International Journal of Solids and Structures 44 (2007) 3738–3755

www.elsevier.com/locate/ijsolstr

Variational asymptotic method for unit cell homogenization

of periodically heterogeneous materials q
Wenbin Yu *, Tian Tang
Department of Mechanical and Aerospace Engineering, Utah State University, Logan, UT 80322-4130, USA

Received 9 January 2006; received in revised form 4 October 2006

Available online 21 October 2006

Abstract

A new micromechanics model, namely, the variational asymptotic method for unit cell homogenization (VAMUCH), is
developed to predict the effective properties of periodically heterogeneous materials and recover the local fields. Consid-
ering the periodicity as a small parameter, we can formulate a variational statement of the unit cell through an asymptotic
expansion of the energy functional. It is shown that the governing differential equations and periodic boundary conditions
of mathematical homogenization theories (MHT) can be reproduced from this variational statement. In comparison to
other approaches, VAMUCH does not rely on ad hoc assumptions, has the same rigor as MHT, has a straightforward
numerical implementation, and can calculate the complete set of properties simultaneously without using multiple loadings.
This theory is implemented using the finite element method and an engineering program, VAMUCH, is developed for
micromechanical analysis of unit cells. Many examples of binary composites, fiber reinforced composites, and particle rein-
forced composites are used to demonstrate the application, power, and accuracy of the theory and the code of VAMUCH.
 2006 Elsevier Ltd. All rights reserved.

Keywords: Homogenization; Unit cell; Heterogeneous; Anisotropic; Variational asymptotic method; VAMUCH

1. Introduction

Along with the increased knowledge and fabrication techniques for materials, more and more structures are
made with heterogeneous materials with engineered microstructures to achieve the ever-increasing perfor-
mance requirements. The increased complexity at the microlevel greatly complicates the analysis of the struc-
tural behavior, which is indispensable for rational designs of these structures. Although it is logically sound to
use the well-established finite element method (FEM) to analyze such structures by meshing all the details of
constituent materials, the size of the finite element model will easily overpower most of the computers we can

q
Parts of this paper appeared in the proceedings of 2005 ASME International Mechanical Engineering Congress and Exposition,
Orlando, Florida, Nov. 5–11, 2005 and the 47th Structures, Structural Dynamics, and Materials Conference, Newport, Rhode Island, May
1–4, 2006.
*
Corresponding author. Tel.: +1 435 7978246; fax: +1 435 7972417.
E-mail address: [email protected] (W. Yu).

0020-7683/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijsolstr.2006.10.020
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3739

Fig. 1. Periodic heterogeneous materials and the corresponding unit cell.

access in the foreseeable future because the macroscopic structural dimensions are usually several orders of
magnitude larger than the characteristic size of constituent materials.
If the structure can be idealized as a periodic assembly of many unit cells (UCs, see Fig. 1), it is possible to
homogenize the heterogeneous UC with a set of effective material properties obtained by a micromechanical
analysis of the UC. As illustrated in Fig. 2, the concept of UC essentially simplifies the original expensive
analysis of structures made with heterogeneous materials using the following three steps:

• Identify the UC and carry out a micromechanical analysis of the UC to obtain effective material properties;
• Analyze the structure with homogenized material properties to study the macroscopic structural behavior;
• Feedback the macroscopic behavior to the micromechanical analysis to calculate local fields such as
displacements, strains, and stresses within the UC, which are only needed for detailed analysis of some
critical zones.

In the past several decades, numerous approaches have been proposed for the micromechanical analysis of
UCs (see Hashin (1983) and references cited therein). These includes the earliest rules of mixture approaches
based on Voigt and Reuss hypotheses. Hill (1952) has shown that Voigt and Reuss assumptions predict the
upper and lower bounds, respectively, for the effective elastic properties of the homogenized UC. For general
heterogeneous materials, the difference between these two bounds could be too large to be of practical use.

Fig. 2. The basic steps of structure analysis with heterogeneous microstructures using the concept of unit cell.
3740 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

Researchers have proposed various techniques to either reduce the difference between the upper and lower
bounds, or find an approximate value between the upper and lower bounds. Typical approaches are the
self-consistent model (Hill, 1965) and its generalizations (Dvorak and Bahei-El-Din, 1979; Accorsi and
Nemat-Nasser, 1986), the variational approach of Hashin and Shtrikman (1962), third-order bounds (Milton,
1981), the method of cells (MOC) (Aboudi, 1982, 1989) and its variants (Paley and Aboudi, 1992; Aboudi
et al., 2001; Williams, 2005b), recursive cell method (Banerjee and Adams, 2004), mathematical homogeniza-
tion theories (MHT) (Bensoussan et al., 1978; Murakami and Toledano, 1990), finite element approaches
using conventional stress analysis of a representative volume element (RVE) (Sun and Vaidya, 1996), and
many others. Hollister and Kikuchi (1992) compared different approaches and concluded that MHT is
preferable over other approaches for periodic composites even when the material is only locally periodic with
a finite periodicity. Although MOC is not compared in Hollister and Kikuchi (1992), MOC expands the local
displacements in terms of global displacements using the Legendre polynomial of different orders according to
the required accuracy. The accuracy of MOC could be comparable to MHT if sufficient terms are used in the
polynomial expansion although the asymptotical correctness cannot be guaranteed. It is interesting to notice
that the author of MOC recently developed a new micromechanical analysis (Aboudi et al., 2001) based on
MHT with the solution procedure borrowed from MOC.
Although different approaches adopt different assumptions in the literature, there are only two essential
assumptions associated with the micromechanical analysis of heterogeneous materials with identifiable
UCs.
Assumption 1. The exact solutions of the field variables have volume averages over the UC. For example, if ui
are the exact displacements within the UC, there exist vi such that
Z
1
vi ¼ ui dX  hui i ð1Þ
X X

where X denotes the domain occupied by the UC and its volume.

