On two micromechanics theories for determining micromacro

relations in heterogeneous solids

Muneo Hori
a
, Sia Nemat-Nasser
b,
*
a
Earthquake Research Institute, University of Tokyo, 1-1-1 Yayoi Bunkyo, Tokyo 113-0032, Japan
b
Department of AMES, Center of Excellence for Advanced Materials, University of California, 0416,
9500 Gilman Drive, La Jolla, San Diego, CA 92093-0416, USA
Received 1 January 1998; received in revised form 1 April 1999
Abstract
The average-eld theory and the homogenization theory are briey reviewed and compared. These theories are often
used to determine the eective moduli of heterogeneous materials from their microscopic structure in such a manner
that boundary-value problems for the macroscopic response can be formulated. While these two theories are based on
dierent modeling concepts, it is shown that they can yield essentially the same eective moduli and boundary-value
problems. A hybrid micromechanics theory is proposed in view of this correspondence. This theory leads to a more
accurate computation of the eective moduli, and applies to a broader class of microstructural models. Hence, the
resulting macroscopic boundary-value problem gives better estimates of the macroscopic response of the material. In
particular, the hybrid theory can account for the eects of the macrostrain gradient on the macrostress in a natural
manner. 1999 Elsevier Science Ltd. All rights reserved.
Keywords: Micromechanics; Homogenization; Average eld theory; Multi-scale anisotropy
1. Introduction
One major objective of micromechanics of
heterogeneous materials is to determine the eec-
tive overall properties by certain microscopic
considerations. The eective properties are then
used to evaluate the response of structural ele-
ments which consist of heterogeneous materials.
Consider two basic approaches for obtaining the
overall response of a heterogeneous medium: (1)
the average-eld theory (or the mean-eld theory)
and (2) the homogenization theory. Roughly
speaking, these are physics- and mathematics-
based theories, respectively. Here, the basic char-
acters of these two theories are interpreted as
follows:
Average-eld theory. This theory is based on the
fact that the eective mechanical properties
measured in experiments are relations between
the volume average of the strain and stress of
microscopically heterogeneous samples. Hence,
macroelds are dened as the volume averages
of the corresponding microelds, and the eec-
tive properties are determined as relations be-
tween the averaged microelds.
Homogenization theory. This theory establishes
mathematical relations between the micro-
elds and the macroelds, using a multi-scale
Mechanics of Materials 31 (1999) 667682
www.elsevier.com/locate/mechmat
*
Corresponding author. Tel.: +1-619-534-4772; fax: +1-619-
534-2727; e-mail: sia@halebopp.ucsd.edu
0167-6636/99/$ - see front matter 1999 Elsevier Science Ltd. All rights reserved.
PII: S 0 1 6 7 - 6 6 3 6 ( 9 9 ) 0 0 0 2 0 - 4
perturbation method. The eective properties
then naturally emerge as consequences of these
relations, without depending on specic physi-
cal measurements.
Nemat-Nasser and Hori (1993) presents a con-
cise list of references related to the average-eld
theory and the homogenization theory; see also
Hill (1963), Mura (1987) for the average-eld
theory. Sanchez-Palencia (1981) and Bakhvalov
and Panasenko (1984) present the homogenization
theory; see also Nuna and Keller (1984), Walker
et al. (1991); see, for instance, Hornung (1996),
Kevorkina and Cole (1996) and Terada et al.
(1996) for recent works.
Thus, fundamentally, the average-eld theory
and the homogenization theory deal with the re-
lation between the macroelds and the microelds
in dierent manners. Hence, it seems that the two
theories are basically dierent. Furthermore, their
modeling of the microstructure is also dierent: the
homogenization theory usually uses a periodic
microstructure as a model, being mainly applied to
composite materials with more or less regularly
arranged microstructure, whereas the average-eld
theory often uses simple microstructure models,
such as an isolated inclusion in an unbounded
body, in order to determine the eective proper-
ties.
It is possible to formulate the homogenization
theory in a setting which produces the eective
properties identical to those obtained through the
application of the average-eld theory. Further-
more, it is possible to combine the two theories to
develop a more general theory which is capable of
rigorously predicting the eective properties even
when the strain gradients are very large and the
average-eld theory may no longer apply.
In this paper, we are primarily concerned with
establishing a link between the average-eld theory
and the homogenization theory, clarifying their
similarities and dierences. A new micromechanics
theory is then proposed as a hybrid of these the-
ories. The content of this paper is as follows: First,
Section 2 claries the problem setting to which the
two micromechanics theories are applied; a heter-
ogeneous elastic body with elastic moduli varying
in a very small length scale is considered as an
example. The formulation of the problem using
the average-eld theory and the homogenization
theory is then presented in Sections 3 and 4, re-
spectively; similar boundary-value problems are
obtained for the two cases even though the un-
derlying assumptions and the method of deriva-
tion are dierent. The resulting boundary-value
problems are compared, and a new micromecha-
nics theory is proposed as the hybrid of these two
theories in Sections 5 and 6.
Symbolic and index notations are used in this
paper; for instance, stress tensor is denoted by ei-
ther r or r
ij
. In the symbolic notation, and : stand
for the rst- and second-order contractions, and
for the tensor product. In the index notation, the
summation convention is employed.
2. Problem setting
We consider a linearly elastic body, B, which
consists of heterogeneous materials; see Fig. 1. Let
D and d be the macro- and micro-length scales; D
is regarded as the dimension of a sample of the
heterogeneous material which is used in experi-
ments, and d is the size of typical micro constitu-
ents of this sample. The length scale of B is
orders
1
of magnitude greater than that of macro-
length scale. We dene a relative length scale pa-
rameter, e, as follows:
e =
d
D
1: (1)
Let C
e
= C
e
(X) be the variable elasticity tensor
of B, where X denotes a continuum point in B.
Here, superscript e emphasizes that the variation
of this elasticity tensor is measured at the relative
scale of e. Displacement, strain, and stress elds of
B are denoted by u
e
= u
e
(X),
e
=
e
(X), and
r

