Available online at www.sciencedirect.

com

Acta Materialia 56 (2008) 52855292

www.elsevier.com/locate/actamat

Delineation of the space of 2-point correlations

in a composite material system
S.R. Niezgoda a, D.T. Fullwood b, S.R. Kalidindi a,*
a

Department of Materials Science and Engineering, Drexel University, 32nd Chesnut Street, Philadelphia, PA 19104, USA
b
Mechanical Engineering Department, Brigham Young University, Provo, UT 84602, USA
Received 4 March 2008; received in revised form 30 June 2008; accepted 1 July 2008
Available online 3 August 2008

Abstract
The complete set of 2-point correlations for a composite material system with a large number of local states (e.g. polycrystalline metals) forms a vast and unwieldy data set containing a large amount of redundant information. The interrelations in these correlations have
been well characterized for composite material systems with two local states, but only a small number have been delineated for the composite systems with many local states. This paper presents an analysis of interrelations between the complete set of 2-point correlations
for composite material systems through their spectral representations via discrete Fourier transforms. These interrelations are used to
delineate a compact and convex space that bounds the set of all physically realizable 2-point correlations called the 2-point correlations
hull. The representation of any given microstructure in this hull, and the techniques to produce a representative volume element are also
explored in this paper.
2008 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
Keywords: 2-Point correlation; Microstructure; Discrete Fourier transform; Pair correlation function; Statistics

1. Introduction
At its core, the eld of materials science and engineering
is concerned with understanding and modeling the relationships between a materials internal structure, its
macro-scale properties and its processing history. Fundamental to establishing these relationships is the quantitative representation of the materials internal structure,
which includes not only an identication of the constituent
local states, but also their spatial placement. As interest is
typically in the micro-scale features, the materials internal
structure is referred to as the microstructure in this paper.
However, the approach and the results presented here can
be applied at any length scale.
It is not usually practical or even necessary, to quantify
completely the microstructure of a large sample in its full
spatial extent. The assumption of statistical homogeneity
*

Corresponding author.
E-mail address: skalidin@coe.drexel.edu (S.R. Kalidindi).

permits one to approximate the relevant statistics of interest in the sample as an average from an ensemble of subdomains extracted randomly from the sample. A rigorous
framework dening the statistics of the microstructure is
available in the literature in the form of n-point correlations or n-point statistics [15]. These correlations provide
a hierarchy of statistical measures of the microstructure.
The simplest of these are the 1-point correlations, f(h),
which essentially reect the volume fractions of the various
distinct constituents (denoted by h and also referred to as
local states; the complete space of local states is referred
to as the local state space H). These are termed 1-point statistics because they reect the probability density associated with nding a specic local state of interest at a
point selected randomly in the microstructure. Expanding
on this basic concept, the 2-point correlations, f h; h0 j r,
capture the probability density associated with nding an
ordered pair of specic local states at the head and tail of
a randomly placed vector r into the microstructure. In a
very similar manner, n-point statistics can be extracted

1359-6454/$34.00 2008 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.actamat.2008.07.005

