disadvantage of this is
to perform a fully resolved finite element simulation of the entire
that truss theory or beam theory cannot be applied to structures that
structure, in fact this is done herein to establish ‘‘true solutions’’ for
do not resemble trusses or beams. It has been shown that for beam
convergence studies. But practical applications involving 𝑂 (100) cells
theory to achieve an error less than 5% (compared to fully resolved
is beyond reach of the typical desktop computer e.g. quad-core 3 GHz
128 GB RAM. Homogenization methods, which replace a detailed finite element simulation) the effective density must be less than 0.5%
model of a unit cell with an effective or apparent material stiffness, in some cases, i.e. very thin struts [3]. Extensions to classic beam
are quite popular for simulation of lattices [2,3]. Numerous commercial theory to enable axially-tapered beams or variable cross section beams

✩ This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory, USA under Contract
DE-AC52-07NA27344.
∗ Corresponding author.
E-mail address: white37@llnl.gov (D.A. White).

https://doi.org/10.1016/j.finel.2022.103855
Received 11 June 2022; Received in revised form 2 September 2022; Accepted 30 September 2022
Available online 23 October 2022
0168-874X/© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

have been developed [10,11], these methods allow for more general notion is that many structural, thermal, or fluidic problems consist of
macrostructures. components that have a distinct input port and output port, and the
Herein, a new method is proposed that is of intermediate complex- complex component can be replace by a simpler input–output model.
ity, it is more efficient than a fully resolved finite element simulation, This approach has been proposed for efficient analysis of finite struc-
but more accurate than other approximations such as beam theory tural 2D lattices [39]. While there are some similarities between [39]
or homogenization. The proposed method is based on Reduced Order and the ROM proposed in Section 3, there are important distinctions,
Modeling (ROM) and is compatible with any finite element simulation namely the new approach explained in Section 3 produces ROMs that
code that supports higher-order elements. The proposed method is pa- are drop-in replacements for higher-order local finite element stiffness
rameterized by the order of the ROM, and converges to the true solution matrices, whereas the approach described in [39] creates ROMs for
as the order is increased. Thus the practicing engineer can choose individual struts and joints for truss-like structures. A key goal of the
the order depending upon the accuracy requirements and available new proposed method is to allow different cellular structures to be
computational resources. assembled together in a manner that ensures continuum compatibility.
The proposed method builds upon much prior research. If a single This is in some sense similar to a component-based ROM combined
computational mesh of a large structure is provided, domain decom- with a mortar contact algorithm [40], although terminology from the
position methods [12–14] can be used to construct an efficient linear
contact literature will not be used here.
solver. The large structure is decomposed into non-overlapping do-
Below is a summary of the main contributions of our paper:
mains, each domain has private Degrees Of Freedom (DOF) on the
interior of the domain and shared DOF on the boundary of the domain. • A ROM based approach for general cellular structures is intro-
These interior and boundary DOF are solved for in an iterative manner duced
until convergence. These solvers build upon the fact that the internal • The proposed method allows unit cells of different topology, mod-
DOF are completely determined by the boundary DOF, a property of eled with different meshes, to be ‘‘glued’’ together in a manner
elliptic partial differential equations, and this concept is used to derive that satisfies continuum compatibility
the proposed ROM method in Section 3. • The proposed method can be used to produce a library of ROMs
There are many applications where a large structure can be de- that can be used with any FEA tool that supports higher order
composed into interacting components. For dynamic analysis of such elements.
structures, Component Mode Analysis (or Component Mode Synthesis) • It is demonstrated that the method converges to the true solu-
constructs a simplified model of each component using a small set tion (fully resolved FEM), and that micro-details e.g. local stress
of component normal modes (eigenvectors, Ritz vectors), these sim-
concentrations, are recovered
ple models are combined to yield an efficient dynamic model of the
large structure [15–18]. In early implementations of Component Mode In Section 2, the finite element method for static linear elasticity is
Analysis the component meshes were required to be conforming, in reviewed to define the model and establish notation. In Section 3, the
later implementations Lagrange multipliers were employed to allow proposed ROM approach is derived. In Section 4, we present a variety
non-conforming meshes of the components [19]. of results that demonstrate convergence of the method for both 2D and
A class of methods known as superelement methods replace a large 3D problems.
collection of small mesh elements with a single large mesh element.
The superelement approach has been applied to standard structural
2. Review of the finite element method for linear elasticity
problems such as L-brackets [20,21]; architectural structures [22,23];
bolt-plate interactions [24]; and truss structures [25,26]. Superele-
ment methods have also been employed to facilitate computational The proposed ROM approach for cellular structures is based upon
design optimization [27]. These superelement methods employ static arbitrary high-order Finite Element Method (FEM). In fact, the overar-
condensation to eliminate internal degrees of freedom. An alterna- ching goal is to produce a ROM that can be ‘‘plugged in’’ to existing
tive approach is to use the boundary element method to form the high order FEM software without any required modifications. In this
superelement [28]. section FEM for linear elasticity is reviewed to establish nomenclature
A class of methods known as Multiscale Finite Element Method used in subsequent sections.
(MsFEM, or FE2 ) [29–31] employ a coarse finite element mesh for The Navier equation of linear elasticity for linear media, neglecting
the global macroscale structure and a second fine-scale mesh for the body force and inertia, is
local microscale structure. One distinguishing feature between various ( )
MsFEM methods is the manner of coupling microscale to macroscale, ∇ ⋅ 𝜎̄ 𝑢⃗ = 0 in 𝛺 (1)
i.e. via effective material properties or via boundary conditions. Ho- 𝑢⃗ = 𝑢⃗𝑝 on 𝛤𝑒𝑠𝑠𝑒𝑛𝑡𝑖𝑎𝑙 (2)
mogenization can be considered a type of MsFEM in which the effective ( )
𝜎̄ 𝑢⃗ ⋅ 𝑛⃗ = ⃗𝑡𝑝 on 𝛤𝑙𝑜𝑎𝑑 (3)
material property of the microscale structure is the coupling mechanism ( )
from miscroscale to macroscale [32]. Another distinguishing feature 𝜎̄ 𝑢⃗ = C∇⃗𝑢 (4)
between MsFEM methods is the manner in which the microstructure
where 𝑢⃗ is the displacement field, 𝑢⃗𝑝
is the prescribed displacement,
response is computed, either concurrently [33] or offline [34]. It should
⃗𝑡𝑝 is the prescribed traction, 𝜎̄ is the stress, 𝑛⃗ is the surface normal,
be noted that a fully resolved structure composed of lattices unit cells
𝛤𝑒𝑠𝑠𝑒𝑛𝑡𝑖𝑎𝑙 is the essential boundary condition surface, 𝛤𝑙𝑜𝑎𝑑 is the traction
can be solved by using a MsFEM-based preconditioner [35].
boundary condition surface, and C is the material stiffness tensor.
While both the superelement approach and the MsFEM approach
result in a type of reduced order model for the microscale solution, in The variational form of (1)–(4) is
this manuscript the term ROM implies a change of basis of the form 𝐀 = Find 𝑢⃗ ∈ 𝑉 (⃗𝑢𝑝 ) ⊂ 𝐻𝑑1 (𝛺) such that
𝐏𝑇 𝐊𝐏, where 𝐊 is the matrix representing a high-fidelity discretization, ∫𝛺 C∇⃗𝑢 ∶ ∇𝑣⃗ 𝑑𝛺 = ∫𝛤 ⃗𝑡𝑝 ⋅ 𝑣⃗ 𝑑𝛤𝑙𝑜𝑎𝑑 (5)
𝑙𝑜𝑎𝑑
𝐀 is a more compact representation, and the matrix 𝐏 is the rectangular for all 𝑣⃗ ∈ 𝑉 (0)
change-of-basis matrix which defines the new basis functions in terms
of the original basis functions. When the ROM change of basis is where 𝐻𝑑1 (𝛺) denotes the standard Sobolev space of order 1, of dimen-
constructed from highly resolved finite element solutions of structural sion 𝑑, meaning each component of 𝑓⃗ ∈ 𝐻𝑑1 (𝛺) satisfies 𝑓𝑖 ∈ 𝐻 1 (𝛺).
{ }
components by applying loads to component interfaces, this is referred The subspace 𝑉 (𝑏) is given by 𝑉 (𝑏) = 𝑎 ∈ 𝐻𝑑1 (𝛺) ; 𝑎 = 𝑏 𝑜𝑛 𝛤𝑒𝑠𝑠𝑒𝑛𝑡𝑖𝑎𝑙 ,
to as a port-based ROM or a component-based ROM [36–38]. The the subspace that satisfies the essential boundary conditions.

