International Journal of Solids and Structures 40 (2003) 1681–1712

www.elsevier.com/locate/ijsolstr

Homogenization of two-phase elasto-plastic

composite materials and structures
Study of tangent operators, cyclic plasticity and numerical
algorithms
I. Doghri *, A. Ouaar
CESAME, Universit

e catholique de Louvain (UCL), B^
atiment Euler, 4 Avenue G., Lema^ıtre, B1348 Louvain-la-Neuve, Belgium
Received 7 March 2002; received in revised form 5 December 2002

Abstract
We develop homogenization schemes and numerical algorithms for two-phase elasto-plastic composite materials and
structures. A Hill-type incremental formulation enables the simulation of unloading and cyclic loadings. It also allows
to handle any rate-independent model for each phase. We study the crucial issue of tangent operators: elasto-plastic (or
‘‘continuum’’) versus algorithmic (or ‘‘consistent’’), and anisotropic versus isotropic. We apply two methods of ex-
traction of isotropic tangent moduli. We compare mathematically the stiffnesses of various tangent operators. All rate
equations are discretized in time using implicit integration. We implemented two homogenization schemes: Mori–
Tanaka and a double inclusion model, and two plasticity models: classical J2 plasticity and ChabocheÕs model with
non-linear kinematic and isotropic hardenings. We consider composites with different properties and present several
discriminating numerical simulations. In many cases, the results are validated against finite element (FE) or experi-
mental data. We integrated our homogenization code into the FE program ABAQUS using a user material interface
UMAT. A two-scale procedure allows to compute realistic structures made of non-linear composite materials within
reasonable CPU time and memory usage; examples are shown.

2002 Elsevier Science Ltd. All rights reserved.

Keywords: Composites; Micro-mechanics; Homogenization; Plasticity; Numerical algorithms; Finite elements

1. Introduction

In this paper, we present homogenization models and robust numerical algorithms for two-phase elasto-
plastic composites. We use the generic term of ‘‘inclusions’’ to designate the reinforcing phase. The in-
clusions can be particles, fibers (long or short) or platelets. Actually, all those shapes can be generated by

*
Corresponding author. Tel.: +32-10-478042/472350; fax: +32-10-472180.
E-mail addresses: [email protected] (I. Doghri), [email protected] (A. Ouaar).

0020-7683/02/$ - see front matter
2002 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0020-7683(03)00013-1
1682 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

setting appropriate values to the aspect ratio of a spheroid (i.e., an ellipsoid with an axis of revolution).
There are many industrial examples of the composites we study in this paper, as illustrated hereafter.

• Polymer matrix composites (PMCs) reinforced with ceramic, glass or Kevlar fibers. Objective: improve
stiffness and strength. Examples: boat hulls, aircraft wings, cars (body frames, hood and door panels),
sporting equipment.
• Polymer matrix with low modulus rubber particles. Objective: improve toughness and impact resistance.
Example: car bumpers.
• Rubber matrix with carbon-black particles. Objective: improve toughness and stiffness. Example: tires.
• Metal matrix composites (MMCs) with ceramic particles or short fibers. Objective: mainly high-tempe-
rature applications. Example: fossil-fuel engine components (e.g., turbochargers).
• Concrete matrix with: metallic fibers (strength in tension or bending), polymer or natural fibers (better
ductility, lower density), rubber inclusions (impact resistance, acoustic isolation).

In all those cases, we wish to predict the influence of the microstructure on the overall properties of the
material or the product. An elegant solution is provided by a micro/macro or two-scale approach with a
macro-scale (that of the body) and a micro-scale (the heterogeneous microstructure). In this paper, tran-
sition between the two scales is made via average-field theories, also known as homogenization models.
One could use another approach: direct FE computation of the boundary-value problem (BVP) at each
representative volume element (RVE). This is feasible for linear elasticity––e.g. (PALMYRA, 2001)––where
the problem is reduced to computing a constant macro-stiffness. The method is also very useful for vali-
dating homogenization models (as we do in this work) or for studying heterogeneous media with complex
material behavior, e.g. (Kouznetsova et al., 2002). However, the direct approach becomes way too ex-
pensive if one or more of the following conditions are met: (1) material or geometric non-linearities; (2)
non-periodic or complex microstructures (e.g., RVEs containing hundreds of fibers in different orienta-
tions); (3) two-scale finite element (FE) simulation of realistic structures (a micro-FE mesh has to be at-
tached to each quadrature point of the macro-mesh). In the latter case, the method requires much more
computer power than what is usually available to engineers or researchers.
For elasto-plastic two-phase composites, a good deal of the literature on homogenization revolves
around the method proposed in (Tandon and Weng, 1988), a Mori–Tanaka (MT) scheme (Mori and
Tanaka, 1973) restricted to J2 elasto-plasticity with a ‘‘secant’’ (or total) deformation formulation. This
precludes the use of other rate-independent models or a simulation of unloading.
Our aim in this work is to develop a formulation and the corresponding numerical algorithms which are
able to simulate within reasonable accuracy, computer time and memory: (1) any rate-independent model
for either phase; (2) cyclic loadings; (3) any multi-axial stress state; (4) structures made of composite
materials.
We address the important issue of tangent operators in elasto-plasticity and their impact on overall
predictions in detail. Indeed, we study anisotropic operators (‘‘continuum’’ and algorithmic), isotropic
moduli (computed with two methods) and the relative stiffnesses of those various operators.
Those issues were not studied in the incremental formulation of MT model which was recently proposed
in (Pettermann et al., 1999). However, that reference deals with thermal strains and contains several in-
teresting numerical simulations for fiber-reinforced materials, while the simulations presented in this paper
are restricted to isothermal conditions and spherical inclusions (validation for other shapes is a work in
progress: (Friebel, 2002; Doghri and Friebel, 2003)).
The paper has the following outline. In Section 2 we present some mathematical notation and results
which are needed in the remainder of the article. We present anisotropic continuum and consistent tangent
operators for two elasto-plastic models (J2 and ChabocheÕs cyclic plasticity) in Section 3. We present two
methods of extraction of the isotropic part of tangent operators in Section 4. A stiffness comparison of
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1683

various tangent operators is carried out in Section 5. We show how to extend homogenization models from
linear elasticity to rate-independent non-linear material behavior in Section 6. The MT model and its
numerical implementation using implicit algorithms are presented in Section 7, while Section 8 gives similar
information for the double inclusion (DI) model. Numerical simulations of three composite materials under
various loadings are presented in Section 9. Two-scale numerical simulations of two composite structures
are presented in Section 10. Finally, conclusions and directions for future work are discussed in Section 11.
In many cases, numerical predictions are compared against FE results or experimental data.

2. Preliminaries

In this section, we define some notations and give some results which are needed later.
Boldface symbols denote tensors, the order of which is indicated by the context. EinsteinÕs summation
convention over repeated indices is used unless otherwise indicated:
X
3
aik bkj
aik bkj
k¼1

Dots and colons are used to indicate tensor products contracted over one and two indices, respectively:
u  v ¼ ui v i ; ða  uÞi ¼ aij uj ;
ða  bÞij ¼ aik bkj ; a : b ¼ aik bki ;
ðC : aÞij ¼ Cijkl alk ; ðC : DÞijkl ¼ Cijmn Dnmkl :
Tensor products are designated by , e.g.,
ðu  vÞij ¼ ui vj ; ða  bÞijkl ¼ aij bkl :
The symbols 1 and I designate the second- and fourth-order symmetric identity tensors, respectively,
1
Iijkl ¼ ðdik djl þ dil djk Þ;
1ij ¼ dij ; ð1Þ
2
where dij is KroneckerÕs symbol,
dij ¼ 1 if i ¼ j; dij ¼ 0 if i 6¼ j: ð2Þ
vol dev
The spherical and deviatoric operators are I and I , respectively,
1
I vol
1  1; I dev
I I vol ; ð3Þ
3
so that for aij ¼ aji we have:
1 1
I vol : a ¼ amm 1; I dev : a ¼ a amm 1
devðaÞ: ð4Þ
3 3
HookeÕs elasticity operator is designated by cel . In the isotropic case, it is given by
cel ¼ 3jI vol þ 2lI dev ; ð5Þ
where j and l are the elastic bulk and shear moduli, respectively.
For C and D any fourth-order tensors, define the following scalar invariant (Bornert et al., 2001b,
p. 178):
C :: D
Cijkl Dlkji ¼ D :: C: ð6Þ
1684 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

Using the definition, the following results are easily obtained:

I vol :: I vol ¼ 1; I dev :: I dev ¼ 5; I vol :: I dev ¼ 0: ð7Þ
Any symmetric and isotropic fourth-order tensor ciso is a linear combination of I and (1  1) or, al-
ternatively, of I vol and I dev . Using the previous results, it is found that ciso can be written as follows (Bornert
et al., 2001b, p. 185):
1
ciso ¼ ðI vol :: ciso ÞI vol þ ðI dev :: ciso ÞI dev : ð8Þ
5
The notation C\ P "D comparing two fourth-order tensors C and D means ‘‘stiffer than’’ and is defined as
follows:
C\ P "D () a : C : a P a : D : a; 8a second-order symmetric tensor ðaij ¼ aji Þ: ð9Þ
In a micro–macro-approach, at each macro-point x, we know the macro-strain ðxÞ and need to
compute the macro-stress rðxÞ, or vice-versa. At micro-level, the macro-point is viewed as the center of a
RVE with domain x and boundary ox. It can be shown (e.g., Section 2 in (Nemat-Nasser and Hori, 1999))
that if linear boundary conditions (BCs) are applied to the RVE, then the problem of relating macro-strains
and stresses  and r can be transformed onto that of relating average strains and stresses hi and hri over
the RVE.
We consider two-phase composites: a number of inclusions (I) in a matrix. The matrix––phase (0)––has
volume V0 and volume fractionP v0 ¼ V0 =V , where V is the volume of the RVE. The inclusions––phase (1)––
have a total volume V1 ¼ I VI and a volume fraction v1 ¼ V1 =V ¼ 1 v0 . Define the following volume
averages:
Z Z
1 1
hf i
f ðx; xÞ dV ; hf ixi
f ðx; xÞ dVi ; i ¼ 0; 1; ð10Þ
V x Vi xi

where integration is carried out w.r.t. micro-coordinates x. In the following, dependence on macro-coor-
dinates x will be omitted for simplicity. It is easy to check that the averages over x (the entire RVE), x0 (the
matrix phase) and x1 (the inclusions phase) are related by
hf i ¼ v1 hf ix1 þ ð1 v1 Þhf ix0 : ð11Þ

The strain averages per phase are related by a strain concentration tensor B as follows:
hix1 ¼ B : hix0 : ð12Þ

Various homogenization models will differ by the expression of B . The per-phase average strains are re-
lated to the macro-strain hi by
1
hix0 ¼ ½v1 B þ ð1 v1 ÞI : hi;
  1
ð13Þ
hix1 ¼ B : ½v1 B þ ð1 v1 ÞI : hi:
Except for the simplest models––Voigt (uniform strain) and Reuss (uniform stress)––the others are based
on the fundamental solution of Eshelby (1957). That solution in turn allows to solve the problem of a single
ellipsoidal inclusion (I) of uniform moduli c1 which is embedded in an infinite matrix of uniform moduli c0 .
Under a remote uniform strain  (i.e., linear boundary displacements), it is found (e.g., Chapter 4 in (Mura,
1987), Section 7 in (Nemat-Nasser and Hori, 1999), Chapter 9 in (Lielens, 1999)) that the strain field inside
the ellipsoid is uniform and related to the remote (or ‘‘macro’’) strain by
ðxÞ ¼ H  ðI; c0 ; c1 Þ :  8x 2 ðIÞ; ð14Þ
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1685

where the single inclusion strain concentration tensor H  has the following expression:
1
H  ðI; c0 ; c1 Þ ¼ fI þ EðI;c0 Þ : ½ðc0 Þ : c1 Ig 1 ; ð15Þ
where EðI; c0 Þ is EshelbyÕs tensor and depends on the geometry of (I) and c0 . For a spheroid (I) and an
isotropic stiffness c0 , E only depends on the aspect ratio and Poisson’s ratio m (Chapter 2 in (Mura, 1987)). If
the moduli are anisotropic, then E is computed numerically (for that purpose, we used a code which was
kindly provided by Lagoudas based on his paper (Gavazzi and Lagoudas, 1990)).
In this paper, we consider two-phase composites where the inclusions (I) have the same shape, orien-
tation and stiffness c1 (more general cases are being implemented and tested: (Friebel, 2002; Doghri and
Friebel, 2003)). For any homogenization model defined by B  ––Eq. (12)––the macro-stiffness c is given by
c ¼ ½v1 c1 : B  þ ð1 v1 Þc0  : ½v1 B  þ ð1 v1 ÞI 1 : ð16Þ
Finally, for time discretization, the symbol D designates an increment over a typical time (or time-like)
interval ½tn ; tnþ1 ,
DðÞ
ðÞnþ1 ðÞn : ð17Þ
For simplicity of notation, the subscript (n þ 1) will be omitted and all variables which do not have this
subscript are computed at tnþ1 .

3. Anisotropic tangent operators

For elasto-plastic materials, one can define at least two tangent operators: a ‘‘continuum’’ and a
‘‘consistent’’ one. Those anisotropic operators can have an important impact on the numerical predictions,
because homogenization models depend explicitly on their expressions.

3.1. Continuum and consistent tangent operators

Using the constitutive equations in rate form, it is possible to relate stress and total strain rates as
follows:
r_ ¼ cep : _ ; ð18Þ
where cep is called the ‘‘continuum’’ (or elasto-plastic) tangent operator. The ‘‘consistent’’ (or algorithmic)
tangent operator is found as follows. First, the rate constitutive equations are discretized in time over each
time interval ½tn ; tnþ1 . Next, the algebraic equations thus found are differentiated w.r.t. all the variables at
tnþ1 and the variations of stress and total strain are related as follows:
drnþ1 ¼ calg : dnþ1 : ð19Þ
Designating by (Dp) the plastic multiplier increment, it is found that:
calg ! cep if Dp ! 0;
but otherwise calg and cep can be quite different. In the context of macro-elasto-plasticity, Simo and Taylor
(1985) have shown that if the global equilibrium equations are solved using NewtonÕs method, then a
quadratic rate of convergence is achieved when using calg instead of cep .
In this paper, we show in Section 5 another result, that is that the elasto-plastic tangent operator cep is
stiffer than the algorithmic tangent calg . This result can have an impact on homogenization results as will be
seen in Section 9.1.
1686 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

3.2. J2 elasto-plasticity

The constitutive equations for classical J2 elasto-plasticity are:

r ¼ cel : ð p Þ; f ¼ req RðpÞ rY 6 0;

of 3 devðrÞ ð20Þ
_ p ¼ p_ N; p_ P 0; p_ f ¼ 0; p_ f_ ¼ 0; N¼ ¼ ;
or 2 req

where rY > 0 is the initial yield stress, p P 0 the accumulated plastic strain, RðpÞ P 0 the hardening stress,
req P 0 the von Mises measure of r and N the normal to the yield surface in stress space. The ‘‘continuum’’
tangent is given by (Section 12.7 in (Doghri, 2000)):
2
ð2lÞ dR
cep ¼ cel N  N; h ¼ 3l þ > 0: ð21Þ
h dp
|{z}
R0 ðpÞ

The ‘‘consistent’’ moduli are given by (Section 12.10.3 in (Doghri, 2000))


alg ep 2 req oN oN 1 3 dev
c ¼c ð2lÞ ðDpÞ tr ; ¼ I N N ; ð22Þ
req or or req 2

with rtreq a trial (elastic predictor) value of req . Note that although the constitutive model is isotropic, both
tensors cep and calg are anisotropic.

3.3. Chaboche’s cyclic plasticity model

We now consider an elasto-plastic model with non-linear kinematic and isotropic hardenings which is
successful in predicting cyclic plasticity of metals, including the Baushinger effect. The model was initially
proposed by Armstrong and Frederick (1966) and later developed and made popular by Chaboche (see
Chapter 5 in (Lemaitre and Chaboche, 1998)). The constitutive equations are:

r ¼ cel : ð p Þ; f ¼ beq RðpÞ rY 6 0;

p
_ ¼ p_ N; X_ ¼ a_ p
bX p_ ;
ð23Þ
of 3 b
p_ P 0; p_ f ¼ 0; p_ f_ ¼ 0; N¼ ¼ ;
or 2 beq

where a > 0 and b P 0 are kinematic hardening parameters, X is a back stress (kinematic hardening),
b
devðrÞ X and beq the von Mises measure of b.
The ‘‘continuum’’ (or elasto-plastic) tangent is given by (Doghri, 1993)
2
ð2lÞ 3
cep ¼ cel N  N; h ¼ 3l þ R0 ðpÞ þ a bN : X: ð24Þ
h 2
The consistent (algorithmic) tangent operator is given by (Doghri, 1993)
2 2
ðDpÞð2lÞ oN ðDpÞð2lÞ b 1 1 3
calg ¼ cmod  Xn ðN : X n ÞN  N; ð25Þ
½1 þ ð3=2Þg ob ½1 þ ð3=2Þg ð1 þ bDpÞ2 beq halg 2
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1687

where cmod is the ‘‘modified’’ continuum tangent operator:

ð2lÞ2
cmod ¼ cel N  N;
halg 
1 3
halg
3l þ R0 ðpÞ þ 2
a bN : X n ; ð26Þ
ð1 þ bDpÞ 2

a Dp
g
2l þ :
1 þ bDp beq
Again, both tangent operators cep and calg are anisotropic.

4. Isotropic part of tangent operators

Numerical experience has shown that good predictions are obtained only when Eshelby’s tensor is
computed not with an anisotropic tangent operator (calg or cep ) but with isotropic moduli (ciso ). The general
expression of the latter is (by design) form-similar to that of HookeÕs operator in isotropic linear elasticity:
ciso ¼ 3jt I vol þ 2lt I dev ; ð27Þ
where jt and lt are ‘‘tangent’’ bulk and shear moduli, resp. Moreover, for a spheroid and isotropic moduli,
EshelbyÕs tensor only needs the ‘‘tangent’’ PoissonÕs ratio mt :
3jt 2lt
mt ¼ ð28Þ
2ð3jt þ lt Þ
The way in which an isotropic part (ciso ) is extracted from an anisotropic operator (cani ) is not unique; we
present hereafter two methods which we tried successfully.