5286

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292

from the microstructure through the placement of a complete set of n-vertex polyhedra.
This paper focuses on the 2-point correlation functions
for composite material systems. Two-point correlations
have been the focus of several investigations in current literature [610]. They capture the rst-order information on
the morphology of the microstructure. There is a tremendous leap in the amount of information contained in the
2-point correlations, when compared with the 1-point statistics. For most real microstructures, the 2-point correlations form a very large and unwieldy set of data. For
example, in single-phase polycrystalline samples, the local
state is dened by the lattice orientation g, which is often
dened by a set of three angles called the BungeEuler
angles [11]. Thus, for this relatively simple class of materials, the 2-point correlations need to be dened in a space
consisting of nine independent variables. The continuous
orientation and spatial spaces spanned by these variables
are often discretized by binning. For example, a coarse discretization of the orientation space into 512 bins produces
262,144 2-point distributions, each of which is a function of
the three-dimensional vector r. Fortunately, this complete
set of 2-point correlations has a large number of interdependencies, only some of which have been outlined by
prior studies in this eld [1214]. This paper presents a rigorous treatment of the interdependencies among the
complete set of 2-point correlations, leading to the establishment of a compact high-dimensional space wherein
each point represents a complete set of 2-point statistics.
This is accomplished largely through spectral representation of these correlations, and exploiting several of the
known properties of discrete Fourier transforms (DFT).
Although, every physically realizable microstructure1 will
have a representation in this high-dimensional space, it is
not true that every point in this space will correspond to
a physically realizable microstructure. The set of points
in this space that correspond to all physically realizable
2-point correlations is referred to in this paper as the 2point correlations hull. In other words, a point inside the
2-point correlation hull will correspond to one or more
physically realizable microstructures, whereas points outside this hull will not correspond to any physically realizable microstructure.
In recent work [1518], a novel mathematical framework
called Microstructure Sensitive Design (MSD) was formulated for establishing invertible structurepropertyprocessing linkages in materials. In a parallel eort, it was
found to be quite ecient to quantify the microstructure
using 2-point correlations [9,10,15]. Second-order structureproperty linkages for elastic response of multi-phase
and polycrystalline microstructures were successfully established. MSD aims to identify the complete set of micro1
Physically realizable microstructures include all those that can be
imagined (or digitally created). It is recognized that a very large fraction of
these have not yet been realized by currently employed materials
processing routes.

structures that are theoretically predicted to exhibit a

designer-specied combination of macro-scale properties
or performance characteristics. A critical element of
MSD is the delineation of the complete space of physically
realizable microstructures, which for several of the problems being currently studied amounts to the delineation
of the space of physically realizable correlations or the 2point correlation hull mentioned earlier. This paper, using
the established interdependencies among the 2-point correlations in a given microstructure, demonstrates the delineation of the 2-point correlation hull for a highly simplied
class of one-dimensional microstructures. To the best of
the authors knowledge, this is the rst report of the successful delineation of the 2-point correlation hull. Although
the example selected for this demonstration deals with a
grossly simplied microstructure, the underlying mathematical framework presented is quite general and easily
extendable to a much broader class of material systems
and microstructures.
The concepts presented in this paper also have signicant implications for the notion of a representative volume
element (RVE). In characterizing the microstructure of a
material system, it is necessary to collect microstructure
information from dierent samples produced with nominally the same processing history, and from dierent locations in each sample. The benet of obtaining multiple
microstructure data sets is that the averaged statistics from
this ensemble of micrographs is expected to represent more
accurately the overall microstructure statistics for the
material system. It should be recognized, however, that it
is not necessarily possible to produce a single instantiation
of the RVE that exhibits the ensemble averaged 2-point
correlations. With a simple example case study, it is demonstrated here that it is often impossible to identify a single
RVE that reects precisely the ensemble averaged 2-point
correlations. Viable alternatives for constructing RVEs
from an ensemble of microstructures are discussed in this
paper.
2. Microstructure function, 2-point correlations and DFT
The application of DFT techniques to the rapid calculation of 2-point statistics has been discussed previously
[19,20] and will be reviewed briey here. Consider a microstructure data set extracted from a material sample. Let the
spatial position in the microstructure be identied by the
vector x, and the local state of interest by h. One can dene
the microstructure function m(x,h) as a distribution on the
local state space for each spatial position as2 [9].

In some prior work, M was used to denote the microstructure function.

In this paper, the authors choose to follow the conventions used in the
signal processing literature and use m for the microstructure function in
real space and the M for the Fourier transform of the microstructure
function. In addition, the Einstein summation convention for repeated
indices is not used. For clarity, all summations will be explicitly indicated
with R notation.