= r

(X), respectively. These elds satisfy

1
To give an insight to the length scale, consider a case when B
is analyzed by the nite element method. In the three
dimensional analysis, B is discretized by using (10
1
~ 10
3
)
3
elements. Each element corresponds to an RVE. Since D
corresponds to the element size, d becomes e(10
1
~ 10
3
) of
the size of B. As is seen, a huge computation capacity is
required to compute the response of B and each microconstit-
uent simultaneously.
668 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682

e
(X) = sym$ u
e
(X); (2)
$ r

(X) = 0; (3)
r

(X) = C
e
(X) :
e
(X); (4)
where sym stands for the symmetric part, i.e.,
sym()
ij
= (()
ij
()
ji
)=2. When B is subjected
to, say, surface displacements, u = u
0
on oB, these
three sets of the eld equations yield the following
boundary-value problem for u
e
:
$ C
e
(X) : ($ u
e
(X)) ( ) = 0 in B;
u
e
(X) = u
0
(X) on oB:
(5)
The boundary-value problem (5) cannot be
solved in its present form, because C
e
changes
within a micro-length scale d while the dimensions
of B are orders of magnitude greater than d, re-
quiring an unreasonable numerical eort. We
consider a micromechanics theory which resolves
this diculty. Since
e
and r

vary within the scale

of d, such a theory usually introduces strain and
stress elds which change within the scale of D. We
call the elds varying within the scale of D mac-
roelds, and those varying within the scale of d
microelds.
3. Average-eld theory
The average-eld theory starts by introducing
a representative volume element, denoted by V, as
a body which models the microstructure of a
given heterogeneous material; see Fig. 2. While
various denitions (See Nemat-Nasser and Hori
(1993)) are possible, we regard a representative
volume element as a model of a sample of the
material to be used to determine the corre-
sponding eective properties experimentally. The
eective properties are given as relations between
the strains and stresses of the sample. These
Fig. 2. Representative volume element as model of material
with microstructure.
Fig. 1. Boundary value problem for body consisting of material
with microstructure.
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 669
strains and stresses are measured from the sam-
ple's surface displacements and tractions, respec-
tively. It is easily shown that these strains and
stresses are actually the volume average of the
corresponding eld variables within the samples.
Indeed, applying the averaging theorems
2
for a
compatible strain and a self-equilibrating stress,
and r, we obtain
)
V
=
1
V
_
oV
symm u dS; (6)
r)
V
=
1
V
_
oV
t x dS; (7)
where )
V
stands for the volume average taken
over V, m is the outer unit normal on the boundary
oV , and u and t are the surface displacement and
traction. These equalities hold for any arbitrary
material with any constitutive properties.
3.1. Macroelds of average-eld theory
Based on Eqs. (6) and (7), the average-eld
theory denes macroeld variables as weighted
averages of the corresponding microled variables
in the representative volume element. The weight
function, denoted by /
V
= /
V
(x), satises
_
/
V
dV = 1, and takes on a constant value of 1=V
within V except for a thin layer near oV where it
decays smoothly from 1=V to 0 on oV ; the size of
V is of the order of D. The macrodisplacement,
macrostrain, and macrostress elds are thus ex-
pressed in terms of /
V
and the corresponding
microelds, as
U
e
(X)
E
e
(X)
R
e
(X)
_

_
_

_
=
_
B
/
V
(X Y)
u
e
(Y)

e
(Y)
r

(Y)
_

_
_

_
dV
Y
: (8)
The averaging with the weight /
V
cancels possible
high oscillations of the eld variables which may
occur within a micro-length scale.
Since /
V
in Eq. (8) is dened as a smooth
function, it follows from Eqs. (2) and (3) that the
macrostrain and macrostress elds satisfy
E
e
(X) = sym$ U
e
(X); (9)
$ R
e
(X) = 0: (10)
Therefore, if a constitutive relation between E
e
and
R
e
is given, a governing equation which determines
the variation of U
e
in B can be obtained from Eqs.
(9) and (10). For instance, if an eective elasticity
tensor of the homogeneous material is found such
that
R
e
(X) = C : E
e
(X); (11)
then, the three sets of the eld equations for the
macroelds, namely, Eqs. (9)(11), yield the fol-
lowing governing equation for U
e
:
$ (C : ($ U
e
(X)) = 0 in B: (12)
Hence, a boundary-value problem for U
e
is ob-
tained by assuming u
e
~ U
e
near oB, i.e., U
e
= u
0
on oB.
The eective elasticity tensor, C, which appears
in Eq. (11) relates the volume average of the mic-
rostrains to that of the microstresses. For a two-
phase composite material, C is expressed in terms
of a strain concentration tensor, A, which relates
the average strain of the inclusion phase, )
I
, to
that of the composite (or, equally, the representa-
tive volume element), )
V
, through )
I
= A :
)
V
, as follows:
C = C
M
f (C
I
C
M
) : A: (13)
Here, C
M
and C
I
are the elasticity of the matrix
and inclusion phases, respectively, and f is the
volume fraction of the inclusion phase; see Ap-
pendix A. In the average-eld theory, the strain
concentration tensor is estimated by using various
averaging schemes such as the dilute distribution
assumption, the self-consistent method, or the
MoriTanaka method; see Nemat-Nasser and
Hori (1993) and Hori and Nemat-Nasser (1993)
for the averaging schemes.
2
The averaging theories are easily proved by applying the
integral by part and the Gauss theorem in computing the
volume average of eld variables.
670 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682
3.2. Statistical homogeneity
The heterogeneous material needs to be statis-
tical homogeneous
3
in order for Eq. (11) to hold at
a point inside of the body B, i.e., in order that the
(weighted) average strain and stress be related
through an eective elasticity tensor which is
common to all points within B.
The eective elasticity tensor, C, which is de-
ned from the relation between the average stress
and the average strain, is also required to give a
relation between the average strain energy and the
average strain, i.e.,
1
2
E
e
(X) : C : E
e
(X)
=
_
B
/
V
(X Y)(
1
2