4.1. First method: spectral decomposition

This first method applies to anisotropic tangent operators which are a linear combination of I vol , I dev and
(N  N), where N is typically a normal to a yield surface in stress space and satisfies:
3
Nij ¼ Nji ; Nmm ¼ 0;
N :N ¼ : ð29Þ
2
For J2 plasticity, both tensors cep and calg satisfy all conditions. For ChabocheÕs cyclic plasticity model, the
conditions apply to cep in all cases and to calg only when b ¼ 0 (linear kinematic hardening). For those
anisotropic operators cani , Ponte Casta~ neda (1996) proposes to re-write them as follows:
cani ¼ 3k1 C ð1Þ þ 2k2 C ð2Þ þ 2k3 C ð3Þ ; ð30Þ
where
2
C ð1Þ ¼ I vol ; C ð3Þ ¼ N  N; C ð2Þ ¼ I dev C ð3Þ : ð31Þ
3
Those tensors satisfy the following conditions:
C ð1Þ þ C ð2Þ þ C ð3Þ ¼ I; C ðiÞ : C ðjÞ ¼ dij C ðiÞ ðno sum over iÞ: ð32Þ
This is why Eq. (30) is known as a spectral decomposition of cani . A basic assumption in the method is the
following:
devðDÞ is collinear with N: ð33Þ
1688 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

It is found then that the incremental stress–strain relation Dr  cani : D reduces to Dr  ciso : D, where ciso
is an isotropic operator defined by
j t ¼ k1 ; lt ¼ k3 : ð34Þ
An application to algorithmic moduli calg of J2 plasticity gives:
" #
3l Dp
jt ¼ j; lt ¼ l 1 ; k2 ¼ l 1 3l tr : ð35Þ
h req

There are some comments to be made. First, the k2 term disappeared from ciso and (Dp) with it, therefore
any advantage of consistent vs. continuum tangent is lost. Second, the basic assumption (33) may lead to
bad results for non-proportional loadings. Finally, for ChabocheÕs model, when b 6¼ 0, calg is such that
Eq. (30) does not apply, and using only the other assumption of the method––Eq. (33)––does not give an
isotropic operator.

4.2. Second method: a general definition

This second method is much more general than the first one and can be applied to any anisotropic
operator cani (even if it does not represent material moduli). By analogy with decomposition (8)––which is
exact for isotropic operators––it is suggested to define an isotropic part ciso of cani as follows (Bornert et al.,
2001a, p. 194):
1
ciso
ðI vol :: cani ÞI vol þ ðI dev :: cani ÞI dev : ð36Þ
5
This definition can be coded as an independent module which acts as a post-processor for any constitutive
box, without knowledge of the particular model or code that the box contains. Indeed, one can check the
following expressions:
1
I vol :: cani ¼ 3jt ¼ cani ;
3 lljj ð37Þ
dev ani 1 ani
I :: c ¼ 10lt ¼ cani
illi c :
3 lljj
An application to moduli calg of J2 plasticity gives:
" #
3 2 1 Dp
jt ¼ j; lt ¼ l l þ 4 tr : ð38Þ
5 h req

The expression of lt is different from that found with the first method, Eq. (35). However, our experience so
far has shown that using one definition or the other does not have a significant impact on homogenization
results.
For ChabocheÕs cyclic plasticity model, the isotropic part of calg is determined by its tangent moduli as
follows:
" #
3 2 1 Dp 1
jt ¼ j; lt ¼ l l þ4 : ð39Þ
5 halg beq 1 þ ð3=2Þg

Note that with this second method, lt is a function of the plasticity increment Dp, and therefore the iso-
tropic parts of the consistent and continuum tangent operators are different, even for J2 elasto-plasticity.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1689

5. Stiffness comparison of various tangent operators

We shall see in Section 9.1 that evaluating EshelbyÕs tensor with anisotropic (cep or calg ) or isotropic (ciso )
tangent operators has a major impact on homogenization results. This might be traced back to the relative
stiffnesses of those operators. Therefore, in this section we set up to prove the following inequalities for J2
plasticity:
cel \ P "cep \ P "calg \ P "ciso ; ð40Þ
where the symbol \ P " means ‘‘stiffer than’’ and is defined in Eq. (9). The first inequality (cel \ P "cep ) just
translates the experimental fact that the elastic stiffness is larger than any elasto-plastic tangent; it implies
that h > 0 (see Section 12.7 in (Doghri, 2000)).
We now prove the second inequality (cep \ P "calg ). From Sections 3.2 and 4.1, it is found that (g desig-
nating any second-order symmetric tensor):
!
2 Dp
g : ðc ep
c Þ : g ¼ 6l tr ðg : C ð2Þ : gÞ
alg
ð41Þ
req

This is the product of two terms, the first of which is always positive. As for the second term, introducing
the following decomposition,
devðgÞ ¼ kN þ g? ð42Þ
it can be checked that:
g : C ð2Þ : g ¼ g? : C ð2Þ : g? : ð43Þ
Using the expressions of C ð2Þ , cel and cep , it is found that:

1 ? R0 ðpÞ
g? : C ð2Þ : g? ¼ g : cel 1þ ðcel cep Þ : g? : ð44Þ
2l 3l
In practice, jR0 ðpÞj=ð3lÞ  1, therefore the final result is:
!
Dp
g : ðc ep
c Þ : g  3l tr ðg? : cep : g? Þ;
alg
ð45Þ
req

where the first term is always positive, and the second also for a strain-hardening material (no softening).
Therefore we proved that cep is stiffer than calg .
We now prove the last inequality (calg \ P "ciso ). Using results from Sections 3.2 and 4.1, it is found that:
!
1 Dp
g : ðc alg iso 2
c Þ : g ¼ ð6l Þ ðg : C ð2Þ : gÞ: ð46Þ
h rtreq

In this product of three scalar factors, we have already proven that the last factor is positive. As for the
second factor, it can be rewritten as follows using Eq. (21)b:
1 Dp ðrtreq 3lDpÞ R0 ðpÞDp
¼ : ð47Þ
h rtreq hrtreq

The denominator is positive (since h > 0 and rtreq > 0). As for the numerator, the return mapping algorithm
used to update the constitutive solution (Section 12.10.2 of (Doghri, 2000)) allows to find the following
expression:
1690 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

ðrtreq 3lDpÞ R0 ðpÞDp ¼ rY þ RðpÞ R0 ðpÞDp: ð48Þ

Now a first-order Taylor expansion of RðpÞ around pn gives

RðpÞ  Rðpn Þ þ ðDpÞR0 ðpn Þ: ð49Þ

Consequently, we obtain:

rY þ RðpÞ R0 ðpÞDp  rY þ Rðpn Þ þ ðDpÞ½R0 ðpn Þ R0 ðpÞ > 0: ð50Þ

This is a sum of positive scalars because rY , Rðpn Þ and Dp are all positive and R0 ðpn Þ > R0 ðpÞ from experi-
mental evidence. Therefore, we proved that calg is stiffer than ciso .

6. Incremental formulation of homogenization models

6.1. Reference moduli

An extension of homogenization models from linear elasticity to rate-independent non-linear behavior

(e.g, elasto-plasticity, non-linear elasticity, etc.) is provided by HillÕs so-called incremental formulation
(Hill, 1965). Indeed, stress and strain rates are related by: r_ ¼ c : _ , which is form-similar to linear elasticity
except that we are using rates and a tangent operator c which should not be confused with HookeÕs operator
cel . Moreover, c will not be uniform, because at any time t, it depends on the state of strain or stress. For
history-dependent models, it also depends on internal variables and has different expressions according to
loading or unloading. Therefore, in the matrix of a composite we have:

r_ ðx; tÞ ¼ c0 ððx; tÞ; tÞ : _ ðx; tÞ 8x 2 x0 :

Consequently, it does not make sense for instance to evaluate EshelbyÕs tensor E using c0 ðx; tÞ. A work
around is to consider a fictitious reference matrix which has uniform tangent moduli ^c0 ðtÞ:

r_ ðx; tÞ ¼ ^c0 ðtÞ : _ ðx; tÞ 8x 2 x0 : ð51Þ

The difficult question of course is how to define the reference material and relate ^c0 ðtÞ to the real
composite. Tandon and Weng (1988) developed a secant (or total deformation) formulation of MT model
in which average strains in the matrix phase are used to compute the response of a reference material. This
approach was followed by many researchers in the field. So similarly, in our incremental formulation of
MT, the constitutive box of the real material in each phase is called with the average strains and strain rates
in that phase, and the moduli that the box computes are taken as the reference moduli for that phase.
However, our formulation and implementation are generic and modular enough to allow other reference
moduli to be used in the future as long as they are computable and the final numerical results agree rea-
sonably well with experimental data or FE simulations.
There is no unique definition for a reference material, and there is still active research in this subject (e.g.,
see (Bornert et al., 2001b; Suquet, 1997; Ponte Casta~neda and Suquet, 1998, 2001) and references therein).
However, the suggestions which are available so far in the literature are either valid for a total deformation
theory (which precludes unloading) or hard-wired to J2 elasto-plasticity or are too complex to implement
efficiently.
In Section 9.2 (Fig. 6) there is a comparison between our results and those obtained by SuquetÕs model in
which a reference material is defined by taking the phase-average of the second moments of the stress.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1691

6.2. Time discretization

Rate relations such as the rate version of (12) are discretized over a time interval ½tn ; tnþ1  as follows:
hDix1 ¼ B nþa : hDix0 ;
where a generalized mid-point rule is used:
ðÞnþa ¼ ðÞðt¼tnþa Þ ; tnþa ¼ ð1 aÞtn þ atnþ1 ; a 2 ½0; 1: ð52Þ
Explicit and implicit integrations correspond to a ¼ 0 and a > 0, respectively, with special cases: a ¼ 1
(backward Euler) and a ¼ 1=2 (mid-point rule). In this paper, good predictions were obtained with a ¼ 1=2
(in most cases) or a ¼ 2=3 (which sometimes allows to take larger time increments for a given accuracy).