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292


V 
;
V x

mx;hdh

mx;hdh 1;

dh 1;

where V denotes the volume of material probed in the

immediate neighborhood of x, and h V denotes the component of V that is associated with the local state h to within
an invariant measure dh.
The microstructure function and the local state space
are assumed to be continuous in the denitions presented
above. Microstructure data, however, are often collected
in a digital form on a discrete grid of spatial locations. A
discrete description of m(x,h) is therefore a natural reection of the available information. Let the local state space
Y be binned into N discrete local states labeled n = 1,2, . . .,
N. Let the spatial domain of the microstructure be binned
into a uniform grid of S cells, whose nodes are enumerated
by the ordered triplet [s1,s2,s3] where each component is
numbered si = 0,1,2, . . ., Si1. For convenience, this
ordered triplet is represented by the vector s, the vector
from the origin to [S11,S21,S31] as S1, and the zero
vector as 0. This discrete microstructure function may be
denoted in a condensed form as n ms . The following properties can be readily established from Eq. (1)
N
X

ms 1;

06

ms ;

n1

S1
X

ms n VS

s0

where n V is the volume fraction of the local state n in the

microstructure.
Next, 2-point correlations f h; h0 jr are dened as
Z
1
f h; h0 jr
mx; hmx r; h0 dx
3
volX X
where X represents the spatial domain of the volume element of the microstructure being investigated, and volX
denotes its volume. In this denition, it is implicitly assumed
that the volume element being investigated is repeated periodically to generate the microstructure of interest. This specic denition is adopted here because of the authors
intention to use DFT representations. In the discretized
representation, the 2-point correlations are expressed as
np

ft

S1
1X
n
ms p mst ;
S s0

where the complete space of vectors r, introduced in Eq.

(3), is conveniently binned using the same uniform grid that
was used for binning the spatial domain X .
The denitions of the 2-point correlations presented in
Eqs. (3) and (4) allow their ecient computation using
established fast Fourier transform (FFT) algorithms. The
DFT of the microstructure function is expressed as
n

M k In ms

S1
X

ms e2pisk=S jn M k jei

nh
k

s0

The term j n M k j will be referred to as the amplitude of the

Fourier transform and n hk as the phase. The DFT of np ft is
computed as

np

F k Inp ft

5287

1 n  p
Mk Mk
S

1
jn M k jjp M k jei
S

nh
k

ei

ph
k

where n M k denotes the complex conjugate of n M k . Taking

n = p in Eq. (6), one arrives at a special set of correlations
termed the autocorrelations. Note that the autocorrelation
is a real-valued even function, and therefore its DFT are
also real. However, n 6 p leads to cross-correlations, and
their DFT are, in general, complex.
In a previous work [8], it was found convenient to classify
microstructures as eigen and non-eigen microstructures. If
n
ms 2 f0; 1g (i.e. they are only allowed to take values of either
0 or 1) for all n, s, that microstructure is referred to as an eigen
microstructure, otherwise it is non-eigen.
The complete set of 2-point correlations for any given
microstructure constitutes a very large data set. If a microstructure of interest comprises N distinct local states, the
complete set of its 2-point statistics contains N2 correlations for each value of the three-dimensional vector r.
The full set of correlations is conveniently visualized as
an N  N array
3
2
11
ft 12 ft . . . . . . 1N ft
6
.. 7
..
7
6 21
6 ft
. 7
.
7
6
6
np
.. 7
..
:
7
ft 6 ...
. 7
.
7
6
7
6 .
.. 7
..
6 .
. 5
.
4 .
N1
ft . . . . . . . . . NN ft
It has long been known that this set exhibits many interdependencies. Frisch and Stillinger [21] showed that only
one of the four correlations is independent for a two-phase
material (i.e. if 11 ft is known, 12 ft , 21 ft and 22 ft can be calculated). More recently, Gokhale et al. [13] showed that, at
most, 12 N N  1 correlations are independent for a material system consisting of N distinct local states. In this
paper, exploiting the known properties of DFT, it is demonstrated that the number of independent correlations is
only N1. By careful consideration of the degrees of freedom of the microstructure and statistics, one can intuitively
arrive at the correct result. It can be seen from Eq. (2) that
the degree of freedom in the microstructure function is
S(N1). As the 2-point statistics are a derived quantity,
the number of degrees of freedom must be less than or
equal to that of the microstructure function. In a previous
work [19], it was shown that any original eigen-microstructure can be exactly reconstructed (to within a linear translation and/or inversion) from its statistics, and hence holds
an equivalent information content. In other words the statistics and microstructure must possess the same number,
S(N1), of degrees of freedom, and therefore only (N1)
correlations can be independent. This constitutes a dramatic reduction in the size of the data set needed to specify
uniquely the complete set of 2-point correlations for any
given microstructure.

