Acta Materialia 179 (2019) 317e327

Contents lists available at ScienceDirect

Acta Materialia
journal homepage: www.elsevier.com/locate/actamat

Full length article

Hierarchical n-point polytope functions for quantitative

representation of complex heterogeneous materials and
microstructural evolution
Pei-En Chen a, Wenxiang Xu b, c, Nikhilesh Chawla b, Yi Ren a, **, Yang Jiao b, *
a
Mechanical Engineering, Arizona State University, Tempe, AZ, 85287, USA
b
Materials Science and Engineering, Arizona State University, Tempe, AZ, 85287, USA
c
College of Mechanics and Materials, Hohai University, Nanjing, 211100, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Effective and accurate characterization and quantification of complex microstructure of a heterogeneous
Received 5 May 2019 material and its evolution under external stimuli are very challenging, yet crucial to achieving reliable
Received in revised form material performance prediction, processing optimization and advanced material design. Here, we
13 August 2019
address this challenge by developing a set of hierarchical statistical microstructural descriptors, which
Accepted 25 August 2019
Available online 28 August 2019
we call the “n-point polytope functions” Pn, for quantitative characterization, representation and
modeling of complex material microstructure and its evolution. These polytope functions successively
include higher-order n-point statistics of the features of interest in the microstructure in a concise,
Keywords:
Quantitative microstructure representation
expressive, explainable, and universal manner; and can be directly computed from multi-modal imaging
(QMR) data. We develop highly efficient computational tools to directly extract the Pn functions up to n ¼ 8 from
N-point polytope functions multi-modal imaging data. Using simple model microstructures, we show that these statistical de-
Heterogeneous materials scriptors effectively “decompose” the structural features of interest into a set of “polytope basis”,
Microstructure evolution allowing one to easily detect any underlying symmetry or emerging features during the structural
Multi-modal imaging data evolution. We apply the Pn functions to quantify and model a variety of heterogeneous material systems,
including particle-reinforced composites, metal-ceramic composites, concretes, porous materials; as well
as the microstructural evolution in an aged lead-tin alloy. Our results indicate that the Pn functions can
offer a practical set of basis for quantitative microstructure representation (QMR), for both static 3D
complex microstructure and 4D microstructural evolution of a wide spectrum of heterogeneous material
systems.
© 2019 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

1. Introduction application of advanced non-destructive in situ imaging techniques,

such as x-ray micro-computed tomography (mCT) [1,2]. The second
Effective and accurate characterization and quantification of the challenge involves the development of efficient mathematical
complex microstructure of heterogeneous materials and micro- frameworks and computational tools for quantitative representa-
structural evolution under external stimuli or during processing are tion, modeling and reconstruction of complex heterogeneous ma-
crucial to processing optimization and advanced material design, terial microstructures and their evolution. In a recent TMS report on
yet are very challenging to achieve. One challenge involves the in advanced computation and data in materials and manufacturing,
situ characterization of the 3D microstructure containing features quantitative microstructure representation (QMR) has been iden-
of interest on multiple length scales as well as the 4D evolution tified as a key technical knowledge gap in core knowledge areas
processes (3D microstructure þ temporal evolution). This challenge “Coupling Simulations and Experiments” and “Digital Representa-
has been partially addressed by the development and successful tion and Visualization” [3].
Recently, a variety of novel techniques have been developed to
address the challenges in QMR, in particular, to devise reduced-
* Corresponding author. dimension quantitative microstructure representations for het-
** Corresponding author. erogeneous material systems based on either complete 3D or lower
E-mail addresses: [email protected] (Y. Ren), [email protected] (Y. Jiao).

https://doi.org/10.1016/j.actamat.2019.08.045
1359-6454/© 2019 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
318 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327

