of the inclusions without spoiling the efficiency of the generation
algorithm: the Fig. 2 shows two such examples, where the surface
of an inclusion is waved or a part of an inclusion is taken out to
produce irregular shapes.
2.2. FFT-based homogenization scheme
Having an efficient scheme of generation of samples we can
now proceed to the computation of effective mechanical
properties. As we have mentioned in the introduction for
evaluating these properties we have adopted the philosophy of
stochastic homogenization. Schematically one can view the
process as follows:
1. Fix the macroscopic parameters of the material (volume fraction
and type of inclusions).
2. Generate a series of samples (RVEs) of a composite material
with these parameters (stochastic part).
3. Perform accurate computation of effective properties on these
RVEs (deterministic homogenization part).
4. Average the computed (macroscopic) characteristics of the
samples.
Let us now describe the mechanical model behind this computation
as well as the main computational method – the homogenization
procedure.
Consider a representative volume element V, and denote uðxÞ
the displacement field defined at any point x 2 V. The system in
mechanical equilibrium is described by the law rðxÞ ¼ @wðxÞ
@eðxÞ , where
eðxÞ ¼ eðuðxÞÞ ¼ 12
ðruðxÞ þruðxÞT Þ – the strain tensor in the
model of small deformations, wðxÞ – the stored mechanical energy,
and rðxÞ – the stress tensor, subject to the condition divrðxÞ ¼ 0.
In the linear case this law simplifies to @wðxÞ
@eðxÞ ¼ cðxÞ : eðxÞ with the
stiffness tensor cðxÞ. Notice that for a composite material the stiffness
tensor does depend on the point x: the dependence is governed
by microscopic geometry of the sample, namely which
phase (matrix or inclusion) the point x belongs to. We suppose that
the averaged strain < e >¼ E is prescribed, and decompose eðxÞ in
two parts: eðuðxÞÞ ¼ E þ eð~uðxÞÞ, which is equivalent to representing
uðxÞ ¼ E:x þ ~uðxÞ, for ~uðxÞ being periodic on the boundary of
V. Thus, the problem we are actually solving reads
rðxÞ ¼ cðxÞ : ðE þ eð~uðxÞÞ; divrðxÞ ¼ 0; ð1Þ
~uðxÞ periodic; rðxÞ:n antiperiodic:
The solution of (1) is the tensor field rðxÞ, we are interested in its
average in order to obtain the homogenized stiffness tensor chom
from the equation
< rðxÞ >¼ chom :< eðxÞ > : ð2Þ
To recover all the components of chom in 3-dimensional space one
needs to perform the computation of < rðxÞ > for six independent
deformations E, which morally correspond to usual stretch and
shear tests.
There is a couple of natural approaches to solving the problem
(1): one can construct a mesh of an RVE V and employ the finite
elements method, or discretize (basically pixelize) the RVE to use
the FFT-based homogenization scheme [3,4]. The major difficulty
arising when applying the former method is that for rather
involved geometry one needs to construct a fine mesh which is a
non-trivial task in its own, and moreover to proceed with finite elements
one requires considerable memory resources. The idea of
the latter is that in the Fourier space the equations of (1) acquire
a rather nice form for which in the case of a homogeneous isotropic
material one can construct a Green operator and basically produce
an exact solution. For a composite material containing possibly
several phases one introduces an artificial reference medium for
which the Green operator is defined, the computation then is an
iterative procedure to approximate the corrections of the microscopic
behavior of the material in comparison to this reference
medium. For the sake of completeness let us present this method
here in details, following essentially the works [13,14].
Introduce the reference medium stiffness tensor c0 and the correction
to it dcðxÞ ¼ cðxÞ c0. Eq. (1) can equivalently be rewritten as
rðxÞ ¼ c0 : eð~uðxÞÞ þ sðxÞ; divrðxÞ ¼ 0; ð3Þ
sðxÞ ¼ dcðxÞ : ðeð~uðxÞÞ þ EÞ þ c0 : E;
~uðxÞ periodic; rðxÞ:n antiperiodic:
The tensor s is called polarization. The periodicity assumptions permit
to rewrite the first line of (3) in the Fourier space. Using the linearity
of the Fourier transform and its property with respect to
derivation, one obtains
^rmjðnÞ ¼ ic0
mjklnl
^~
ukðnÞ þ ^smjðnÞ; i^rmjnj ¼ 0; ð4Þ
where ^ denotes the Fourier image of and nj’s are the coordinates
in the Fourier space. The key observation is that in the Fourier space
there is a relation between the polarization and the deformation
tensors, namely