e
(Y) : C
e
(Y) :
e
(Y)dV
Y
:
(14)
To obtain conditions for Eq. (14) to be valid, we
use the averaging theorem for the strain energy,
: r)
V
=
1
V
_
oV
t u dS; (15)
which leads to
: r)
V
)
V
: r)
V
=
1
V
_
oV
(u x )) (t m r)) dS: (16)
It is shown that if V is a subdomain of a material
which consists of more or less similar microstruc-
tures, the right side of Eq. (16) decays to 0 as the
size of V increases. Since the average weighted by
/
V
is almost the same as the unweighted volume
average, we can assume that Eq. (14) holds when a
suciently large representative volume element is
taken from a composite with a statistically ho-
mogeneous microstructure.
For the homogeneous strain or stress boundary
conditions, u = x
0
or t = m r
0
on oV , with
constant
0
or r
0
, the right side of Eq. (16) van-
ishes even if V is not statistically homogeneous.
However, the strain and stress elds within V will
not be the same for dierent boundary conditions.
This means that the eective elasticity tensor can
vary depending on the displacement or traction
boundary data on the boundary oV . Indeed, the
following exact universal inequalities
4
(Nemat-
Nasser and Hori, 1993) hold for strain elds of a
common volume average

R
: C :
R
)
V
6
G
: C :
G
)
V
6
E
: C :
E
)
V
;
(17)
where
E
,
G
, and
R
are the strain elds with a
common volume average when V is subjected to
homogeneous strain boundary conditions, general
(possibly mixed) boundary conditions, and ho-
mogeneous stress boundary conditions, respec-
tively; see Appendix B for the proof of Eq. (17).
Fig. 3 presents a schematic view of the universal
inequalities. Therefore, the representative volume
element must be such that
E
: C :
E
)
V

R
: C :
R
)
V
is negligibly small, in order to
uniquely dene the eective elasticity tensor.
4. Homogenization theory
The homogenization theory considers the gov-
erning equations in Eq. (5) for u
e
, and in order to
express the changes in C
e
within a micro-length
scale, it replaces this elasticity tensor eld C
e
by
C
e
(X) ~ C(x); (18)
with x = e
1
X. Usually,
5
it is assumed that C(x)
is spatially periodic, and a periodic structure is
used as a model of the microstructure; see Fig. 4.
3
In a narrower sense, the statistical homogeneity means that
a probability of nding a phase at a point does not depend on
the point.
4
As shown by Nemat-Nasser and Hori (1993), the universal
theorems play a key role in rendering rigorous the classical
bounds on the eective moduli, developed by Hashin and
Shtrikman (1962); see also Walpole (1969), Willis (1977),
Francfort and Murat (1986), Milton and Kohn (1988), and
Torquato (1991), and Nemat-Nasser and Hori (1995).
5
See, for instance, Oleinik et al. (1992) for the homogeniza-
tion theory of non-periodic media.
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 671
The dimensions of the unit cell, U, of the periodic
structure are of the same order as D.
4.1. Singular perturbation of homogenization theory
For a periodic C, the homogenization theory
considers the following multi-scale or singular
perturbation representation
6
of u
e
:
u
e
(X) ~

n=0
e
n
u
n
(X; x); (19)
where each u
n
has the same periodicity as C with
respect to x. Since $ is now replaced by
$
X
e
1
$
x
, substitution of Eq. (19) into the gov-
erning equations of (5) yields
e
2
$
x
(C(x) : ($
x
_
u
0
(X; x)))
_
e
1
$
X
(C(x) : ($
x
_
u
0
(X; x)))
$
x
C(x) : ($
X
u
0
(X; x) $
x
u
1
(X; x))
_

n=0
e
n
$
X
(C(x) : ($
X
u
n
(X; x)
$
x
u
n1
(X; x))
$
x
(C(x) : ($
X
u
n1
(X; x)
$
x
u
n2
(X; x))
_
= 0: (20)
To solve (20) up to O(e
0
), the homogenization
theory rst assumes that u
0
is a function of only
X and that u
1
admits the representation u
1
(X; x) =
v
1
(x) : ($
X
u
0
(X)), where a third-order tensor
v
1
is periodic with respect to x. Then, terms of
O(e
2
) vanish, and terms of O(e
1
) become
6
In Eq. (19), the perturbation parameter e is dened as the
ratio of the micro-length scale over the macro-length scale.
While this parameter is simple, it neglects the change in the
magnitude of C; it is intuitively expected that the perturbation
parameter should decrease as, say, the ratio of the maximum
and minimum values of each component of C decreases.
Fig. 4. Periodic structure as model of material with micro-
structure.
Fig. 3. Universal inequalities.
672 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682
$
x
(C(x) : ($
x
_
v
1
(x) 1
(4s)
))
_
: ($
X
u
0
(X));
where 1
(4s)
is the symmetric fourth-order identity
tensor. In order for these terms to vanish identi-
cally for both X and x, v
1
must satisfy the fol-
lowing governing equation with the periodic
boundary conditions
$
x
(C(x) : ($
x
v
1
(x) 1
(4s)
)) = 0: (21)
It is seen that v
1
: ($
X
u
0
) is a microscopic dis-
placement eld in the presence of the stress eld,
C : ($
X
u
0
). Since this stress eld does not sat-
isfy equilibrium, oscillating microstrains and as-
sociated stresses are produced. In another word, v
1
corresponds to the microscale response which ac-
commodates the strain eld sym$
X
u
0
which
produces non-equilibrating stresses. Note that v
1
satises the symmetry, v
1
ikl
= v
1
ilk
.
Once v
1
is determined, terms of O(e
0
) in Eq. (20)
become
$
X
C(x) : ($
x
_ _
v
1
(x) 1
(4s)
) : ($
X
u
0
(X)
__
$
x
C(x) : ($
X
_ _
u
1
(X; x)
$
x
u
2
(X; x))
__
:
If the volume average over the unit cell is taken,
terms varying with x drop out, and the governing
equation for u
0
is obtained as
$
X
(C
0
: ($
X
u
0
(X)) = 0 in B; (22)
with
C
0
=
1
U
_
U
C(x) : ($ v
1
(x) 1
(4s)
) dV : (23)
Therefore, a boundary-value problem for u
0
is
obtained if u
0
~ u
e
is assumed and u
0
= u
0
is pre-
scribed as the boundary conditions on oB.
4.2. Macroelds of homogenization theory
Since the leading term of the perturbation ex-
pansion, u
0
, is a function of only X, it corresponds
to the macrodisplacment in the average-eld the-
ory. The next term, eu
1
, contributes little, as it is of
the order of O(e
1
). Indeed, the volume average of
u
1
= v
1
: ($
X
u
0
) taken over U vanishes since v
1
is periodic.
A singular perturbation expansion similar to
Eq. (19) is applicable to the strain and stress elds,
i.e.,
e
; r

n=0
e
n

n
; r
n
. The rst terms of
these expansions are expressed in terms of u
0
and
v
1
as