Assumption 2. The effective material properties obtained from the micromechanical analysis of the UC are
independent of the geometry, the boundary conditions, and loading conditions of the macroscopic structure,
which means that effective material properties are assumed to be the intrinsic properties of the material when
viewed macroscopically.
Of course, the micromechanical analysis of the UC is only needed and appropriate if g= h/l  1, with h as
the characteristic size of the UC and l as the characteristic wavelength of the deformation of the structure. All
the others assumptions such as particular geometry shape and arrangement of the constituents, specific boundary
conditions, and prescribed relations between local fields and global fields are convenient but not essential.
In this study, the variational asymptotic method (VAM) of Berdichevsky (1979) will be used to develop a
new unit cell homogenization technique invoking these two essential assumptions. VAM simplifies the proce-
dure of solving physical problems that can be formulated in terms of a variational statement involving one or
more small parameters. In contrast to conventional asymptotic methods, VAM carries out asymptotic analysis
of the variational statement, synthesizing both merits of variational methods (viz., systematic, simple, and easy
to be implemented numerically) and asymptotic methods (viz., without ad hoc assumptions). VAM has been
used extensively to construct efficient high-fidelity structure models for composite beams (Yu et al., 2002c),
composite and smart plates (Yu et al., 2002a; Yu and Hodges, 2004a,b), and composite and smart shells
(Yu et al., 2002b, 2005), achieving an excellent compromise between accuracy and efficiency. VAM has also
been used to homogenize isotropic material with periodic cavities (Berdichevsky, 1977), which laid a founda-
tion for the present work.
First, we extend the work of Berdichevsky (1977) to extract a variational statement for the micromechanical
analysis of the UC from the three-dimensional (3D) continuum formulation of the periodically heterogeneous,
anisotropic materials. This variational statement can be solved using VAM asymptotically to find the relation
between the local displacements and global displacements for the purpose to predict effective material
properties and local fields. Then we will implement this theory using FEM to develop an engineering code,
VAMUCH, to uniformly handle general heterogeneous microstructures including one-dimensional (1D),
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3741

Fig. 3. Coordinate systems for 2D heterogeneous materials.

two-dimensional (2D), or 3D UCs. Finally, many examples including binary composites, fiber reinforced com-
posites, and particle reinforced composites, are used to demonstrate the application, power, and accuracy of
the present theory and the companion code VAMUCH.

2. A variational statement for unit cells

As shown in Fig. 3, to facilitate our formulation, we need to setup three coordinate systems: two cartesian
coordinates x = (x1, x2, x3) and y = (y1, y2, y3), and an integer-valued coordinate n = (n1, n2, n3). We use xi as
the global coordinates to describe the macroscopic structure and yi parallel to xi as the local coordinates to
describe the UC (Here and throughout the paper, Latin indices assume 1, 2, and 3 and repeated indices are
summed over their range except where explicitly indicated). If the UC is a cube with dimensions as di, we chose
the local coordinates yi in such a way that y i 2 ½d i =2; d i =2. Since the heterogeneous material composes of
many countable UCs, it is also convenient to introduce integer coordinates ni to locate each individual UC.
The integer coordinates are related to the global coordinates in such a way that ni = xi/di (no summation over
i). If the material is uniform in one of the directions, such as the fiber reinforced composites in the fiber direc-
tion, the dimension of that direction can be chosen to be an arbitrary length different from zero.
To formulate a variational statement for UCs, we need to use one of the energy principles, such as the prin-
ciple of minimum total potential energy. The second assumption implies that we could obtain the same effec-
tive material properties from an imaginary unbounded and unloaded heterogeneous material with the same
microstrucutre as the loaded and bounded one. Hence we could derive the micromechanical analysis from
a heterogeneous material which could completely occupy the 3D space R and composes of infinite many
UCs. For elastic material,1 the total potential energy is equal to the summation of the strain energy stored
in all the UCs, which is:
X1 Z
1
P¼ C ijkl ðy 1 ; y 2 ; y 3 Þij kl dX ð2Þ
n¼1 X
2

1
Although it is possible to use the present methodology to obtain inelastic properties of heterogeneous materials, we will focus on elastic
properties in this study.
3742 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

where Cijkl are the components of the periodically varying fourth-order elasticity tensor and ij are the com-
ponents of the 3D strain tensor defined for linear theory as
" #
1 oui ðn; yÞ ouj ðn; yÞ
ij ðn; yÞ ¼ þ ð3Þ
2 oy j oy i

Here ui(n;y) are functions of the integer coordinates and the local coordinates for each UC. In view of the fact
that the infinite many UCs form a continuous heterogeneous material, we need to enforce the continuity of the
displacement field ui on the interface between adjacent UCs, which can be written as follows for a UC with
integer coordinates (n1, n2, n3):
ui ðn1 ; n2 ; n3 ; d 1 =2; y 2 ; y 3 Þ ¼ ui ðn1 þ 1; n2 ; n3 ; d 1 =2; y 2 ; y 3 Þ
ui ðn1 ; n2 ; n3 ; y 1 ; d 2 =2; y 3 Þ ¼ ui ðn1 ; n2 þ 1; n3 ; y 1 ; d 2 =2; y 3 Þ ð4Þ
ui ðn1 ; n2 ; n3 ; y 1 ; y 2 ; d 3 =2Þ ¼ ui ðn1 ; n2 ; n3 þ 1; y 1 ; y 2 ; d 3 =2Þ
According to the principle of minimum total potential energy, the exact solution will minimize the energy in
Eq. (2) under the constraints in Eqs. (1) and (4). Although correctly formulated, this problem is very difficult
to solve due to discrete integer arguments. To take advantage of well-developed analytical techniques for con-
tinuous functions, we need to transform the formulation into a more convenient format using the idea of quas-
icontinuum (Kunin, 1982). The basic idea is to associate a function of integer arguments defined in the integer
space with a continuous function defined in R. Following the procedures spelled out in Berdichevsky (1977),
we can reformulate Eqs. (2)–(4), respectively, as:
Z  
1
P¼ C ijkl ij kl dR ð5Þ
R 2
" #
1 oui ðx; yÞ ouj ðx; yÞ
ij ðx; yÞ ¼ þ  uðijjÞ ð6Þ
2 oy j oy i