dimensional imaging data set [4e24], and to devise the associated In particular, we develop highly efficient computational tools to
virtual material generation/reconstruction methods [25e48]. Ex- directly extract these statistical descriptors from multi-modal im-
amples of established microstructure representation schemes aging data, and investigate the information content of the functions
include random field models, statistical descriptor-based repre- for quantifying different types of structural features and their
sentations, and abstract image-based decompositions obtained via evolution. Furthermore, we apply the Pn functions quantify and
machine-learning, to name but a few. Among the descriptor-based model a variety of heterogeneous material systems, including
representations, a recently developed framework for Hierarchical particle-reinforced composites, metal-ceramic composites, con-
Materials Informatics, based on complete 2-point statistics and its cretes, porous materials; as well as the microstructural evolution in
lower dimensional projections [49e51], can directly yield accurate an aged lead-tin alloy. Our results indicate that the Pn functions can
estimates of material properties [42,52e59] and thus, has been offer a practical set of basis for quantitative microstructure repre-
incorporated into various integrated computational material design sentation (QMR), for both static 3D complex microstructure and 4D
frameworks. microstructural evolution of a wide spectrum of heterogeneous
In general, it is reasonable to expect that an ideal set of material material systems.
representations (QMR) shall acquire the following crucial proper-
ties: (1) conciseness to facilitate tractable sampling and inverse 2. The n-point polytope functions pn
design within the low-dimensional representation space; (2)
expressiveness to enable high-fidelity reconstruction of complex 2.1. Definition of the n-point polytope functions pn
microstructures with heterogeneous morphologies, and accurate
predictions of material properties of interest; (3) interpretability to In this section, we introduce the n-point polytope functions Pn.
allow easy and intuitive understanding of the key morphological Without loss of generality, we consider a heterogeneous micro-
features in the material systems and their evolution from the rep- structure (either in 3D or a 2D slice) in which the different struc-
resentations. We note this last property “interpretability” is crucial tural features are segmented and grouped into different “phases”. A
for microstructural evolution representation, and that the limita- simple example is a composite microstructure contains a “matrix
tion of the state-of-art frameworks, including the data-driven phase” and a “particle phase”, see Fig. 1a. Note this microstructure
machine learning techniques, is mainly the lack of “interpret- can be a snapshot from an evolution process. The definition of Pn is
ability”. For example, the representations obtained in these then given as follows:
frameworks (e.g., the image-based decompositions of latent space
in convolutional neural network models [60]) are usually too ab- Pn(r) ≡ Probability that all of the n vertices of a randomly selected
stract and very challenging to visualize and interpret physically. regular n-point polytope with edge length r fall into the phase of
However, we emphasize that these existing representations are interest
very effective and successful in achieving conciseness, expressive-
ness and computational efficiency. Base on this definition, one can derive two sets of the Pn func-
Another class of widely used descriptor-based representations tions. The first set involves n-point regular polygons, for which the
involves the standard n-point correlation (or probability) function vertex (edge) number n can take any positive integer values; and in
Sn, which provides the probability of occurrence of specific n-point the limit n/∞, the shape becomes a circle (in this limiting case, the
configurations in the material microstructure [61e64]. The com- quantity r is the radius of the circle, instead of the edge length). We
plete set of Sn with n ¼ 1, 2, 3, …∞ provides a complete quantitative note that the n-point polygon functions can be computed from both
microstructure representation, and thus, determines the physical 2D slices and full 3D microstructure. The other set involves 3D
properties and performance of the material system under consid- polyhedra whose edges are of the same length. Only a small
eration. It has been shown that even the lower order function S2 can number of 3D polyhedra satisfy this condition, including the five
be employed to model a wide spectrum of different heterogeneous Platonic solids (i.e., the regular polyhedra: tetrahedron, octahedron,
material systems [65e72]. In practice, it is very challenging to uti- dodecahedron, icosahedron, and cube) and the thirteen Archime-
lize Sn with n  3, for which one needs to enumerate all distinct n- dean solids (i.e., the semi-regular polyhedra) (see Fig. 1b) [81]. In the
point configurations and efficiently compute and store their prob- case of n ¼ 2, the 2-point polytope function P2 is identical to the
ability of occurrence. The resulting statistical data sets are typically standard 2-point correlation function S2, which is also equivalent to
much larger in size than the original imaging data for the material the lineal-path function if the material only contains isolated
system. It has been shown that two-point statistics alone might not compact clusters [82e85]. The higher-order Pn functions can be
be sufficient to represent certain complex microstructures [73e76]. considered as special one-dimensional projection (or “slice”) of the
An alternative approach is to employ non-standard lower-order full standard n-point correlation function Sn, which themselves are
correlation functions such as the cluster functions [77,78] or surface very difficult to directly utilize and compute. Using the Pn functions
functions [79,80], which encode partial higher-order n-point sta- allows us to employ the partial higher order structural statistics in
tistics. This method is very effective in capturing specific morpho- the Sn functions, which could provide better quantification of
logical features (e.g., clustering) [35] but cannot allow to certain subtle structural features that could not be captured by
systematically incorporate hierarchical higher order structural in- lower order functions as illustrated in Sec. 3.
formation for microstructure quantification. Fig. 1a schematically illustrates the stochastic events that
In this paper, we develop a set of hierarchical statistical micro- contribute to the Pn functions in the case of regular polygons. For
structural descriptors, called the “n-point polytope functions” Pn, r ¼ 0, the polygon reduces to a single point, and Pn(r ¼ 0) gives the
which can simultaneously achieve conciseness, expressiveness, and volume fraction 4 of the phase of interest (i.e., the probability a
interpretability, for quantitative characterization, representation randomly selected point falling into the phase of interest). For finite
and modeling of microstructural evolution during processing. r values, Pn(r) provides n-point spatial correlations in the phase
These polytope functions successively include higher-order n-point (feature) of interest. For very large r values (e.g., r/∞), the prob-
statistics of the features of interest in the microstructure, and can abilities of finding the vertices in the phase of interest are almost
be directly computed from multi-modal imaging data, including x- independent of one another, thus, we have Pn(r/∞) z 4n, where 4
ray tomographic radiographs, optical/SEM/TEM micrographs, and is the volume fraction of the phase of interest in the system. These
EBSD color maps for quantification of different features of interest. asymptotic behaviors allow us to introduce a convenient re-scaled
 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327 319

Fig. 1. (a) Schematic illustration of stochastic events contributing to the Pn functions in the case of regular polygons. (b) Examples of regular polygons (triangle, square, pentagon,
hexagon, and decagon) and polyhedra (including the Platonic solids P1eP5, and Archimedean solids A1-A13) for computing the Pn functions.

form of the Pn functions, i.e., full Pn(r) function.

We note that for a digitized representation of a material
fn ðrÞ ¼ ½Pn ðrÞ  4n  = ½4  4n  (1) microstructure (i.e., an image), a point falls into the phase of in-
terest if it fall into a pixel (or voxel in 3D) of that phase. Therefore,
with fn(r ¼ 0) ¼ 1 and fn(r/∞) ¼ 0. Finally, we note that one can the edge length of the pixel/voxel defines the smallest distance in
define “cross-correlation” polytope functions, by requiring a subset the system and provides a natural unit for measuring the distance.
of the vertices falling into different phases (features) in the In addition, besides placing the polytope template with random
microstructure. In this paper, we will mainly focus on the “auto” orientations (i.e., isotropic sampling), one can fix the orientation of
polytope functions defined with 2D polygons, in which all of the the polytopes while placing them at randomly selected locations,
vertices of the polygons fall into the same phase of interest. We also see Fig. 2b. We refer to this later case as “directional sampling”.
note that unlike the Sn function, which naturally includes all of the
lower-order functions S1, S2, …, Sn-1 as degenerate cases (i.e., when
a subset of points coincide together), each Pn function contains 3. Information content of pn functions
unique statistical information of n-polytope correlations that is
complementary to that contained in the lower-order P1, P2, …, Pn-1 In this section, we investigate the information content of the Pn
functions. For example, the P8 function quantifies the octagonal functions using simple model microstructures. As shown in Fig. 3a
correlations in the system, and contains no information of hep- and Fig. 4a, the model microstructures are composed of congruent
tagonal correlations contained in the P7 function nor the hexagonal equilateral triangle particles arranged on a square lattice and
correlations contained in P6, etc. Therefore, P8 is complementary to triangular lattice, respectively. Within each model microstructure,
P7, and all the other lower-order Pn functions. As illustrated in Sec. 3, the orientations of all triangles are the same; and different micro-
an interesting property of these Pn functions is that the higher order structures contain triangles with different orientations.
functions stat to pick up the underlying n-fold symmetry from the Fig. 3bec shows the directional 3-point polytope functions P3
system, which is often difficult to capture otherwise. for the model microstructures shown in Fig. 3a. In particular, the
directional P3 functions are computed by fixing the orientation of
the equilateral triangular sampling templates vertically and hori-
2.2. Extracting pn functions from imaging data zontally respectively (see the insets of the figures). It can be seen
that the P3 functions for all microstructures start with the same
The probability-based definition of the Pn functions allow us to value at r ¼ 0, which is the volume fraction of the particles; and
easily compute these function from microstructural data, including monotonically decrease as one moves away from r ¼ 0. This
both 2D images and 3D digital representations of the material behavior is well known in the case of two-point correlation func-
microstructure. For example, in order to compute the value of Pn(r) tion S2^P3. This is because for small r, the majority of contributions
with r ¼ r*, the following procedure is used: to the correlation functions are from the events in which the
sampling templates entirely fall into a single particle. Clearly, the
(i) We first generate a regular n-polytope (e.g., a n-polygon) success of such events rapidly decreases as the size of the template
with edge length r*; increases. As we will see later, this monotonic decaying behavior is
(ii) This n-polygon is then placed in the material microstructure the case for all Pn functions due to the same reason. We note that
with randomly selected center location and random orien- when the orientation of the triangular particles is consistent with
tation, for M times (see Fig. 2a); the orientation of the sampling triangular template, the resulting P3
(iii) Each time the polygon is placed in the system, we check function exhibits a slower decay for small r. This is because in such
whether all of its vertices fall into the phase of interest (i.e., a cases, the sampling templates with larger sizes (i.e., associated with
“success” event), and we count the total number of success larger r values) can still entirely fall into a single particle.
event Ms out of a total of M trials; Interestingly, it can be seen that beyond the initial decay, the P3
(iv) We compute Pn(r ¼ r*) ¼ Ms/M, which is the probability that functions exhibit significant peaks at certain intermediate and
a randomly selected n-polygon having all its vertices fall into event large r values (e.g., r z 180, 290, 340, 430 pixels). These peaks
the phase of interest. indicate the existence strong triangular correlations (i.e., hidden
triangular patterns/symmetry) on the length scales as defined by
This procedure is repeated for different r values to compute the the associated r values. We note it is very surprising to detect the
320 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327

