Acta Materialia 129 (2017) 428e438

Contents lists available at ScienceDirect

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

Full length article

Reduced-order structure-property linkages for polycrystalline

microstructures based on 2-point statistics
Noah H. Paulson a, Matthew W. Priddy b, David L. McDowell a, c, Surya R. Kalidindi a, d, *
a
George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332-0405, USA
b
Department of Mechanical Engineering, Mississippi State University, Mississippi State, MS 39762, USA
c
School of Materials Science and Engineering, Georgia Institute of Technology, Atlanta, GA 30332-0245, USA
d
School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA

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

Article history: Computationally efficient structure-property (S-P) linkages (i.e., reduced order models) are a necessary
Received 30 November 2016 key ingredient in accelerating the rate of development and deployment of structural materials. This need
Received in revised form represents a major challenge for polycrystalline materials, which exhibit rich heterogeneous micro-
26 February 2017
structure at multiple structure/length scales, and exhibit a wide range of properties. In this study, a novel
Accepted 6 March 2017
Available online 8 March 2017
framework is described for extracting S-P linkages in polycrystalline microstructures that are obtained
using 2-point spatial correlations (also called 2-point statistics) to quantify the material's microstructure,
and principal component analysis (PCA) to represent this information in a reduced dimensional space.
Keywords:
2-Point correlations
Additionally, it is demonstrated that the use of generalized spherical harmonics (GSH) as a Fourier basis
Microstructure for functions defined on the orientation space leads to a compact and computationally efficient repre-
Computational model sentation of the desired S-P linkages. In this study, these novel protocols are developed and demon-
Crystal plasticity strated for elastic stiffness and yield strength predictions for a
Ti microstructures using a dataset
Reduced-order model produced through microscale finite element simulations.
© 2017 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

1. Introduction for rapid down selection [1,6,7,16e18]. In this work, we specifically

focus on reduced order structure-property (S-P) linkages for poly-
Materials development and deployment efforts require evalua- crystalline microstructures that exhibit rich heterogeneity in the
tion of the properties of large sets of potential material internal spatial distribution of the crystal lattice orientation at the meso-
structures (generically referred to as microstructures; this scale (also referred to as grain-scale). These microstructures exhibit
description also includes information regarding chemical compo- strongly anisotropic elastic and inelastic properties at the macro-
sition at the relevant length scales) [1e13]. Currently, experiments scale [5,16,19e31].
and simulations are used for such analyses at a high cost, severely It is important to pursue the development of reduced order
limiting the extent of the materials design spaces explored [14,15]. models within the context of established physically-based theories.
In addition, most of the current approaches (e.g., finite element There is indeed a rich literature on homogenization theories for
models) employed in modeling and simulation are not readily polycrystalline microstructures that can be used to guide the
invertible, i.e., it isn't easy to determine a candidate microstructure development of surrogate models. For elastic stiffness or compli-
given a set of desired properties. Reduced order models are critical ance components, the foundational concepts can be traced back to
to the acceleration of materials discovery and development, as they the elementary bounds established by Voigt [32], Reuss [33], and
enable high throughput exploration of the materials design spaces Hill [34], as well as the more sophisticated self-consistent ap-
proaches of Kroner [35], which leverage Eshelby's solution [36]. For
inelastic properties, analogous foundations can be traced to the
works of Taylor [37] and Sachs [38]. Detailed discussion of these
* Corresponding author. George W. Woodruff School of Mechanical Engineering, theories cast in a modern continuum mechanics framework can be
Georgia Institute of Technology, Atlanta, GA 30332-0405, USA.
found in Nemat-Nasser and Mori [39], Milton [40], Mura [41], Qu
E-mail addresses: [email protected] (N.H. Paulson), [email protected].
edu (M.W. Priddy), [email protected] (D.L. McDowell), surya. and Cherkaoui [42] and Roters [43].
[email protected] (S.R. Kalidindi). The central limitation of the theories mentioned above is that

http://dx.doi.org/10.1016/j.actamat.2017.03.009
1359-6454/© 2017 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438 429

they can only incorporate a very limited amount of information state. However, this approach requires a very large number of
regarding the material microstructure. To date, microstructural orientation bins in order to obtain a representation of satisfactory
information included in the application of established composite fidelity. In this work, we will develop an extension to the existing
theories to polycrystalline microstructures has largely been limited MKS framework for homogenization S-P linkages by using gener-
to the orientation distribution (also called texture) in the sample, alized spherical harmonics (GSH) [80] for the functional compo-
and in some cases, averaged shape factors (based on idealizing the nents over the orientation space. We will specifically demonstrate
grains as ellipsoids) [31,44]. The main exception to these state- that this approach produces computationally efficient, highly
ments comes in the form of statistical continuum theories formu- compact representations of reduced-order S-P linkages for poly-
lated originally by Brown [45] for electrical properties of materials, crystalline microstructures. Note that GSH have already been used
and subsequently introduced by Volkov and Klinskikh [46], Loma- successfully in related problems [19,29,57,81,82], including certain
kin [47], and Beran and Molyneux [48] for mechanical properties. MKS localization linkages [19,66,69]. The efficiency and efficacy of
These ideas have been further refined in later years by Beran the new approach developed and presented in this work is spe-
[49,50], Kro € ner [51e53], Torquato [54e56], Adams [16,57,58], and cifically demonstrated through the predictions of the elastic stiff-
others [59e63]. A unique feature of this formulation is that the ness and yield strengths in a variety of a-titanium microstructures.
effective property is expressed as a series expansion, where each It is emphasized that S-P linkages can only attempt to match the
term represents a contribution of a specific microstructural attri- physics of whatever process was used to produce the set of cali-
bute expressed in the formalism of n-point statistics [57,58,64,65]. bration data. Consequently, the linkages in this work are calibrated
In some cases, the pre-factor for each term in the series can be using the results of CPFE simulations on synthetic microstructures,
evaluated using known Green's functions [16,59]. Most impor- and the success of this approach is evaluated by comparing against
tantly, this approach allows one to include as much detail of the results of that same model.
material microstructure as one wishes in arriving at the homoge-
nized properties of interest.
The sophisticated higher-order homogenization theories 2. Current theoretical framework
mentioned above have encountered some difficulties in their
implementation. First, the Green's functions required for the In order to construct S-P linkages, microstructural information
implementation of these theories are only available for a limited must be presented in a form that allows for a rigorous quantitative
number of physical phenomena. Second, the convergence of the comparison. Unfortunately, images of microstructure cannot fulfill
series is rather slow for high contrast composites [54,64]. Over the this role in their raw representations. In other words, a pixel by
past decade, Kalidindi and coworkers [1,16,19,66e76] have devel- pixel comparison of two random microstructures, either from the
oped a novel data science approach that still utilizes the basic same sample or from different samples, will not reveal useful in-
mathematical formulation of the statistical continuum theories, formation. Traditionally, metrics such as grain size distributions
and circumvents the central impediments mentioned above by and phase volume fractions have been used to capture the salient
employing modern data science tools. Mainly, it has been shown microstructural information [83,84]. However, these approaches do
that the elusive pre-factors (related to Green's functions) in the not consider the complexities of the microstructure geometry with
series expansion can be calibrated to results from numerical sim- sufficient fidelity for many properties of interest. In recent years,
ulations (e.g., performed using finite element models). This new Kalidindi and co-workers have employed 2-point spatial correla-
framework for establishing S-P linkages has been referred as ma- tions to evaluate microstructural distances [64], cluster micro-
terials knowledge systems (MKS) and has addressed both homog- structures [74,85], and to provide a basis for the construction of S-P
enization (bottom-up) [68,70,73,74] and localization (top-down) linkages [68,70,73]. This higher order statistical representation of
problems [19,66,67,69,71,72,75,76]. It is clarified here that the term microstructure precludes the need to manually select microstruc-
“MKS framework” in our prior work was used exclusively with the ture features (and potentially introduce bias) in the construction of
latter linkages. However, since both the homogenization and S-P linkages; the 2-point spatial correlations already include many
localization linkages developed in this new data science framework of the conventionally defined metrics of the microstructure.
share the same foundations (they both come from the same sta- Following the framework developed by Adams et al. [5,16,57],
tistical continuum theories discussed earlier), we will henceforth the microstructure in hierarchical materials systems can be math-
refer to both homogenization and localization linkages established ematically captured through the function mðh; xÞ, which reflects the
using this approach as MKS linkages. probability density of finding local state h (within an incremental
In the MKS framework, the homogenization S-P linkages Dh around h) at the spatial location x. In this formalism, h and x are
(reduced-order models) [1,68,70,73,74] connect the microstructure both treated as continuous variables. However, in practice it is
descriptors (inputs) to the effective macroscale properties (out- much more useful to employ a discretized description of the
puts). These linkages are typically expressed in the form of a microstructure function. Moreover, most experimental character-
polynomial series. The input microstructure descriptors are ization techniques recover only a discretized description of the
generally based on principal component analysis (PCA) [77] of the microstructure (as a consequence of the resolution limits of the
2-point spatial correlations obtained from all the microstructures equipment). Recognizing this, Adams et al. [5,16,57] have proposed
in the ensemble studied. The outputs have frequently included a discretized description of the microstructure function Msn based
elastic stiffness or compliance components and the yield strength. on primitive binning of both the spatial domain and the local state
Much of the prior work has been limited to materials with multiple space, i.e.,
distinct phases [1,68,70,73,74]. N X
X S
The main goal of this work is to extend the MKS homogenization mðh; xÞDhz Msn cn ðhÞcs ðxÞ (1)
framework to polycrystalline microstructures, where the local n¼1 s¼1
material state in each voxel is characterized by a crystal lattice
orientation (defined by an ordered set of three Euler angles). Prior where the indicator function ci ð Þ is defined such that it is equal to
attempts [78,79] in this direction have binned the local state space one if and only if the argument belongs to the bin labeled i and zero
(i.e., the orientation space) and treated the material as a composite; otherwise, and Dh denotes the size of the local state bin employed.
each bin in the orientation space corresponded to a distinct local An important consequence of this definition is that Msn has a very
430 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438