and
ui ðx1 ; x2 ; x3 ; d 1 =2; y 2 ; y 3 Þ ¼ ui ðx1 þ d 1 ; x2 ; x3 ; d 1 =2; y 2 ; y 3 Þ
ui ðx1 ; x2 ; x3 ; y 1 ; d 2 =2; y 3 Þ ¼ ui ðx1 ; x2 þ d 2 ; x3 ; y 1 ; d 2 =2; y 3 Þ ð7Þ
ui ðx1 ; x2 ; x3 ; y 1 ; y 2 ; d 3 =2Þ ¼ ui ðx1 ; x2 ; x3 þ d 3 ; y 1 ; y 2 ; d 3 =2Þ
Using the technique of Lagrange multipliers, we can pose the variational statement of the micromechanical
analysis of UC as a stationary value problem of the following functional:
Z
1
J¼ ½hC ijkl uðijjÞ uðkjlÞ i þ ki ðhui i  vi ÞdR
2 R
Z Z
þ bi1 ½ui ðx1 ; x2 ; x3 ; d 1 =2; y 2 ; y 3 Þ  ui ðx1 þ d 1 ; x2 ; x3 ; d 1 =2; y 2 ; y 3 ÞdS 1 dR
ZR ZS 1
þ bi2 ½ui ðx1 ; x2 ; x3 ; y 1 ; d 2 =2; y 3 Þ  ui ðx1 ; x2 þ d 2 ; x3 ; y 1 ; d 2 =2; y 3 ÞdS 2 dR
ZR ZS 2
þ bi3 ½ui ðx1 ; x2 ; x3 ; y 1 ; y 2 ; d 3 =2Þ  ui ðx1 ; x2 ; x3 þ d 3 ; y 1 ; y 2 ; d 3 =2ÞdS 3 dR ð8Þ
R S3

where ki and bij are Lagrange multipliers introducing constraints in Eqs. (1) and (7), respectively, and Si are
the surfaces with ni = 1. Because vi is unvarying for the UC, our problem is to find the displacement field ui
vanishing the first variation of J, which is solved asymptotically using VAM in the following section.

3. Variational asymptotic method for unit cell homogenization

In view of Eq. (1), it is natural to express the exact solution ui as a sum of the volume average vi plus the
difference, such that
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3743

ui ðx; yÞ ¼ vi ðxÞ þ wi ðx; yÞ ð9Þ

where hwii = 0 according to Eq. (1). The very reason that the heterogeneous material can be homogenized
leads us to believe that wi should be asymptotically smaller than vi, i.e.,

wi  gvi ð10Þ

Substituting Eq. (9) into Eq. (8) and making use of Eqs. (6) and (10), we can obtain the leading terms of the
functional as:
Z
1
J1 ¼ ½hC ijkl wðijjÞ wðkjlÞ i þ ki hwi idR
2 R
Z Z  
ovi
þ bi1 wi ðx; d 1 =2; y 2 ; y 3 Þ  wi ðx; d 1 =2; y 2 ; y 3 Þ  d 1 dS 1 dR
R S1 ox1
Z Z  
ovi
þ bi2 wi ðx; y 1 ; d 2 =2; y 3 Þ  wi ðx; y 1 ; d 2 =2; y 3 Þ  d 2 dS 2 dR
R S2 ox2
Z Z  
ovi
þ bi3 wi ðx; y 1 ; y 2 ; d 3 =2Þ  wi ðx; y 1 ; y 2 ; d 3 =2Þ  d 3 dS 3 dR: ð11Þ
R S3 ox3

Although it is possible to carry out the variation of J1 and find the Euler-Lagrange equations and associ-
ated boundary conditions for wi, which corresponds to homogenous governing differential equations along
with inhomogeneous boundary conditions. It is more convenient to use change of variables to reformulate
the same problem so that inhomogeneous governing differential equations along with homogeneous boundary
conditions can be obtained. Considering the last three terms in Eq. (11), we use the following change of
variables:

ovi
wi ðx; yÞ ¼ y j þ vi ðx; yÞ ð12Þ
oxj

with v termed as fluctuation functions. Notice, we still have hvii = 0 if the origin of the local system is chosen
to be the center of UC. Then from the functional J1 in Eq. (11), we can obtain the following functional defined
over a UC:

1
J 1 ¼ hC ijkl ½ij þ vðijjÞ ½kl þ vðkjlÞ i þ ki hvi i
2
Z Z
þ bi1 ½vi ðx; d 1 =2; y 2 ; y 3 Þ  vi ðx; d 1 =2; y 2 ; y 3 ÞdS 1 þ bi2 ½vi ðx; y 1 ; d 2 =2; y 3 Þ  vi ðx; y 1 ; d 2 =2; y 3 ÞdS 2
S1 S2
Z
þ bi3 ½vi ðx; y 1 ; y 2 ; d 3 =2Þ  vi ðx; y 1 ; y 2 ; d 3 =2ÞdS 3 ð13Þ
S3

where ij ¼ vði;jÞ will be shown later to be the components of the global strain tensor for the structure with
homogenized effective material properties. The functional J 1 in Eq. (13) forms the backbone of the present
theory, variational asymptotic method for unit cell homogenization (VAMUCH). Performing the variation,
we can obtain conditions for J 1 to be stationary as:

o
C ijkl ðij þ vðijjÞ Þ ¼ 0 in X ð14Þ
oy l
vi ðx; d 1 =2; y 2 ; y 3 Þ ¼ vi ðx; d 1 =2; y 2 ; y 3 Þ ð15Þ
3744 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

vi ðx; y 1 ; d 2 =2; y 3 Þ ¼ vi ðx; y 1 ; d 2 =2; y 3 Þ ð16Þ

vi ðx; y 1 ; y 2 ; d 3 =2Þ ¼ vi ðx; y 1 ; y 2 ; d 3 =2Þ ð17Þ
C ijkl ðij þ vðijjÞ Þjy 1 ¼d 1 =2 ¼ C ijkl ðij þ vðijjÞ Þjy 1 ¼d 1 =2 ð18Þ
C ijkl ðij þ vðijjÞ Þjy 2 ¼d 2 =2 ¼ C ijkl ðij þ vðijjÞ Þjy 2 ¼d 2 =2 ð19Þ
C ijkl ðij þ vðijjÞ Þjy 3 ¼d 3 =2 ¼ C ijkl ðij þ vðijjÞ Þjy 3 ¼d 3 =2 ð20Þ
hvi i ¼ 0 ð21Þ

where Eq. (14) is the governing differential equations, Eqs. (15)–(17) are the periodic boundary conditions
for fluctuation functions, and Eqs. (18)–(20) are the periodic boundary conditions for local stresses. All
these equations are identical to those of MHT, as listed in Manevitch et al. (2002). Although VAMUCH
can reproduce the results of MHT as expected, VAMUCH is different from MHT in the following
aspects:

• The periodic boundary conditions are derived in VAMUCH, while they are assumed a priori in MHT.
• The fluctuation functions are determined uniquely in VAMUCH due to Eq. (21), while they can only be
determined up to a constant in MHT.
• VAMUCH has an inherent variational nature which is convenient for numerical implementation, while vir-
tual quantities should be carefully chosen to make MHT variational as shown in Guedes and Kikuchi
(1990).