5288

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292

The denition of the discrete 2-point correlations in Eq.

(4) requires that np ft pn ft . In the Fourier space, this
requirement translates to
np

1
n
p
F k jn M k j jp M k j ei hk ei hk pn F k :
S

np

Furthermore, the product of two of the DFT of the 2-point

correlations exhibits the following property:
np

1 p
n
q
j M k j jp M k j jn M k j ei hk jq M k j ei hk
S2
pp F k nq F k :

pn

Fk

F k pq F k
:
pp F
k

10

Eq. (10) is a key result of this paper. This equation

implies that, if pn F k (and by extension pn ft ) are known for
any one choice of p and all n or vice versa (i.e. either a column or a row in the array in Eq. (7) is known), all the other
correlations can be calculated.
In addition to the redundancies described above
between the various 2-point correlations, a number of
other constraints or bounds on the values of np F k exist.
The following relationships are easily seen from the denitions in Eqs. (2)(6)
np

F 0 S n V p V ; 0 6 np F 0 6 S; jnp F k j 6 S 2

11

These bounds allow one to dene an additional redundancy; Eq. (2) (together with Eqs. (4)(6)) requires that
 p
N
X
S V ; if k 0
np
Fk
12
0;
if k 6 0
n1
Since p V is already specied by pp F 0 (Eq. (11)), it follows
that there are only (N1) independent correlations in the
complete set shown in Eq. (7). Although the interdependencies shown in Eqs. (10) and (12) are formulated between
the DFT of the 2-point correlations, the same relationships
can be expressed in terms of their continuous Fourier
transforms or classical Fourier series. However, as the plan
is to exploit the signicant advantages associated with FFT
calculation of DFT, all further discussion in this paper will
be restricted to DFT.
Substituting Eq. (11) into Eq. (12) allows Eq. (12) to be
recast completely in terms of the DFT as
( p
N
X
S pp F 0 ; if k 0
np
Fk
13
0;
if k 6 0
n1
For eigen microstructures (n ms 2 f0; 1g), one additional
redundancy can be established by summing over all frequencies in Fourier space as
X
pp
np
14
F k dnp S nn F 0
k

where dnp denotes the Kronecker delta.