6.3. Tangent operators

In Sections 7.2 and 8.2, it is important to notice that only Eshelby’s tensor is computed with the isotropic
part of the reference moduli, all other computations are performed with the anisotropic algorithmic moduli.
This insures a good convergence of the whole procedure, i.e.: (i) the homogenization algorithms (Sections
7.2 and 8.2), (ii) the algorithm for macro-stress constraints (Section 6.4) and (iii) the FE computations at
macro-scale (Section 10).

6.4. Macro-stress constraints

Macro-stress constraints are handled as in Section D.2 of (Doghri, 2000). Assume for instance that we
need to simulate a macro-tension/compression test in the one-direction. In this case, at each time tnþ1 , 11 is
known but 22 and 33 need to be computed from the following conditions:
r22 ð22 ; 33 Þ ¼ 0; r33 ð22 ; 33 Þ ¼ 0: ð53Þ
This is a system of two non-linear scalar equations, which is solved iteratively using NewtonÕs method. For
each iteration (m), we have:
   
ðmÞ ðmÞ ðmþ1Þ ðmÞ ðmÞ ðmþ1Þ ðmÞ
rii þ cii22 22 22 þ cii33 33 33 ¼ 0; no sum; ii ¼ 22; 33: ð54Þ
Convergence is achieved if jr22 j and jr33 j are smaller than a tolerance. This means that there is an outer loop
around the homogenization algorithms of Sections 7.2 and 8.2 which for each time interval ½tn ; tnþ1  and
ðmÞ
each iteration (m) passes macro-strain values ij .

7. Mori–Tanaka model

7.1. Formulation

The MT model was proposed by Mori and Tanaka (1973) and is such that the strain concentration
tensor––Eq. (12)––has the following expression (see also Chapter 9 in (Lielens, 1999)):
B ¼ H  ðI; c0 ; c1 Þ; ð55Þ
where operator H  is that of the single inclusion problem, Eq. (15). Consequently, Benveniste (1987) has
proposed the following useful interpretation of the MT model: each inclusion (I) behaves like an isolated
inclusion in the matrix seeing hix0 as a far-field strain. The incremental extension of MT to non-linear rate-
independent models gives:

h_ ix1 ¼ H ðI;^c0 ;^c1 Þ : h_ ix0 : ð56Þ

1692 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

BenvenisteÕs interpretation of this result would be that each inclusion (I) behaves like an isolated in-
clusion in the reference matrix seeing h_ ix0 as a far-field strain rate. The macro-tangent moduli c in MT
model will be found as follows:
hr_ i ¼ c : h_ i;
h i h i 1 ð57Þ
c ¼ v1^c1 : H ðI;^c0 ;^c1 Þ þ ð1 v1 Þ^c0 : v1 H ðI;^c0 ;^c1 Þ þ ð1 v1 ÞI :

7.2. Numerical implementation

Consider a time interval ½tn ; tnþ1 . The data are: macro-total strains n and D, and history variables at tn .
The problem is to compute macro-stress r and macro-tangent moduli c.

• Initialization: hDix1 D.

• Iterations (i) (upper index (i) omitted for simplicity):
11. Call constitutive box of real inclusions material with average strains in the inclusions phase hn ix1
and hDix1 as arguments. Take the algorithmic (and anisotropic) moduli ^c1 that the box returns
as reference moduli for the inclusions phase.
12. Compute average strain in matrix phase:
D v1 hDix1
hDix0 ¼ :
1 v1
13. Call constitutive box of real matrix material with average strains in the matrix phase hn ix0 and
hDix0 as arguments. Take the algorithmic (and anisotropic) moduli ^c0 that the box returns as refe-
rence moduli for the matrix phase.
14. Extract isotropic part ^ciso
0 from reference matrix moduli ^
c0 .
15. Compute EshelbyÕs tensor with those isotropic moduli: EðI; ^ciso
0 Þ.
16. Compute values at tnþa of the anisotropic moduli ^c0 and ^c1 :
^crðnþaÞ ¼ ð1 aÞ^crðnÞ þ a^cr ; r ¼ 0; 1; a 20; 1
(Our experience so far showed that robust and accurate numerical simulations are obtained with
a ¼ 1=2 (mid-point rule) or a ¼ 2=3.)
17. Compute strain concentration tensor
1 1
B  ¼ fI þ E : ½ð^c0ðnþaÞ Þ : ^c1ðnþaÞ Ig :
18. Check compatibility of average strain in inclusions phase by computing residual:
1
R ¼ B  : ½v1 B  þ ð1 v1 ÞI : D hDix1 :
19. If jRj 6 TOL, then exit the loop.
10. Else: new iteration (go to step 1) with new hDix1 :
hDix1 hDix1 þ nR; n 20; 1
(n is a line search parameter; all the numerical simulations that we ran so far used a value of n ¼ 1.)
• After convergence, compute macro-moduli cnþa and macro-stress increment Dr:
1
cnþa ¼ ½v1^c1ðnþaÞ : B þ ð1 v1 Þ^c0ðnþaÞ  : ½v1 B  þ ð1 v1 ÞI ;

Dr ¼ cðnþaÞ : D:
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1693

8. Double inclusion model

8.1. Formulation

The DI model was proposed by Nemat-Nasser and Hori (1999, Section 10)––see also Chapter 9 in
(Lielens, 1999)––and supposes that each spheroidal inclusion (I)––of stiffness c1 ––is wrapped with a hollow
inclusion (I0 ) of stiffness c0 . The outer material has a stiffness cR . Inclusions (I) and (I0 ) have the same aspect
ratio, symmetry axis and center, and their volume ratio equals that of inclusions and matrix in the actual
composite (v1 =v0 ).
By changing the stiffness cR of the outer material, one can retrieve many homogenization models. The
choice cR ¼ c gives the self-consistent model. A second choice is cR ¼ c0 , the stiffness of the real matrix
material. In this case, it can be shown that the strain concentration tensor––Eq. (12)––is given by
B ¼ H  ðI; c0 ; c1 Þ
B l ; ð58Þ
i.e., the MT model. A third choice is cR ¼ c1 , the stiffness of the real inclusions (I). In this case, it is found
that:
1
B ¼ ½H  ðI; c1 ; c0 Þ
B u ; ð59Þ

and this can be called the inverse MT model, as it corresponds to MT for a composite where the material
properties of the inclusions and the matrix are permutated. It can be shown that Bl and B u correspond to
lower and upper stiffness estimates, respectively, which are closely related to the Hashin–Shtrikman bounds
(Hashin and Shtrikman, 1963). Consequently, Lielens (Chapter 9 in (Lielens, 1999)) proposed a homo-
genization model which is based on the following interpolation:
1 1 1
B ¼ ½ð1 fðv1 ÞÞðBl Þ þ fðv1 ÞðB u Þ  ; ð60Þ
where fðv1 Þ is a smooth interpolation function which satisfies:
df
fðv1 Þ > 0; ðv1 Þ > 0; lim fðv1 Þ ¼ 0; lim fðv1 Þ ¼ 1: ð61Þ
dv1 v1 !0 v1 !1

Lielens proposed the following expression for fðv1 Þ:

1
fðv1 Þ ¼ v1 ð1 þ v1 Þ: ð62Þ
2
In the remainder of the paper, ‘‘DI’’ means LielensÕ interpolation model thus defined.

8.2. Numerical implementation

The implementation of DI model is very similar to that of MT, and the only changes in the algorithm of
Section 7.2 are in steps 4, 5 and 7 which now become as follows:

4. Extract isotropic parts ^ciso ciso

0 and ^1 from reference moduli of the matrix and inclusions, ^
c0 and ^c1 , respec-
tively.
5. Compute EshelbyÕs tensors with those isotropic moduli:
E0
EðI; ^ciso
0 Þ; E1
EðI; ^ciso
1 Þ:

7. Compute strain concentration tensors:

1 1
Bl ¼ fI þ E0 : ½ð^c0ðnþaÞ Þ : ^c1ðnþaÞ Ig ;
1694 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

1
B u ¼ I þ E1 : ½ð^c1ðnþaÞ Þ : ^c0ðnþaÞ I;

1 1 1
B  ¼ ½ð1 fðv1 ÞÞðB l Þ þ fðv1 ÞðBu Þ  :

9. Numerical simulations of RVEs

We developed a code called DIGIMAT (for ‘‘Digital Materials’’) in which we implemented two homo-
genization schemes: MT and DI and two material models: J2 elasto-plasticity and ChabocheÕs cyclic plas-
ticity. Those models can be used for any phase of a composite material. The inclusions can be spheres or
spheroids with any aspect ratio. All RVE simulations with DIGIMAT reported in this paper took less than
1 s of CPU time on an ordinary PC. For validation purposes, we conducted direct FE computations on a
unit cell using ABAQUS (2001)––see Appendix A.