The difficulty of solving the variational problem in Eq. (13) is tantamount to 3D anisotropic elasticity prob-
lems. Closed-form solutions exist only for very simple cases. For general cases we need to turn to numerical
techniques such as FEM for approximate solutions.

4. Finite element implementation of VAMUCH

It is possible to formulate the FEM solution based on Eq. (13), however, it is not the most
convenient and efficient way. First, Lagrange multipliers will increase the number of unknowns. Second,
the linear system will have zeros on the diagonal of the coefficient matrix eluding the use of common
LU decomposition technique. Considering the problem governed by Eqs. (14)–(21), the last equation,
Eq. (21), will not affect the solution obtained by the rest of equations, which means the variational
statement in Eq. (13) can be reformulated as seeking the minimum value of the following functional
PX
Z
1
PX ¼ C ijkl ½ij þ vðijjÞ ½kl þ vðkjlÞ dX ð22Þ
2X X
under the constraints in Eqs. (15)–(17). The constraints in Eq. (21) do not affect the minimum value of PX but
help uniquely determine vi. In practice, we can constrain the fluctuation functions at an arbitrary node to be
zero and later use these constraints to recover the unique fluctuation functions. It is fine to use penalty func-
tion method to introduce the periodic boundary conditions in Eqs. (15)–(17), as shown in Hollister and Kiku-
chi (1992). However, this method introduces additional approximation and the robustness of the solution
depends on the choice of large penalty number. Here, we choose to make the nodes on the positive boundary
surface (i.e., yi = di/2) slave to the nodes on the opposite negative boundary surface (i.e., yi = di/2). By
assembling all the independent active degrees of freedom, we can implicitly and exactly incorporate the peri-
odic boundary conditions in Eqs. (15)–(17). In this way, we also reduce the total number of unknowns in the
linear system.
Introduce the following matrix notations
T
 ¼ b11 212 22 213 223 33 c ð23Þ
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3745
8 ov 9 2 3
>
>
1
>
> o
0 0
>
>
oy 1 >
> oy 1
>
> ov1 ov2 >> 6 o 7
>
> þ >
oy 1 >
6 o
07
>
> oy 2 >
> 6 oy 2 oy 1 78 9
>
> >
> 6 7 > v1 >
< ov 2 = 60 o
07 < =
oy 2 6 oy 2 7
¼ 6 o 7
o 7> 2 >
v  Ch v ð24Þ
>
>
ov1 ov3 >
þ oy > 6 oy 0 oy 1 7: ;
>
> oy 3 1 >
> 6 3 v
>
> ov3 >
> 6 o 7
3
>
> ov2
þ >
> 60 o 7
>
> oy oy >
2 > 4 oy 3 oy 2 5
>
>
3
>
>
: ov3 ; 0 0 o
oy 3
oy 3

where Ch is an operator matrix and v is a column matrix containing the three components of the fluctuation
functions. If we discretize v using the finite elements as
vðxi ; y i Þ ¼ Sðy i ÞX ðxi Þ ð25Þ
where S representing the shape functions (in the assembled sense disregarding the constrained node and slave
nodes) and X a column matrix of the nodal values of the fluctuation functions for all active nodes. Substitut-
ing Eqs. (23)–(25) into Eq. (22), we obtain a discretized version of the functional as
1
PX ¼ ðX T EX þ 2X T Dh  þ T D Þ ð26Þ
2X
where
Z Z Z
T T
E¼ ðCh SÞ DðCh SÞdX Dh ¼ ðCh SÞ DdX D ¼ DdX ð27Þ
X X X

with D as the 6 · 6 material matrix condensed from the fourth-order elasticity tensor Cijkl. Since the periodic
constraints have already been incorporated through assembly in Eq. (25), our problem becomes to minimize
PX in Eq. (26), which gives us the following linear system
EX ¼ Dh  ð28Þ
It is clear that X will linearly depend on , which means it is unnecessary to assign values to  (even 1’s and 0’s
as in common practice), and they can be treated as symbols without entering the computation. The solution
can be written as
X ¼ X 0 ð29Þ
Substituting Eq. (29) into Eq. (26), we can calculate the energy storing in the UC
1 T T 1 
PX ¼  ðX 0 Dh þ D Þ  T D ð30Þ
2X 2
Clearly D  is the so-called effective (or homogenized) material matrix and  the global strains. The effective
medium has an energy density PX, which can be used to carry out macroscopic analyses.
If the local fields within the UC are of interest, we can recover those fields based on the global displace-
ments v, global strains , and the fluctuation functions v. To this end, we need to uniquely determine the
fluctuation functions first, otherwise, we could not uniquely determine the local displacement field. Recall,
we fixed an arbitrary node and made nodes on the positive boundary surfaces slave to facilitate the solution
for the fluctuation functions. First, we need to construct a new array X ~ 0 from X 0 by assigning the values
for slave nodes according to the corresponding active nodes and assign zeros to the fixed node. Obviously,
X
~ 0 still yields the minimum value of PX in Eq. (22) under constrains in Eqs. (15)–(17). However, X ~ 0 may
not satisfy the constraints in Eq. (21) because it is different from the real solution by a constant. To find
the real solution, denoting as X 0 , we need to construct a discretized version of Eq. (21). Let us rewrite
Eq. (21) as
Z
vT wdX ¼ 0 ð31Þ
X
3746 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

with w as the 3 · 3 identity matrix. It can be easily verified that w is the kernel matrix of Ch in Eq. (24). Both v
and w can discretized using the finite elements as

v ¼ SX w ¼ SW ð32Þ

with W as the discretized kernel matrix. Substituting Eq. (32) to Eq. (31), we can obtain the discretized version
of the constraints as:

X
 TH W ¼ 0 ð33Þ
R
with H ¼ X S T SdX. Please note that we can always normalize the kernel matrix so that WTHW = w.
The real solution X
 can be expressed as

X
 ¼X
~ 0  þ Wk ð34Þ

where k are constants to be determined. Substituting the above relation into Eq. (33), we can solve k as

k ¼ WT H X
~ 0 ð35Þ

Then, the real solution is

X ¼ ðI  WWT H ÞX
~ 0  X 0 ð36Þ

with I as the n · n identity matrix and n is the total number of degrees of freedom.
After the fluctuation functions are determined uniquely, we can recover the local displacement based on
Eqs. (9), (12), and (32) as
2 ov ov ov 3
1 1 1 8 9
ox ox2 ox3 > y 1 >
6 1 7< =
6 2 ov2 ov2 7
u ¼ v þ 6 ov 7 y2 þ S X 0 ð37Þ
4 ox1 ox2 ox3 5>: > ;
ov3 ov3 ov3 y3
ox1 ox2 ox3

with u as the column matrix of ui and v as the column matrix of vi. The local strain field can be recovered using
Eqs. (6) and (24) as