simple interpretation - it is essentially the volume fraction of local orthogonal and linear combinations of the original features such
states belonging to the bin n (in the local state space) present in the that they are ordered from highest to lowest contributors to the
bin s (in the spatial domain). It is important to recognize that both n variance in the dataset [77]. Let j ¼ 1; 2; …; J index the individual
and s are simply indices that point to a specific local state and elements (i.e., microstructure
 exemplars) in the dataset being
ðjÞ 
spatial bin (i.e., voxel), respectively. studied. Let fFt t ¼ 1; 2; ::; Rg denote a vectorized set of all spatial
Although Msn has a simple interpretation and provides a useful statistics of interest for a single microstructure exemplar. As an
description of the microstructure, it alone does not provide the example, this set might include all of the 2-point statistics
salient information on the topological features of the microstruc- computed from Eq. (4) for all combinations of (t, n, p). This set can
ture. For this, one has to compute the spatial correlations. The most also include higher-order spatial statistics (e.g., 3-point statistics), if
systematic approach to quantifying the spatial correlations in a deemed essential for the specific study. R denotes the total number
given microstructure is through the use of the n-point spatial cor- of features (i.e., microstructure measures) included in the analysis.
ðjÞ
relations framework [54,60,65,73,78,86e88]. The most basic spatial With this notation, PCA allows an orthogonal decomposition of Ft
correlations can be computed in the form of 2-point statistics as
formally expressed as
X
minððJ
1Þ;ðRÞÞ
Z ðjÞ
Ft ¼ aðjÞ 4it þ F t (5)
0 1 0 i
f ðh; h jrÞ ¼ mðh; xÞmðh ; x þ rÞdx (2) i¼1
VolðUr Þ
Ur ðjÞ
where 4it are the PC basis vectors, ai are the transformed co-
In Eq. (2), f ðh; h0 jrÞ
denotes the 2-point statistics reflecting the ordinates of the exemplar in the space spanned by the PC basis
ðjÞ
conditional probability density associated with finding the local vectors, and F t is ensemble average of Ft . Note that the maximum
state h at the spatial location x and the local state h0 at the spatial number of principal components is limited by the smaller value of R
location x þ r. As before, f ðh; h0 jrÞ½Dh2 denotes the corresponding and ðJ
1Þ. However, in most practical applications, only a handful
ðjÞ
conditional probability. Note that the volume examined in evalu- of ai (called PC scores) are employed in the reduced-order rep-
ating the integral (denoted Ur ) is expected to depend on the choice resentation of the  original dataset. This representation may be
ðjÞ  ~ where the truncation level R ~ is
of r (this is because of the need to know the local states at both written as fai i ¼ 1; 2; …; Rg,
spatial locations x and x þ r). For the same reasons that were objectively decided based on the variance explained from the
mentioned earlier, a much more practical form of the 2-point original dataset. The low-dimensional PC scores of the micro-
spatial correlations is a discretized version expressed as structures have been successfully utilized in building searchable
microstructure databases [85] as well as in the construction of
N X
X N X robust and reliable reduced-order S-P linkages [68,70].
f ðh; h0 jrÞ½Dh2 z Ft cn ðhÞcp ðh0 Þct ðrÞ
np
(3) The introduction of PC representations of the n-point statistics
p¼1 n¼1 t
in the series expressions developed in statistical continuum the-
Once again, the discretized version Ftnp has a simple physical ories for effective properties produces a surprisingly simple
interpretation - it captures the conditional probability of finding expression [1,68], i.e.,
ordered local states belonging to bins n and p (in the local state ~
space) in spatial bins whose centers are separated by the set of X
R
Peff zA0 þ Ai a i (6)
vectors indexed by t. Further details and a rigorous treatment of
i¼1
this framework can be found in Ref. [57]. The discretized forms of
the microstructure function and the 2-point spatial correlations are where Ai capture the underlying physics of the problem, Peff is the
indeed related by the following simple relation: ~ is the truncation level selected
effective property of interest, and R
for the PC representations. Indeed, Eq. (6) expresses a linear rela-
np 1 X St
p tionship between Peff and the scores of a chosen set of PC basis
Ft ¼ Mn M (4)
jS t j s¼1 s sþt vectors (which themselves are extremely rich sets of microstruc-
tural features). Although the above equation is easily derived for
Specific attention needs to be afforded to several details of Eq. elastic response of composite material systems, it can be extended
(4) before one endeavors to implement it in a computational al- to nonlinear effective properties with suitable modifications
gorithm: (i) the mapping of the vector r to the index t needs to [1,40,59,60,68,70,73,74]. In the MKS framework, the Ai coefficients
follow specific protocols such as those outlined in Ref. [57], (ii) the are established through regression using datasets generated by
earlier discussion regarding Ur translates to S t to identify the set of numerical simulation tools (e.g., finite element models) on selected
valid trials involved in evaluating Ftnp , which depends on the vector microstructure exemplars.
index t [89], (iii) it is often convenient to employ multidimensional
index arrays for s and t to naturally index the spatial bins in 2-D and
3-D microstructures [1], and (iv) jS t j denotes the total number of 3. Extensions to the theoretical framework
spatial bins involved when s is used as a multidimensional array
index. Local states for the general class of polycrystalline metals may
np
For a complete representation of f ðh; h0 jrÞ from equation (3), Ft include descriptors such as crystal symmetry, dislocation density,
must be calculated for all t, n and p. Depending on the size of the and chemical composition (among others). However, for single-
microstructure and choice of basis functions, this set can be large phase annealed polycrystals, the main local state of interest at the
and unwieldy. For example, Ftnp contains close to one-million fea- grain-scale is the crystal lattice orientation, g. The corresponding
tures for a 3-dimensional microstructural sample with 21 elements local state space is simply the orientation space. Based on the
per edge and 10 local state bins. At this stage, data-science tech- concepts discussed thus far, a simple approach to address poly-
niques can be applied to reduce the dimensionality of the problem. crystalline microstructures is to bin the fundamental zone of
One approach is to use principal component analysis (PCA), which orientation space [57]. As a specific example, the fundamental zone
is a distance-preserving linear transformation. PCA finds of orientation space for hexagonal crystals is expressed as
 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438 431