Fig. 2. Schematic illustration of different sampling templates for computing Pn functions (in the case of n ¼ 3) from images: (a) Isotropic sampling, in which both the location and
orientation of the polygon template are randomly selected. (b) Directional sampling, in which only the location of the template is randomly selected.

Fig. 3. (a) Model microstructures composed of congruent equilateral triangle particles with different orientations arranged on a square lattice. The resolution of the images is 1048
by 1048 pixels. (b) Directional 3-point polytope functions P3 for the model microstructures associated with vertically oriented triangular sampling templates. (c) Directional 3-point
polytope functions P3 for the model microstructures associated with horizontally oriented triangular sampling templates. (d) 4-point polytope functions P4 for the model mi-
crostructures. (e) 6-point polytope functions P6 for the model microstructures. The unit of distance is one pixel length.

emergence of large-scale triangular correlation (symmetry) in efficient in capturing hidden order and spatial correlations in the
packing arrangements based on square lattice. This example also microstructure.
indicates that the statistical descriptors we developed are very Fig. 3d shows the 4-point polytope functions P4 for the model
 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327 321

Fig. 4. (a) Model microstructures composed of congruent equilateral triangle particles with different orientations arranged on a triangular lattice. The resolution of the images is
1048 by 1048 pixels. (b) Directional 3-point polytope functions P3 for the model microstructures associated with vertically oriented triangular sampling templates. (c) Directional 3-
point polytope functions P3 for the model microstructures associated with horizontally oriented triangular sampling templates. (d) 4-point polytope functions P4 for the model
microstructures. (e) 6-point polytope functions P6 for the model microstructures. The unit of distance is one pixel length.

microstructures. A series of peaks (i.e., oscillations) beyond the as r increases. These peaks are associated with triangular correla-
initial decay can be clearly observed. These peaks correspond to the tions on different length scales, resulted from the underlying
square correlation/symmetry on different length scales in the triangular packing lattice. Fig. 4d shows the 4-point polytope
packings, which are resulted from the square symmetry of the functions P4 for the model microstructures. The hidden square
underlying packing lattice. The P4 functions are almost identical for correlations on large length scales (e.g., r z 275, 340 pixels) are
all four microstructures (see Fig. 3a), indicating that this function is again picked up and manifested as the associated peaks. Fig. 4e
not sensitive to the local particle shape, but only depends on the shows the 6-point polytope functions P6 for the model micro-
packing arrangements of the particles. structures. Similar to the P3 functions, the P6 functions exhibit
Fig. 3e shows the 6-point polytope functions P6 for the model significant peaks and strong oscillations, indicating the hidden
microstructures. It can be seen that the initial decay of P6 is much hexagonal correlations resulted from the triangular packing
faster than that in P3 and P4. This is because it is more difficult to arrangements.
entirely fit a hexagon in the particle compared to a triangle or a We can see from these examples that the Pn functions can
square with the same edge length. The P6 functions also exhibit provide a systematic “decomposition” of the features of interest in
significant peaks on certain intermediate and large length scales the microstructure in terms of a series of polytope basis and very
(e.g., r z 180, 290, 340, 430 pixels), indicating hidden hexagonal effective in capturing hidden symmetry/spatial orders in the sys-
symmetry/correlations on that length scale. These length scales are tem. Therefore, the set of Pn functions offers a systematic way to
completely consistent with the length scales associated with the devise more accurate microstructure representations by succes-
triangular peaks in P3. sively incorporating higher order morphological information.
Fig. 4bec shows the directional 3-point polytope functions P3
respectively associated with vertically and horizontally oriented 4. Quantification of complex heterogeneous materials and
triangular sampling templates, for the model microstructures microstructural evolution
shown in Fig. 4a. The P3 functions initially decay and exhibits a
series of significant peaks (oscillations) with decaying magnitude In this section, we employ the Pn functions to quantify and
322 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327

