2. Sample generation and computational techniques
As we have outlined in the introduction, this section is devoted
to a brief description of the methods that we have used to perform
computations and the reasons to choose these concrete methods.
2.1. RVE generation
The generation of samples for computation is an important step
in the process of modeling of the behavior of composite materials.
Since the morphology of composites may be quite complex this can
be a very challenging task. On the one hand it is important to be
able to approximate rather involved geometries, on the other hand
the method should be fast and reliable; in the ideal case the stage
of generation should be much shorter than the computation itself.
There has been a number of works where the inclusions were
represented by simple geometric objects like spheres or ellipsoids
(see for example, [5–8]). If one considers more complicated geometry,
the problem of managing the intersection of inclusions arises
immediately. Among the established approaches of dealing with
it, one can mention two important families: random sequential
adsorption (RSA) type algorithms and molecular dynamics (MD)
based methods. The RSA [9] is based on sequential addition of
inclusions verifying for each of them the intersection; the main
idea of the MD [10–12] is to make the inclusions move, until they
reach the desired configuration. Let us mention that the first
method needs an efficient algorithm of verification of intersection
between the geometric shapes, and the second one an algorithm of
predicting the time to the intersection of moving objects, which
exists for a very limited class of shapes and often amounts to a difficult
minimization problem. In [1], we have described the classical
RSA and a time-driven version of MD applied to the mixture of
inclusions of spherical and cylindrical shapes. The key ingredient
for both of the approaches was the explicit formulation of algebraic
conditions of intersection of a cylinder with a sphere and of two
cylinders. To be more specific, we recapitulate the ideas of these
algorithms (Algorithms 1 and 2).
We have observed that the RSA approach is extremely efficient
for relatively small volume fractions of inclusions (up to 30%),
where it permits to generate a sample in fractions of a second.
The MD-based method is powerful for higher volume fractions
(of order 50–60%): it generates a configuration in about a second
while the RSA can get stuck. An example of a sample with a
mixture of non-intersecting spherical and cylindrical inclusions is
presented on the Fig. 1.
The outcome of these algorithms is a list of inclusions in the
‘‘vector’’ form, i.e. a list of coordinates of centers, radii, and eventually
axes of symmetry of inclusions. This is perfectly suitable
for various computational techniques: FFT-based homogenization
procedures applied to the pixelized samples, as well as finite
element computations on the mesh constructed from this pixelization.
In addition[2] we are able to introduce various imperfections