components prior to performing the PCA. Both of these approaches

FZh ¼ fg ¼ ðf1 ; F;f2 Þ j 0  f1 <2p; 0  F  p=2; 0  f2  p=3g (7) would be completely equivalent, especially given that half of the
FtLK coefficients are complex conjugates of the other half in our
The main impediment to primitive binning of the orientation
application. In order to construct the S-P linkage, we would
space (as implied in Eq. (1)) comes essentially from the fact that a
continue to use the model form shown earlier in Eq. (6) and cali-
very large number of bins are needed to capture the orientation
brate the physically-based coefficients Ai with numerical simula-
information accurately in the microstructure function and the 2-
tion results obtained from finite element models for a selected
point spatial correlations. As an example, if one were to use one
number of microstructure exemplars.
degree bins for each of the three Bunge-Euler angles involved in the
definition of the orientations (cf. Eq. (7)), one would end up with
4. Case study
1,944,000 local state bins. This large number of bins makes the
computation, storage, and utilization of the 2-point spatial corre-
The purpose of this case study is to demonstrate the viability
lations highly unwieldy.
and potential of the data-driven protocols developed in the previ-
As an alternative, we explore here the utilization of generalized
ous sections for the creation of S-P linkages. In these approaches,
spherical harmonics [5,80,90,91] as an efficient Fourier basis for the
the physics capturing coefficients of Eq. (6) will be determined
computation of the 2-point spatial correlations in polycrystalline
through the following steps: (i) An ensemble of microstructure
microstructures. In this approach, we retain the primitive binning
exemplars will be generated which exhibit a rich and diverse set of
for the spatial variables and only change the basis for the orienta-
topological features (these might include grain size, shape, and
tion variables. With these changes, the microstructure function can
orientation distributions) covering a range of microstructures of
now be expressed as
interest for the intended application of the S-P linkages being
P P
S produced. (ii) Crystal plasticity finite element (CPFE) simulations
mn mn
mðg; xÞdgz Mls Tl ðgÞcs ðxÞdg are performed on each microstructure exemplar in the ensemble to
l;m;n s¼1 extract the effective (bulk) properties of interest. (iii) Each micro-
Z Z (8)
mn 2l þ 1 structure exemplar is quantified using 2-point spatial statistics
Mls ¼ mðg; xÞTlmn ðgÞcs ðxÞdxdg
jDxj with GSH basis functions. (iv) PCA is performed on the 2-pt sta-
FZh Ux tistics representations of the microstructure exemplars in the
ensemble to produce a low-dimensional representation of each
where Tlmn ðgÞ denote symmetrized GSH functions (in the present exemplar microstructure in the form of their PC scores. (v) Multi-
work, the symmetry of interest is the hexagonal-triclinic symmetry, variate linear regression is performed on the assembled dataset
where the first symmetry refers to crystal symmetry and the sec- (from steps (i)-(iv)) to calibrate the physics capturing coefficients of
ond one to the sample symmetry), * denotes a complex conjugate, Eq. (6). This calibrated model may then be validated by comparing
jDxj is the uniform size (volume) of a spatial bin, and Mls mn are the
the CPFE computed properties for new microstructure exemplars
GSH Fourier coefficients. As a special case, when there is a single against the predictions of the reduced-order model. Each of the
crystal of lattice orientation g0 in a spatial bin s, the corresponding steps described above will be detailed in the following sections.
GSH Fourier coefficients are simply given by
4.1. Generation of microstructure ensemble
mn
Mls ¼ ð2l þ 1ÞTlmn ðg0 Þ (9)
In this work synthetic microstructure volume elements (MVEs)
For simplicity of notation, we will map every distinct combi-
are generated both for establishing the S-P linkages of interest and
nation of ðl; m; nÞ to a single index L in all of the ensuing equations.
mn will be henceforth denoted for validating them. The MVEs must be sufficiently large to capture
As a result of this simplification, Mls
the rich microscale interactions between grains of different sizes,
simply as MsL . Extending the concept above, the description of the
shapes, and orientations. In prior work, it was shown that these
2-point spatial correlations of orientations in the polycrystalline
microscale interactions are most efficiently captured through in-
microstructure is expressed as
fluence kernels [19,66,67,69,71,72,75,76]. More specifically, the
XXX decay in these Green's function-based influence kernels provides
f ðg; g 0 jrÞdgdg 0 z FtLK TL ðgÞTK ðg0 Þct ðrÞdgdg 0 (10)
guidance for the minimum size of the MVEs. It is generally noted
K L t
that the size of the MVEs should be larger for material systems with
However, if one were to start with the formal definition of the 2- higher contrast in the local properties of the microscale constitu-
point spatial correlation (Eq. (2)) and substitute the GSH repre- ents. Based on our prior experience [19,66,69], we employ MVEs of
sentations of the microstructure function from Eq. (8), we can size 21  21  21 voxels in this work, where each voxel has a side
derive the following relationship between the GSH Fourier co- length of 10mm.
efficients in Eqs. (8) and (10): The MVEs used in this work are generated using the open-
source DREAM.3D package [92], which accepts as inputs micro-
1 X St
structure measures such as grain size, grain shape, orientation and
FtLK ¼ ML MK (11)
jS t j s¼1 s sþt misorientation distributions. In this first study on polycrystalline
microstructures, we will focus our attention on the distribution of
The implicit convolution in Eq. (11) (same as in Eq. (4)) allows grain orientations (this is the main reason for using the GSH rep-
the exploitation of the FFT algorithms in computing the 2-point resentations described earlier). Consequently, we restrict our
spatial correlations [64,85e87]. attention to MVEs with different textures, but exhibiting a common
As mentioned earlier, the set of 2-point statistics computed grain size distribution. In part, this choice is also made due to lack of
using Eq. (11) will be large and unwieldy. As before, we will seek a a significant collection of experimental characterizations of poly-
lower dimensional representation of this set using PCA. In this case, crystalline titanium materials accessible to us. In future work, it
however, the Fourier coefficients computed using Eq. (11) are would be desirable to expand the range of the generated MVEs
complex. One can either perform the PCA directly on these complex using more sophisticated approaches [93].
representations, or reshape them into separate real and imaginary The targeted grain size distribution for the synthetically
432 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438