F k np F Sk

15

for all k 6 0. In other words, about half the transforms are

simply complex conjugates of the other half.
3. Hull of 2-point correlations

F k pq F k

Using Eq. (8), Eq. (9) can be recast into a more convenient form as
nq

As a nal note, the reader is reminded that, as all np ft are

real valued, there is an implicit symmetry in their DFT that
may be expressed as

Having established the interdependencies in a given set

of 2-point correlations, attention is now turned to identifying the complete set of theoretically feasible sets of 2-point
correlations for a given material system. This problem has
been tackled in various guises in the past, but never completely solved. For example, using the WienerKhinchin
theorem, it has been demonstrated [22,23] that a semi-positive denite 2-point correlation in one-dimensional space
relates to a real random process. The reader is referred to
the work of Torquato [12,14] and Gokhale [13] for other
important results obtained to date using such approaches.
Jiao et al. [24] have recently proposed a novel approach
to exploring a certain sub-space of the 2-point correlation
space described here. The space explored by these authors
is called the S2 space and captures only the auto-correlations of eigen microstructures. In particular, for composites
with more than two distinct local states, the S2 space will be
missing the important phase information that is only present in the cross-correlations (i.e. the n hk described earlier).
Previous work [18] demonstrated the advantages of visualizing in Fourier spaces the set of all theoretically feasible
1-point statistics in a given material system. In these constructs, the 1-point statistics of any microstructure are visualized as a point in the Fourier space whose coordinates are
identied by the Fourier coecients of the selected 1-point
statistics. More specically, it was demonstrated that the
space corresponding to the set of all theoretically feasible
1-point statistics in a given material system is compact
and convex. These constructs have been broadly referred
to as microstructure hulls in previous work, and have been
found to be of tremendous value in designing components
for optimized performance characteristics [16]. It is particularly emphasized here that, in material systems with a
continuous local state space, the hulls for 1-point statistics
are most easily established in the Fourier space. This is
because the values of the distribution (probability density
function) used to dene the 1-point statistics could be
unbounded for many microstructures. For example, in
describing the 1-point statistics in polycrystalline microstructures, the values taken by the orientation distribution
function (ODF) have to be positive, but are otherwise
unbounded. In fact, the ODF for a single crystal microstructure would be a Dirac-delta function. However, the
Fourier coecients corresponding to any theoretically feasible texture occupy a bounded region, making it very convenient to visualize the complete set of all theoretically
possible 1-point statistics in a given material system. In

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292

the present work, because of the use of discrete sampling of

the microstructure in both the spatial domain and the local
state space, hulls for 2-point correlations could be theoretically delineated either in the vector space of pn ft or in spectral space of pn F k . In fact, the discrete sampling employed
in dening pn ft is tantamount to the use of a primitive Fourier basis called the indicator functions or characteristic
functions [25], and therefore both descriptions can be interpreted as Fourier representations. However, as shown in
the previous section, the use of DFT allows many of the
redundancies that exist in the set of pn F k to be identied
much more easily than in pn ft . Therefore, this work focuses
exclusively on the delineation of the hulls of 2-point correlations in the DFT space.
The concept of eigen microstructures described earlier,
where all n ms values are either 0 or 1, is central to the construction of the microstructure hulls. Let pn F~ k denote the
DFT of the 2-point statistics of an eigen microstructure
~ denote the set of
of a material system of interest. Let H
pn ~
all theoretically feasible F k . Then the convex region
~ represents the hull of 2-point
dened by the elements of H
correlations sought here. Eqs. (10)(15) explicitly dene
the hull in a 2N  3dS 1=2e -dimensional space
(where dxe denotes the ceiling function which returns
the smallest integer P x). The 2N  3 term comes from
the fact that (N1) correlations are necessary to describe
a N local state material system (Eq. (10)) and the observation that the Fourier terms in cross correlations are complex valued and real valued in autocorrelations. The
dS 1=2e term comes from Eq. (15) and the symmetry
of each Fourier transform. It is anticipated that the eigen
microstructures occupy an extremely small portion of the
points in the hull. The points in the hull that are not occupied by eigen microstructures can be interpreted in two different ways. One option is to interpret them as belonging to
non-eigen microstructures. However, investigations have
revealed that it may not be possible to assign a (single) specic non-eigen microstructure, of spatial extent S, to every
point in the hull. It is important to note that the inability to
generate a single instantiation is the byproduct of the discreteness in space. It is expected that, in the continuous
limit, a single non-eigen instantiation could be found for
all points in the hull. The second option is to interpret these
points as the ensemble averaged 2-point statistics over multiple instantiations of the microstructure drawn either from
a single sample or from multiple samples with nominally
the same processing history. The authors prefer the second
interpretation for the following two main reasons: (i) dening a representative microstructure in terms of the ensemble average of statistics from multiple instantiations
(rather than from a single instantiation) assures one that
the microstructure statistics for the sample are captured
as accurately as possible; (ii) ensemble averaging also
assures one that the hull is convex and compact, which in
turn makes it possible to search for microstructures with
optimal combinations of properties using the available
computationally ecient quadratic programming algo-