0
(X; x) = sym$
X
u
0
(X)
sym$
x
v
1
(x) : ($
X
u
0
(X));
(24)
r
0
(X; x) = C(x) : ($
x
v
1
(x) 1
(4s)
)
: ($
X
u
0
(X)); (25)
where sym$
x
v
1
stands for (ov
1
ikl
=ox
j

ov
1
jkl
=ox
i
)=2. Since v
1
is periodic, the volume av-
erages of
0
and r
0
taken over U become

0
)
U
(X) = sym$
X
u
0
(X);
r
0
)
U
(X) = C
0
:
0
)
U
(X); (26)
where C
0
is dened by Eq. (23). These volume
averages correspond to the macrostrain and mac-
rostress of the average-eld theory. That is, if the
strain and stress of O(e
0
) are regarded as the mic-
roelds, the homogenization theory denes the
macroelds as their volume averages, as does the
average-eld theory. Note that
0
and r
0
are ex-
pressed in terms of
0
)
U
and r
0
)
U
as

0
(X; x) =
0
)(X) sym$
x
v
1
(x) :
0
)(X);
r
0
(X; x) = r
0
)(X) (C(X) : (sym$
x
: v
1
(x)
1
(4s)
) C
0
) :
0
)(X):
5. A hybrid micormechanics theory
While the homogenization theory is based on
the singular perturbation of the microelds, the
resulting elds of O(e
0
) and their averages taken
over U correspond to the microelds and the
macroelds of the average-eld theory; see Table 1
for the comparison of the eld variables of O(e
0
)
with the eld variables of the average-eld theory.
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 673
There are, however, two major dierences between
these two theories. The rst dierence is the
modeling of the microstructure: the homogeniza-
tion theory uses a unit cell of the periodic struc-
ture, while the average-eld theory considers a
representative volume element of a statistically
homogeneous body. The second dierence is that
the homogenization theory is able to treat the
macromicro relations more rigorously, allowing
higher-order terms in the singular perturbation
expansion. These dierences are not essential, i.e.,
the homogenization theory can be applied to ma-
terials with a non-periodic microstructure and
higher order terms can be still computed with the
aid of suitable microstructure models using the
average-eld theory. In this section, we propose a
micromechanics theory which is a hybrid of the
homogenization and average-eld theories. Since
tensors of higher order appear, index notation is
mainly used in this section; in particular, $
X
and
$
x
are replaced by D
i
= o=oX
i
and d
i
= o=ox
i
,
respectively.
5.1. More general modeling of microstructure
We consider an elasticity tensor eld C = C(x)
which is not necessarily periodic. The singular
perturbation of u
e
, Eq. (19), is still applicable,
and the assumptions of u
0
i
= v
0
im
u
0
m
(X) and u
1
i
=
v
1
imp
(x)(D
p
u
0
m
(X)) make terms of O(e
2
) and
O(e
1
) vanish, if
7
v
0
ij
= d
ij
and v
1
satises Eq. (21).
Furthermore, assuming that u
2
is given by
u
2
i
= v
2
impq
(D
p
D
q
u
0
m
), where v
2
is a fourth-order
tensor depending only on x, we rewrite terms of
O(e
0
) in Eq. (20) as follows:
C
0/
ijkl
D
i
D
l
u
0
k
(X) R
0
j
(X; x); (+)
where, at this stage, C
0/
is as yet an arbitrary
fourth-order tensor, and R
0
is dened as
R
0
j
(X; x) = d
i
C
ijkl
(x)(d
l
v
2
kmpq
(x)
_ _
v
1
kmq
(x)d
lp
)
_
C
pjkl
(x)(d
l
v
1
kmq
(x) v
0
km
d
lq
) C
/0
pjmq
_
D
p
D
q
u
0
m
(X)
_ _
: (27)
The rst term in (*) yields a governing equation
for u
0
and the second term, R
0
j
, is a residue of this
governing equation.
It is possible to enforce R
0
= 0 and to satisfy
Eq. (20) up to O(e
0
) for any arbitrary C
0/
, since v
2
can be determined such that R
0
vanishes. Ac-
cording to the average-eld theory, however, the
most suitable
8
C
/
is probably given by taking the
volume average of C
ijpq
(d
q
v
1
pkl
v
0
pk
d
lq
), over a
domain in which v
1
is dened.
Let V be a domain in which v
1
is dened. Al-
though V is not necessarily a unit cell of a periodic
structure, we seek to determine v
1
or C
0/
which
relates the average strain to the average strain
energy as well as the average stress, independently
7
Here, v
0
ik
= d
ik
is introduced to make the expression of u
0
consistent with other u
n
s.
8
From mathematical point of view, the most suitable C
0/
can
be dened by minimizing a suitable measure of the perturbation
error of O(e
1
).
Table 1
Comparison of average-eld theory and homogenization theory
Average eld theory Homogenization theory
Microdisplacement u u
0
ev
1
: ($
X
u
0
)
Microstrain = sym$
x
u
0
= 1
(4s)
sym$
x
v
1

_ _
: sym$
X
u
0

Microstress r = C : ($
x
u) r
0
= C : (1
(4s)
$
x
v
1
) : ($
X
u
0
)
Macrodisplacement U =
_
/
V
u dV u
0
Macrostrain E = sym$
x
U
0
) = sym$
X
u
0