generated MVEs follows a log-normal distribution with an average the number of discrete orientation bins needed to attain compa-
grain size of 30mm and a 15mm standard deviation. Each MVE is rable accuracy in the desired S-P linkages. In prior work (cf. [19,69]),
comprised of an average of 215 distinct grains. Seven unique tex- this economy in the representation of S-P linkages was largely
tures are selected as targets for the synthetic microstructure gen- attributed to the fact that the physics-capturing kernels and co-
eration. These textures are inspired by those seen in our prior work efficients (based on Green's functions) operating on the micro-
as well as in literature [94,95]. The target textures are entered into structure descriptors (e.g., n-point spatial correlations) could be
DREAM.3D as lists of orientations expressed as triplets of Euler captured in a small number of terms in GSH-based expansions. We
angles with associated weights and spreads. DREAM.3D is then anticipate similar economy in the present application. For example,
utilized to randomly generate thirty polycrystalline MVEs for each it can be shown theoretically that only l ¼ ð0; 2; 4Þ terms in the GSH
microstructure class through an ellipsoidal packing algorithm. expansions should appear in homogenization linkages for effective
Then, DREAM.3D matches the target texture information for each elastic properties [5,16,25,108]. However, we do not know a priori
microstructure class. Fig. 1 shows example MVEs and (0002) pole the acceptable truncation level on L for capturing the S-P linkages
figures for each microstructure class (corresponding to each for the effective inelastic properties. Therefore, we treat the trun-
selected texture). This figure only enables a qualitative comparison cation level in the GSH coefficients as a hyper-parameter in the
of microstructures. However, in later sections, the 2-pt spatial establishment of the S-P linkages (we will return to this parameter
correlations and PCA will be used to quantify and visualize the subsequently). It should be noted that the truncation on L should be
differences between these microstructures. carried out such a way that the complete set of terms corre-
sponding to any selected value of l are included as a group. In other
4.2. CPFE simulations words, if one truncates at l ¼ 4, then it is important to include all
GSH terms corresponding to all combinations of m and n for l  4
In this work, CPFE simulations are used to extract the homog- [1,16]. For the hexagonal-triclinic GSH basis employed in this work,
enized properties for each MVE. The combination of each MVE and adoption of this consideration results in acceptable truncation
its associated properties constitutes one data point for the desired levels of L at values of 6, 15, 41, and 90 (these correspond to l values
S-P linkages. The crystal plasticity framework used here was truncated at 2, 4, 6 and 8, respectively).
developed in prior work for a-phase grains in Ti6-4 [66,96e102]. Once a truncation level has been selected for L, the Fourier co-
The model parameters relevant to the predictions of the effective efficients of the 2-point statistics may be computed using Eq. (11).
(bulk) elastic stiffness and yield strength are summarized in Table 1. Let us denote this truncation value as ~L. In employing Eq. (11) one
These include the five independent components of the transversely can take advantage of a few strategies known to dramatically
isotropic elastic stiffness tensor and the critical resolved shear reduce the computational cost involved. First, the number of 2-
2
strengths (CRSS) of relevant slip families. These elastic and inelastic point spatial correlations is ~L (this includes all possible pairings
parameters were obtained in prior work through fitting to experi- of indices L and K in Eq. (11)). In prior work on the multiphase
mental results on Ti6-4 [94,98,103]. As mentioned in the intro- microstructures [87], analytic relationships were derived between
duction, the S-P linkages resulting from this work learn the physics the spatial correlations. Appendix A presents these interrelation-
of the model used to generate the calibration data. So long as the ships for the 2-point statistics computed here on polycrystalline
results of the S-P linkage and CPFE simulations match, the goals of microstructures, where it is shown that the number of independent
this work are achieved. Therefore, beyond the calibration of the correlations is 2~ L
1. It is important to recognize that these de-
crystal plasticity model, no further attempt is made to experi- pendencies are not all linear. Furthermore, since PCA automatically
mentally validate the simulated properties; these homogenized removes the linearly dependent correlations, we do not need to
properties are taken as ground truth in the remainder of this work. worry about the linearly dependent correlations. By repeated trials,
Periodic boundary conditions are applied to all surfaces of the the optimal set of spatial correlations to represent each MVE was
three-dimensional MVEs to simulate the loading conditions needed selected (further details are provided in Appendix A). It was
to extract the effective properties of interest for this study [104]. determined through these trials that the specific set of 2~ L
1 cor-
Displacement-controlled loading is imposed on the X-faces of the relations identified in Appendix A produces the best compromise
cuboidal MVEs, with zero net traction lateral faces to mimic the between accuracy and computational cost for the S-P linkages
uniaxial stress state commonly observed in uniaxial tensile tests produced in this work.
[97]. The loading is performed to 1.5% strain, a sufficient level to
capture the elastic-plastic transition at the macroscale. The ABA- 4.4. Reduced dimensional microstructure representation
QUS [105] mesh used in this work contains 9261 C3D8R elements
(one per voxel in each MVE); these element types have been used in Once spatial statistics are computed for all MVEs in the selected
previous work [94,106,107] with good agreement as compared to ensemble, PCA is performed to obtain reduced dimensional rep-
C3D8 elements for the determination of bulk properties. The sim- resentations of the microstructure statistics. In this study, there are
ulations produce engineering stress-strain curves; the elastic far more features (coefficients in the Fourier representation of the
stiffness of each digital microstructure is measured from the initial spatial statistics) than samples (MVEs). Therefore, the maximum
elastic portion of the engineering stress-strain curve and the yield number of PC scores that can be computed is 209 (one less than the
strength is determined using a 0.2% offset with the calculated number of MVEs used to build the PC space; see Eq. (5)) [74].
elastic stiffness. However, a much smaller number of PCs are required to satisfac-
torily represent the spatial statistics for this case study. Fig. 2 shows
4.3. Computation of 2-point spatial correlations the cumulative variance captured by the inclusion of additional PCs.
~ increases, the cumulative explained variance (representing the
As R
Equation (10) allows for the computationally efficient calcula- 2-point statistics information) asymptotically approaches 100%. In
tion of the 2-point spatial correlations in polycrystalline micro- addition, as the number of GSH basis functions used in the Fourier
structures through the use of GSH representations for the representation of the 2-point statistics increases (improving the
functional dependencies on the orientation space. This is mainly fidelity of the microstructure representation), so does the R ~
because the number of GSH coefficients (determined by the trun- required to achieve a given explained variance.
cation level used on L) required is expected to be far smaller than PCA also serves as an effective method to cluster MVEs and to
 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438 433

Fig. 1. Sample MVEs (displayed using Grain ID maps) and (0002) pole figures for the generated microstructures classes A through G.