 ¼  þ Ch SX 0  ð38Þ

Finally, the local stress field can be recovered straightforwardly as

r ¼ D ð39Þ

It can be easily verified that

hri ¼ D ð40Þ

which is expected because the effective material matrix can be equivalently defined through the energy density
of the UC or the relation between the average stress and average strain of the UC. In comparison to common
numerical micromechanical simulations in the literature, the present implementation is unique in the following
aspects:

• No external load is necessary to perform the simulation and the complete set of material properties can be
predicted within one analysis.
• The fluctuation functions and local displacements can be determined uniquely;
• The effective material properties and recovered local fields are calculated directly with the same accuracy of
the fluctuation functions. No postprocessing type calculations which introduces more approximations are
needed.
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3747

• The dimensionality of the problem is determined by that of the periodicity of the UC. A complete
set of 3D material properties can be obtained using a 1D analysis of microstructures with 1D peri-
odicity such as binary composites. It is noted that the macroscopic analysis of a structure made with
material having 1D periodicity could be 3D, which requires the complete set of effective material
properties.

We have coded the above formulation using Fortran 90/95 into a program named VAMUCH. To demon-
strate the application, accuracy, and efficiency of this theory and code, we will analyze several material systems
using VAMUCH in the next section.

5. Validation of VAMUCH

VAMUCH provides a unified analysis for general 1D, 2D, or 3D UCs. First, the same code VAMUCH can
be used to homogenize binary composites (modeled using 1D UCs), fiber reinforced composites (modeled
using 2D UCs), and particle reinforced composites (modeled using 3D UCs). Second, VAMUCH can repro-
duce the results for lower-dimensional UCs using higher-dimensional UCs. That is, VAMUCH will predict the
same results for binary composites using 1D, 2D or 3D UCs, and for fiber reinforced composites using 2D or
3D UCs.

5.1. Binary composites

First, let us consider a periodic binary composite formed by orthotropic layers and the material axes are the
same as the global coordinates xi so that the material is uniform in the x1  x2 plane and periodic along x3
direction. A typical UC can be identified as shown in Fig. 4, the dimension along y3 is h and dimensions along
y1 and y2 can be arbitrary. Let /1 and /2 denote the volume fractions of the first phase and the second phase,
respectively, and we have /1 + /2 = 1. This problem has been solved analytically in Yu (2005). The strain
energy density of the effective material can be obtained as:
8 9T 2  38 9
>
> 11 > > c11 0 c13 0 0 c16 > > 11 > >
>
> >
> 6 7>> >
>
>
> 2 >
> 6 0 c 
0 0 0 0 7 >
> 2 >
>
>
> 12 >
> 6 22 7>>
> 12 >
>
>
> >
> 6 7 > >
>
>
< >
= 6    7 >
< >
=
1  22 c
6 13 0 c 33 0 0 c 36 7  22
PX ¼ 6 7 ð41Þ
2>> 213 > > 6 0 0 0 c44 0 0 7 >
> 213 > >
>
> >
> 6 7 > >
>
> >
> 6 7>> >
>
>
> 2 >
> 6 0 0 0 0 c 
0 7>>
> 2
>
>
>
>
> 23 >
> 4 55 5 >
> 23 >
>
>
: >
; >
: >
;
  
33 c16 0 c36 0 0 c66 33
It can be observed that the homogenized material properties still have the same orthotropic symmetry for this
special case, although in general the homogenized material could be anisotropic, which means a fully popu-
lated 6 · 6 stiffness matrix. The expressions of effective material properties cij are listed here.

Fig. 4. Sketch of a binary composite.

3748 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

 2
ð2Þ ð1Þ
/1 /2 c16  c16
c11 ¼ hc11 i  ð2Þ ð1Þ
/1 c66 þ /2 c66
  
ð2Þ ð1Þ ð2Þ ð1Þ
/1 /2 c16  c16 c36  c36
c13 ¼ hc13 i  ð2Þ ð1Þ
/1 c66 þ /2 c66
ð1Þ ð2Þ ð2Þ ð1Þ
/1 c16 c66 þ /2 c16 c66
c16 ¼ ð2Þ ð1Þ
/1 c66 þ /2 c66 ð42Þ
 2
ð2Þ ð1Þ
/1 /2 c36  c36
c33 ¼ hc33 i  ð2Þ ð1Þ
/1 c66 þ /2 c66
ð1Þ ð2Þ ð2Þ ð1Þ
/1 c36 c66 þ /2 c36 c66
c36 ¼ ð2Þ ð1Þ
/ c þ /2 c66
 1 66
    
 1  1  1
c66 ¼ 1= c55 ¼ 1= c44 ¼ 1= c22 ¼ hc22 i
c66 c55 c44
where the superscripted quantities are those from each phase of the composite. It can be observed that even
for this simple case, only c22 is the same as the rule of mixture based on the Voigt hypothesis, and
c44 ; c55 ; c66 are the same as the rule of mixture based on the Reuss hypothesis. All the other components
are different from these two rules of mixture. The effective material properties of the present theory repro-
duce those of a mathematical homogenization theory in Manevitch et al. (2002). When both layers are
made of isotropic material, having Lamé properties k1, l1 for layer 1 and k2, l2 for layer 2, the formulas
in Eq. (42) reproduce the well-known exact expressions listed on Page 140 of Christensen (1979) which
was obtained by Postma (1955).
Yu (2005) shows that the fluctuation functions are piecewise linear functions for binary composites, which
means we can use 2-noded line elements for 1D UC in VAMUCH to exactly, in the numerical sense, reproduce
the analytical solution. We have also obtained the same results using 2D and 3D UCs numerically. For the
sake of saving space, such results are not presented here. Suffice to state that the numerical results from
VAMUCH are exactly the same as the analytical solution within the machine precision.