2
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

The displacement 𝑢⃗ is approximated via the basis function expansion mesh and its FEM space, (2) the fine microscale unit cell meshes
and their FEM spaces. The macroscopic structure 𝛺 is meshed with
∑ a coarse conforming mesh such that each macroscopic element 𝛺𝐸
𝑢⃗(𝑥) = ⃗ 𝑖 (𝑥),
𝑈𝑖 𝛹 (6) engulfs a single unit cell 𝜔𝑒 ∈ 𝛺. To retain sufficient accuracy, the
𝑖
∑ macroscopic elements 𝛺𝐸 are high-order elements. Each of the unit
Expressing 𝑣⃗ = 𝑖 𝑉𝑖 𝛹 ⃗ 𝑖 (𝑥) similarly to 𝑢⃗ in (6), inserting these expan- cells 𝜔𝑒 are meshed with their own distinct fine meshes; the fine meshes
sions into (5) and utilizing the arbitrariness of 𝑣⃗ and hence 𝐕 results in are conforming within the unit cells, but nonconforming across their
a sparse symmetric positive semi-definite 𝐿×𝐿 linear equation 𝐊𝐔 = 𝐅. interfaces. In reality, we do not mesh every unit cell, rather only one
The components of the load vector 𝐅 are given by mesh is generated for each ‘‘type’’ of unit cell, e.g. quadrilateral with
a hole, tee structure, . . . cf. Fig. 1; and the meshes are independent of
𝐹𝑖 = ⃗𝑡𝑝 ⋅ 𝛹
⃗ 𝑖 𝑑𝛤𝑙𝑜𝑎𝑑 , (7) one another, i.e. each unit cell type is meshed as if it were a stand-alone
∫𝛤𝑙𝑜𝑎𝑑
structure.
and the components of the stiffness matrix 𝐊 are given by There is no notion of an apparent or effective material property in
⃗ 𝑖 ∶ ∇𝛹
⃗ 𝑗 𝑑𝛺 the ROM. Instead the unit cell 𝜔𝑒 that is engulfed by the macroscopic
𝐾𝑖𝑗 = C∇𝛹 𝑖 ∈ {1, ..., 𝐿} , 𝑗 ∈ {1, ..., 𝐿} (8)
∫𝛺 element 𝛺𝐸 is modeled by an 𝑀 ×𝑀 element stiffness matrix ̂𝐤𝑒 matrix
(with ̂𝐤𝑒 defined below). The matrices ̂𝐤𝑒 have the same units and
⃗ 𝑖 (𝑥) will be defined in terms of local functions
The basis functions 𝛹
dimension 𝑀 as a macroscopic element stiffness matrix 𝐊𝐸 , and are
𝑁⃗ 𝑖 (𝑥) defined on mesh elements. The computational mesh may in gen-
assembled to form the global stiffness matrix 𝐊 of the macroscopic
eral consist of any geometric polytype, but for simplicity of exposition, structure 𝛺. This assumes the finite element simulation software sup-
we limit our discussion to quadrilaterals in 2D and hexahedrals in 3D. ports higher order basis functions. In principle these ̂𝐤𝑒 matrices could
For the case of a quadrilateral element 𝛺𝐸 , the shape functions 𝑁 ⃗ 𝑖 are
( )2 be obtained by experimentally measuring the influence coefficients on
the tensor products of 1𝐷 Lagrange polynomials 𝑃 𝑘 = 𝑃 𝑘 (𝑥)⊗𝑃 𝑘 (𝑦). the unit cells, however here they will be computed using a sequence
The shape functions of order 𝑘 require (𝑘 + 1)2 interpolation points, of FEM simulations using the highly resolved meshes for the given unit
i.e. nodes, that are distributed within the element according to cell types.
4 at vertices The ROM process begins by meshing each unit cell 𝜔𝑒 with a fine
4 (𝑘 − 1) on edges (9) mesh and computing its global high fidelity stiffness matrix 𝐤𝑒 . This
1 (𝑘 − 1)2 in interior is done by following the process described above, cf. (8). However,
to distinguish the macroscopic structure{𝛺 from the unit } cell structure
The dimension of the finite element basis within the element 𝛺𝐸 is
𝜔𝑒 we replace the space 𝛹ℎ with 𝜙ℎ = 𝜙⃗1 , 𝜙⃗2 , … , 𝜙⃗𝑁 . The unit cell
two times that of the polynomial space, to model the 𝑥, 𝑦 components
meshes are highly resolved conforming finite element meshes of the
of the displacement field. For the case of a hexahedral element 𝛺𝐸 ,
⃗ 𝑖 are the tensor products of 1𝐷 polynomials solid material within the unit cell and any type of mesh element may
the shape functions 𝑁
( 𝑘 )3 be used, cf. e.g. Fig. 1.
𝑃 = 𝑃 𝑘 (𝑥) ⊗ 𝑃 𝑘 (𝑦) ⊗ 𝑃 𝑘 (𝑧). The shape functions of order 𝑘 require
The next step of the ROM process introduces a linear change of
has (𝑘 + 1)3 nodes that are distributed within the element according to
basis. The displacement 𝑢⃗ over the fine unit cell mesh is given by