Table 1
Model parameters used in CPFE simulations.

Parameter Value Parameter Value

c11 172.8 GPa tbasal

CRSS
350 MPa
c12 97.91 GPa tprism 275 MPa
CRSS
c13 73.43 GPa tpyr〈a〉
CRSS
470 MPa
c33 192.3 GPa tpyr〈aþc〉
CRSS
570 MPa
c44 49.70 GPa

evaluate the distances between them [68,70,73,74,85]. It is impor-

tant to note that PCA results in unsupervised classification as the
MVEs are not identified or labeled in any manner. Fig. 3a demon-
strates this natural clustering of the MVEs in the first two PC di-
mensions (the dimensions with the largest and the second-largest
variances) for GSH spatial statistics at a truncation level of ~
L ¼ 15.
Remarkably, even in just two dimensions, the MVEs distinctively ~ and ~L.
Fig. 2. Percentage PCA explained structural variance versus R
cluster by their associated microstructure class (here differentiated
exclusively by target texture). In other words, even though the PCA
is not informed explicitly about the textures in the MVEs, the first microstructure classes with strong basal texture components are
two PC scores effectively cluster the MVEs based on their differ- located towards the left of Fig. 3a, while microstructure classes with
ences in texture (note that the texture information is included in strong transverse texture components are located towards the right
the 2-point statistics). Furthermore, it is seen that the intraclass of this figure. Although such observations provide intuition into the
variance in each cluster is not the same. For example, the MVEs meanings of the different PCs, it should be noted that the complete
belonging to class A exhibit more differences among themselves information captured by each PC is actually stored in the vectors,
compared to the MVEs of class G in the first two PCs. This is, of 4it , obtained in PCA (see Eq. (5)). Given the large size of these
course, a consequence of how each MVE is randomly generated for vectors (for a truncation level of ~L ¼ 15 each PC has a length of
target microstructure statistics in the DREAM.3D software. 268,569 in the present study), it is not straightforward to establish
Comparing Figs. 1 and 3a, it is also seen that microstructures with their precise physical meaning.
similar textures are nearby in the PC space, while microstructures Fig. 3b shows the relative distances between the centers of the
with dramatically different textures are well separated. Notice that microstructure clusters in PC space, presented by way of a
434 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438

Fig. 3. Visualizations of microstructures enabled by PCA (for ~L ¼ 15). (a) Clustering of MVEs by microstructure type in the first two PCs (b) Dendrogram showing the distances
between microstructures and clusters of microstructures in PC space.

dendrogram. The advantage of this representation is that the dis- representations are highly consistent with prior related work
played distances are computed over all available dimensions (in the [29,69,80]. Considering the effect of the value of R, ~ the most dra-
present work this is 209). From the dendrogram, it can be seen that matic decrease in error occurs within the inclusion of the first
the first two PC dimensions adequately capture the relative dis- 10 PCs. This is consistent with the observation in Fig. 2 that the
tances between the microstructure classes. This is indeed the po- explained structural variance quickly approaches 100% with
wer of PC representations; they provide high value low increasing numbers of PCs.
dimensional representations of practical worth in quantifying the Fig. 4a and b illustrate the potential for overfitting the S-P
microstructure statistics. As such, this is an important component linkages. The calibration error is usually not a good indicator for
in the effort to extract robust and reliable S-P linkages. overfitting as it always decreases for increasing values of hyper-
parameters such as R. ~ In prior work, the Leave-One-Out Cross
Validation (LOOCV) error metric was employed as an indicator of
4.5. Extraction of structure-property linkages
overfitting [68,70]. In this approach, the S-P linkage is calibrated
using the data for all MVEs in the ensemble except for one, and the
The low dimensional representation of each MVE and its asso-
error in the prediction of the property for the excluded MVE is
ciated properties established using the CPFE model described
recorded. The entire procedure is repeated for each MVE in the
earlier constitute the calibration dataset for extracting the desired
ensemble. The LOOCV error is computed as the mean of the pre-
S-P linkages. More specifically, we intend to calibrate the physics-
diction errors for each excluded MVE in the ensemble and
capturing coefficients of Eq. (6) using a simple multivariate linear
normalized the same way as the calibration error. The normalized
regression. Prior to calibration, two critical hyper-parameters, ~ L and
~ must be selected. These hyper-parameters cannot be arbitrarily LOOCV error is plotted for elastic stiffness and yield strength link-
R
ages in Fig. 4c and d, respectively. For both properties, the LOOCV
chosen; they must be selected individually for each property of
~ The LOOCV error, however,
error drops quickly for small values of R.
interest by systematically exploring a large number of potential
begins to increase at higher values of R ~ indicating overfitting. It is
values (for both parameters) and evaluating their efficacy in pro-
also observed that the minimum LOOCV error occurs at a higher
ducing the most reliable S-P linkages.
value of R~ for the higher GSH truncation levels. This is once again
In order to accomplish this step, it is necessary to select the
consistent with the observation in Fig. 2 that a higher R ~ is required
metrics by which we objectively evaluate each potential S-P linkage
to achieve the same explained variance for at higher values of ~ L.
produced in this work. One obvious metric, called the calibration
The optimal selection of ~ L and R~ depends on the desired char-
error, is simply defined as the mean of the absolute error between
the MKS predicted and the CPFEM simulated values of the property acteristics of the S-P linkages to be produced. If the minimization of
error is paramount, the linkage with the lowest LOOCV error should
for all MVEs, normalized by the mean value of the simulated
property. Fig. 4 shows the calibration error for R ~ up to 60 and four be chosen. Such a linkage is expected to perform the best for new
MVEs. However, if compactness is desired, then the linkage which
different ~ L levels for elastic stiffness (Fig. 4a) and yield strength
(Fig. 4b). It is clear from these plots that the gains in accuracy has the minimum ~ L and R~ yet has a LOOCV error below some
~ and ~L. This acceptable threshold should be selected. In this study, reduced-
quickly diminish with increasing values of both R
observation attests to the power of the PC and GSH representations order (low-dimensional) linkages are prioritized to demonstrate
their applicability for inverse materials design problems. Through
employed in this work. Note also that the average error values are
the consideration of Fig. 4c a linkage for elastic stiffness is selected
remarkably low - only about 0.2% for the elastic stiffness prediction
with 3 PCs and 15 GSH basis functions. From Fig. 4d a linkage for
and about 1.5% for the yield strength prediction. The fact that the
yield strength is selected with 6 PCs and 41 GSH basis functions.
prediction error for the yield strength is higher than that for the
This means that the S-P linkages produced in this study for elastic
elastic stiffness is quite reasonable and expected. Indeed, prior
stiffness and yield strength only have 4 and 7 coefficients, respec-
work has shown that the elastic response of polycrystalline mi-
tively, in Eq. (6).
crostructures can be represented accurately by a limited set of GSH
basis functions [29,69]. In the present work, we observe no sig-
nificant improvement in the accuracy of the elastic S-P linkage 4.6. Validation of structure-property linkages
beyond ~ L ¼ 15 (corresponding to l  4). For the predictions of the
yield strength, higher values of ~ L did provide a modest improve- In the previous sections, a novel data science framework is
ment in the accuracy. These observations of the efficacy of the GSH developed and implemented for the extraction of highly accurate
 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438 435