9.1. Time-integration and tangent operators

Figs. 1–3 show numerical simulations with MT of macro-tension tests under imposed macro-strain for a
MMC with a J2 elasto-plastic matrix and v1 ¼ 20% of elastic spherical inclusions (complete material data
are given in Section 9.2). Figs. 1 and 2 are obtained with EshelbyÕs tensor computed with tangent moduli
^calg
0 for two values of the integration parameter: a ¼ 1 (backward Euler) in Fig. 1 and a ¼ 1=2 (mid-point
rule) in Fig. 2. It is seen that in the latter case, the results are insensitive to the value of the macro-strain
increment. Fig. 3 compares the results obtained using EshelbyÕs tensor computed with the anisotropic
tangent moduli ^cep calg
0 (elasto-plastic) or ^ 0 (algorithmic) to those obtained from FE computation. It is seen
that the overall response is much stiffer than it should. However, Fig. 3 shows that if Eshelby’s tensor is

MMC, vol. frac. = 20%, alpha = 1, anisotropic consistent tangent

400

350

300
Macro stress [MPa]

250

200

150 strain inc. = 1.e-3

1.e-3
2.e-3
100

50

0
0 0.005 0.01 0.015 0.02 0.025
Macro strain

Fig. 1. MMC (v1 ¼ 20%) under tension. MT results when a ¼ 1 and EshelbyÕs tensor is computed with anisotropic consistent tangent.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1695

MMC, vol. frac. = 20%, alpha = 0.5, anisotropic consistent tangent

400

350

300
Macro stress [MPa]

250

200

150 strain inc. = 1.e-3

1.e-3
2.e-3
100

50

0
0 0.005 0.01 0.015 0.02 0.025
Macro strain

Fig. 2. MMC (v1 ¼ 20%) under tension. MT results when a ¼ 0:5 and EshelbyÕs tensor is computed with anisotropic consistent tangent.

MMC, vol. frac. = 20%, alpha = 0.5, macro strain inc. = 3.e-3
500

450 anisotropic continuum tangent

anisotropic consistent tangent
isotropic tangent
400 finite elements (ABAQUS)

350
Macro stress [MPa]

300

250

200

150

100

50

0
0 0.02 0.04 0.06 0.08 0.1 0.12
Macro strain

Fig. 3. MMC (v1 ¼ 20%) under tension. MT results when a ¼ 0:5 and EshelbyÕs tensor is computed with: (1) anisotropic continuum
tangent, (2) anisotropic consistent tangent and (3) isotropic tangent. Comparison with target FE unit cell results.
1696 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

computed with an isotropic part of ^calg

0 , then a good agreement with FE results is found. Those facts have
already been recorded in the literature (e.g., Bornert et al., 2001b; Gonzalez and Llorca, 2000; Masson
et al., 2000; Ponte Casta~neda and Suquet, 2001; Suquet, 1997 and references therein), but there is still no
satisfying explanation for neither the problem nor the work around. It is difficult to tackle this problem
because the overall stiffness is a non-linear function of the strain concentration tensor, which itself is a non-
linear function of EshelbyÕs tensor, and the latter can only be computed numerically if the moduli are
anisotropic.
Our contribution to this issue is the following. We proved in Section 5 that ^cep
0 \ P "^calg ciso
0 \ P "^0 , and it is
seen from Fig. 3 that the predictions of the overall response follow the exact same order. So there is perhaps
a link that one could establish mathematically in the future.
Note that from now on, all elasto-plastic results which are reported in the following sections and figures
were obtained under the following conditions: algorithmic (consistent) tangent operators are used every-
where except in EshelbyÕs tensor which is computed with an isotropic part of the tangent; time-integration
parameter is either a ¼ 1=2 (in most cases) or a ¼ 2=3 (for PMC, this sometimes allows to use larger time
increments without loss in accuracy).

9.2. A metal matrix composite

A metallic matrix (aluminum alloy) has the following properties: E0 ¼ 75 GPa, m0 ¼ 0:3, rY ¼ 75 MPa,
with power-law isotropic hardening (RðpÞ ¼ kpm ) with k ¼ 416 MPa and m ¼ 0:3895. The matrix is
reinforced with spherical ceramic inclusions (E1 ¼ 400 GPa, m1 ¼ 0:2) of volume fraction v1 . This
MMC has also been studied by Suquet (1997, p. 254), Gonzalez and Llorca (2000) and Segurado
et al. (2002).

MMC with spherical inclusions. Macro elastic stiffness with different models.
400

350 Voigt
Double inclusion
Mori-Tanaka
Finite elements (ABAQUS)
300 Reuss
Macro Young’s modulus [GPa]

250

200

150

100

50
0 0.2 0.4 0.6 0.8 1
Volume fraction of inclusions

Fig. 4. MMC: macro-YoungÕs modulus predictions with different models: (1) Voigt (uniform strain), (2) double inclusion, (3) Mori–
Tanaka, (4) finite elements (unit cell) and (5) Reuss (uniform stress).
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1697

Fig. 4 shows predictions of macro-YoungÕs modulus obtained with various homogenization models and
with direct FE computations on a unit cell. Both MT and DI give excellent predictions, although the latter
performs better for higher values of v1 . Note that in Figs. 4 and 10, values of v1 higher than the maximum
packing concentration ðv1 Þmax are not physically meaningful. For spherical inclusions of uniform size,
ðv1 Þmax  0:75.
Fig. 5 shows numerical simulations with MT of macro-tension tests with imposed macro-strain for
various values of v1 . The same tests were simulated with direct FE computations and the results are also
shown in Fig. 5. There is good agreement between the two sets of results.
The same MMC under tension was simulated in (Segurado et al., 2002) for v1 ¼ 30% using several
formulations and the results are plotted in Fig. 6 together with ours. The target results (label 2) were
obtained in (Segurado et al., 2002) by performing 3D FE computations on a cube containing 30 spheres. A
typical mesh had about 60,000 elements and 90,000 nodes. Twelve meshes were generated corresponding to
different arrangements of the spheres and BCs; the average response is plotted (label 2). Results obtained
with our incremental MT formulation are plotted with label 1 and those found with a secant (total-
deformation) formulation of MT have label 3. Finally, results with label 4 were obtained using a formu-
lation proposed by Suquet (1997) in which reference materials are defined by taking the phase-averages of
the second-order moments of the stress. It was shown that this approach is equivalent to Ponte Casta~ nedaÕs
variational formulation (Suquet, 1997; Ponte Casta~ neda, 1996; Ponte Casta~ neda and Suquet, 1998, 2001).
Comparison between the target results (label 2) and the others shows that our predictions (label 1)
overestimate the non-linear entry but then as hardening develops, they become very close to the target. On
the other hand, predictions with SuquetÕs formulation (label 4), match the non-linear curvature closer but
diverge from the target as the strains increase. Finally, the classical secant formulation (label 3) overesti-
mates the stresses at all levels of plastic strains.

MMC under tension: Mori-Tanaka (M.-T.) and finite element (F.E.) results
300
FE 30
MT 30

250 MT 20 FE 20

MT 10 FE 10
MT 5
200 FE 5
Macro stress [MPa]

MT 0 FE 0

150

100 matrix: M.-T. (MT 0), F.E. (FE 0)

v1 = 5%, M.-T. (MT 5), F.E. (FE 5, 204.01 sec)
v1 = 10%, M.-T. (MT 10), F.E. (FE 10,206.32 sec)
v1 = 20%, M.-T. (MT 20), F.E. (FE 20, 207.07 sec)
v1 = 30%, M.-T. (MT 30), F.E. (FE 30, 205.90 sec)
50

0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Macro strain

Fig. 5. MMC under macro-tension. Comparaison between MT and unit cell FE predictions for various volume fractions of inclusions
v1 .
1698 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

MMC (v1 = 30%) under tension: comparison with results reported by Segurado et al. 2002
300 3
1
2
4
250

200
Macro stress [MPa]

150

100 1: M.-T. incremental, this work

2: Average 3D F.E.
3: M.-T. secant, classical
4: M.-T. secant, Suquet
50

0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Macro strain

Fig. 6. MMC (v1 ¼ 30%) under macro-tension. Comparaison of different predictions: (1) incremental MT (this work), (2) average 3D
FEs (Segurado et al., 2002), (3) classical secant MT and (4) secant MT with (Suquet, 1997).

In conclusion, it appears that our incremental MT formulation gives acceptable predictions and that the
accuracy increases with hardening.
In Fig. 7, for v1 ¼ 30%, a cyclic tension/compression test under imposed macro-strain (peak values:
4%) is simulated with three methods: MT, DI and FE. It is seen that the two homogenization models
perform remarkably well, with MT doing better in the first monotonic stage, and DI giving a closer
agreement with FE afterwards. We have done similar simulations for all other values of v1 and the con-
clusions are the same. It is important to notice that our incremental formulation allows to simulate cyclic
plasticity while a secant formulation does not.
In order to further illustrate the robustness of our algorithms, we simulate macro-shear tests in Fig. 8 and
macro-bi-axial plane stress loads in Fig. 9. In the latter case, the initial yield envelope is plotted, and it was
defined as follows. Proportional macro-strain histories 11 ðtÞ and 22 ðtÞ are imposed such that their ratio is
constant (33 ðtÞ is computed by our DIGIMAT code so that r33 ðtÞ ¼ 0). When the macro-stress ratio
r11 ðtÞ=r22 ðtÞ is no longer constant, than initial macro-yield is predicted (this criterion is probably too strong).