Macrostress R = C : E r
0
) = C
0
:
0
)
Eective elasticity C = C
M
f (C
I
C
M
) : A C
0
= C : (1
(4s)
$
x
v
1
))
674 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682
of the boundary conditions prescribed on oV .
According to the universal inequalities (17), the
following two inequalities hold
_
V

0R
: C :
0R
dV
x
6
_
V

0G
: C :
0G
dV
x
6
_
V

0E
: C :
0E
dV
x
;
(28)
where
0E
,
0G
, and
0R
, respectively, are the strain
elds of O(e
0
) when homogeneous strain, general,
and homogeneous stress boundary conditions are
prescribed on oV in such a manner that they
produce the same average strain. These v
1E
, v
1G
,
and v
1R
are given by

0(:)
(X; x) = sym($
x
v
1(:)
(x) 1
(4s)
)
: ($
X
u
0
(X);
and give the micro-scopic strain whose volume
average coincides with sym$
X
u
0
, i.e.,
sym($
x
v
1(:)
1
(4s)
) : ($
X
u
0
))
V
(X)
= sym$
X
u
0
(X): (29)
Hence, the average response of V does not depend
on the prescribed boundary conditions, if
9
value
of average strain energy of V under the homoge-
neous strain boundary conditions, e
E
)
V
, is close to
the value of the average strain energy under the
homogeneous stress boundary conditions, e
R
)
V
,
i.e.,
e
E
)
V
e
R
)
V
e
E
)
V
1; (30)
where
e
(:)
)
V
=
1
2

0(:)
: C :
0(:)
)
V
: (31)
The eective elasticity tensor, C
0/
, can now be
uniquely determined as
C
0/
~ C : ($
x
v
1E
1
(4s)
))
V
or C : ($
x
_
v
1R
1
(4s)
))
V
_
:
(32)
The eective elasticity tensor, C
0
, which is given
by Eq. (23), corresponds to a case when periodic
boundary conditions are prescribed on the boun-
dary of a parallelepiped V = U; see Fig. 5. This
eective elasticity tensor is bounded by two ef-
fective elasticity tensors, C : ($
x
v
1E
1
(4s)
))
V
and C : ($
x
v
1R
1
(4s)
))
V
, which are deter-
mined when homogeneous strain and stress
boundary conditions are prescribed on the paral-
lelepiped U.
For a given (constant) average strain,
0
)
V
=
sym$
X
u
0
, the boundary conditions for v
1E
is
prescribed as v
1E
= 0 on oV . Boundary conditions
for v
1R
, however, are not easily dened; zero
traction boundary conditions, m (C : ($
x
v
1R
))
= 0, are not suitable since the resulting v
1R
does
not satisfy Eq. (29). Taking advantage of the
9
These two boundary data produce elds satisfying Eq. (29),
and hence the average strains are the same.
Fig. 5. Unit cell used in hybrid micromechanics theory.
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 675
linearity, however, we can determine v
1R
using the
solution for the homogeneous stress boundary
conditions. Indeed, let u
R
= u
R
(x; R) be the dis-
placement eld when V is subjected to t = m R on
oV , where R is constant. The volume average of
the associated strain,
R
)
V
= sym$
x
u
R
)
V
, is
linearly related to the prescribed stress, R, which
is the same as the average stress. Hence, the in-
verse of a fourth-order tensor which relates R to

R
)
V
is the eective elasticity tensor, and v
1R
is
given by
v
1R
(x) :
R
)
V
= u
R
(x; R) x
R
)
V
:
This v
1R
satises Eq. (29), if R is chosen such that

R
)
V
=
0
)
V
or sym$
X
u
0
.
5.2. Treatment of higher-order terms
It is of interest to examine the eects of the
higher order terms; in particular, the second-order
terms which, as shown later, are associated with
the eects of the macroscopic strain gradients. It
should be noted that when the material is linearly
elastic, higher order terms in the perturbation ex-
pansion may not be signicant, since the volume
averages of these terms taken over V vanish. In
non-linear cases, however, there are cases when
such higher order terms make signicant contri-
butions.
The homogenization theory usually assumes
that the nth order term, u
n
i
, is given by
u
n
i
(X; x) = v
n
imp
1
p
2
...p
n
(x)D
p
1
D
p
2
. . . D
p
n
u
0
m
(X);
and derives a boundary-value problem for v
n
from
the terms of O(e
n2
) in Eq. (20) using periodic
boundary conditions; see Appendix B for a brief
summary of the treatment of such higher order
terms in the homogenization theory. For instance,
when u
2
i
= v
2
impq
(x)D
p
D
q
u
0
m
(X) is included, the
terms of O(e
1
) are written as
C
1
ijmpq
D
i
D
p
D
q
u
0
m
(X) R
1
j
(X; x);
where C
1
is
C
1
ijmpq
= C
ijkl
(d
l
v
2
kmpq
v
1
kmp
d
lq
))
U
; (33)
and R
1
is similar in form to R
0
. In order to satisfy
Eq. (20) up to O(e
1
), u
0
must satisfy
C
0
ijkl
D
i
D
l
u
0
k
(x) eC
1
ijklm
D
i
D
l
D
m
u
0
k
(x) = 0 in B:
The homogenization theory takes another (regu-
lar) perturbation expansion for u
0
,
u
0
(X) ~ U
0
(X) eU
1
(X) ; (34)
and obtains a set of governing equations for U
0
and U
1
as follows:
C
0
ijkl
D
i
D
l
U
0
k
(X) = 0 in B; (35)
C
0
ijkl
D
i
D
l
U
1
k
(X) C
1
ijklm
D
i
D
l
D
m
U
0
k
(X) = 0 in B:
(36)
Since u
e
~ u
0
eu
1
is replaced by U
0
e(U
1
v
1
:
($
X
U
0
)), the boundary conditions for U
0
and
U
1
are U
0
= u
0
and U
1
= 0 on oB. Hence, once U
0
is obtained by solving Eq. (35), U
1
is determined
by solving Eq. (36) with U
1
= 0 on oU.
As shown in the preceding subsection, the pe-
riodicity of C is not essential. The governing
equations for the higher order terms, v
n
s, remain
the same if the domain V is used as a microstruc-
ture model. For instance, the governing equation
for v
2
is derived from R
0
= 0. In view of Eq. (27),
they are expressed in terms of v
1
as follows
d
i
C
ijkl
(x)d
l
v
2
kmpq
(x)
_ _
d
i
(C
ijkp
(x)v
1
kmq
(x))
_
C
pjkl
(x)(d
l
v
1
kmq
(x) C
pjmq
(x) C
0/
pjmq
_
= 0:
(37)
Note that C
0
is replaced by C
0/
which is given by
Eq. (32).
If m, p, and q are xed, v
2
kmpq
is the displacement
eld of V when body forces are prescribed by the
terms in the parenthesis. Since V is chosen such
that the dependence of the average response on
prescribed boundary conditions is negligibly small,
we can determine v
2
kmpq
assuming either zero
displacements (v
2
kmnpq
= 0) or zero tractions
(n
i
C
ijkl
d
l
v
2
kmpq
= 0) on oV . Such v
2E
or v
2R
replaces
C
1
in Eq. (33) which is computed for a unit cell;
676 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682
for instance, when zero displacement boundary
conditions are used,
C
1/
ijmpq
= C
ijkl
(d
l
v
2E
kmpq
v
1E
kmp
d
lq
))
V
: (38)
Here, superscript E emphasizes that v
1
and v
2
are
for zero displacement boundary conditions.
Once C
0/
and C
1/
are determined, the governing
equations for U
0
and U
1
are rewritten by replacing
C
0
and C
1
with C
0/
and C
1/
, respectively. That is,
C
0/
ijkl
D
i
D
l
U
0
k
(X) = 0 in B; (39)
C
0/
ijkl
D
i
D
l
U
1
k
(X) C
1/
ijklm
D
i
D
l
D
m
U
0
k
(X) = 0 in B:
(40)
While it is possible to solve these equations in
a recursive manner and to determine u
e
~ U
0