Fig. 4. Error in the property prediction versus number of PCs included in the S-P linkage for different selections of the number of GSH basis functions. (a) Error in the prediction of
elastic stiffness for MVEs in calibration set. (b) Error in prediction of yield strength for MVEs in calibration set. (c) LOOCV error in prediction of elastic stiffness for MVEs in calibration
set. (d) LOOCV error in prediction of yield strength for MVEs in calibration set.

and robust S-P linkages (note the very low error values in Fig. 4). predict the properties of completely new microstructures. As
We now seek to further validate these linkages using completely before, thirty MVEs are generated for each of the five new micro-
new MVEs (those that have not been used in the calibration). For structure classes. Furthermore, thirty additional MVEs are gener-
this purpose, a new set of MVEs are generated and their properties ated for each of the existing microstructure classes. CPFEM
are simulated using the same protocols described earlier in this simulations are performed for all MVEs in the validation ensemble
paper. However, in an effort to make these MVEs different from the to extract their yield strength and elastic stiffness.
calibration MVEs, five additional target textures are employed. For each MVE in the validation ensemble, the 2-point spatial
Example MVEs and (0002) pole figures for each of the new correlations are computed up to the GSH truncation levels selected
microstructure classes are presented in Fig. 5a. Microstructure class for each property in Sec. 4.5. The spatial statistics for each new MVE
H is inspired from literature [95], while classes I through L are are then transformed into the reduced dimensionality PC space
created using different combinations of texture components pre- already established in Sec. 4.4. This validation ensemble is dis-
sent in classes A through H. The new microstructure classes are played in Fig. 5b in the PC1-PC2 space. For better visualization, the
intended to be interpolations between existing classes. The S-P MVEs from microstructure classes H through L are highlighted in
linkages may give good results for small extrapolations beyond the color in this figure, while MVEs from the original microstructure
bounds of the calibration MVE ensemble, but cannot be expected to classes are shown in gray. The properties of each MVE in the

Fig. 5. (a) Sample MVEs (with Grain ID displayed) and (0002) pole figures for microstructure classes H through L. (b) The ensemble of validation MVEs in PC1 and PC2 (for ~L ¼ 15).
MVEs from previously existing microstructure classes are colored gray.
436 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438

Fig. 6. Predicted versus simulated response for (a) elastic stiffness with 3 PCs and (b) yield strength with 6 PCs. Grey markers are for validation MVEs generated with target
microstructure statistics from the original seven microstructure classes.

validation ensemble are then predicted using the calibrated S-P development of this work.
linkages obtained in Sec. 4.5. Fig. 6 presents parity plots for each S-P
linkage, where the predicted property is plotted versus the simu- Appendix
lated property (via CPFE simulations) for each MVE in the validation
ensemble. The maximum prediction errors for elastic stiffness and In previous work, Niezgoda et al. [87] identified redundancies in
yield strength in the calibration and validation ensembles are the full set of 2-point spatial statistics. Knowledge of these re-
highly consistent with each other, and are below 1% and 5%, dundancies is crucial for identifying an efficient yet comprehensive
respectively. The success of both linkages in accurately predicting set of spatial statistics for the representation of microstructure.
properties for these new microstructure classes demonstrates the Specifically, it was shown that the discrete Fourier transforms
np
(DFTs) of the 2-point statistics, F~j ¼ F ðFt Þ, exhibit the following
np
robustness of the S-P linkage produced in this work.
interrelationships:
5. Conclusions
pn pq
np pq pp nq np pn nq F~j F~ j
In this work, novel data science protocols for the construction of F~j F~ j ¼ F~ j F~j ; F~j ¼ F~j ; F~j ¼ pp (12)
F~
j
reduced-order structure-property linkages in polycrystalline ma-
terials were developed and validated for the low crystal symmetry
where * denotes the complex conjugate. It is convenient to visualize
a-titanium materials system. This is the first use of GSH basis np np
F~j in the array ½F~j NN , where N is the number of discrete local
functions in the MKS framework for the computation of spatial
states (in conventional binning of the local state space using indi-
statistics and for the construction of homogenization linkages. np
cator functions). Eq. (12) implies that if F~j is known for a row or
Protocols to predict both elastic stiffness and yield strength were
column of the correlation array, then the remainder of the array
developed, each in the form of polynomial equations with a small
may be calculated. Furthermore, due to the properties of the indi-
number of coefficients; it was demonstrated that only four and
cator basis in the representation of the local state, only N
1 of the
seven coefficients were required to predict elastic stiffness and
correlations in a given row or column are independent. Therefore
yield strength respectively. These linkages provided massive
there are only N
1 independent correlations in the set of the 2
computational savings versus traditional protocols. CPFE simula-
point spatial correlations for an N-phase microstructure.
tions for the prediction of yield strength in a single MVE took two
When the GSH basis is selected to describe the functional
hours on four processors on a super-computing cluster, while the
dependence of the grain orientation in the microstructure function,
reduced-order S-P linkage required only three seconds with one
FtKL (see Eq. (11)) does not have a simple interpretation as a prob-
processor. This represents nearly four orders of magnitude reduc-
ability. As a result, the conclusions of Eq. (12) need to be suitably
tion in processing time. The simplicity and efficiency of these
modified. Recognizing that T1 ðgÞ ¼ 1 it can be shown that
robust structure-property linkages are well suited for inverse
design protocols and will be instrumental in future materials Q1 1Q 11 Q1 1Q
design efforts. F~j¼0 ¼ F~j¼0 ¼ CQ ; F~ j¼0 ¼ jSj; F~js0 ¼ F~ js0 ¼ 0 (13)

Acknowledgments where CQ are related to the 1-point statistics of the microstructure

for single-phase polycrystals and contain the same information as
This work was supported by the National Science Foundation the ODF. At this point it is useful to introduce two different arrays of
under Grant No. CMMI-1333083. Any opinions, findings, and con- GSH-based 2-pt correlations:
clusions or recommendations expressed in this material are those h LK i h LK i
of the authors and do not necessarily reflect the views of the Na- A ¼ F~j ~ ; B ¼ F~j ~ with L ¼ 1 and K
L~
L L
1~
L
1
tional Science Foundation. NHP would like to thank Dipen Patel and
¼ 1 excluded (14)
David Brough for many illuminating conversations throughout the
 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438 437