9.3. A polymer matrix composite

A polymer matrix (epoxy) has the following properties: E0 ¼ 3:16 GPa, m0 ¼ 0:35, rY ¼ 75:86 MPa, and
power-law isotropic hardening with k ¼ 32:18 MPa and m ¼ 0:26. The properties of the reinforcing
spherical elastic inclusions (in silica) are E1 ¼ 73:1 GPa and m1 ¼ 0:18. This composite was studied by
Tandon and Weng (1988) who used a secant (total deformation) theory; experimental results are also re-
ported in that reference.
Fig. 10 shows predictions of macro-YoungÕs modulus obtained with various homogenization models
and with FE computations. It is seen that the DI model gives excellent agreement with FE, and that the
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1699

MMC (vol. frac = 30%) under cyclic strain. Comparison between different results.
500

400

300

200
Macro stress [MPa]

100

0
Mori-Tanaka
Finite elements (ABAQUS, CPU = 185 sec)
-100 Double inclusion

-200

-300

-400

-500
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
Macro strain

Fig. 7. MMC (v1 ¼ 30%) under cyclic macro-strain (macro-uniaxial stress). Comparaison of different predictions: (1) MT model,
(2) finite elements (unit cell) and (3) interpolative double inclusion model.

MMC under shear. DIGIMAT results with Mori-Tanaka.

160

140

120
Macro shear stress [MPa]

100

80

60
matrix
vol. frac. = 5%
40 10%
20%
30%
20

0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Macro engineering shear strain

Fig. 8. MMC under macro-shear strain (macro-plane stress). MT predictions for various values of inclusionsÕ volume fraction v1 .
1700 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

Metal matrix composite (v = 0.20). Initial yield surface under bi-axial loads.
100

80

60

40
22-stress [MPa]

20

-20

-40 composite (macro)

matrix alone (v = 0)
-60

-80

-100
-100 -80 -60 -40 -20 0 20 40 60 80 100
11-stress [M Pa]

Fig. 9. MMC (v1 ¼ 20%): MT prediction of initial yield surface for proportional bi-axial macro-strains under macro-plane stress
conditions.

difference between MT and DI predictions for high values of v1 is more pronounced than for MMC (the
stiffness ratio E1 =E0 of the two phases is 23.13 for PMC and 5.33 for MMC).
Figs. 11 and 12 show the results of macro-tension tests under imposed macro-strain. Our MT results are
compared against those of Tandon and Weng (1988) in Fig. 11 and excellent agreement is found. In Fig. 12,
our MT predictions are plotted against experimental data, and as in (Tandon and Weng, 1988) an excellent
match is found in all cases except v1 ¼ 0:52.
In Fig. 13, for v1 ¼ 42%, a cyclic tension/compression test under imposed macro-strain (peak values:
4%) is simulated with three methods: MT, DI and FE. In the first monotonic phase of the loading history,
it seen that DI agrees perfectly with FE for the linear elastic response, but predicts an overly stiff plastic
response. The MT model, however, gives a poor prediction of the elastic stiffness but a satisfying simulation
of the plastic response. This observation continues to hold for the subsequent phases of the loading, and
overall MT predicts the plastic response much better than DI. We have done similar simulations for all
other values of v1 and the conclusions are the same. Again, note that the secant formulation is not able to
simulate cyclic plasticity, since in essence it is a non-linear elastic formulation.
Finally, macro-shear tests were simulated and the results obtained with MT are plotted in Fig. 14.

9.4. Chaboche’s cyclic plasticity model

The matrix material considered in this section is a low-carbon (AISI 1010) steel. The experimental data
were first identified with ChabocheÕs cyclic plasticity model with the parameter values reported in (Doghri,
1993): E0 ¼ 210 GPa, m0 ¼ 0:3, rY ¼ 200 MPa, exponential-law isotropic hardening with saturation
(RðpÞ ¼ R1 ½1 expð mpÞ) with R1 ¼ 2 GPa and m ¼ 0:26, non-linear kinematic hardening with para-
meters a ¼ 17 GPa and b ¼ 21.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1701

PMC with spherical inclusions. Macro elastic stiffness with different models.
80

70
Voigt
Double inclusion
Mori-Tanaka
60 Finite elements (ABAQUS)
Macro Young’s modulus [GPa]

Reuss

50

40

30

20

10

0
0 0.2 0.4 0.6 0.8 1
Volume fraction of inclusions

Fig. 10. PMC: macro-YoungÕs modulus predictions with different models: (1) Voigt (uniform strain), (2) double inclusion, (3) Mori–
Tanaka, (4) finite elements (unit cell) and (5) Reuss (uniform stress).

PMC under tension. Results of two Mori-Tanaka formulations.

160

140

120
Macro stress [MPa]

100

80

60 This work, v = 0.52

v = 0.47
v = 0.42
v = 0.35
40 v = 0.15
v = 0.
Tandon Weng (1988)
20

0
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Macro plastic strain

Fig. 11. PMC under macro-tension for various values of inclusionsÕ volume fraction v1 . Predictions obtained with two MT formu-
lations: (1) incremental (this work) and (2) secant (Tandon and Weng, 1988).
1702 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

PMC under tension. Mori-Tanaka and experimental results.

180

160

140

120
Macro stress [MPa]

100

80

This work, v = 0.52

60 v = 0.47
v = 0.42
v = 0.35
40 v = 0.15
v = 0.
Experiment
20

0
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Macro plastic strain

Fig. 12. PMC under macro-tension for various values of inclusionsÕ volume fraction v1 . Comparison between: (1) predictions obtained
with Mori–Tanaka (this work) and (2) experimental results.

PMC (vol. frac = 42%) under cyclic strain. Comparison between different results.
200

150 Mori-Tanaka
F.E. (ABAQUS, CPU = 1075 sec)
Double inclusion
100
Macro stress [MPa]

50

-50

-100

-150

-200
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
Macro strain

Fig. 13. PMC (v1 ¼ 42%) under cyclic macro-strain (macro-uniaxial stress). Comparaison of different predictions: (1) MT model,
(2) finite elements (unit cell) and (3) interpolative double inclusion model.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1703

PMC under shear. DIGIMAT results with Mori-Tanaka.

90

80

70
Macro shear stress [MPa]

60

50

40

30 matrix
vol. frac. = 15%
35%
20 42%
47%
52%
10

0
0 0.02 0.04 0.06 0.08 0.1
Macro engineering shear strain

Fig. 14. PMC under macro-shear strain (macro-plane stress). MT predictions for various values of inclusionsÕ volume fraction v1 .

Two plasticity models fitting the same monotonic stress-strain curve.

1400

1200

1000

800

600
Stress [MPa]

400

200

0 Isotropic hardening
Chaboche’s kinematic hardening
-200

-400

-600

-800
-0.02 0 0.02 0.04 0.06 0.08 0.1
Strain

Fig. 15. Two plasticity models, one with isotropic hardening only and the other with combined non-linear isotropic and kinematic
hardenings (ChabocheÕs model). The two models fit the same monotonic stress–strain curve up to 10% strain, but their responses differ
under cyclic load history (results for 1% cyclic strain shown).
1704 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

Metal-matrix composite (v1 = 30%). Two plasticity models for the matrix.
1500

1000

Isotropic hardening
Chaboche’s kinematic hardening
500
Macro stress [MPa]

-500

-1000

-1500
-0.015 -0.01 -0.005 0 0.005 0.01
Macro strain

Fig. 16. MMC (v1 ¼ 30%) under cyclic macro-strain (macro-uniaxial stress). MT predictions using two plasticity models fitting the
same monotonic stress–strain curve of the matrix: (1) isotropic hardening only and (2) non-linear isotropic and kinematic hardenings
(ChabocheÕs model).

Metal-matrix composite (v1 = 30%) under cyclic shear. Two plasticity models for the matrix.
800
Isotropic hardening
Chaboche’s kinematic hardening
600

400
Macro shear stress [MPa]

200

-200

-400

-600

-800
-0.025 -0.02 -0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02
Macro engineering shear train

Fig. 17. MMC (v1 ¼ 30%) under cyclic macro-shear strain (macro-plane stress). MT predictions using two plasticity models fitting the
same monotonic stress–strain curve of the matrix: (1) isotropic hardening only and (2) non-linear isotropic and kinematic hardenings
(ChabocheÕs model).
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1705

Next, the matrix material was modeled with classical J2 elasto-plasticity (no kinematic hardening) and a
new set of parameters was identified so that to obtain the same stress–strain curve in a monotonic tension
test up to 10% of total strain. The following parameters were found for isotropic hardening with saturation
(the values of E0 , m0 and rY are unchanged): R1 ¼ 1286:68 MPa and m ¼ 19:9325.
Fig. 15 shows that the monotonic responses of the two models up to 10% strain are indeed identical.
However, for a cyclic tension/compression test under imposed strain (peak values 1%), it is seen in Fig.
15 that upon unloading and re-entry into plasticity, the answers become very different as ChabocheÕs
model predicts a Baushinger effect while classical J2 finds opposite yield stresses in tension and com-
pression.
A hypothetical composite was designed by reinforcing the matrix with spherical elastic ceramic inclu-
sions (E1 ¼ 400 GPa and m1 ¼ 0:2).
Results of macro-uniaxial tension/compression of the composite (v1 ¼ 30%) are shown in Fig. 16 using
MT and the two material models for the matrix discussed above. Remarks made for the matrix alone (Fig.
15) hold for the composite (Fig. 16).
The same composite was simulated under macro-cyclic shear and the results are plotted in Fig. 17.
Again, the results and the comments are similar to those of Figs. 15 and 16.