5.2. Fiber reinforced composites

To show the predictive capability of VAMUCH for unidirectional fiber reinforced composites, we choose a
few examples extensively studied in the literature. For the first two examples, comparisons are made between
FEM (Sun and Vaidya, 1996), method of cell (MOC) (Aboudi, 1982), generalized method of cell (GMC)
(Paley and Aboudi, 1992), high-fidelity generalized method of cell (HFGMC) (Aboudi et al., 2001), and elas-
ticity-based cell method (ECM) (Williams, 2005b). The FEM approach of Sun and Vaidya (1996) is estab-
lished on a rigorous mechanics foundation and 3D RVEs with periodic boundary conditions are used for
homogenization. The MOC and its variants (GMC, HFGMC, and ECM) expand the local displacements
in terms of global displacements using the Legendre polynomials of different orders. ECM starts from this
assumption and solve the equations of continuum mechanics in a strong form. In contrast, MOC, GMC,
and HFGMC invokes additional ad hoc assumptions such as that the interfacial continuous conditions and
periodic boundary conditions are only satisfied in the integral sense. FEM results are directly taken from
Sun and Vaidya (1996), MOC and ECM results from Williams (2005b), GMC and HFGMC results from
Aboudi et al. (2001).
The first example is a boron/aluminum composite. Both constituents are isotropic with Young’s modulus
E = 379.3 GPa and Poisson’s ratio m = 0.1 for boron fibers, and E = 68.3 GPa and m = 0.3 for aluminum
matrix. The fiber is of circular shape and arranged in a square array (see the sketch in the middle of
Fig. 1) and the fiber volume fraction is 0.47. The effective material properties predicted by different approaches
are listed in Table 1. It can be observed that MOC and GMC significantly underpredict the shear moduli G12
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3749

Table 1
Effective material properties of boron/aluminum composites
Models E11 (GPa) E22 (GPa) G12 (GPa) G23 (GPa) m12 m23
VAMUCH 215.3 144.1 54.39 45.92 0.195 0.255
FEM 215 144 57.2 45.9 0.19 0.29
MOC 215 142.6 51.3 43.7 0.20 0.25
GMC 215.0 141.0 51.20 43.70 0.197 0.261
HFGMC 215.4 144.0 54.34 45.83 0.195 0.255
ECM 215 143.4 54.3 45.1 0.19 0.26

Table 2
Effective material properties of graphite/epoxy composites
Models E11 (GPa) E22 (GPa) G12 (GPa) G23 (GPa) m12 m23
VAMUCH 142.9 9.61 6.10 3.12 0.252 0.350
FEM 142.6 9.60 6.00 3.10 0.25 0.35
MOC 143 9.6 5.47 3.08 0.25 0.35
GMC 143.0 9.47 5.68 3.03 0.253 0.358
HFGMC 142.9 9.61 6.09 3.10 0.252 0.350
ECM 143 9.6 5.85 3.07 0.25 0.35

and G23 while FEM overpredicts the longitudinal shear modulus G12. The closest correlation for all the values
is found between VAMUCH and HFGMC.
The second example is graphite/epoxy composites. Graphite fiber is transversely isotropic with E11 =
235 GPa, E22 = 14 GPa, G12 = 28 GPa, m12 = 0.2, and m23 = 0.25. Epoxy matrix is isotropic with
E = 4.8 GPa and m = 0.34. The fiber is circular and arranged in a square array and the fiber volume fraction
is 0.6. The results from different approaches are listed in Table 2. Again the closest correlation is found
between VAMUCH and HFGMC. Considering the fact that HFGMC uses the governing equations of
MHT and that VAMUCH can reproduce MHT, it is not surprising to find out that HFGMC has an excellent
agreement with VAMUCH. The FEM and ECM predictions are also very close to VAMUCH results for this
case. It is noted that the ECM results listed in Tables 1 and 2 are obtained from the 3rd order model. If the 5th
order theory is used, the correlation between ECM and VAMUCH might be improved as shown in the fol-
lowing two examples.
In the following two examples, VAMUCH is compared with MOC, ECM (both 3rd order and 5th order),
Green’s function based approach (G-F) (Walker et al., 1993) and FEM. The results of MOC, ECM, and G-F
are directly taken from Williams (2005b), while FEM results are calculated using ANSYS following the
approach proposed in Sun and Vaidya (1996). To be consistent with Williams (2005b), we use 2D UCs having
square inclusions in the center for VAMUCH. As pointed out in Williams (2005b), square inclusions provide a
stringent test of correct modeling the local and global behavior of heterogeneous materials due to strong
gradients in the local fields induced by the corners. The two material systems we consider are tungsten/copper
composite and void/copper composite. Both tungsten and copper are assumed to be isotropic with
E = 395.0 GPa and m = 0.28 for tungsten, and E = 127.0 GPa and m = 0.34 for copper.
For both material systems, we calculate the effective transverse Young’s modulus at different inclusion vol-
ume fractions of 0.0204, 0.1837, 0.5102, and 0.7511. The results are listed in Table 3 for tungsten/copper com-
posite and Table 4 for void/copper composite. It is verified that the FEM approach of Sun and Vaidya has no
size effects for E22, which means this approach will provide the most accurate prediction of E22 with a con-
verged mesh. As one can observe from Tables 3 and 4, MOC and 3rd order ECM underpredict this value
up to 1.6% for tungsten/copper composite and 8.6% for void/copper composite. VAMUCH, G-F, and 5th
order ECM have excellent agreement with FEM, with VAMUCH having the closest correlations.
To show the effect of shape of inclusions, we predict the effective transverse Young’s modulus using UC
with circular inclusions arranged in a square array. As shown in Tables 3 and 4, the shape effects of inclusions
become more and more significant and cannot be neglected with large volume fraction of inclusions, partic-
3750 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

Table 3
E22 (GPa) of W/Cu composites varying with fiber volume fraction
Models 0.0204 0.1837 0.5102 0.7511
VAMUCH 129.92 156.51 229.72 300.99
FEM 129.92 156.51 229.71 301.0
G-F 129.87 156.18 229.09 300.70
MOC 129.50 154.40 226.20 299.00
ECM (3rd order) 129.50 154.60 226.60 299.10
ECM (5th order) 129.80 156.50 229.50 300.80
VAMUCH (circular) 129.81 155.19 226.94 298.12
FEM (circular) 129.82 155.20 226.97 298.14