5289

rithms [26]. All the concepts described above will be claried through a very simple case study in the next section.
4. Case study: one-dimensional two-phase microstructure
As a simple case study that is mainly intended for clarication of the various concepts presented earlier, a set of
hypothetical one-dimensional two-phase microstructures
extracted from a larger two-dimensional sample shown in
Fig. 1 is investigated. These represent multiple instantiations of eigen microstructures extracted from a hypothetical sample. The two distinct phases present in these
microstructures are shown as black and white, respectively.
Furthermore, the spatial domain in these microstructures
was divided into only 10 bins, as shown in Fig. 1. As there
are only two phases, only one 2-point correlation is needed
to represent the microstructure. Let 11 F k denote the DFT
selected for representation in a 2-point hull. As S has been
selected as 10 in this simple case study, the relevant DFT
space here is six-dimensional (taking advantage of Eq.
(15) and the fact that the autocorrelation coecients 11 F k
are all real). This is the main reason for selecting a highly
simplied idealized microstructure for this case study. As
one can see, the dimensionality of the space of the 2-point
hulls described in this work grows approximately proportionally with the product SN. However, it is only possible
to view projections of this hull in, at most, a selected
three-dimensional subspace. Although the mathematical
framework presented above can be employed on microstructures with much larger values of both S and N, the
visualization of their 2-point hulls in three-dimensional
subspaces is not particularly insightful. Nevertheless, the
reader might be interested to know that the mathematical
framework presented here has been used successfully on
microstructures with S values of about two million and N
values of about ve hundred [19].

Fig. 1. A set of 10 one-dimensional two-phase samples extracted from a

very large hypothetical sample. It is expected that the large sample can be
represented as an ensemble average of these smaller samples.

5290

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292

Fig. 2 shows the projections of the computed 2-point

hull for the selected one-dimensional material system in
the 11 F 0 , 11 F 1 ; 11 F 2 subspace and the 11 F 0 , 11 F 4 ; 11 F 5 subspace. The starred points inside the hull in Fig. 2 denote
the set of all possible eigen microstructures in this material
system. Although one might expect a total of 210 eigen
microstructures (based on S = 10 and N = 2), the number
of eigen microstructures shown in Fig. 2 is substantially
less. Two factors can help explain this discrepancy. (i) A
very large number of eigen microstructures share the exact
same representation in the 2-point hull shown in Fig. 2. In
fact, in the DFT representation 2-point correlations are
invariant under a translation and/or an inversion of the
microstructure. It is easy to see from the denitions in
Eqs. (3) and (4) that the microstructure data sets n ms and
n
msa would produce identical 2-point correlations, where
a denotes a translation of the microstructure data set by an
integer number of grid points on the spatial domain of the
microstructure. (ii) The projection of the six-dimensional
hull into a three-dimensional subspace causes some of the
distinct points in the six-dimensional space to occupy the
same location in the smaller three-dimensional space.
The ensemble of eigen microstructures shown in Fig. 1
have been identied as black squares in Fig. 3. These are
a subset of the starred points shown in Fig. 2. Note that
eigen microstructures 1 and 3 are related to each other by
a translation and therefore occupy the same location in
Fig. 3. The ensemble averaged 2-point statistics, denoted
h11 F k i, for the entire set of microstructures shown in Fig.
1 is shown as a lled circle in Fig. 3. The ensemble averaged
statistics in this case did not correspond to any of the possible eigen microstructures for the selected material system.
Given the relatively small number of points in the hull
occupied by the eigen microstructures, it should be
expected that there is only an exceedingly low chance that
the ensemble averaged 2-point statistics would correspond
to an eigen microstructure. As discussed earlier, all the
points in the hull that do not belong to an eigen micro-