10. Two-scale numerical simulations of composite structures

10.1. General considerations

We integrated our homogenization code DIGIMAT into the FE program ABAQUS through its user-
defined material interface UMAT. We adopted the following two-scale approach. A classical FE analysis is
carried out at macro-scale, and for each time interval ½tn ; tnþ1  and each iteration of the global equilibrium
equations at macro-scale, and at each quadrature point of the macro-FE mesh, the homogenization module
UMAT/DIGIMAT is called. The data that are passed to it by ABAQUS are the total macro-strains n and
 (as well as material constants and history information at tn ). The DIGIMAT code returns the macro-
stress r and macro-tangent moduli c at tnþ1 . The microstructure is not ‘‘seen’’ by ABAQUS but only by
DIGIMAT, which considers each quadrature point to be the center of a RVE which contains the hetero-
geneous microstructure.
As we shall see in the next subsections, this two-scale procedure allows to compute structures made of
composite materials within reasonable CPU time and memory usage on an ordinary workstation. The
results prove that the whole procedure converges and does so rapidly. The procedure is comprised of the
following algorithms:

ii(i) The homogenization algorithms (Sections 7.2 and 8.2), with as a subset the algorithms for elasto-plas-
ticity (time-integration and computation of consistent moduli).
i(ii) The algorithm for macro-stress constraints (Section 6.4) because the two examples use shell elements
which must satisfy macro-plane stress conditions.
(iii) The FE computations at macro-scale. The results show that the macro-tangent moduli c and the macro-
stresses r which are returned by the homogenization module DIGIMAT/UMAT allow ABAQUS to
converge rapidly to a solution in the FE sense.

Finally, we point out that our homogenization-based approach gives results with acceptable accuracy
while being much cheaper than a direct method (FE computations at macro- and micro-scales) on all
aspects: CPU time, memory usage and user time.
1706 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

Fig. 18. MMC spoon built-in at both ends, under cyclic displacement at mid-span. Two-scale procedure: FE program ABAQUS at
macro-scale and homogenization code DIGIMAT at micro-scale (with an instantaneous UMAT interface between ABAQUS and
DIGIMAT).

10.2. MMC spoon under cyclic bending

A spoon made of the MMC material of Section 9.2 is built-in at both ends and subjected to a cyclic
displacement (peak values 4 mm) at mid-span, see Fig. 18. The mesh used 800 shell elements of type S4R5
and 891 nodes.
The load–displacement curves are shown in Fig. 19 for two cases: a homogeneous matrix material
(ABAQUS alone) and the composite material (v1 ¼ 30%, ABAQUS with UMAT/DIGIMAT interface).
The effect of the reinforcing phase on the overall response of the structure is obtained in a very cost-effective
way: the two-scale approach only takes 2889 s of CPU time on an average Compac (ex-Digital) DEC Alpha
workstation, and this is just 3.95 times more expensive than the homogeneous case.

10.3. PMC bottle under cyclic torsion

A bottle made of the PMC material of Section 9.3 is built-in at one end and subjected to cyclic torsion
under imposed axial rotation at the other end (peak values 0.5 rad), see Fig. 20. The mesh used 3290 shell
elements of type S4R and 3292 nodes.
The torque–rotation curves are shown in Fig. 21 for two cases: a homogeneous matrix material
(ABAQUS alone) and the composite material (v1 ¼ 42%, ABAQUS with UMAT/DIGIMAT interface).
The two-scale computation only takes 3980 s of CPU time on the same Compac DEC Alpha workstation,
and this is only 3.44 times the CPU cost of the homogeneous calculation.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1707

MMC spoon built-in at both ends under cyclic displacement at mid-span.

400
MMC, v1 = 30%, ABAQUS w. UMAT/DIGIMAT, CPU = 2889.0 SEC
Matrix only, ABAQUS alone, CPU = 730.63 SEC
300

200

100
Force [N]

-100

-200

-300

-400

-500
-4 -3 -2 -1 0 1 2 3 4
Displacement [mm]

Fig. 19. MMC spoon built-in at both ends, under cyclic displacement at mid-span. Force–displacement curves in two cases:
(1) composite (v1 ¼ 30%) using a two-scale procedure with an instantaneous UMAT interface between FE program ABAQUS and
homogenization code DIGIMAT (with MT model) and (2) matrix only material (v1 ¼ 0) with ABAQUS.

Fig. 20. PMC bottle under cyclic torsion. Two-scale procedure: finite element program ABAQUS at macro-scale and homogenization
code DIGIMAT at micro-scale (with an instantaneous UMAT interface between ABAQUS and DIGIMAT).
1708 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

PMC bottle under cyclic torsion

300000
PMC, v1 = 42%, ABAQUS w/ UMAT/DIGIMAT, 3980.3 sec
Matrix only: ABAQUS alone, 1156.7 sec
200000

100000
Total reaction moment [N mm]

-100000

-200000

-300000

-400000
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Rotation [rad]

Fig. 21. PMC bottle under cyclic torsion. Moment–rotation curves in two cases: (1) composite (v1 ¼ 42%) using a two-scale procedure
with an instantaneous UMAT interface between FE program ABAQUS and homogenization code DIGIMAT (with MT model) and
(2) matrix only material (v1 ¼ 0) with ABAQUS.

11. Conclusions and future work

In this paper, we studied micro/macro-modeling and numerical simulation of two-phase elasto-plastic

composite materials and structures.
We extended to non-linear, rate-independent models two homogenization schemes: MT and a DI. The
latter is LielensÕ version (Chapter 9 in (Lielens, 1999)) of the original Nemat-Nasser and Hori (1999, Section
10).
The extension of homogenization schemes is based on a Hill-type incremental formulation which allows
the simulation of unloading and cyclic loadings.
The formulation allows to handle any rate-independent model for each phase. It was tested with two
different material models: classical J2 elasto-plasticity and ChabocheÕs model with non-linear kinematic and
isotropic hardenings.
We studied the crucial issue of tangent moduli: elasto-plastic (or ‘‘continuum’’) versus algorithmic (or
‘‘consistent’’), and anisotropic versus isotropic. We applied two methods of extraction of isotropic parts of
anisotropic operators. We compared mathematically the stiffnesses of various tangent moduli, as this might
give an explanation for the overall predictions obtained when EshelbyÕs tensor is computed with anisotropic
(continuum or consistent) or isotropic moduli.
All rate equations were discretized in time using implicit schemes at two levels: (i) the constitutive models
and (ii) the homogenization schemes. In the latter case, we showed the importance of the time-integration
parameter (a) on the accuracy of the overall response. The results that we reported in this paper were all
obtained with a ¼ 1=2 (mid-point rule) or a ¼ 2=3.
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1709

We tested the robustness of the formulation and its implementation by considering at least two com-
posites with completely different properties: a MMC and a PMC, and running several discriminating nu-
merical simulations: cyclic macro-tension/compression, cyclic macro-shear and macro-bi-axial loadings.
In many cases, the results were compared with those of direct computations in which the BVP for each
RVE is solved directly using FEs. For MMC, our results were also compared to predictions obtained with
other formulations and reported in (Segurado et al., 2002), namely: 3D FE simulations, a classical secant
MT formulation and SuquetÕs formulation where reference materials are defined with the phase-averages of
the second-order moments of stress. Good agreement was found between our predictions and the 3D FE
targets, and the accuracy increases with increasing hardening.
The results showed that DI gives an excellent prediction of the elastic stiffness but that the plastic res-
ponse is generally better predicted by MT.
Finally, our homogenization code DIGIMAT was integrated into the FE program ABAQUS using a
user material interface UMAT. A two-scale method was used: a FE model at macro-scale, and at each
quadrature point of the macro-FE mesh, the homogenization module UMAT/DIGIMAT is called. The
procedure allows to compute real-life structures made of composite materials within reasonable CPU time
and memory usage. Two examples were shown: a MMC spoon built-in at both ends and subjected to cyclic
bending and a PMC bottle under cyclic torsion.
The results reported in this paper are based on two assumptions: (1) reference moduli are computed with
average strains in each phase; and (2) EshelbyÕs tensor is computed with an isotropic part of the reference
matrix moduli.
Although good results were obtained, there is a need for a better understanding of the two issues, and it
would be interesting to implement some other proposals from the literature and compare their predictive
capabilities.
This paper only dealt with rate-independent small-strain composite materials, with inclusions having the
same shape, material properties and orientation. Numerical results were only presented for spherical in-
clusions.
Firstly, we are currently studying the influence of the shape and orientation of inclusions on the overall
properties of two-phase elasto-plastic composites (Friebel, 2002; Doghri and Friebel, 2003).
Secondly, we intend to extend the formulation and algorithms to finite-strain elasto-plastic composites.
Hill (1972) and Nemat-Nasser (1999) showed that the finite-strain formulation follows the same lines as the
infinitesimal-strain case provided that one uses the following ingredients: (i) work with the reference
configuration; (ii) use the nominal stress PT as a stress measure; (iii) use the rate of deformation gradient F_
as a strain rate measure.
Thirdly, we will turn our attention to elasto-viscoplastic composites. The earlier formulations of
Hutchinson (1976) and Molinari et al. (1987), although successful, are based on several restrictive as-
sumptions. We are looking for a general formulation which would work with any elasto-viscoplastic model
under general loadings (including cyclic ones). With this perspective, it seems worthwhile to investigate the
recent proposal of Masson et al. (2000). In the authorsÕ so-called ‘‘affine method’’, homogenization models
are formulated using instantaneous ‘‘linear thermoelastic’’ solids.