model a variety of heterogeneous materials with distinct micro- n-polytope with edge length r. Fig. 6c shows the corresponding
structures, including particle reinforced composites, bi-phase rescaled fn functions highlighting the spatial correlations (c.f. Eq.
interpenetrating composites and porous materials. In addition, (1)).
we also apply the Pn functions to quantify and model the micro- Fig. 7 shows the quantification of a 2D concrete microstructure
structural evolution (e.g., coarsening development) in a lead-tin composed of reinforcement rocks (shown in black) and the cement
alloy aged at 175  C up to 120 h. paste (shown in white) [35] (see Fig. 7a). The rock particles possess
complex polygonal morphologies and a wide size distribution. The
4.1. Quantification of complex microstructures size of the image is 400 by 400 pixels. Fig. 7b shows the Pn functions
for the rock phase with n ¼ 2, 3, 4, 6 and 8. Similar to the previous
Fig. 5 shows the quantification of a 2D microstructure composed systems, all Pn functions initially decay from the volume fraction
of equal-sized hard spheres in a matrix [31], i.e., a packing (see 4 ¼ 0.48 as r increases from 0. The r value associated with the first
Fig. 5a). The sphere packing is generated using Monte Carlo simu- minimum in the functions provides the average particle size in the
lations [81]. The size of the image is 1024 by 1024 pixels and the system, i.e., ~40 pixels. After the initial decay, the lower-order Pn
diameter of the particle is D ¼ 65 pixels. Fig. 5b shows the Pn functions (e.g., n  4) exhibit weak oscillations, reflecting the
functions for the particle phase with n ¼ 2, 3, 4, 6 and 8. Similar to spatial correlations resulted from the mutual exclusion effects of
the cases of the model microstructures discussed in Sec. 3, all Pn the rock particles. The correlations are much weaker compared
functions initially decay from the volume fraction 4 ¼ 0.545 as r those in hard-sphere systems, mainly due to the anisotropy and
increases from 0. The positions of the first minimum in the Pn size polydispersity of the rock particles. Fig. 7c shows the corre-
functions for small n values roughly correspond to the linear size of sponding rescaled fn functions highlighting the spatial correlations
the particle (~65 pixels). After the initial decay, all Pn functions (c.f. Eq. (1)).
studied here except for n ¼ 8 exhibit strong oscillations and the first Fig. 8 shows the quantification of a 2D microstructure of a
peaks in different Pn functions occur at approximately the same r Fontainebleau sandstone composed of the rock phase (shown in
values. These oscillations respectively indicates strong pair, trian- black) and the pore phase (shown in white) [34] (see Fig. 8a).
gle, square and hexagonal correlations on different length scales in Similar to the concrete microstructure, the rock particles possess
the system. Fig. 5a illustrates examples of such correlations, which complex shapes and a wide size distribution. In addition, the par-
are all associated with the mean nearest neighbor separate dis- ticles are densely packed and compressed such that their bound-
tance, i.e., the distance associated with the first peak in P2. These aries are fused and cannot be clearly distinguished. The size of the
correlations result from the tendency for the particles to self- image is 512 by 512 pixels. Fig. 8b shows the Pn functions for the
organize on a triangular lattice at high densities. The P8 is almost rock phase with n ¼ 2, 3, 4, 6 and 8. Similar to the previous systems,
flat after the initial decay, indicating the system does not possess all Pn functions initially decay from the volume fraction 4 ¼ 0.82 as r
any octagonal correlations on intermediate and large length scales. increases from 0. The r value associated with the first minimum in
Fig. 5c shows the corresponding rescaled fn functions highlighting the functions provides the average particle size in the system, i.e.,
the spatial correlations (c.f. Eq. (1)). ~32 pixels. After the initial decay, the Pn functions exhibit very weak
Fig. 6 shows the quantification of a 2D microstructure composed oscillations for small and intermediate r values. This behavior is
of equal-sized overlapping spheres in a matrix [31] (see Fig. 6a). The similar to that observed for the overlapping sphere system, as the
spheres are randomly placed in the matrix without any built-in rocks are compressed and fused which can be effectively consid-
spatial correlations. The size of the image is 512 by 512 pixels and ered as “overlapped” in the boundary regions. Fig. 8c shows the
the diameter of the particle is D ¼ 45 pixels. Fig. 6b shows the Pn corresponding rescaled fn functions highlighting the spatial corre-
functions for the particle phase with n ¼ 2, 3, 4, 6 and 8. Similar to lations (c.f. Eq. (1)).
the previous systems, all Pn functions initially decay from the vol- Fig. 9 shows the quantification of a 2D an interpenetrating
ume fraction 4 ¼ 0.50 as r increases from 0. After the initial decay, metal-ceramic composite composed of the boron-carbide phase
all Pn functions are virtually flat, indicating that the particles (shown in black) and the aluminum phase (shown in white) [34]
possess no spatial correlations of any symmetry on any length (see Fig. 9a). This microstructure contains “ligaments” of similar
scales beyond the diameter of the particles. We note that for the width instead of “particles”. The size of the image is 512 by 512
totally random overlapping sphere system, the Pn functions possess pixels. Fig. 8b shows the Pn functions for the rock phase with n ¼ 2,
the analytical expression PnI ¼ exp[-rvn(r; R)], where r is the 3, 4, 6 and 8. Similar to the previous systems, all Pn functions
number density of the spheres in the system (i.e., number of initially decay from the volume fraction 4 ¼ 0.65 as r increases from
spheres per unit volume) and vn(r; R) is the volume of the union of 0. The r value associated with the first minimum in the functions
n spheres with radius R with centers placed at the vertices of a provides the average ligament width in the system, i.e., ~27 pixels.

Fig. 5. (a) 2D microstructure composed of equal-sized hard spheres in a matrix. (b) The Pn functions for the particle phase with n ¼ 2, 3, 4, 6 and 8. (c) The corresponding rescaled fn
functions highlighting the spatial correlations (c.f. Eq. (1)).
 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327 323

Fig. 6. (a) 2D microstructure composed of equal-sized overlapping spheres in a matrix. (b) The Pn functions for the particle phase with n ¼ 2, 3, 4, 6 and 8. (c) The corresponding
rescaled fn functions highlighting the spatial correlations (c.f. Eq. (1)).

Fig. 7. (a) 2D image of a concrete microstructure composed of reinforcement rocks (shown in black) and the cement paste (shown in white). (b) The Pn functions for the rock (black)
phase with n ¼ 2, 3, 4, 6 and 8. (c) The corresponding rescaled fn functions highlighting the spatial correlations (c.f. Eq. (1)).

Fig. 8. (a) 2D image of a Fontainebleau sandstone microstructure composed of the rock phase (shown in black) and the pore phase (shown in white). (b) The Pn functions for the
rock (black) phase with n ¼ 2, 3, 4, 6 and 8. (c) The corresponding rescaled fn functions highlighting the spatial correlations (c.f. Eq. (1)).