e(U
1
v
1
: ($
x
U
0
)), we derive instead a gov-
erning equation for u
e
which accounts for the
eects of higher order terms. To this end, assuming
that C
0/
is invertible, we rst rewrite the left side of
Eq. (40) as
D
i
C
0/
ijhp
D
t
d
pt
U
1
h
_
((C
0/
)
1
: C
1/
)
hprst
(D
s
U
0
r
)
_
:
We seek to obtain a suitable function B which
satises D
i
(C
0/
ijhp
D
t
B
hpt
) = 0. This B gives U
1
as
d
pt
U
1
h
= B
hpt
((C
0/
)
1
: C
1/
)
hprst
D
s
U
0
r
, or, sum-
ming over p and t,
U
1
i
(X) =
1
3
B
ipp
(X)
_
((C
0/
)
1
: C
1/
)
ipklp
(D
l
U
0
k
(X))
_
: (41)
According to the standard procedure of the
homogenization theory, B should be determined
such that the boundary conditions prescribed for
U
1
are satised. However, according to the aver-
age-eld theory which regards a macroscopic eld
variable as the local average of the corresponding
microscopic variable, it is more natural to locally
relate U
1
to the gradient of U
0
, neglecting B in Eq.
(41). That is, to set
U
1
i
(X) ~ N
ikl
(D
l
U
0
k
(X)); (42)
where N is a third-order constant tensor given by
N
ikl
= ((C
0/
)
1
: (C
1/
))
ipklp
= (C
0/
)
1
ipmn
C
1/
mnklp
: (43)
Since C
0/
and C
1/
are determined from v
1
and v
2
,
this
10
N can be explicitly determined once the
microstructure model is established.
In view of Eq. (42), we obtain a relation be-
tween u
e
and U
0
, which accounts for the second-
order terms, i.e.,
u
e
i
(X) ~ U
0
i
(X) e N
ikl
v
1
ikl
(x)
_ _
(D
l
U
0
k
(X)) :
(44)
It is straightforward to obtain the inverse relation
between u
e
and U
0
up to O(e), as
U
0
i
(X) ~ u
e
i
(X) e N
ikl
_
v
1
ikl
(x)
_
(D
l
u
e
k
(X) :
(45)
In this equation, v
1
is the only term which changes
within a micro-length scale, or the only function
which depends on x. If, say, the volume average
of v
1
is taken over V, then, the right side of
this equation is approximated by u
e
i
e(N
ikl

v
1
ikl
))(D
l
u
e
k
), a function of X only. Hence, the
following equations are obtained from Eq. (39):
D
i
C
0/
ijkl
D
l
u
e
k
(X)
_
eC
/
ijklmn
D
l
D
m
u
e
k
(X)
_
= 0 in B;
(46)
where C
/
is dened by
C
/
ijklm
= C
0/
ijpl
(N
pkm
v
1
pkm
)
V
); (47)
and satises the symmetry property,
C
/
ijklm
= C
/
ijkml
= C
/
ijlkm
; C
/
ijklm
= C
/
jiklm
:
While eC
/
ijklm
D
l
D
m
u
e
k
in Eq. (46) are mathemat-
ically derived from the perturbation analysis, these
terms can be interpreted as stress due to the strain
gradient; the left side of Eq. (46) is rewritten as
D
i
r

ij
, where r

ij
is given by
r

ij
(X) = C
/
ijkl

e
kl
(X) eC
/
ijklm

e/
klm
(X): (48)
10
By denition, N satises the symmetry property,
N
ikl
= N
ilk
.
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 677
Here,
e
ij
is the strain associated with u
e
(symD
i
u
e
j
), and
e/
ijk
is the gradient of the strain
given by