8 at vertices ∑
𝑁
𝑢⃗ (𝑥) = 𝑢𝑖 𝜙⃗𝑖 (𝑥) (13)
12 (𝑘 − 1) on edges 𝑖=1
(10)
6 (𝑘 − 1)2 on faces
where the 𝑢𝑖 are displacement DOF. We define another basis function
(𝑘 − 1)3 in interior
expansion of the displacement as
The dimension of the finite element basis within the element 𝛺𝐸 is

𝑀 ∑
𝑀 ∑
𝑁
three times that of the polynomial space, to model the 𝑥, 𝑦, 𝑧 compo- 𝑢⃗ (𝑥) =
̂ 𝑢𝑗 𝜃⃗𝑗 (𝑥) =
̂ ̂
𝑢𝑗 𝑃𝑖𝑗 𝜙⃗𝑖 (𝑥) (14)
nents of the displacement field. 𝑗=1 𝑗=1 𝑖=1
Knowledge of the element shape functions allows us to compute the with 𝑀 ≪ 𝑁. As seen above the ̂ 𝑢𝑖 are new displacement DOF and the
𝑀 × 1 element 𝛺𝐸 force vector 𝐅𝐸 and 𝑀 × 𝑀 element 𝛺𝐸 stiffness new basis functions 𝜃⃗𝑗 (𝑥) are linear combinations of the original basis
matrix 𝐊𝐸 with entries
functions 𝜙⃗𝑖 (𝑥),
𝐹𝑖𝐸 = ⃗ 𝑖 ⋅ ⃗𝑡𝑝 𝑑𝛺
𝑁 𝑖 ∈ {1, ..., 𝑀} (11) ∑
𝑁
∫𝜕𝛺𝐸
𝜃⃗𝑗 (𝑥) = 𝑃𝑖𝑗 𝜙⃗𝑖 (𝑥) , 𝑗 = 1, 2, … , 𝑀 (15)
𝑖=1
𝐾𝑖𝑗𝐸 = ⃗ 𝑖 ∶ ∇𝑁
C∇𝑁 ⃗ 𝑗 𝑑𝛺 (12) { }
∫𝛺𝐸
because of this, the new space 𝜃ℎ = 𝜃⃗1 , 𝜃⃗2 , … , 𝜃⃗𝑀 is a subspace of
𝑖 ∈ {1, ..., 𝑀} , 𝑗 ∈ {1, ..., 𝑀}
the original space 𝜙ℎ . Upon combining (13) and (14) we have 𝐮 = 𝐏 ̂ 𝐮,
i.e. we replace the 𝑁 DOF displacement 𝑢⃗ of (13) by its projection
3. The reduced order model 𝑢⃗ onto the lower 𝑀 dimensional space 𝜃ℎ . The reduced unit cell 𝜔𝑒
̂
stiffness matrix ̂𝐤𝑒 is given by (8) using 𝜃𝑖 instead of 𝛹𝑖 , i.e.
In the conventional FEM we evaluate the element force vectors 𝐅𝐸
and element stiffness matrices 𝐊𝐸 and assemble them into the global 𝑘𝑒𝑖𝑗 =
̂ C ∇𝜃⃗𝑖 ⋅ ∇𝜃⃗𝑗 𝑑𝑣
∫𝜔𝑒
force vector 𝐅 and stiffness matrix 𝐊. In the ROM we form 𝐅 and 𝐊 by
replacing the 𝐅𝐸 and 𝐊𝐸 with their ROM unit cell 𝜔𝑒 counterparts 𝐟̂ 𝑒 ∑
𝑁
= C 𝑃𝑘𝑖 ∇𝜙⃗𝑘 ⋅ 𝑃𝑙𝑗 ∇𝜙⃗𝑙 𝑑𝑣
and ̂𝐤𝑒 . 𝑘,𝑙=1
∫𝜔𝑒


𝑁
3.1. The ROM element stiffness matrix ̂𝐤𝑒 = 𝑃𝑘𝑖 C ∇𝜙⃗𝑘 ⋅ ∇𝜙⃗𝑙 𝑑𝑣 (16)
∫𝜔𝑒
𝑘,𝑙=1

The proposed ROM approach involves several meshes and their ∑


𝑁
𝑃𝑙𝑖 = 𝑃𝑘𝑖 𝑘𝑒𝑘𝑙 𝑃𝑙𝑗 (17)
respective finite element spaces, namely (1) the coarse macroscale 𝑘,𝑙=1

3
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

Fig. 1. Illustration of three different unit cell structures. Each structure is defined to fit within a unit square, but internal structure is different. Subfigure (a) is square with a
hole, Subfigure (b) is a simple Tee structure, and Subfigure (c) is a negative Poisson structure.

Fig. 2. Illustration of basis functions #5, and #14 (out of 16 total, for order 𝑘 = 2), visualized as mesh displacement. These basis functions are the solutions to the microscale
unit cell problem. By construction these basis functions match on the boundaries of the unit cell, but differ in the interior.

where we also used (15). Summarizing the high fidelity 𝑁 × 𝑁 𝐤𝑒 and are chosen to define the new basis. There are many variations of this
reduced 𝑀 × 𝑀 ̂𝐤𝑒 stiffness matrices for the unit cell 𝜔𝑒 are related via method e.g. greedy methods, adaptive methods, etc. These methods
are appropriate if the goal is to efficiently simulate a macrostructure
̂𝐤𝑒 = 𝐏𝑇 𝐤𝑒 𝐏 composed of identical unit cells, but a key goal of the proposed method
(18)
is to allow macrostructures 𝛺 composed of different unit cells 𝜔𝑒 , e.g. cf.
This reduced stiffness matrix ̂𝐤𝑒 is used as the macroscopic element 𝛺𝐸 Figs. 1 and 3. When different unit cells are used within the same
stiffness matrix that is assembled to form the global stiffness matrix 𝐊 macrostructure the displacement of adjacent unit cells must conform
of the macroscopic structure 𝛺. The critical task that remains is the across their interfaces. Clearly the eigenmodes of different unit cell
computation of the projection matrix 𝐏.
structures would not conform across the interface. POD bases computed
A change of basis of this form is not new. For some applications
for these adjacent cell structures using a random, greedy, or adaptive
the new basis functions 𝜃⃗𝑖 could be eigenmodes of 𝐤𝑒 , selecting 𝑀
algorithms would likewise not conform across such interfaces. Thus
eigenmodes corresponding to the 𝑀 smallest eigenvalues for example.
standard POD methods would require some sort of additional constraint
The well known Proper Orthogonal Decomposition (POD) [41], which
is related to Principle Components Analysis and Karhunen–Loeve Ap- mechanism to satisfy continuity. This continuity problem can be solved
proximation [42], is another linear change of basis. In the POD method by considering all possible pairs of distinct unit cells and generating
the system 𝐤𝑒 𝐮 = 𝐟 is solved over and over again for different right a constrained basis for each pair [39], but this approach prohibits
hand sides 𝐟, and the solutions 𝐮 are stored as columns of a matrix 𝐐. the development of a library of independent ROMs. Instead, herein a
The singular value decomposition of 𝐐 is performed and the first 𝑀 change of basis is proposed that guarantees displacement compatibility
left-singular vectors (corresponding to the largest 𝑀 singular values) across the unit cells interfaces by construction.