Fig. 9. (a) 2D image of an interpenetrating metal-ceramic composite microstructure composed of the boron-carbide phase (shown in black) and the aluminum phase (shown in
white). (b) The Pn functions for the boron-carbide (black) phase with n ¼ 2, 3, 4, 6 and 8. (c) The corresponding rescaled fn functions highlighting the spatial correlations (c.f. Eq. (1)).
324 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327

After the initial decay, the Pn functions exhibit weak oscillations for interest, which are a subset of the full n-point correlation function
small and intermediate r values, characterizing the exclusion ef- Sn. We have shown that the Pn functions allow one to systematically
fects between the ligaments. Fig. 9c shows the corresponding incorporate higher order spatial correlations for microstructure
rescaled fn functions highlighting the spatial correlations (c.f. Eq. quantification in a concise, expressive, and interpretable fashion,
(1)). and provide an effective decomposition of the features of interest
into polytope correlations. The utility of the Pn functions has been
4.2. Quantification of microstructure evolution demonstrated by applying them to quantify and model a variety of
complex heterogeneous material systems as well as microstruc-
We now employ the Pn functions to quantify microstructure tural evolution. Our results suggest that the n-point polytope
evolution (e.g., coarsening development) in a lead-tin alloy aged at functions offer a practical set of basis for quantitative microstruc-
175  C up to 120 h, see Fig. 10. In Ref. [65], we model the evolution ture representation (QMR), for both static 3D complex micro-
process using time-dependent 2-point correlation function S2, structure and 4D microstructural evolution of a wide spectrum of
which is equivalent to P2. We showed that when properly scaling heterogeneous material systems. In addition, we note that the
with the time-dependent increasing length scale (e.g., the average polytope functions can be readily incorporated into material
width of the ligaments), the S2 functions correspond to different reconstruction algorithms for fast virtual 3D microstructure gen-
aging time all collapse approximately onto a universal curve, eration and also allow fast material property prediction via
capturing the intrinsic density fluctuations in the system. Here, we analytical structure-property mappings. The reconstruction can be
investigate the higher order functions. either formulated as a stochastic optimization [31,34,35] or con-
Fig. 11 shows the Pn functions (with n ¼ 2, 3, 4, 6 and 8) asso- strained neural network models [60,86]. This would potentially
ciated with the PbeSn alloy microstructures at different aging enable in-processing material property monitoring and real-time
times as shown in Fig. 10. It can be seen that for Pn functions, the processing optimization.
initial decay as r increases from 0 becomes slower as aging time In order to achieve the aforementioned utility, an important
increases. In other words, the r value associated with the first question is that how many Pn functions one should consider for a
minimum in the functions increases as aging time increases. As given material system of interests? To rigorously determine the set
discussed above, this r value characterizes the average width of the of correlation functions that are sufficient and necessary to quantify
“ligaments” in the microstructure, which coarsens as aging time a given material system, the density of states (i.e., number of mi-
increases. crostructures compatible with the given set of constrained func-
Moreover, at large aging time (e.g., t ¼ 120 h), we can see that tions up to a small tolerance) and microstructural degeneracy need
significant triangle and square correlations emerge in the micro- to be computed [73], which is very computationally challenging. A
structure, which are manifested as the relatively strong oscillations less rigorous but more practical method is to incorporate the target
in the corresponding P3 and P4 functions. Such higher order cor- set of functions into inverse material reconstruction [35]. The in-
relations are not observed for small aging time and might be formation content of the correlation functions can then be ascer-
resulted from the additional correlation induced due to coarsening tained from the accuracy of the reconstructed materials. Our
development. We note that the P2 functions also exhibit stronger previous experience also indicates that if the a particular correla-
oscillations as coarsening proceeds. It is now clear that the tion function Pn does not possess significant oscillations beyond the
observed pair correlations are associated with the emergence of initial decay as the distance increases, the system typically does not
higher order correlations due to coarsening. This example illus- possess strong n-polytope correlations. Therefore, that particular Pn
trates the importance of incorporating higher-order statistics for function might not be necessary for quantifying the system. For
accurate quantification and representation of complex micro- most of material systems studied in the paper (such as the metal-
structural evolution. ceramic composite, concrete, sandstone and PbeSn alloy), we can
see that the octagonal function P8 does not possess significant os-
5. Conclusions and discussion cillations, which indicates that for these systems, it might not be
necessary to go up to n ¼ 8. However, for the model composite
In this paper, we have introduced a set of hierarchical statistical microstructure (i.e., a packing of equal-sized spheres in a matrix), P8
descriptors, i.e., the n-point polytope functions Pn for quantitative clearly shows oscillations for small and intermediate distances,
representation of complex heterogeneous materials and micro- indicating P8 is required in order to accurately capture octagonal
structure evolution. The Pn functions provide the probability of correlations on these length scales.
finding regular n-polytopes with vary sizes in the material phase of The set of Pn functions can be employed to estimate linear

Fig. 10. Representative microstructures of the lead-tin alloy (Pb37Sn63) at 175  C for different times. The dark region represents lead-rich phase and white region represents tin-
rich phase. The linear size of the microstructure is ~100 mm. The resolution of the image is 600 by 600 pixels.
 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327 325

Fig. 11. The Pn functions associated with the PbeSn alloy microstructures at different aging times.

properties of the material systems, such as thermal/electrical con- accurate estimates by successively incorporating higher order Pn
ductivity, effective diffusivity, and elastic moduli, via the so-call functions. Finally, we expect that the Pn functions can also be linked
strong contrast expansion formalism [87e89]. In this formalism, to certain nonlinear properties (e.g., brittle fracture strength).
the effective material properties can be expressed as an infinite However, for these cases no analytical procedure is available. One
series involving integrals of the n-point correlation functions. Ap- possible approach is to employ supervised deep learning (SDL),
proximations of the properties can be obtained by truncating the which utilizes the Pn functions as basis to correlate the properties of
infinite series at some finite order. For example, consider a two- interest based on properly selected training data.
phase material, the 3rd order approximation of the bulk modulus
Ke, obtained by truncating the series expansion at “3-point” term Acknowledgement
involving P3 is given by
This work is supported by ACS Petroleum Research Fund under
к 10G1 к21 m21 Grant No. 56474-DNI10 (Program manager: Dr. Burtrand Lee). Y.J.
F2 21 ¼ 1  F z (2)
кe1 3ðK1 þ 2G1 Þ 1 2 and N.C. acknowledge the Office of Knowledge Enterprise Devel-
opment (OKED) and the Fulton Schools of Engineering for a seed
where F1 andF2 are respectively the volume fractions of the phases, grant.
and
Appendix A. Supplementary data
K K G2  G1 Ke  K1
к21 ¼ 2 4 1 m21 ¼ к ¼
G1 ½3K1 =2þ3G1 =4 e1
(3)
K1 þ 3G1 4 Supplementary data to this article can be found online at
G1 þ K1 þ2G1
K1 þ 3G1
https://doi.org/10.1016/j.actamat.2019.08.045.
K1 and G1 are the bulk and shear modulus of phase 1 and K2 and
G2 are the bulk and shear modulus of phase 2. z2 in (Eq. (2)) is the 3- References
pt parameter given by
[1] J. Williams, K. Yazzie, E. Padilla, N. Chawla, X. Xiao, F. De Carlo, Understanding
fatigue crack growth in aluminum alloys by in situ X-ray synchrotron to-
ð
∞ ð
∞ ð1 "
9 dr ds    mography, Int. J. Fatigue 57 (2013) 79e85.
ð2Þ
z2 ¼ dðcosqÞLðcosqÞ P 3 √ r 2 þ s2 þ t 2 [2] S. Singh, J. Williams, M. Lin, X. Xiao, F. De Carlo, N. Chawla, In situ investigation
2F1 F2 r s of high humidity stress corrosion cracking of 7075 aluminum alloy by three-
0 0 1 dimensional (3D) X-ray synchrotron tomography, Mater. Res. Lett. 2 (4)
# (2014) 217e220.
 ð2Þ ð2Þ
P 2 ðrÞP 2 ðsÞ [3] D. McDowell, V. Sundaraghavan, J.A. Scott, Advanced Computation and Data in
 