e/
ijk
(X) =
1
3
D
i
D
j
u
e
k
(X)
_
D
j
D
k
u
e
i
(X)
D
k
D
i
u
e
j
(X)
_
: (49)
Fig. 6 summarizes the hybrid micromechanics
theory presented in this section.
It should be noted that both C
0/
and C
/
are
fully determined from the microstructure model.
The eects of C
/
, however, can be neglected as
they are proportional to e. In addition to this, C
/
is
small for linear problems, because: (1) N is small
since it gives the overall displacement when a lin-
early varying strain eld with zero volume average
is prescribed; and (2) v
1
)
V
is small since it gives
the average displacement when the microstructure
model is subjected to zero displacements (or
traction) boundary conditions. Therefore, it is
concluded that the eects of C
/
or the strain gra-
dient can be omitted unless the microstructure
model is non-linear and large strain gradients are
presented.
Fig. 6. Summary of hybrid micromechanics theory.
678 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682
6. Periodic boundary conditions
When the microstructure model V is given,
there arises the question of what boundary con-
ditions should be used to determine the eective
moduli of the rst and higher orders. As shown in
the preceding section, we may use the homoge-
neous stress and strain boundary conditions such
that the solutions of the resulting macroeld
equations can overestimate or underestimate
11
macroscopic responses which are obtained by
using other eective moduli, since these boundary
conditions provide upper and lower bounds for
all possible eective moduli. If the dierence of
these upper and lower bounds are large, however,
we may use the periodic boundary conditions.
This is because they provide the eective moduli
which represent those obtained by using a certain
class of boundary conditions. This property of the
periodic boundary conditions is shown in this
section.
For simplicity, consider a cubic representative
volume element, denoted by U, with the edge
length 2p. Let u
P
and t
P
be the surface displace-
ment and traction when periodic boundary con-
ditions are prescribed, and the resulting average
strain, stress, and strain energy are denoted by E,
R, and e
P
)
U
. Due to the uniqueness of the solu-
tion, displacement boundary conditions of u = u
P
on oU or traction boundary conditions of t = t
P
on oU yield the same response when the periodic
boundary conditions are prescribed.
First, consider a disturbance in displacement
boundary conditions, u = u
P
du on oU. The
strain energy due to du is expressed as
e
P
)
U
de)
U
d
2
e)
U
;
where de)
U
and d
2
e)
U
are functions of du, de-
ned by
de)
U
(du) =
1
U
_
oU
t
P
du dS; (50)
d
2
e)
U
(du) =
1
2U
_
oU
t(du) du dS; (51)
with t(du) being the surface tractions when the
boundary conditions are u = du on oU. Due to
the periodicity of t
P
, the functional de)
U
vanishes
for the disturbance displacement of the following
form
du
P
(x) =

M
m
1
;m
2
;m
3
=M
du
m
exp(m x) on oU;
where du
m
s are arbitrary constants and M is an
arbitrary number. Indeed, at, say, (p; x
2
; x
3
) and
(p; x
2
; x
3
) on oU, the integrand satises
(t
P
u
P
)(p; x
2
; x
3
) = (t
P
u
P
)(p; x
2
; x
3
);
since the stress and displacement take the same
value due to the periodicity but the unit normals
are in the opposite direction; see Nemat-Nasser
and Hori (1993), Section 12. This du
P
does not
change the average strain since it is periodic. It
follows from d
2
e)
U
> 0 that the displacement
boundary conditions of u = u
P
on oU yield the
minimum strain energy among the class of dis-
placement boundary conditions of u = u
P
du
P
on oU.
Next, consider a class of traction boundary
conditions which disturb traction boundary con-
ditions, t = t
P
on oU. These boundary conditions
produce the same eld variables as the periodic
boundary conditions. Let dt be a disturbance,
and t = t
P
dt on oU be the traction boundary
conditions. The strain energy is given as e
P
)
U

de
/
)
U
d
2
e
/
)
U
, where
de
/
)
U
(dt) =
1
U
_
oU
dt u dS; (52)
d
2
e
/
)
U
(dt) =
1
2U
_
oU
dt u(dt) dS; (53)
with u(dt) being the surface displacements when
the boundary conditions are t = dt (rigid-body
motion is excluded). The periodicity of u
P
x E
leads to de
/
)
U
= 0 for dt of the following form
dt
P
(x) = m(x)

M
m
1
;m
2
;m
3
=M
dr
m
exp(m x)
_ _
on oU;
where dr
m
s are arbitrary except for dr
0
= 0. This
dt
P
does not change the average stress. Hence, the
11
See Munashinghe et al. (1996).
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 679
boundary condition t = t
P
on oU yields the mini-
mum strain energy for this class of traction-
boundary conditions. Equivalently, t = t
P
yields
the maximum strain energy of the class of traction
boundary conditions t = R
/
t
P
on oU, where
R
/
is chosen such that the resulting average strain
is E.
Now, denote by e
E
)
U
and e
R
)
U
the average
strain energy due to the homogeneous strain and
stress boundary conditions which produce the
same average strain, E. It then follows from the
two universal theorems that for the above arbi-
trarily disturbed boundary conditions which pro-
duce the average strain E, the resulting average
strain energy is bounded as follows
e
R
)
U
6e
P
)
U
d
2
e
/
)(dt
P
) 6e
P
)
U
6e
P
)
U
d
2
e)(du
P
) 6e
E
)
U
; (54)
see Fig. 7 for the schematic view of these in-
equalities. In this sense, we regard the periodic
boundary conditions as the boundary conditions
that represent
12
the two classes of the boun-
dary conditions, u = u
P
du
P
and t = t
P
dt
P
on oU.
7. Concluding remarks
It is shown that the two micromechanics theo-
ries, the average-eld theory and the homogeni-
zation theory, can be related to each other, even
though they are based on dierent concepts. In
particular, the rst order terms in the expanded
strain and stress elds of the homogenization
theory correspond to the average strain and stress
considered in the average-eld theory. Taking
advantage of this correspondence, we combine the
two theories to obtain a hybrid micromechanics
theory, which is applicable to a representative
volume element, while, at the same time, it allows
us to compute the overall properties more accu-
rately than possible with the average-eld theory.
The new resulting eld equations for the macro-
displacement in this hybrid theory include the ef-
fects of the macrostrain gradient in a natural
manner.
Appendix A. Strain concentration of average-eld
theory
Suppose that V is a multi-phase composite,
consisting of N phases, each with a distinct elas-
ticity tensor C
a
, and a matrix phase of the elas-
ticity tensor C
M
. If V is statistically homogeneous,
and
E
: C :
E
)
V