Equations (12) and (13) imply that the independent set of cor- Integrated Computational Materials Engineering: a Transformational Disci-
pline for Improved Competitiveness and National Security, Technical Report,
relations consists of the first row or column of A in addition to any
on Integrated Computational Materials Engineering, National Research
row or column of B (for 2~ L
1 total correlations). Council (US). Committee, 2008.
To select the optimal set of correlations for this study, the errors [16] B.L. Adams, S.R. Kalidindi, D.T. Fullwood, Microstructure-sensitive Design for
associated with the S-P linkages calibrated using different sets of Performance Optimization, Butterworth-Heinemann, 2012.
[17] Z. Li, B. Wen, N. Zabaras, Computing mechanical response variability of
correlations are compared. These test linkages are developed for polycrystalline microstructures through dimensionality reduction tech-
both elastic stiffness and yield strength at a GSH truncation level of niques, Comput. Mater. Sci. 49 (2010) 568e581.
~
L ¼ 6. The following list describes a selection of correlation sets and [18] C. Suh, A. Rajagopalan, X. Li, K. Rajan, The application of Principal Component
Analysis to materials science data, Data Sci. J. 1 (2002) 19e26.
the quality of the resultant linkages. [19] Y.C. Yabansu, S.R. Kalidindi, Representation and calibration of elastic locali-
zation kernels for a broad class of cubic polycrystals, Acta Mater 94 (2015)
2
1. All correlations in A (~
L correlations): This results in the lowest 26e35.
[20] R.A. Lebensohn, A.K. Kanjarla, P. Eisenlohr, An elasto-viscoplastic formulation
error of all sets. based on fast Fourier transforms for the prediction of micromechanical fields
2
2. The lower triangular correlations of A (12 ð~ L þ~ LÞ correlations): in polycrystalline materials, Int. J. Plast. 32e33 (2012) 59e69.
The error response is similar to set 1. [21] F. Roters, P. Eisenlohr, L. Hantcherli, D.D. Tjahjanto, T.R. Bieler, D. Raabe,
Overview of constitutive laws, kinematics, homogenization and multiscale
3. The first columns of A and B and the diagonal elements of B methods in crystal plasticity finite-element modeling: theory, experiments,
(3ð~
L
1Þ correlations): This results in an error slightly higher applications, Acta Mater 58 (2010) 1152e1211.
than sets 1 and 2. [22] J.B. Shaffer, M. Knezevic, S.R. Kalidindi, Building texture evolution networks
4. The first column and diagonal elements of A (2~ L
1 correla- for deformation processing of polycrystalline fcc metals using spectral ap-
proaches: applications to process design for targeted performance, Int. J.
tions): This results in an error response similar to set 3. Plast. 26 (2010) 1183e1194.
5. The first columns of A and B (2~ L
1 correlations): This results in [23] H.F. Al-Harbi, M. Knezevic, S.R. Kalidindi, Spectral approaches for the fast
an error response worse than sets 1 and 2 but slightly better computation of yield surfaces and first-order plastic property closures for
polycrystalline materials with cubic-triclinic textures, Comput. Mater. Con-
than 3 and 4. tin. 15 (2010) 153e172.
6. The first column of A (~ L correlations): This results in higher er- [24] D. Banabic, F. Barlat, O. Cazacu, T. Kuwabara, Advances in anisotropy and
rors than in the preceding sets. Furthermore, error does not formability, Int. J. Mater. Form. 3 (2010) 165e189.
[25] T. Fast, M. Knezevic, S.R. Kalidindi, Application of microstructure sensitive
decrease beyond the inclusion of the first several PCs. This in- design to structural components produced from hexagonal polycrystalline
dicates that the ODF information is important, but not sufficient metals, Comput. Mater. Sci. 43 (2008) 374e383.
for high-quality predictive capability. [26] M. Knezevic, S.R. Kalidindi, R.K. Mishra, Delineation of first-order closures for
7. The first column of B (~L
1 correlations): This results in
plastic properties requiring explicit consideration of strain hardening and
crystallographic texture evolution, Int. J. Plast. 24 (2008) 327e342.
extremely high errors (the linkages had no predictive utility). It [27] S. Graff, W. Brocks, D. Steglich, Yielding of magnesium: from single crystal to
is hypothesized that the correlations in B do not contain any polycrystalline aggregates, Int. J. Plast. 23 (2007) 1957e1978.
[28] P. Van Houtte, A.K. Kanjarla, A. Van Bael, M. Seefeldt, L. Delannay, Multiscale
information from the ODF, which is of first-order importance. modelling of the plastic anisotropy and deformation texture of poly-
crystalline materials, Eur. J. Mech. A/Solids 25 (2006) 634e648.
As a result of this investigation the first columns of A and B are [29] G. Proust, S.R. Kalidindi, Procedures for construction of anisotropic elastic-
plastic property closures for face-centered cubic polycrystals using first-
selected as a good compromise between computational efficiency
order bounding relations, J. Mech. Phys. Solids 54 (2006) 1744e1762.
and S-P linkage quality. [30] S.R. Kalidindi, S.E. Schoenfeld, On the prediction of yield surfaces by the
crystal plasticity models for fcc polycrystals, Mater. Sci. Eng. A 293 (2000)
120e129.
References [31] R.A. Lebensohn, C.N. Tome , A self-consistent anisotropic approach for the
simulation of plastic deformation and texture development of polycrystals:
[1] S.R. Kalidindi, Hierarchical Materials Informatics, Butterworth-Heinemann, application to zirconium alloys, Acta Metall. Mater 41 (1993) 2611e2624.
2015. [32] W. Voigt, Lehrbuch der Kristallphisic, 1928.
[2] H. Xu, Y. Li, C. Brinson, W. Chen, A descriptor-based design methodology for [33] A. Reuss, Berechnung der Fließgrenze von Mischkristallen auf Grund der
developing heterogeneous microstructural materials system, J. Mech. Des. Plastizit€atsbedingung für Einkristalle, ZAMM Appl. Math. Mech./Z. Angew.
136 (2014) 206e216. Math. Mech 9 (1929) 49e58.
[3] J.H. Panchal, S.R. Kalidindi, D.L. McDowell, Key computational modeling is- [34] R. Hill, The elastic behaviour of a crystalline aggregate, Proc. Phys. Soc. Sect. A
sues in integrated computational materials engineering, Comput. Des. 45 65 (1952) 349e354.
(2013) 4e25. [35] E. Kro€ner, Berechnung der elastischen Konstanten des Vielkristalls aus den
[4] D. L. McDowell, D. Backman, Simulation-assisted Design and Accelerated Konstanten des Einkristalls, Z. Phys. 151 (1958) 504e518.
Insertion of Materials, Springer US, Boston, MA, pp. 617e647. [36] J.D. Eshelby, The determination of the elastic field of an ellipsoidal inclusion,
[5] D.T. Fullwood, S.R. Niezgoda, B.L. Adams, S.R. Kalidindi, Microstructure sen- and related problems, Proc. R. Soc. A Math. Phys. Eng. Sci. 241 (1957)
sitive design for performance optimization, Prog. Mater. Sci. 55 (2010) 376e396.
477e562. [37] G.I. Taylor, Plastic strains in metals, J. Inst. Met. 62 (1938) 307e325.
[6] D.L. McDowell, J.H. Panchal, H. Choi, C. Seepersad, J.K. Allen, F. Mistree, In- [38] G. Sachs, Zur Ableitung einer Fließbedingung, in: Mitteilungen der Dtsch.
tegrated Design of Multiscale, Multifunctional Materials and Products, But- Mater. Sonderh. IX Arb. aus dem Kaiser Wilhelm-Institut für Met. und
terworth-Heinemann, 2009. demStaatlichen Mater. zu Berlin-Dahlem, Springer Berlin Heidelberg, Berlin,
[7] V. Sundararaghavan, N. Zabaras, A statistical learning approach for the design Heidelberg, 1929, pp. 94e97.
of polycrystalline materials, Stat. Anal. Data Min. 1 (2009) 306. [39] S. Nemat-Nasser, M. Hori, Micromechanics: Overall Properties of Heteroge-
[8] C.C. Seepersad, B. Dempsey, J.K. Allen, F. Mistree, D.L. McDowell, Design of neous Materials, Elsevier Science Publishers, Amsterdam, 1993.
multifunctional honeycomb materials, AIAA J. 42 (2004) 1025e1033. [40] G.W. Milton, The Theory of Composites, vol. 6, Cambridge University Press,
[9] F. Mistree, C. C. Seepersad, B. M. Dempsey, D. L. McDowell, J. K. Allen, Robust Cambridge, 2002.
concept exploration methods in materials design, in: 9th AIAA/ISSMO Symp. [41] T. Mura, Micromechanics of Defects in Solids, Martinus Nijhoff Publishers,
Multidiscip. Anal. Optim., September, pp. 1e11. 1982.
[10] O. Sigmund, S. Torquato, Design of smart composite materials using topology [42] Q. Jianmin, M. Cherkaoui, Wiley: Fundamentals of Micromechanics of Solids,
optimization, Smart Mater. Struct. 8 (1999) 365e379. Wiley Hoboken, 2006.
[11] G.B. Olson, Systems design of hierarchically structured materials : advanced [43] F. Roters, P. Eisenlohr, T.R. Bieler, D. Raabe, Crystal plasticity Finite Element
steels, Steel Res. 4 (1998) 143e156. Methods: in Materials Science and Engineering, John Wiley & Sons, 2011.
[12] S. Ramakrishna, Microstructural design of composite materials for crash- [44] H. Wang, P.D. Wu, C.N. Tome , Y. Huang, A finite strain elastic-viscoplastic
worthy structural applications, Mater. Des. 18 (1997) 167e173. self-consistent model for polycrystalline materials, J. Mech. Phys. Solids 58
[13] V.K. Ganesh, S. Ramakrishna, S.H. Teoh, N.K. Naik, Microstructural design of (2010) 594e612.
textile composites, Mater. Des. 18 (1997) 175e181. [45] W.F. Brown, Solid mixture permittivities, J. Chem. Phys. 23 (1955) 1514.
[14] Technical Report, Materials Genome Initiative for Global Competitiveness, [46] S.D. Volkov, N.A. Klinskikh, Theory of the elastic properties of polycrystals,
National Science and Technology Council, 2011. Phys. Met. Metallogr. 19 (1965) 24.
[15] T.M. Pollock, J.E. Allison, D.G. Backman, M.C. Boyce, M. Gersh, E. Holm, [47] V.A. Lomakin, Deformation of microscopically nonhomogeneous elastic
R. LeSar, M. Long, A.C.I. Powell, J.J. Schirra, D.D. Whitis, C. Woodward, bodies, J. Appl. Math. Mech. 29 (1965) 1048e1054.
438 N.H. Paulson et al. / Acta Materialia 129 (2017) 428e438