F2 Materials and Manufacturing: Core Knowledge Gaps and Opportunities, TMS
Technical Report, 2018.
(4) [4] Y. Gueguen, J. Dienes, Transport properties of rocks from statistics and
percolation, Math. Geol. 21 (1) (1989) 1e13.
[5] M.E. Frary, C.A. Schuh, Correlation-space description of the percolation tran-
where t ¼ ðr 2 þs2 2rscosqÞ 1=2 (i.e., r, s, t are three edge lengths of a sition in composite microstructures, Phys. Rev. 76 (4) (2007), 041108.
triangle), and L(x) is the Legendre polynomials of order two, and the [6] M. Groeber, S. Ghosh, M.D. Uchic, D.M. Dimiduk, A framework for automated
superscript “(2)” in the Pn functions indicates they are associated analysis and simulation of 3d polycrystalline microstructures.: Part 1: statis-
tical characterization, Acta Mater. 56 (6) (2008) 1257e1273.
with phase 2, which is selected as the phase of interest without loss [7] S. Niezgoda, D. Fullwood, S. Kalidindi, Delineation of the space of 2-point
of generality. It is reasonable to expect that one can achieve more correlations in a composite material system, Acta Mater. 56 (18) (2008)
326 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327

5285e5292. 34 (10) (2011) 1256e1267.