Table 4
E22 (GPa) of Void/Cu composites varying with void volume fraction
Models 0.0204 0.1837 0.5102 0.7511
VAMUCH 120.22 81.73 39.75 18.25
FEM 120.22 81.70 39.75 18.25
G-F 120.63 83.50 40.48 18.40
MOC 110.20 75.27 38.22 17.99
ECM (3rd) 110.20 75.38 38.23 17.99
ECM (5th) 118.90 80.97 39.64 18.20
VAMUCH (circular) 120.34 82.67 39.08 10.31
FEM (circular) 120.34 82.64 39.08 10.31

ularly for void/copper composites. For example, E22 of void/copper composite with square holes is 80% larger
than the composite with circular holes, when the void volume fraction reaches 0.7511. It is interesting to note
that E22 of the W/Cu composite with square inclusions is slightly larger than that with circular inclusions, with
the difference getting bigger with larger fiber volume fraction. However, for the case of Void/Cu, material with
square voids is slightly smaller than that with circular voids for small void volume fractions. As the void vol-
ume fraction getting bigger the trend is reversed. A parametric study is carried out to find out the point where

Fig. 5. Change of Young’s modulus of material with square voids and circular voids with respect to void volume fractions.
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3751

Fig. 6. Effective Young’s modulus of glass/epoxy composite with spherical inclusions.

the trend is reversed. As observed from Fig. 5, when the void volume fraction is greater than zero and less than
0.45 (approximate), E22 of materials having square voids are slightly smaller than those having circular voids.
When the void volume fraction is greater than 0.45 (approximate), materials with square voids have bigger E22
that those with circular voids.

5.3. Particle reinforced composites

Due to special arrangements of constituents of particle reinforced composites, 3D UCs are required to
accurately model the microstructures. We are going to use VAMUCH to analyze several particle reinforced
composites to validate its 3D capability. In previous section, we have shown that the prediction of MOC
and GMC is not accurate for fiber reinforced composites and one could infer that they can not provide very
accurate prediction for particle reinforced composites either. Although HFGMC and G-F provide excellent
prediction for fiber reinforced composites, we could not find 3D examples analyzed by these two approaches.
It is easy to verify that Sun and Vaidya’s FEM approach is equally applicable to particle reinforced compos-
ites. Two other approaches we believe will provide critical evaluations for VAMUCH are the 3D version of
ECM (Williams, 2005a) and an approach based on mathematical homogenization theory and finite element
method (Banks-Sills et al., 1997) (later we follow Williams (2005a) to name this approach as HFE).
The first example is to predict the effective Young’s modulus for a glass/epoxy composite. The UC of this
composite is composed of glass spheres embedded in a triply periodic cubic array. Both constituents are iso-
tropic with Young’s modulus E = 76.00 GPa and Poisson’s ratio m = 0.23 for glass, and Young’s modulus
E = 3.01 GPa and Poisson’s ratio m = 0.394 for epoxy.2 We plot the change of effective Young’s modulus with
respect to the inclusion volume fraction in Fig. 6. In comparison to HFE, VAMUCH outperforms ECM (both
3rd order and 5th order). We are surprised to find out that it is counter intuitive that the predictions of 5th
order ECM are worse than 3rd order ECM for this particular case. It is worthy to point out that the data of
HFE and ECM are provided independently by the author of Williams (2005a), where ECM data are calculat-
ed and HFE data are directly picked out from the plots in Banks-Sills et al. (1997).

2
The isotropic assumption is convenient for comparing with available results in the literature. VAMUCH can deal with constituents
with full anisotropy with material properties characterized as many as 21 independent constants.
3752 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

Effective Young's Modulus (Gpa)

Fig. 7. Effective Young’s modulus of Al2O3/Al composites with cubic inclusions.

The second example is a Al2O3/Al composite with cubic inclusions in a cubic array. Both constituents are
isotropic with Young’s E = 350.00 GPa and Poisson’s ratio m = 0.30 for aluminum oxide, and Young’s mod-
ulus E = 70.00 GPa and Poisson’s ratio m = 0.30 for aluminum. The effective Young’s modulus and Pois-
son’s ratio are plotted in Figs. 7 and 8, respectively. It can be observed that both VAMUCH and 5th
order ECM have an excellent agreement with HFE while the predictions of 3rd order ECM are not as
accurate.
The last example is a Al2O3/Al composite having rectangular parallelepiped inclusions with the ratio
between the three dimensions as a1/a2/a3 = 2/1/2 where ai is the dimension of the inclusion along the corre-

Fig. 8. Effective Poisson’s ratio of Al2O3/Al composites with cubic inclusions.

 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3753

Fig. 9. Effective Young’s modulus E33 of Al2O3/Al composites with rectangular parallelepiped inclusions.

Fig. 10. Effective Poisson’s ratio m12 of Al2O3/Al composites with rectangular parallelepiped inclusions.

sponding yi direction. The effective material properties of this composite is not isotropic any more. For the
sake of saving space, we only plot the effective Young’s modulus E33 and effective Poisson’s ratio m12 as func-
tions of inclusion volume fraction in Figs. 9 and 10, respectively. Again VAMUCH and 5th order ECM have
excellent agreements with HFE and outperform 3rd order ECM.
3754 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755

6. Conclusion

A new micromechanics model, the variational asymptotic method for unit cell homogenization
(VAMUCH), has been developed to homogenize heterogeneous materials and recover the local fields within
the microstructure after obtaining the global responses of the material. VAMUCH provides a uniform anal-
ysis for microstructures which can be described using 1D, 2D, or 3D UCs, such as binary composites, fiber-
reinforced composites, and particle-reinforced composites. In comparison to existing micromechanics
approaches, VAMUCH has the following major advantages:

(1) VAMUCH adopts the variational asymptotic method as its mathematical foundation. It has the same rig-
or as MHT without even assuming periodic fluctuation functions and boundary conditions. VAMUCH
uses only assumptions inherent in micromechanics without invoking any additional ad hoc assumptions.
(2) VAMUCH has an inherent variational nature and its numerical implementation is shown to be
straightforward.
(3) VAMUCH handles 1D/2D/3D unit cells uniformly. The dimensionality of the problem is determined by
that of the periodicity of the unit cell.
(4) VAMUCH can obtain different material properties in different directions simultaneously, which is more
efficient than those approaches requiring multiple runs under different loading conditions.
(5) VAMUCH calculates effective properties and local fields directly with the same accuracy as the fluctu-
ation functions. No postprocessing calculations such as stress averaging and strain averaging are needed.

The companion code, VAMUCH, is extensively validated using various examples including binary compos-
ites, fiber reinforced composites, and particle reinforced composites. We can confidently conclude that
VAMUCH provides a versatile and convenient tool for engineers to efficiently yet accurately design and ana-
lyze heterogeneous materials.