Acknowledgements

A. Ouaar acknowledges an assistantship from UCL and partial support of the IUAP P5/08 project
‘‘From microstructure towards plastic behaviour of single- and multi-phase materials’’ of the Federal Office
for Scientific, Technical and Cultural Affairs, Belgian State Prime MinisterÕs Office. The authors are grateful
to F. Lani, G. Lielens and T. Pardoen for helpful discussions.
1710 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

Fig. 22. Unit cell used for FE computations assuming a periodic microstructure and axisymmetric loadings.

MMC (v1 = 30%) under tension: comparison between unit cell and 3D F.E. computations
300

250

200
Macro stress [MPa]

150

100 Unit cell, this work

Average 3D F.E., Segurado et al. 2002

50

0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Macro strain

Fig. 23. MMC (v1 ¼ 30%) under macro-tension. Comparaison of finite element results obtained with two models: (1) a unit cell as-
suming a periodic microstructure (this work) and (2) average 3D FE computations of a RVE containing randomly located spheres
(Segurado et al., 2002).
 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712 1711

Appendix A. Unit cell for FE computations

Some macro-tension/compression results were validated against FE computations of a unit cell, as-
suming that the composite has a periodic microstructure with a square arrangement. The cell has a plane
geometry in the (r; z) plane, where r is the radial coordinate and z the coordinate along the symmetry axis.
The cell geometry is shown in Fig. 22; z- and r-coordinate lines correspond to vertical and horizontal
segments, respectively. Along the symmetry axis (left vertical segment DA), radial displacement is zero.
Along the bottom horizontal line AB, the vertical displacement is zero. Vertical displacements are imposed
on the top horizontal side CD while the right vertical side BC is constrained to have the same radial
displacement. The mesh used 3002 nodes and 963 ABAQUS axisymmetric CAX8R elements.
For the MMC studied in Section 9.2 (with v1 ¼ 30%) a macro-tension test was simulated in (Segurado
et al., 2002) by performing 3D FE computations on a cube containing 30 spheres. A typical mesh contained
about 60,000 elements and 90,000 nodes. Twelve meshes were generated corresponding to different ar-
rangements of the spheres and BCs and the average response is plotted in Fig. 23, together with the FE
curve obtained with our unit cell. The figure shows that the unit cell over-predicts the macro-response by
comparison with the more accurate 3D FE computations on a cubic RVE containing many spheres without
a periodic distribution.

References

ABAQUS, a general-purpose finite element program, Hibbitt, Karlsson & Sorensen, Inc., Pawtucket, RI, USA, 2001.
Armstrong, P.J., Frederick, C.O., 1966. A mathematical representation of the multiaxial Baushinger effect. Report no. RD/B/N731,
General Electricity Generating Board, Berkeley Nuclear Laboratoires.
Benveniste, Y., 1987. A new approach to the application of Mori-TanakaÕs theory in composite materials. Mech. Mater. 6, 147–
157.
Bornert, M., Bretheau, T., Gilormini, P., (Eds.) 2001a. Homog
en
eisation en m
ecanique des mat
eriaux, 1. Mat
eriaux al
eatoires

elastiques et milieux p
eriodiques. HERMES Science, Paris, 2001.
Bornert, M., Bretheau, T., Gilormini, P., (Eds.) 2001b. Homog
en
eisation en m
ecanique des mat
eriaux, 2. Comportements non lin
eaires
et problemes ouverts, HERMES Science, Paris, 2001.
Doghri, I., 1993. Fully implicit integration and consistent tangent modulus in elastoplasticity. Int. J. Numer. Meth. Engng. 36, 3915–
3932.
Doghri, I., 2000. Mechanics of Deformable Solids- Linear, Nonlinear, Analytical and Computational Aspects. Springer-Verlag, Berlin.
Doghri, I., Friebel, C., 2003. Effective elasto-plastic properties of inclusion-reinforced composites. Study of shape, orientation and
cyclic response, in preparation.
Eshelby, J.D., 1957. The determination of the elastic field of an ellipsoidal inclusion, and related problems. Proc. Roy. Soc. London
Ser. A 241, 376–396.
Friebel, C., 2002. Mod
elisation et simulation micro-macro de mat
eriaux composites. Etude de lÕinfluence des renforts (mat
eriau, forme
et orientation). Graduation report in applied mathematics, UCL/FSA, Louvain-la-Neuve, Belgium.
Gavazzi, A.C., Lagoudas, D.C., 1990. On the numerical evaluation of EshelbyÕs tensor and its application to elastoplastic fibrous
composites. Comput. Mech. 7, 13–19.
Gonzalez, C., Llorca, J., 2000. A self-consistent approach to the elasto-plastic behaviour of two-phase materials including damage.
J. Mech. Phys. Solids 48, 675–692.
Hashin, Z., Shtrikman, S., 1963. A variational approach to the theory of the elastic behaviour of multiphase materials. J. Mech. Phys.
Solids 11, 127–140.
Hill, R., 1965. A self-consistent mechanics of composite materials. J. Mech. Phys. Solids 13, 213–222.
Hill, R., 1972. On constitutive macro-variables for heteregeneous solids at finite strain. Proc. R. Soc. London A 326, 131–147.
Hutchinson, J.W., 1976. Bounds and self-consistent estimates for creep of polycrystalline materials. Proc. Roy. Soc. London. A 348,
101–127.
Kouznetsova, V., Geers, M.G.D., Brekelmans, W.A.M., 2002. Multi-scale constitutive modeling of heterogeneous materials with a
gradient-enhanced computational homogenization scheme. Internat. J. Numer. Meth. Engng. 54, 1235–1260.
Lemaitre, J., Chaboche, J.-L., 1998. Mechanics of Solid Materials. Cambridge University Press, England.
Lielens, G., 1999. Micro-macro modeling of structured materials. Ph.D. dissertation, UCL/FSA, Louvain-la-Neuve, Belgium.
1712 I. Doghri, A. Ouaar / International Journal of Solids and Structures 40 (2003) 1681–1712

Masson, R., Bornert, M., Suquet, P., Zaoui, A., 2000. An affine formulation for the prediction of the effective properties of nonlinear
composites and polycrystals. J. Mech. Phys. Solids 48, 1203–1227.
Molinari, A., Canova, G.R., Ahzi, S., 1987. A self-consistent approach of the large deformation polycrystal viscoplasticity. Acta
Metall. 35 (12), 2983–2994.
Mori, T., Tanaka, K., 1973. Average stress in matrix and average elastic energy of materials with misfitting inclusions. Acta Metall. 21,
571–574.
Mura, T., 1987. Micromechanics of Defects in Solids, second ed. Martinus Nijhoff Publishers, Dordrecht, The Netherlands.
Nemat-Nasser, S., 1999. Averaging theorems in finite deformation plasticity. Mech. Mater. 31, 493–523.
Nemat-Nasser, S., Hori, M., 1999. Micromechanics: Overall Properties of Heterogeneous Materials, second ed. Elsevier Science
publishers, Amsterdam.
PALMYRA, 2001. A finite-element software for heterogeneous micro-structures. Materials Simulation GmbH, Zurich, Switzerland.
Pettermann, H.E., Planskensteiner, A.F., Bohm, H.J., Rammerstorfer, F.G., 1999. A thermo-elasto-plastic constitutive law for
inhomogeneous materials based on an incremental Mori–Tanaka approach. Comput. Struct. 71, 197–214.
Ponte Casta~ neda, P., 1996. Exact second-order estimates for the effective mechanical properties of nonlinear composite materials.
J. Mech. Phys. Solids 44 (6), 827–862.
Ponte Casta~ neda, P., Suquet, P., 1998. Adv. Appl. Mech. 34, 171–302.
Ponte Casta~ neda, P., Suquet, P., 2001. Nonlinear composites and microstructure evolution. In: Mechanics for a New Millenium,
Proceedings of ICTAM 2000-Chicago. Kluwer Academic Publishers, Dordrecht, The Netherlands, pp. 253–273.
Segurado, J., Llorca, J., Gonzalez, C., 2002. On the accuracy of mean-field approaches to simulate the plastic deformation of
composites. Scr. Mater. 46, 525–529.
Simo, J.C., Taylor, R.L., 1985. Consistent tangent operators for rate-independent elastoplasticity. Comput. Meth. Appl. Mech. Engng.
48, 101–118.
Suquet, P., (Ed.), 1997. Continuum micromechanics, CISM Lecture notes, Udine, Italy, Springer-Verlag, Berlin, 1997.
Tandon, G.P., Weng, R.L., 1988. A theory of particle-reinforced plasticity. J. Appl. Mech. Trans. ASME 55, 126–135.