4
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

Fig. 3. A 2D cantilever beam example, the left end is fixed, and uniform traction in the −𝑦 direction is applied to the right end. The macroscale ROM mesh (red) is shown in
(a) and consists of a 6 × 3 array of unit cells. The fine scale structure (black) is shown just for illustrative purposes, this fine scale mesh is replaced by the ROMs of the specific
microstructure. Since there is no exact solution to this problem, a true solution is constructed using a fully conforming unstructured mesh as shown in (b).

To begin our ROM development we denote the standard Lagrange applications the ̂𝐤𝑒 can be translated, rotated, and scaled in accordance
interpolatory shape function {basis of order 𝑘} defined on each macro- to how the unit cells are arranged in the macroscopic structure 𝛺. And
scopic element 𝛺𝐸 by 𝑁𝐸 = 𝑁 ⃗ 1, 𝑁
⃗ 2, … , 𝑁
⃗ 𝐿 with thus, the ̂𝐤𝑒 only have to be computed for each type of unit cell 𝜔𝑒 in the
macroscopic structure 𝛺, e.g. three different ̂𝐤𝑒 matrices are computed
2𝐷 ∶ 𝐿= 2 (𝑘 + 1)2 to analyze the structure illustrated in Fig. 3.
(19)
3𝐷 ∶ 𝐿= 3 (𝑘 + 1)3 The computational complexity of generating a ROM stiffness matrix
̂𝐤𝑒 of a unit cell is as follows. Let 𝑁 denote the number of DOF in
We define a subset of this basis as
{ } the fine mesh unit cell displacement vector 𝐮. The computational cost
𝑁𝐸 = 𝑁 ⃗ 1, 𝑁
⃗ 2, … , 𝑁
⃗ 𝑀 ⊂ 𝑁𝐸 (20) of computing each displacement solution 𝐮𝑙 is 𝑂 (𝑁) when a scalable
𝑅𝑂𝑀
multigrid solver is employed [43]. Computing the change of variables
with
matrix 𝐏 requires 𝑀 such linear solves, with 𝑀 given by (21). Note that
2𝐷 ∶ 𝑀 = 2 (4 + 4 (𝑘 − 1)) on a parallel computer these 𝑀 solves can be computed in parallel.
( ) (21)
3𝐷 ∶ 𝑀 = 3 8 + 12 (𝑘 − 1) + 6 (𝑘 − 1)2 Also, if the unit cell exhibits symmetries, the number and size of the
{ } linear solves can be reduced significantly. Finally, computing ̂𝐤𝑒 =
The set 𝑁 ⃗ 1, 𝑁
⃗ 2, … , 𝑁
⃗ 𝑀 consists only of the basis functions 𝑁
⃗ 𝑖 that
𝐏𝑇 𝐤𝑒 𝐏 requires 𝑂 (𝑀𝑁) operations. Clearly, the 𝑂 ((𝑀 + 1) 𝑁) cost of
are non-zero on the surface 𝜕𝛺𝐸 of the macroscopic element 𝛺𝐸 . computing the ROM ̂𝐤𝑒 is expensive for large ROM order 𝑘. However,
To generate the ROM change of basis matrix 𝐏 the equilibrium it is essential to note that this is a one-time ‘‘offline’’ cost. The idea is
equation (5) is repeatedly solved for the displacement 𝑢⃗ over the fine that a library of ROM ̂𝐤𝑒 matrices corresponding to different unit cell
scale unit cell mesh 𝜔𝑒 . The loading in these problems consists of a types 𝜔𝑒 are computed just once and then used repeatedly for the finite
nonzero prescribed displacement. In regard to (1)–(4), 𝛺 = 𝜔𝑒 , 𝑢⃗𝑝 = element analyzes of different macroscopic structures 𝛺 that are in-filled
(𝑁⃗ 𝑗 ⋅ 𝑒⃗𝑘 ) 𝑒⃗𝑘 , 𝛤𝑒𝑠𝑠𝑒𝑛𝑡𝑖𝑎𝑙 = 𝜕𝛺𝐸 ∩ 𝜕𝜔𝑒 , ⃗𝑡𝑝 = 0 and 𝛤𝑙𝑜𝑎𝑑 = 𝜕𝜔𝑒 ⧵ 𝛤𝑒𝑠𝑠𝑒𝑛𝑡𝑖𝑎𝑙 with the same types of unit cells 𝜔𝑒 .
for 𝑗 = 1, 2, … , 𝑀∕𝑑 and 𝑘 = 1, … , 𝑑. And in regard to (7) 𝑢𝑝𝑚 =
𝑁⃗ 𝑗 (𝑥𝑖 )⋅ 𝑒⃗𝑘 . The columns of 𝐏 are then simply the finite element solutions 3.2. The ROM element force vector ̂
𝐟𝑒
𝐮𝑙 for 𝑙 = 1, 2, … , 𝑀. The ROM basis functions 𝜃⃗ (𝑥) given by (15) are
unique to each unit cell microstructure 𝜔𝑒 . But by construction the ROM To apply the ROM, the local macroscopic coarse element matrix
bases 𝑁𝐸𝑅𝑂𝑀 for different unit cells conform across their cell interfaces, 𝐊𝐸 is simply replaced by its associated ROM stiffness matrix ̂𝐤𝑒 in
since 𝜃ℎ = 𝑁𝐸𝑅𝑂𝑀 on 𝜕𝛺𝐸 for every macroscopic element 𝛺𝐸 . This the assembly of the macroscopic stiffness matrix 𝐊. However, there
is illustrated by examining the ROM basis functions shown in Fig. 2, is one subtlety that must be dealt with, namely the computation of
clearly these basis functions are different within the interiors of the unit the macroscopic load vector 𝐅, cf. (7). Without loss of generality,
cells, but they are equal over the unit cell boundaries. Further, since we assume 𝑢⃗𝑝 = 0 in the macroscopic analysis, and thus the only
𝜃ℎ = 𝑁𝐸𝑅𝑂𝑀 on 𝜕𝛺𝐸 the ROM basis functions inherit the Kronecker contribution to 𝐅 is due to the traction ⃗𝑡𝑝 .
property and by design are well suited to represent the deformation As in the computation of 𝐊, we replace the macroscopic element 𝛺𝐸
modes of the individual unit cells 𝜔𝑒 . Finally, it should be noted that force vector 𝐅𝐸 in (12) with the unit cell 𝜔𝑒 ROM force vector ̂
𝐟 𝑒 . We
while the ROM stiffness matrix ̂𝐤𝑒 is defined on a unit cell 𝛺𝐸 , for ̂
compute 𝐟 by first evaluating the global high fidelity force vector 𝐟 𝑒
𝑒