Fig. 3. The projection of the ensemble of microstructures from Fig. 1 into

the 2-point correlation hull. The average statistics for the ensemble are
denoted by the black circle.

structure can be interpreted as representing the ensemble

averaged 2-point statistics of a set of microstructures
extracted from one or more samples with nominally the
same processing history.
One method of visualizing the points in the 2-point hull
is through reconstruction techniques. Recent advances in
reconstruction techniques using phase retrieval algorithms
have resulted in rapid restoration of microstructures from
a given set of 2-point statistics. The procedures used for
the reconstructions have been described in other papers
from the authors group [19,20] and have been highly successful in non-eigen reconstructions and exact reconstructions of eigen microstructures (up to an arbitrary
translation or inversion). However, eorts to reconstruct
a non-eigen microstructure corresponding to the ensemble
averaged statistics in Fig. 3 have resulted in only limited
success. In general, reconstructions from ensemble averaged statistics result in multiple solutions where the statistics of the reconstructed microstructures are close to the
ensemble averaged statistics, but the normalized error is
still outside the bounds of what would be considered a
successful reconstruction [19,24]. This leads one to

Fig. 2. Projections of the computed 2-point correlation hull for the selected one-dimensional material system. The starred points inside the hull represent
the set of all possible eigen microstructures in this material system.

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292

Fig. 4. Best possible reconstruction of a non-eigen microstructure

corresponding to the ensemble averaged statistics shown in Fig. 3. The
statistics of the reconstructed microstructure dier from the ensemble
averaged statistics by 1%.

believe that there are points in the 2-point hull that would
not correspond to any single non-eigen microstructure
(dened at the adopted spatial resolution). An example of
the reconstructed non-eigen microstructure that comes closest to capturing the ensemble averaged statistics is shown in
Fig. 4. This inability to capture the ensemble averaged statistics is a direct result of the limited size of the spatial discretization used in this example. Given the extremely
limited spatial discretization (only 10 spatial points in the
sample) it would be impossible to build a microstructure
for each point in the convex hull. In principle, if one is willing to rene the spatial grid indenitely, it will become possible to nd a non-eigen microstructure with the statistics of
the ensemble structures to a pre-dened approximation.
Any non-eigen microstructure will eventually become an

5291

eigen microstructure (to a given approximation) as the spatial grid on which it is discretized is made ner. Alternatively, the non-eigen microstructures can be thought of as
an eigen microstructure on a coarse spatial grid.
Instead of the above approach, the authors prefer to
represent the ensemble averaged 2-point statistics for the
set of microstructures shown in Fig. 3 using a weighted
set of RVE. For this purpose, a Euclidian distance between
any two points in the 2-point correlation hull, normalized
by the size of the 2-point hull (i.e. the largest distance of
the hull vertices from the origin) is dened and used to
identify a set of eigen microstructures that are close to
the ensemble averaged statistics of the given set of microstructures. Fig. 5 describes examples of how a weighted
set of eigen microstructures can approach the ensemble
averaged 2-point statistics. As expected, the more microstructures that can be used in the RVE set, the closer the
ensemble averaged statistics can be approximated.
5. Summary and conclusions
n-Point statistics are an important method of describing
material structure in a statistical manner. This paper has
described an ecient manner for representing the complete
hull of 2-point statistics an essential step towards explor-

Fig. 5. RVE for a large hypothetical sample represented as the weighted fractions of ve eigen-microstructures.