[8] L. Jiang, N. Chawla, M. Pacheco, V. Noveski, Three-dimensional (3D) micro- [37] M. Blacklock, H. Bale, M. Begley, B. Cox, Generating virtual textile composite
structural characterization and quantification of reflow porosity in Sn-rich specimens using statistical data from micro-computed tomography: 1D tow
alloy/copper joints by X-ray tomography, Mater. Char. 62 (10) (2011) representations for the Binary Model, J. Mech. Phys. Solids 60 (3) (2012)
970e975. 451e470.
[9] S.R. Kalidindi, S.R. Niezgoda, A.A. Salem, Microstructure informatics using [38] R.G. Rinaldi, M. Blacklock, H. Bale, M.R. Begley, B.N. Cox, Generating virtual
higher-order statistics and efficient data-mining protocols, JOM (J. Occup. textile composite specimens using statistical data from micro-computed to-
Med.) 63 (4) (2011) 34e41. mography: 3D tow representations, J. Mech. Phys. Solids 60 (8) (2012)
[10] J. Sosa, S. Koduri, V. Dixit, P. Collins, S. Niezgoda, S. Kalidindi, H. Fraser, 1561e1581.
Application of novel techniques to the three-dimensional characterization of [39] P. Tahmasebi, M. Sahimi, Cross-correlation function for accurate reconstruc-
microstructural features in aþ beta titanium alloys, in: Minerals, Metals and tion of heterogeneous media, Phys. Rev. Lett. 110 (7) (2013), 078002.
Materials Society/AIME, 420 Commonwealth Dr., P. O. Box 430 Warrendale PA [40] H. Xu, M.S. Greene, H. Deng, D. Dikin, C. Brinson, W.K. Liu, C. Burkhart,
15086 United States.[np], Feb, 2011. G. Papakonstantopoulos, M. Poldneff, W. Chen, Stochastic reassembly strategy
[11] S.E. Wilding, D.T. Fullwood, Clustering metrics for two-phase composites, for managing information complexity in heterogeneous materials analysis
Comput. Mater. Sci. 50 (7) (2011) 2262e2272. and design, J. Mech. Des. 135 (10) (2013) 101010.
[12] P.G. Callahan, Quantitative Characterization and Comparison of Precipitate [41] K.M. Gerke, M.V. Karsanina, R.V. Vasilyev, D. Mallants, Improving pattern
and Grain Shape in Ni-Base Superalloys Using Moment Invariants, Carnegie reconstruction using directional correlation functions, EPL (Europhysics Let-
Mellon University, 2012. ters) 106 (6) (2014) 66002.
[13] A. Cecen, E. Wargo, A. Hanna, D. Turner, S. Kalidindi, E. Kumbur, 3-D micro- [42] H. Xu, Y. Li, C. Brinson, W. Chen, A descriptor-based design methodology for
structure analysis of fuel cell materials: spatial distributions of tortuosity, void developing heterogeneous microstructural materials system, J. Mech. Des. 136
size and diffusivity, J. Electrochem. Soc. 159 (3) (2012) B299eB307. (5) (2014), 051007.
[14] A. Salem, D. Turner, D. Satko, J. Shaffer, S. Niezgoda, S. Kalidindi, Quantifying [43] K.M. Gerke, M.V. Karsanina, Improving stochastic reconstructions by
Ti-6 Al-4 V bimodal microstructure using microstructure informatics, in: AIP weighting correlation functions in an objective function, EPL (Europhysics
Conference Proceedings, American Institute of Physics, 2012. Ste. 1 NO 1 Letters) 111 (5) (2015) 56002.
Melville NY 11747-4502 United States. [44] X. Liu, V. Shapiro, Random heterogeneous materials via texture synthesis,
[15] M. Wang, J. Williams, L. Jiang, F. De Carlo, T. Jing, N. Chawla, Three dimen- Comput. Mater. Sci. 99 (2015) 177e189.
sional (3D) microstructural characterization and quantitative analysis of so- [45] R. Bostanabad, A.T. Bui, W. Xie, D.W. Apley, W. Chen, Stochastic microstruc-
lidified microstructures in magnesium-based alloys, Metallography, ture characterization and reconstruction via supervised learning, Acta Mater.
Microstructure, and Metallography. Microstruct. Anal. 1 (1) (2012) 7e13. 103 (2016) 89e102.
[16] D. Brandon, W.D. Kaplan, Microstructural Characterization of Materials, John [46] D.M. Turner, S.R. Kalidindi, Statistical construction of 3-D microstructures
Wiley & Sons, 2013. from 2-D exemplars collected on oblique sections, Acta Mater. 102 (2016)
[17] S.R. Niezgoda, J. Glover, Unsupervised learning for efficient texture estimation 136e148.
from limited discrete orientation data, Metall. Mater. Trans. A 44 (11) (2013) [47] M.V. Karsanina, K.M. Gerke, Hierarchical optimization: fast and robust mul-
4891e4905. tiscale stochastic reconstructions with rescaled correlation functions, Phys.
[18] S.R. Niezgoda, A.K. Kanjarla, S.R. Kalidindi, Novel microstructure quantifica- Rev. Lett. 121 (26) (2018) 265501.
tion framework for databasing, visualization, and analysis of microstructure [48] J. Feng, Q. Teng, X. He, X. Wu, Accelerating multi-point statistics reconstruc-
data, Integrating. Mater. Manuf. Innovat. 2 (1) (2013) 1e27. tion method for porous media via deep learning, Acta Mater. 159 (2018)
[19] S. Ratanaphan, Y. Yoon, G.S. Rohrer, The five parameter grain boundary 296e308.
character distribution of polycrystalline silicon, J. Mater. Sci. 49 (14) (2014) [49] S.R. Kalidindi, Hierarchical Materials Informatics: Novel Analytics for Mate-
4938e4945. rials Data, Elsevier, 2015.
[20] T. Hardin, T. Ruggles, D. Koch, S. Niezgoda, D. Fullwood, E. Homer, Analysis of [50] S.R. Kalidindi, J.A. Gomberg, Z.T. Trautt, C.A. Becker, Application of data science
traction-free assumption in high-resolution EBSD measurements, J. Microsc. tools to quantify and distinguish between structures and models in molecular
260 (1) (2015) 73e85. dynamics datasets, Nanotechnology 26 (34) (2015) 344006.
[21] A. Cecen, T. Fast, S.R. Kalidindi, Versatile algorithms for the computation of 2- [51] P. Steinmetz, Y.C. Yabansu, J. Ho €tzer, M. Jainta, B. Nestler, S.R. Kalidindi, An-
point spatial correlations in quantifying material structure, Integrating. Mater. alytics for microstructure datasets produced by phase-field simulations, Acta
Manuf. Innovat. 5 (1) (2016) 1e15. Mater. 103 (2016) 192e203.
[22] A. Choudhury, Y.C. Yabansu, S.R. Kalidindi, A. Dennstedt, Quantification and [52] D.T. Fullwood, S.R. Niezgoda, B.L. Adams, S.R. Kalidindi, Microstructure sen-
classification of microstructures in ternary eutectic alloys using 2-point spatial sitive design for performance optimization, Prog. Mater. Sci. 55 (6) (2010)
correlations and principal component analyses, Acta Mater. 110 (2016) 477e562.
131e141. [53] A. Jain, J.R. Errington, T.M. Truskett, Dimensionality and design of isotropic
[23] Z. Ma, S. Torquato, Precise algorithms to compute surface correlation func- interactions that stabilize honeycomb, square, simple cubic, and diamond
tions of two-phase heterogeneous media and their applications, Phys. Rev. 98 lattices, Phys. Rev. X 4 (3) (2014), 031049.
(1) (2018), 013307. [54] V. Sharma, C. Wang, R.G. Lorenzini, R. Ma, Q. Zhu, D.W. Sinkovits, G. Pilania,
[24] H. Malmir, M. Sahimi, Y. Jiao, Higher-order correlation functions in disordered A.R. Oganov, S. Kumar, G.A. Sotzing, Rational design of all organic polymer
media: computational algorithms and application to two-phase heteroge- dielectrics, Nat. Commun. 5 (2014).
neous materials, Phys. Rev. 98 (6) (2018), 063317. [55] A.F. Baldwin, T.D. Huan, R. Ma, A. Mannodi-Kanakkithodi, M. Tefferi, N. Katz,
[25] A.P. Roberts, Statistical reconstruction of three-dimensional porous media Y. Cao, R. Ramprasad, G.A. Sotzing, Rational design of organotin polyesters,
from two-dimensional images, Phys. Rev. 56 (3) (1997) 3203. Macromolecules 48 (8) (2015) 2422e2428.
[26] M. Pilotti, Reconstruction of clastic porous media, Transp. Porous Media 41 (3) [56] A. Gupta, A. Cecen, S. Goyal, A.K. Singh, S.R. Kalidindi, Structureeproperty
(2000) 359e364. linkages using a data science approach: application to a non-metallic inclu-
[27] D.M. Saylor, J. Fridy, B.S. El-Dasher, K.-Y. Jung, A.D. Rollett, Statistically sion/steel composite system, Acta Mater. 91 (2015) 239e254.
representative three-dimensional microstructures based on orthogonal [57] A. Kaczmarowski, S. Yang, I. Szlufarska, D. Morgan, Genetic algorithm opti-
observation sections, Metall. Mater. Trans. A 35 (7) (2004) 1969e1979. mization of defect clusters in crystalline materials, Comput. Mater. Sci. 98
[28] H. Okabe, M.J. Blunt, Pore space reconstruction using multiple-point statistics, (2015) 234e244.
J. Pet. Sci. Eng. 46 (1) (2005) 121e137. [58] H. Xu, R. Liu, A. Choudhary, W. Chen, A machine learning-based design rep-
[29] V. Sundararaghavan, N. Zabaras, Classification and reconstruction of three- resentation method for designing heterogeneous microstructures, J. Mech.
dimensional microstructures using support vector machines, Comput. Des. 137 (5) (2015), 051403.
Mater. Sci. 32 (2) (2005) 223e239. [59] S. Kirklin, J.E. Saal, V.I. Hegde, C. Wolverton, High-throughput computational
[30] A. Brahme, M. Alvi, D. Saylor, J. Fridy, A. Rollett, 3D reconstruction of micro- search for strengthening precipitates in alloys, Acta Mater. 102 (2016)
structure in a commercial purity aluminum, Scr. Mater. 55 (1) (2006) 75e80. 125e135.
[31] Y. Jiao, F. Stillinger, S. Torquato, Modeling heterogeneous materials via two- [60] R. Cang, Y. Xu, S. Chen, Y. Liu, Y. Jiao, M.Y. Ren, Microstructure representation
point correlation functions: basic principles, Phys. Rev. 76 (3) (2007), 031110. and reconstruction of heterogeneous materials via deep belief network for
[32] D. Fullwood, S. Kalidindi, S. Niezgoda, A. Fast, and N. Hampson, Gradient- computational material design, J. Mech. Des. 139 (7) (2017), 071404.
based microstructure reconstructions from distributions using fast Fourier [61] S. Torquato, G. Stell, Microstructure of two-phase random media. I. The n-
transforms, Mater. Sci. Eng. A 494 (1) (2008) 68e72. point probability functions, J. Chem. Phys. 77 (4) (1982) 2071e2077.
[33] D.T. Fullwood, S.R. Niezgoda, S.R. Kalidindi, Microstructure reconstructions [62] S. Torquato, G. Stell, Microstructure of two-phase random media. III. The n-
from 2-point statistics using phase-recovery algorithms, Acta Mater. 56 (5) point matrix probability functions for fully penetrable spheres, J. Chem. Phys.
(2008) 942e948. 79 (3) (1983) 1505e1510.
[34] Y. Jiao, F. Stillinger, S. Torquato, Modeling heterogeneous materials via two- [63] S. Torquato, G. Stell, Microstructure of two-phase random media. V. The n-
point correlation functions. II. Algorithmic details and applications, Phys. point matrix probability functions for impenetrable spheres, J. Chem. Phys. 82
Rev. 77 (3) (2008), 031135. (2) (1985) 980e987.
[35] Y. Jiao, F. Stillinger, S. Torquato, A superior descriptor of random textures and [64] S. Torquato, Random Heterogeneous Materials: Microstructure and Macro-
its predictive capacity, Proc. Natl. Acad. Sci. 106 (42) (2009) 17634e17639. scopic Properties, vol.16, Springer Science & Business Media, 2013.
[36] A. Hajizadeh, A. Safekordi, F.A. Farhadpour, A multiple-point statistics algo- [65] Y. Jiao, E. Padilla, N. Chawla, Modeling and predicting microstructure evolu-
rithm for 3D pore space reconstruction from 2D images, Adv. Water Resour. tion in lead/tin alloy via correlation functions and stochastic material
 P.-E. Chen et al. / Acta Materialia 179 (2019) 317e327 327