5
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

Fig. 4. Subfigure (a) shows the reconstructed ROM solution as a displaced mesh, note that even though the individual unit cell meshes are non-conforming, the ROM solution is
continuous. Subfigure (b) shows the relative error in the ROM solution compared to the true solution, the maximum error is 0.08%.

𝑗
via the first integral in (7) for the unit cell 𝜔𝑒 . As a reminder, in (7) the each edge (𝑑 = 2) or face (𝑑 = 3) 𝜕𝜔𝑒 ∩ 𝜕𝛺𝐸 with 𝑗 ∈ {1, … , 2 𝑑} of
unit cell structure 𝜔𝑒 replaces the macroscopic structure 𝛺, 𝜙ℎ replaces unit cell 𝜔𝑒 , we require
𝛹ℎ and ⃗𝑡𝑝 is the restriction of ⃗𝑡𝑝 to 𝜕𝜔𝑒 ∩ 𝛤𝑙𝑜𝑎𝑑 . The change-of-basis (15)

𝑀 ∑
𝑀
is then used to define ⃗𝑡𝑝 (𝑥) = 𝑡𝑝𝑖 𝜃⃗𝑖 (𝑥) = 𝑡𝑝𝑖 𝑁
⃗ 𝑖 (𝑥) . (24)
𝑖=1 𝑖=1
𝑓̂𝑖𝑒 = 𝜃⃗𝑖 ⋅ ⃗𝑡𝑝 𝑑𝑎
∫𝜕𝜔𝑒 ∩𝛤𝑙𝑜𝑎𝑑 where we recall that 𝜃⃗𝑖 = 𝑁 ⃗ 𝑖 on 𝜕𝛺𝐸 . Notably, traction may not be

𝑁 applied to the ‘‘interior’’ surfaces of the unit cell, i.e. over the surfaces
= 𝑃𝑘𝑖 𝜙⃗𝑘 ⋅ ⃗𝑡𝑝 𝑑𝑎 𝜕𝜔𝑒 ∖𝜕𝛺𝐸 .
∫𝜕𝜔𝑒 ∩𝛤𝑙𝑜𝑎𝑑
𝑘=1
Substituting (24) into (22) gives
∑𝑁
= 𝑃𝑘𝑖 𝜙⃗𝑘 ⋅ ⃗𝑡𝑝 𝑑𝑎 ∑
𝑁 ∑
𝑀 ∑
𝑁 ∑
𝑀
∫𝜕𝜔𝑒 ∩𝛤𝑙𝑜𝑎𝑑 𝑓̂𝑖𝑒−𝑗 = 𝑃𝑘𝑖 𝜙⃗𝑘 ⋅ 𝑡𝑝𝑙 𝑁
⃗ 𝑙 𝑑𝑎 = 𝑗 𝑝
𝑃𝑘𝑖 𝑀𝑘𝑙 𝑡𝑙 (25)
𝑘=1
∫ 𝑗
𝜕𝜔𝑒 ∩𝜕𝛺𝐸
𝑘=1 𝑙=1 𝑘=1 𝑙=1
∑𝑁
= 𝑃𝑘𝑖 𝑓𝑘𝑒 (22) 𝑗
where the components of the nonsymmetric ‘‘surface 𝜕𝛺𝐸 mass’’ matrix
𝑘=1 𝑗
𝐌 are
This renders the following relationship between the high fidelity 𝑁 × 1
𝑗 ⃗ 𝑘 ⋅ 𝜃⃗𝑙 𝑑𝛤𝑅𝑂𝑀 = ⃗𝑘 ⋅ 𝑁
⃗ 𝑖 𝑑𝛤𝑅𝑂𝑀
𝐟 𝑒 and reduced 𝑀 × 1 ̂𝐟 𝑒 element load vectors 𝑀𝑘𝑙 =
∫𝜕𝜔𝑒 ∩𝜕𝛺𝑗
𝛹
∫𝜕𝜔𝑒 ∩𝜕𝛺𝑗
𝛹 (26)
𝐸 𝐸
̂
𝐟 𝑒 = 𝐏𝑇 𝐟 𝑒 (23) 𝑖 ∈ {𝑘, ..., 𝑁} 𝑙 ∈ {1, ..., 𝑀}
A dilemma with this approach is that we have to compute the for 𝐟𝑒 Since both 𝐏 and 𝐌𝑗
are independent of the applied traction ⃗𝑡𝑝 the
𝑗
all of the unit cell type — traction ⃗𝑡𝑝 distribution combinations. If the matrix product 𝐏𝑇 𝐌𝑗 can be pre-computed for each face 𝜕𝜔𝑒 ∩ 𝜕𝛺𝐸 of
⃗𝑡𝑝 were identical for each type of unit cell, then this would be a one- each distinct unit cell type and stored. Hence the ROM element force
time off-line computation, similar to that for ̂𝐤𝑒 . Unfortunately, this is vector
an unrealistic expectation as the force distributions applied to different
̂
𝐟 𝑒−𝑗 = 𝐏𝑇 𝐌𝑗 𝐭 𝑝 (27)
cellular structures are generally different.
The solution to our force vector dilemma is to restrict the allowable can be readily evaluated for any nodal traction vector acting on the 𝐭𝑝
traction such that ⃗𝑡𝑝 ∈ 𝑁𝐸𝑅𝑂𝑀 i.e. to those that can be represented macroscopic element boundary 𝜕𝛺𝐸 𝑗
.
exactly by the element 𝛺𝐸 ROM basis functions 𝜃⃗𝑖 . In this way all The force vector defined above has to be interpreted carefully. The
possible element load vectors can be easily computed. Indeed, over magnitude of the applied load |̂
𝐟 𝑒−𝑗 | depends on the area of the surface