Acknowledgements

This study was supported by NSF under Grant DMI-0522908. The views and findings contained herein are
those of the authors and should not be interpreted as necessarily representing the official policies or endorse-
ment, either expressed or implied, of NSF. The authors want to thank Dr. Todd O. Wlliams from Los Alamos
National Laboratory for technical discussions and providing the data of ECM and HFE for particle rein-
forced composites. The authors also want to thank Dr. Victor Berdichevsky of Wayne State University,
the author of Variational Asymptotic Method, for technical discussions and advise.

References

Aboudi, J., 1982. A continuum theory for fiber-reinforced elastic-visoplastic composites. International Journal of Engineering Science 20
(5), 605–621.
Aboudi, J., 1989. Micromechanical analysis of composites by the method of cells. Applied Mechanics Reviews 42 (7), 193–221.
Aboudi, J., Pindera, M.J., Arnold, S.M., 2001. Linear thermoelastic higher-order theory for periodic multiphase materials. Journal of
Applied Mechanics 68, 697–707.
Accorsi, M.L., Nemat-Nasser, S., 1986. Bounds on the overall elastic and instantaneous elastoplastic moduli of periodic composites.
Mechanics of Materials 5 (3), 209–220.
Banerjee, B., Adams, D.O., 2004. On predicting the effective elastic properties of polymer bonded explosives using the recursive cell
method. International Journal of Solids and Structures 41 (2), 481–509.
Banks-Sills, L., Leiderman, V., Fang, D., 1997. On the effect of particle shape and orientation on elastic properties of metal matrix
composites. Composites Part B: Engineering 28B, 465–481.
Bensoussan, A., Lions, J., Papanicolaou, G., 1978. Asymptotic Analysis for Periodic Structures. North-Holland, Amsterdam.
Berdichevsky, V.L., 1977. On averaging of periodic systems. PMM 41 (6), 993–1006.
Berdichevsky, V.L., 1979. Variational-asymptotic method of constructing a theory of shells. PMM 43 (4), 664–687.
Christensen, R.M., 1979. Mechanics of Composite Materials. Wiley-Interscience, New York.
Dvorak, G.J., Bahei-El-Din, Y.A., 1979. Elastic-plastic behavior of fibrous composites. Journal of Mechanics and Physics of Solids 27, 51–72.
Guedes, J.M., Kikuchi, N., 1990. Preprocessing and postprocessing for materials based on the homogenization method with adaptive finite
element method. Computer Methods in Applied Mechanics and Engineering 83, 143–198.
 W. Yu, T. Tang / International Journal of Solids and Structures 44 (2007) 3738–3755 3755

Hashin, Z., 1983. Analysis of composite materials-a survey. Applied Mechanics Review 50, 481–505.
Hashin, Z., Shtrikman, S., 1962. A variational approach to the theory of the elastic behaviour of polycrystals. Journal of Mechanics and
Physics of Solids 10, 343–352.
Hill, R., 1952. The elastic behavior of crystalline aggregate. Proceedings of the Physical Society of London A65, 349–354.
Hill, R., 1965. Theory of mechanical properties of fibre-strengthened materials-iii. Self-consistent model. Journal of Mechanics and
Physics of Solids 13, 189–198.
Hollister, S.J., Kikuchi, N., 1992. A comparison of homogenization and standard mechanics analyses for periodic porous composites.
Computational Mechanics.
Kunin, I., 1982 Theory of Elastic Media with Microstructure, Vol. 1 and 2. Springer Verlag, Berlin.
Manevitch, L.I., Andrianov, I.V., Oshmyan, V.G., 2002. Mechanics of Periodically Heterogeneous Structures. Springer, Berlin.
Milton, G.W., 1981. Bounds on the electromagnetic, elastic and other properties of two component composites. Physics Review Letters 46
(8), 542–545.
Murakami, H., Toledano, A., 1990. A higher-order mixture homogenization of bi-laminated composites. Journal of Applied Mechanics
57, 388–396.
Paley, M., Aboudi, J., 1992. Micromechanical analysis of composites by the generalized cells model. Mechanics of Materials 14, 127–139.
Postma, G.W., 1955. Wave propagation in a stratified medium. Geophysics 20, 780.
Sun, C.T., Vaidya, R.S., 1996. Prediction of composite properties from a representative volume element. Composites Science and
Technology 56, 171–179.
Walker, K.P., Freed, A.D., Jordan, E.H., 1993. Accuracy of the generalized self-consistent method in modeling the elastic behavior of
periodic composite properties of unidirectional fiber-reinforced composites and their sensitivity coefficients. Philosophical Transactions
of the Royal Society of London A345, 545–576.
Williams, T.O., 2005a. A three-dimensional, higher-order, elasticity-based micromechanics model. International Journal of Solids and
Structures 42, 971–1007.
Williams, T.O., 2005b. A two-dimensional, higher-order, elasticity-based micromechanics model. International Journal of Solids and
Structures 42, 1009–1038.
Yu, W., 2005. A variational-asymptotic cell method for periodically heterogeneous materials. In: Proceedings of the 2005 ASME
International Mechanical Engineering Congress and Exposition. ASME, Orlando, Florida.
Yu, W., Hodges, D.H., 2004a. An asymptotic approach for thermoelastic analysis of laminated composite plates. Journal of Engineering
Mechanics 130 (5), 531–540.
Yu, W., Hodges, D.H., 2004b. A simple thermopiezoelastic model for composite plates with accurate stress recovery. Smart Materials and
Structures 13 (4), 926–938.
Yu, W., Hodges, D.H., 2005. Mathematical construction of an engineering thermopiezoelastic model for smart composite shells. Smart
Materials and Structures 14 (1), 43–55.
Yu, W., Hodges, D.H., Volovoi, V.V., 2002a. Asymptotic construction of Reissner-like models for composite plates with accurate strain
recovery. International Journal of Solids and Structures 39 (20), 5185–5203.
Yu, W., Hodges, D.H., Volovoi, V.V., 2002b. Asymptotic generalization of Reissner-Mindlin theory: accurate three-dimensional recovery
for composite shells. Computer Methods in Applied Mechanics and Engineering 191 (44), 5087–5109.
Yu, W., Hodges, D.H., Volovoi, V.V., Cesnik, C.E.S., 2002c. On Timoshenko-like modeling of initially curved and twisted composite
beams. International Journal of Solids and Structures 39 (19), 5101–5121.