reconstruction, Acta Mater. 61 (9) (2013) 3370e3377. [77] S. Torquato, J. Beasley, Y. Chiew, Two-point cluster function for continuum
[66] E.-Y. Guo, N. Chawla, T. Jing, S. Torquato, Y. Jiao, Accurate modeling and percolation, J. Chem. Phys. 88 (10) (1988) 6540e6547.
reconstruction of three-dimensional percolating filamentary microstructures [78] E. Cinlar, S. Torquato, Exact determination of the two-point cluster function
from two-dimensional micrographs via dilation-erosion method, Mater. Char. for one-dimensional continuum percolation, J. Stat. Phys. 78 (3e4) (1995)
89 (2014) 33e42. 827e839.
[67] Y. Jiao, N. Chawla, Three dimensional modeling of complex heterogeneous [79] S. Prager, Interphase transfer in stationary two-phase media, Chem. Eng. Sci.
materials via statistical microstructural descriptors, Integrating. Mater. Manuf. 18 (4) (1963) 227e231.
Innovat. 3 (1) (2014) 1e19. [80] S. Torquato, Interfacial surface statistics arising in diffusion and flow problems
[68] Y. Jiao, N. Chawla, Modeling and characterizing anisotropic inclusion orien- in porous media, J. Chem. Phys. 85 (8) (1986) 4622e4628.
tation in heterogeneous material via directional cluster functions and sto- [81] S. Torquato, Y. Jiao, Dense packings of the platonic and archimedean solids,
chastic microstructure reconstruction, J. Appl. Phys. 115 (9) (2014), 093511. Nature 460 (7257) (2009) 876.
[69] S. Chen, H. Li, Y. Jiao, Dynamic reconstruction of heterogeneous materials and [82] J. Mering, D. Tchoubar, Interpretation de la diffusion centrale des rayons X par
microstructure evolution, Phys. Rev. 92 (2) (2015), 023301. les systemes poreux. I, J. Appl. Crystallogr. 1 (3) (1968) 153e165.
[70] S. Chen, A. Kirubanandham, N. Chawla, Y. Jiao, Stochastic multi-scale recon- [83] B. Lu, S. Torquato, Lineal-path function for random heterogeneous materials.
struction of 3D microstructure consisting of polycrystalline grains and II. Effect of polydispersivity, Phys. Rev. 45 (10) (1992) 7292.
second-phase particles from 2D micrographs, Metall. Mater. Trans. A 47 [84] C. Burger, W. Ruland, Analysis of chord-length distributions, Acta Crystallogr.
(2016) 1e11. A: Found. Crystallogr. 57 (5) (2001) 482e491.
[71] H. Li, S. Singh, N. Chawla, Y. Jiao, Direct Extraction of Spatial Correlation [85] N. Stribeck, Extraction of domain structure information from small-angle
Functions from Limited X-Ray Tomography Data for Microstructure Quanti- scattering patterns of bulk materials, J. Appl. Crystallogr. 34 (4) (2001)
fication. In Preparation, 2016. 496e503.
[72] H. Li, P.-E. Chen, Y. Jiao, Accurate reconstruction of porous materials via sto- [86] R. Cang, H. Li, H. Yao, Y. Jiao, Y. Ren, Improving direct physical properties
chastic fusion of limited bimodal microstructural data, Transp. Porous Media prediction of heterogeneous materials from imaging data via convolutional
(2017) 1e18. neural network and a morphology-aware generative model, Comput. Mater.
[73] C. Gommes, Y. Jiao, S. Torquato, Density of states for a specified correlation Sci. 150 (2018) 212e221.
function and the energy landscape, Phys. Rev. Lett. 108 (8) (2012), 080601. [87] S. Torquato, Exact expression for the effective elastic tensor of disordered
[74] C.J. Gommes, Y. Jiao, S. Torquato, Microstructural degeneracy associated with composites, Phys. Rev. Lett. 79 (4) (1997) 681.
a two-point correlation function and its information content, Phys. Rev. 85 (5) [88] S. Torquato, Effective stiffness tensor of composite mediadI. Exact series
(2012), 051140. expansions, J. Mech. Phys. Solids 45 (9) (1997) 1421e1448.
[75] Y. Jiao, F. Stillinger, S. Torquato, Geometrical ambiguity of pair statistics: point [89] D. Pham, S. Torquato, Strong-contrast expansions and approximations for the
configurations, Phys. Rev. 81 (1) (2010), 011105. effective conductivity of isotropic multiphase composites, J. Appl. Phys. 94
[76] Y. Jiao, F.H. Stillinger, S. Torquato, Geometrical ambiguity of pair statistics. II. (10) (2003) 6591e6602.
Heterogeneous media, Phys. Rev. 82 (1) (2010), 011106.