6
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

𝐸 = 1.0, and Poisson’s ratio of 𝜈 = 0.29. All results were obtained using
the public domain Modular Finite Element Method (MFEM) library [44]
which supports higher-order finite element basis functions, quadrature
rules, matrix assembly, linear system solution, etc.
A 2D inhomogeneous cantilever beam problem is shown in Fig. 3.
The beam consists of three different types of unit cells. The red mesh
in Fig. 3-(a) is the macroscale mesh, the black mesh in Fig. 3-(a) is
for illustration only, these microscale meshes will be replaced by the
specific ROM for each distinct unit cell. The true solution mesh is shown
in Fig. 3-(b) and is for testing purposes only, this is a highly resolved
conforming triangular mesh. A zero-displacement boundary condition
is applied to the left side of the beam, a downward traction is applied
to the right side of the beam.
It is possible to quantify the error of the ROM solution by comparing
it to the true solution. This is done by defining a relative 𝐿2 error as
( )2 1∕2
⎡∫ ⃗ − 𝑢̂ ⎤
𝛺𝐸 𝑢
⎢ ⎥ (28)
⎢ ∫ (𝑢) ̂ 2 ⎥⎦
⎣ 𝛺𝐸

where 𝑢̂ is true solution and 𝑢⃗ is the reconstructed fine-scale ROM


solution. The ROM solution using order 𝑘 = 6 is shown in Fig. 4, with
subfigure (a) showing the displacement and subfigure (b) showing the
displacement error. Careful examination of Fig. 4-(a) shows that the
displacement is fully continuous even though the individual unit cell
meshes are non-conforming, as expected. The maximum error in Fig. 4-
(b) is 0.03%. Given the ROM solution, it is straightforward to compute
the local Von-Mises stress in each unit cell. This stress, as well as the
error in the stress, is shown in Fig. 5. The maximum error in the stress
Fig. 5. Subfigure (a) shows the reconstructed ROM stress magnitude, and Subfigure
(b) shows the relative error stress compared to the true solution. The maximum error
is 0.5%
is 6%.
4.2. 3D results

𝜕𝜔𝑒 ∩ 𝜕𝛺𝐸 𝑗
. And hence the same nodal traction 𝐭 𝑝 applied to different As a 3D example consider an L-bracket macrostructure, composed
unit cells 𝜔𝑒 will yield different net forces. e.g., the same 𝐭 𝑝 will have of octet truss unit cells. In many practical applications of additively
a much larger net force if it is applied to the quadrilateral with a hole manufactured lattice structures it is common to utilize a continuous
rather than the tee structure. This is, of course, exactly what would surface skin to distribute traction loads or facilitate bonding to other
happen in a laboratory experiment when a traction is applied directly structures. In this example, the skin is simply a thin layer of solid
to the surface of a lattice structure. If one desires to apply a ‘‘net force’’ material, and by construction the proposed ROM is easily connected
boundary condition some care must be taken to account for the lattice to solid structures. The octet truss unit cell is shown in Fig. 6-(a),
surface area. All this said, many cellular structures are surrounded by because of the complexity of this structure a tetrahedral mesh is used.
a solid ‘‘skin’’, and as such no traction is applied to the individual unit The macroscale L-bracket is shown in Fig. 6-(b), the red mesh is the
cells. macroscale mesh that is used for the actual calculation, the black mesh
As stated above, traction can only be applied to outer boundary is shown for illustrative purposes only, as these fine scale unit cells are
𝜕𝜔𝑒 ∩ 𝜕𝛺𝐸 of the unit cell 𝜔𝑒 , so e.g. a pressure cannot be applied to replaced by the ROM.
the circular face in the quadrilateral with a hole unit cell. This could be As in the 2D results above, the 3D ROM result can be compared
mitigated by forming yet another mass matrix to accommodate internal to a highly resolved conforming mesh true solution. The displacement
pressures. Similar procedures could be used to apply uniform body results are shown in Fig. 7, with Fig. 7-(a) showing the true solution
loads to model, e.g. self weight. displacement and Fig. 7-(b) showing the magnitude of the relative

It has been noted on many occasions that 𝑢⃗(𝑥) = 𝑀 𝑢 𝑁⃗ (𝑥) for error in the ROM displacement for a ROM of order 𝑘 = 6. As in the
∑ 𝑖 𝑖𝑝 ⃗
𝑖=1
𝑥 ∈ 𝜕𝜔𝑒 ∩𝜕𝛺𝐸 . Because of this we necessarily have 𝑢⃗𝑝 (𝑥) = 𝑀 𝑖=1 𝑢𝑖 𝑁𝑖 (𝑥)
2D example, the error is computed by first reconstructing microscale
which is a restriction in the form of the prescribed displacement and we solution from the ROM solution via (15), and using (28) to define the
can only prescribe displacement over the outer boundary 𝜕𝜔𝑒 ∩ 𝜕𝛺𝐸 of error. The maximum error is 0.05%.
the unit cell 𝜔𝑒 , so e.g., we cannot constrain the displacement on the Since the microscale solution can be reconstructed from the ROM
circular face in the quadrilateral with a hole unit cell. Typically this solution it is possible to compute the local Von-Mises stress. This is
⋃ 𝑗
poses no issues as we often have 𝑢⃗𝑝 = 0 and 𝛤𝑒𝑠𝑠𝑒𝑛𝑡𝑖𝑎𝑙 = 𝑗,𝐸 𝜕𝜔𝑒 ∩ 𝜕𝛺𝐸 . shown in Fig. 8, with the true solution stress shown in Fig. 8-(a), and
the reconstructed ROM stress shown in Fig. 8-(b). The maximum stress
4. Results error is 0.5%.
It is important to note that the ROM result is being compared to a
4.1. 2D results highly resolved conforming tetrahedral mesh solution (using quadratic
basis functions) and not to an exact analytical solution. As the order
In this section the ROM algorithm is verified by comparing ROM so- of the ROM is increased, the ROM result improves, as is shown in
lutions to ‘‘true solutions’’ which are FEM solutions of the fully resolved Table 1. Naturally, as the order of ROM is increased the number of
cellular structure. While most practical applications of the ROM are for finite element DOF increases, this is also shown in Table 1. Note that
three-dimensional structures, two-dimensional test problems are useful the number of finite element DOF for the true solution is 2878113. This
for researchers with limited computational resources. In all of these table also shows the required CPU time for the ROM solution relative
examples, the isotropic material properties are Young’s modulus of to the true solution. For this particular example, a ROM for order 𝑘 = 3

7
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

Fig. 6. A 3D L-bracket example. Subfigure (a) is the fine-scale mesh of an octet truss unit cell. Subfigure (b) shows the macro-scale mesh of the L-bracket. The top of the structure
(blue lines) is a zero-displacement boundary condition, and the square blue patch shows the location of the traction boundary condition. Note that the L-bracket utilizes a solid
material ‘‘skin’’ to support constraints and tractions.