5292

S.R. Niezgoda et al. / Acta Materialia 56 (2008) 52855292

ing the spaces of n-point statistics for material analysis and

design. The framework presented above made use of spectral representation through DFT techniques to examine
and enumerate many of the interdependencies in the complete set of 2-point microstructure correlation functions for
composite material systems with multiple local states. The
interdependencies presented demonstrate that the number
of independent correlations necessary to dene an N local
state eigen-microstructure completely is N1.
These interdependencies in Fourier space have been
used to delineate the bounded space of all physically meaningful 2-point correlations, termed the 2-point correlations
hull. Visualizations of the hull have been presented for a
simplied one-dimensional two-phase microstructure.
It was found that it is often impossible to capture exactly
the ensemble averaged statistics in a single representative
eigen or non-eigen microstructure. Instead, it was found
possible to approach the ensemble averaged statistics using
a small weighted set of RVE. The utility of the hull in producing such an RVE set for a bulk sample was demonstrated. The case study presented showed that each point
in the hull maps to the ensemble averaged statistics from
multiple microstructure maps, presumably obtained for
samples with nominally the same processing history. This
concept needs further rigorous treatment and development.
Acknowledgments
The authors acknowledge nancial support for this
work from the Oce of Naval Research, Award No.
N000140510504 (Program Manager: Dr Julie Christodoulou). SRN has been supported by the National Science
Foundation Drexel/UPenn IGERT Program in Nanoscale
Science & Engineering NSF Award No. DGE-0221664.

References
[1] Torquato S, Stell G. J Chem Phys 1982;77:2071.
[2] Torquato S, Stell G. J Chem Phys 1985;82:980.
[3] Ohser J, Mucklich F. Statistical analysis of microstructures in
materials science. Chichester: John Wiley; 2000.
[4] Roberts AP, Knackstedt MA. Phys Rev E 1996;54:2313.
[5] Quintanilla J, Torquato S. Phys Rev E 1997;55:1558.
[6] Tewari A, Gokhale AM, Spowart JE, Miracle DB. Acta Mater
2004;52:307.
[7] Gokhale AM, Singh H. Microsc Microanal 2005;11:1632.
[8] Gao X, Przybyla C, Adams B. Metall Mater Trans A 2006;37:2379.
[9] Adams BL, Gao X, Kalidindi SR. Acta Mater 2005;53:3563.
[10] Fullwood DT, Adams BL, Kalidindi SR. Comput Mater Sci
2007;38:788.
[11] Bunge HJ. Texture analysis in materials science. London: Butterworth;
1982.
[12] Torquato S. Ind Eng Chem Res 2006;45:6923.
[13] Gokhale AM, Tewari A, Garmestani H. Scripta Mater 2005;53:989.
[14] Uche OU, Stillinger FH, Torquato S. Phys A Stat Mech Appl
2006;360:21.
[15] Kalidindi SR, Binci M, Fullwood D, Adams BL. Acta Mater
2006;54:3117.
[16] Kalidindi SR, Houskamp JR, Lyons M, Adams BL. Int J Plast
2004;20:1561.
[17] Wu X, Proust G, Knezevic M, Kalidindi SR. Acta Mater
2007;55:2729.
[18] Proust G, Kalidindi SR. J Mech Phys Solids 2006;54:1744.
[19] Fullwood DT, Niezgoda SR, Kalidindi SR. Acta Mater 2008;56:942.
[20] Fullwood DT, Kalidindi SR, Niezgoda SR, Fast A, Hampson N. Mat
Sci Eng A, in press. doi:10.1016/j.msea.2007.10.087.
[21] Frisch HL, Stillinger FH. J Chem Phys 1963;38:2200.
[22] Torquato S. Random heterogeneous materials. New York: Springer
Verlag; 2002.
[23] Cohen L. Signal processing letters. IEEE 1998;5:292.
[24] Jiao Y, Stillinger FH, Torquato S. Phys Rev E Stat Nonlinear and
Soft Matter Phys 2007;76:031110.
[25] Folland GB. Real analysis: modern techniques and their applications. New York: Wiley-Interscience; 1999.
[26] Boggs PT, Tolle JW. J Comput Appl Math 2000;124:123.