R
: C :
R
)
V
is negligibly
small for the same average strain
E
) =
R
), the
eective elasticity tensor C is determined by con-
sidering a response of V subjected to, say, the
homogeneous strain boundary conditions. By
Fig. 7. Dependence of strain energy on boundary data.
12
The strict mean of the strain energy is dened as
(e
E
)
U
e
R
)
U
)=2. Depending on the microstrcture, e)
U
pre-
cisely satises e)
U
= (e
E
)
U
e
R
)
U
)=2, though this equality
does not always hold.
680 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682
denition, C gives the relation between the average
strain and stress, and the following exact equality
holds
(C C
M
) : )
V
=

N
a=1
f
a
(C
a
C
M
) : )
a
;
where f
a
and )
a
are the volume fraction and the
volume average of the a-th phase. Hence, if the
strain concentration of the ath phase is given by
)
a
= A
a
: )
V
, then the eective elasticity tensor
becomes
C = C
M

N
a=1
f
a
(C
a
C
M
) : A
a
:
Appendix B. Higher order terms in homogenization
theory
If u
n
m
(= v
n
mkp
1
...p
n
(D
p
1
. . . D
p
n
u
0
m
))s are used, terms
of O(e
n
) in the singular perturbation expansion of
the governing equation become
C
n
ijmp
1
...p
n1
D
i
D
p
1
. . . D
p
n1
u
0
m
R
n
j
;
where
C
n
ijmp
1
...p
n1
= C
ijkl
(d
l
v
n1
kmp
1
...p
n1
v
n
kmp
1
...p
n
d
lp
n1
))
U
R
n
j
= d
i
C
ijkl
(d
l
v
n2
kmp
1
...p
n2
_ _
v
n1
kmp
1
...p
n1
d
lp
n2
)
_
d
ip
n2
C
ijkl
(d
l
v
n1
kmp
1
...p
n1
_
v
n
kmp
1
...p
n
d
lp
n1
)
_
d
ip
n2
C
n
p
n2
jp
1
...p
n1
_
D
p
1
. . . D
p
n2
u
0
m
:
The (n 4)th order tensor C
n
gives a governing
equation for u
0
up to O(e
n
), i.e.,

n
m=0
e
m
C
m
ijnp
1
...p
m1
D
i
D
p
1
. . . D
p
m1
u
0
n
= 0:
Terms in the parenthesis of the right side of R
n
s
equations yield a governing equation for v
n2
, and
recursive formulae for v
n
s are naturally obtained.
It is easily seen that v
n
s have symmetry properties
for their suces; for instance, v
1
satises
v
1
mkp
= v
1
mpk
as it gives a displacement component in
the x
m
-direction when an (x
k
,x
p
)-component of
macrostrain is presented. Similarly, v
2
satises
v
2
mkpq
= v
2
mkqp
= v
2
mpkq
.
References
Bakhvalov, N., Panasenko, G., 1984. Homogenization: Aver-
aging Processes in Periodic Media. Kluwer, New York.
Francfort, G.A., Murat, F., 1986. Homogenization and optimal
bounds in linear elasticity. Archive Rat. Mech. and
Analysis 94, 307334.
Hashin, Z., Shtrikman, S., 1962. On some variational principles
in anisotropic and nonhomogeneous elasticity. J. Mech.
Phys. Solids 10, 335342.
Hill, R., 1963. Elastic properties of reinfoced solids: Some
theoretical principles. J. Mech. Phys. Solids 11, 357372.
Hori, M., Nemat-Nasser, S., 1993. Double-inclusion model and
overall moduli of multi-phase composites. Mech. Mat. 14,
189206.
Hornung, U. (Ed.), 1996. Homogenization and Porous Media.
Springer, Berlin.
Kevorkina, J., Cole, J.D., 1996. Multiple Scale and Singular
Perturbation Methods. Springer, Berlin.
Milton, G.W., Kohn, R., 1988. Variational bounds on the
eective moduli of anisotropic composites. J. Mech. Phys.
Solids 43, 63125.
Munashinghe, H.M.S., Hori, M., Enoki, Y., 1996. Applica-
tion of HashinShtrikman variational principle for
computing upper and lower approximate solutions of
elastoplastic problems. In: Proceedings of the Interna-
tional Conference on Urban Engineering in Asian Cities.
pp. 16.
Mura, T., 1987. Micromechanics of Defects in Solids. Martinus
Nijho, New York.
Nemat-Nasser, S., Hori, M., 1993. Micromechanics: Overall
Properties of Heterogeneous Materials. North-Holland,
London.
Nemat-Nasser, S., Hori, M., 1995. Universal bounds for overall
properties of linear and nonlinear heterogeneous solids.
Trans. ASME 117, 412422.
Nuna, K.C., Keller, J.B., 1984. Eective elasticity tensor of a
periodic composite. J. Mech. Phys. Solids 32, 259280.
Oleinik, O.A., Shamaev, A.S., Yosian, G.A., 1992. Mathe-
matical Problems in Elasticity and Homogenization.
North-Holland, New York.
Sanchez-Palencia, E., 1981. Non-homogeneous Media and
Vibration Theory. Lecture Note in Physics 127, Springer,
Berlin.
Terada, K., Miura, T., Kikuchi, N., 1996. Digital image-based
modeling applied to the homogenization analysis of
composite materials. Comput. Mech., 188202.
Torquato, S., 1991. Random heterogeneous media: Microstruc-
ture and improved bounds on eective properties. Appl.
Mech. Rev. 42 (2), 3776.
M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682 681
Walker, K.P, Freed, A.D., Jordan, E.H., 1991. Microstress
analysis of periodic composites. Composite Engrg. 1, 29
40.
Walpole, L.J., 1969. On the overall elastic moduli of composite
materials. J. Mech. Phys. Solids 17, 235251.
Willis, J.R., 1977. Bounds and self-consistent estimates for the
overall properties of anisotropic composites. J. Mech.
Phys. Solids 25, 185202.
682 M. Hori, S. Nemat-Nasser / Mechanics of Materials 31 (1999) 667682