achieves an error of 0.38% with a speedup of over 1400×, whereas Table 1


increasing to order 𝑘 = 6 yields a maximum error of 0.08%, and is 10× Performance of the ROM for the octet truss L-bracket problem. The first row is the order
of the ROM solution. The second row is the number of DOF for the ROM solution. The
more efficient than computing the full true solution. Thus the proposed third row is the maximum relative error, the forth row is the ROM CPU time relative
ROM approach gives the practicing engineer the ability prioritize either to the CPU time for the true solution.
accuracy and efficiency. ROM order 1 2 3 4 5 6
It is interesting to compare the ROM solution to homogenization for # of DOF 576 3381 10200 22815 43008 72561
the octet truss L-bracket problem. The homogenized material properties Relative error (%) 8.8 1.1 0.38 0.26 0.091 0.049
are computed for the octet truss unit cell shown in Fig. 6 using the Relative time 0.0 0.0001 0.0007 0.0043 0.015 0.1
procedure defined in [3]. The structure has cubic symmetry, and the
resulting homogenized materials properties are 𝐸 = 0.024802, 𝜈 =
0.32947, and 𝜇 = 0.01692. Given these properties, the finite element
solution is simply computed on the same macroscale mesh as the ROM homogenization and ROM is identical, as both simulations are using
solution, with the same FEM order, and the same boundary conditions the same order of FEM basis functions, with the same dimension finite
and tractions. The total compliance and maximum stress values are element stiffness matrix, etc.
tabulated in Table 2. Note that the full FEM true solution has total
compliance of 758 and a maximum stress of 110, thus the homogenized 5. Conclusions
solution overestimates the compliance (i.e. is too soft) whereas the
ROM solution underestimates the compliance (i.e. is too stiff). Note, Due to advances in manufacturing, in particular additive manufac-
however, that the ROM solution is converging to the true solution turing, there is an increased need for accurate and efficient simula-
whereas the homogenized solution is not. Also, Table 2 shows that the tion capability for microarchitected cellular structures. A ROM based
homogenized solution underestimates the maximum stress by almost method is proposed, this method bypasses the notion of apparent or
two orders of magnitude, this is because the homogenized solution effective material properties of a lattice and instead creates an 𝑀 × 𝑀
computes only average displacements, strains, and stresses and cannot matrix 𝐀 that resembles a higher-order local finite element stiffness
recover the actual stress on the fine scale unit cell mesh. Finally, matrix. This 𝑀 × 𝑀 ROM stiffness matrix 𝐀 can be ‘‘plugged in’’ to
it must be noted that in this comparison the computational cost of any finite element software package that supports standard high order

8
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

Fig. 7. Subfigure (a) is the displacement true solution for the L-bracket problem,
Subfigure (b) is the magnitude of the displacement error for the ROM solution of
order 𝑘 = 6. The maximum error is 0.03%. Fig. 8. Subfigure (a) is the magnitude of the stress true solution for the L-bracket
problem, Subfigure (b) is the magnitude of the stress error for the ROM solution of
order 𝑘 = 6. The maximum stress error is 0.5%.

Table 2
Comparison of Homogenization versus ROM for the octet truss L-bracket problem. The
first row is the order of the finite element basis functions used on the macroscale mesh. and stress within each microstructure, if desired. One disadvantage of
The second and third rows are the total compliance, for ROM and homogenization, the proposed ROM method is that for very thin rod structures e.g. radius
respectively. The fourth and fifth rows are the maximum stress, for ROM and
homogenization, respectively.
of 0.025 or volume fraction of around 5%, the ROM system of equations
Basis order 1 2 3 4 5 6
becomes ill-conditioned. Thus, the proposed ROM is not presented as
an alternative to beam theory for very thin rod metamaterials. In fact,
ROM compliance 380 416 551 598 743 750
HOM compliance 769 850 865 868 870 871 beam theory compares well to fully resolved FEM for lattices with
ROM max stress 44.82 48.99 68.32 60.83 62.24 67.34 volume fraction of 5% or less [3], thus beam theory and the proposed
HOM max stress 0.74 1.90 2.62 3.22 3.85 4.41
ROM are quite complimentary methods.

𝐻 1 Lagrange interpolatory basis functions. The ROM stiffness matrix Declaration of competing interest
is parameterized by the order 𝑘, which defines the polynomial order
of the displacement of the boundary of the unit cell. The approach is The authors declare that they have no known competing finan-
verified by both 2D and 3D simulations that demonstrate the proposed cial interests or personal relationships that could have appeared to
ROM converges to the true solution as the order 𝑘 is increased. It is
influence the work reported in this paper.
also observed that the cost of the using the ROM increases significantly
as the order 𝑘 is increased, thus there is a trade off between accuracy
and efficiency. For the 3D example, it appears that using a ROM of
Data availability
order 𝑘 = 3 gives good accuracy while being 1000x faster than a
fully resolved FEM simulation. Another benefit of the proposed ROM
approach is that it is possible to reconstruct the fine-scale displacement Data will be made available on request.

9
D.A. White et al. Finite Elements in Analysis & Design 214 (2023) 103855

References [23] A. Fooladi, M.R. Banan, A super-element based on fimnite element method for
latticed columns, Int. J. Civil Eng. 13 (2) (2014) 202–212.
[1] C. Pan, Y. Han, J. Liu, Design and optimization of lattice structures: A review, [24] N.L. Pedersen, P. Pedersen, Bolt-plate contact assemblies with prestress and
Appl. Sci. 10 (18) (2020) 676–700. external loads: Solved with super element method, Comput. Struct. 87 (2009)
[2] G. Allaire, Shape Optimization By the Homogenization Method, Springer-Verlag, 1374–1383.
2002. [25] M. Živković, M. Kojić, R. Slavković, N. Grujović, A general beam finite element
[3] S. Watts, W. Arrighi, J. Kudo, D.A. Tortorelli, D.A. White, Simple, accurate with deformable cross-section, Comput. Methods Appl. Mech. Engrg. 190 (20)
surrogate models of the elastic response of three-dimensional open truss micro- (2001) 2651–2680.
architectures with applications to multiscale topology design, Struct. Multidisp. [26] A. Li, C. Liu, Lightweight design of a crane frame under stress and stiffness
Opt. 60 (2019) 1887–1920. constraints using super-element technique, Adv. Mech. Eng. 9 (8) (2018) 1–15.
[4] J. Groen, O. Sigmund, Homogenization-based topology optimization for high- [27] K. Qiu, W. Zhang, M. Domaszewski, D. Chamoret, Topology optimization of
resolution manufacturable microstructures, Internat. J. Numer. Methods Engrg. periodic cellular solids based on a superelement method, Eng. Optim. 41 (3)
113 (8) (2018) 1148–1163. (2009) 225–239.
[5] S. Pecullan, L.V. Gibiansky, S. Torquato, Sacle effects on the elastic behavior of [28] K. Kohno, T. Tsunada, M. Tanaka, Hybrid stress analysis of boundary and finite
periodic and hierachical and two-dimensional composites, J. Mech. Phys. Solids elements by a super-element method, Finite Elem. Anal. Des. 7 (1991) 279–290.
47 (1999) 1509–1542. [29] T. Hou, X.H. Wu, A multiscale finite element method for elliptic problems in
[6] M. Jiang, I. Jasiuk, M. Ostoja-Starzewski, Apparent elastic and elastoplastic composite materials and porous media, J. Comp. Phys. 134 (1997) 169–189.
behavior of periodic composites, Int. J. Solids Struct. 39 (2002) 199–212. [30] Y. Efendiev, T. Hou, Multiscale Finite Element Methods. Theory and Applica-
[7] W. Zhang, S. Sun, Scale-related topology optimization of cellular materials and tions, in: Surveys and Tutorials in the Applied Mathematical Sciences, vol. 4,
structures, Internat. J. Numer. Methods Engrg. 68 (2006) 993–1011. Springer-Verlag, 2009.
[8] V. Kouznetsova, M.G.D. Geers, W.A.M. Brekelmans, Multi-scale constitutive [31] G. Allaire, R. Brizzi, A multiscale finite element method for numerical
modeling of heterogeneous materials with a gradient-enhanced computa- homogenization, Multiscale Model. Simul. 4 (3) (2005) 790–812.
tional homogenization scheme, Internat. J. Numer. Methods Engrg. 54 (2002) [32] J. Yvonnet, Computational Homogenization of Heterogeneous Materials with Fi-
1235–1260. nite Elements, in: Solid Mechanics and its Applications, vol. 258, Springer-Verlag,
[9] A.F. Bower, Applied Mechanics of Solids, CRC Press, 2010. 2019.
[10] S. Li, J. Hu, C. Zhai, L. Xie, A unified method for modeling of axially and/or [33] V.B.C Tan, K. Raju, H.P. Lee, Direct FE2 for concurrent multilevel modelling
transversally functionally graded beams with variable cross-section profile, Mech. of heterogeneous structures, Comput. Methods Appl. Mech. Engrg. 360 (2020)
Based Design Struc. Mach. 41 (2013) 168–188. 112694.
[11] N.-T. Nguyen, N.-I. Kim, I. Cho, Q.T. Phung, J. Lee, Static analysis of transversely [34] R. Xu, J. Wang, W. Yan, Q. Huang, G. Giunta, S. Belouettar, H. Zahrouni, T.B.
or axially functionally graded tapered beams, Mat. Res. Innov. 18 (sup2) (2014) Zineb, H. Hu, Data-driven multiscale finite element method: From concurrence
S2–260–S2–264. to separation, Comput. Methods Appl. Mech. Engrg. 363 (2020) 112893.
[12] U. Langer, O. Steinbach, Coupled boundary and finite element tearing and inter- [35] J. Alexandersedn, B.S. Lazarov, Topology optimisation of manufacturable mi-
connecting methods, in: Lecture Notes in Computational Science and Engineering, crostructural details without length scale separation using a spectral coarse basis
Springer, 2003, pp. 83–97. preconditioner, Comput. Methods Appl. Mech. Engrg. 290 (2015) 156–182.
[13] C. Farhat, F.X. Roux, A method of finite element tearing and interconnecting and [36] D.B.P. Huynh, D.J. Knezevic, A.T. Patera, A static condensation reduced basis
its parallel solution algorithm, Internat. J. Numer. Methods Engrg. 32 (1991) element method: Complex problems, Comput. Methods Appl. Mech. Engrg. 259
1205–1227. (2013) 197–216.
[14] P. Le Tallec, Y.H. De Roeck, M. Vidrascu, Domain-decomposition methods for [37] D.B.P. Huynh, D.J. Knezevic, A.T. Patera, A static condensation reduced basis
large linearly elliptic three dimensional problems, J. Comput. Appl. Math. 34 element method: Approximation and a-posteriori error estimation, ESAIM Math.
(1991) 93–117. Model. Numer. Anal. 47 (2013) 213–251.
[15] R.H. MacNeal, Hybrid method of component mode synthesis, Comp. Struct. 1 [38] J. Elfang, A.T. Patera, A port-reduced static condensation reduced basis ele-
(1971) 581–601. ment method for large component-synthesized structures: Approximation and
[16] R.R. Craig, M.C.C. Bampton, Coupling of substructures for dynamic analysis, a-posteriori error estimation, Adv. Model. Simul. Eng. Sci. 1 (3) (2014) 1–49.
AIAA J. 6 (7) (1968) 1313–1320. [39] S. McBane, Y. Choi, Component-wise reduced order model lattice type structure
[17] R.R. Craig, Z. Ni, Component mode synthesis for model order reduction of non design, Comput. Methods Appl. Mech. Engrg. 381 (2021) 113813.
classically damped systems, J. Guid. Control Dyn. 12 (1989) 577–584. [40] T. Laursen, M.A. Puso, J. Sanders, Mortar contact formulations for deformable–
[18] E. Balmes, Optimal Ritz vectors for component mode synthesis using the singular deformable contact: Past contributions and new extensions for enriched and
value decomposition, AIAA J. 34 (6) (1996) 1256–1261. embedded interface formulations, Comput. Methods Appl. Mech. Engrg. 205
[19] C. Farhat, M. Geradin, On a component mode synthesis method and its (2012) 3–15.
application to incompatible substructures, Comput. Struct. 51 (5) (1994) [41] A. Chatterjee, An introduction to the proper orthogonal decomposition, Current
459–473. Sci. 78 (7) (2000) 808–817.
[20] Z. Yosibash, B. Schiff, A superelement for two-dimensional singular boundary [42] Y.C. Liangh, P. Lee, S.P. Lim, Z. Lin, K.H. Lee, C.G. Wu, Proper orthogonal
value problems, Int. J. Fract. 62 (1993) 325–340. decomposition and its application-Part 1: Theory, J. Sound Vib. 252 (2002)
[21] M. Galanin, S. Lazareva, E. Savenkov, Numerical investigtion of the finite 527–544.
superelement method for the 3D elasticity problems, Math. Model. Anal. 12 (1) [43] R.D. Falgout, U.M. Yang, Hypre: A library of high performance preconditioners,
(2007) 39–50. in: Lecture Notes in Computer Science, Springer, 2002, pp. 632–641.
[22] Y. He, X. Zhou, P. Hou, Combined method of super element and substructure [44] MFEM: Modular finite element methods. mfem.org.
for analysis of ILTDBS reticulated mega-structure with single-layer latticed shell
substructures, Finite Elem. Anal. Des. 46 (2010) 563–570.

10

You might also like