[48] M. Beran, J. Molyneux, Use of classical variational principles to determine (2008) 942e948.
bounds for effective bulk modulus in heterogeneous media, Q. Appl. Math. [79] X. Gao, C.P. Przybyla, B.L. Adams, Methodology for recovering and analyzing
24 (1966) 107e118. two-point pair correlation functions in polycrystalline materials, Metall.
[49] M.J. Beran, T.A. Mason, B.L. Adams, T. Olsen, Bounding elastic constants of an Mater. Trans. A Phys. Metall. Mater. Sci. 37 (2006) 2379e2387.
orthotropic polycrystal using measurements of the microstructure, J. Mech. [80] H. Bunge, Texture Analysis in Materials Science, Butterworths, London,
Phys. Solids 44 (1996) 1543e1563. Boston, Sydney, Wellington, Durban, Toronto, 1982.
[50] M. Beran, Statistical Continuum Theories, vol 9, 1965. [81] M. Knezevic, S.R. Kalidindi, Fast computation of first-order elastic-plastic
[51] E. Kro€ner, Statistical modelling, in: J. Gittus, J. Zarka (Eds.), Model. Small closures for polycrystalline cubic-orthorhombic microstructures, Comput.
Deform. Polycrystals, Springer Netherlands, Dordrecht, 1986, pp. 229e291. Mater. Sci. 39 (2007) 643e648.
[52] E. Kro € ner, Bounds for effective elastic moduli of disordered materials, [82] B.L. Adams, A. Henrie, B. Henrie, M. Lyon, S.R. Kalidindi, H. Garmestani,
J. Mech. Phys. Solids 25 (1977) 137e155. Microstructure-sensitive design of a compliant beam, J. Mech. Phys. Solids 49
[53] E. Kro€ner, Statistical Continuum Mechanics vol. 53, Springer Vienna, 1971. (2001) 1639e1663.
[54] S. Torquato, Random Heterogeneous Materials: Microstructure and Macro- [83] E.O. Hall, The deformation and ageing of mild steel: III discussion of results,
scopic Properties, Interdisciplinary Applied Mathematics, Springer New Proc. Phys. Soc. Sect. B 64 (1951) 747.
York, 2002. [84] N.J. Petch, The cleavage strength of polycrystals, J. Iron Steel Inst. 174 (1953)
[55] S. Torquato, Random heterogeneous media: microstructure and improved 25e28.
bounds on effective properties, Appl. Mech. Rev. 44 (1991) 37. [85] S.R. Niezgoda, A.K. Kanjarla, S.R. Kalidindi, Novel microstructure quantifica-
[56] S. Torquato, Effective electrical conductivity of two-phase disordered com- tion framework for databasing, visualization, and analysis of microstructure
posite media, J. Appl. Phys. 58 (1985) 3790e3797. data, Integr. Mater. Manuf. Innov. 2 (2013) 3.
[57] B.L. Adams, X. Gao, S.R. Kalidindi, Finite approximations to the second-order [86] S.R. Niezgoda, D.M. Turner, D.T. Fullwood, S.R. Kalidindi, Optimized structure
properties closure in single phase polycrystals, Acta Mater 53 (2005) based representative volume element sets reflecting the ensemble-averaged
3563e3577. 2-point statistics, Acta Mater 58 (2010) 4432e4445.
[58] B.L. Adams, G.R. Canova, A. Molinari, A statistical formulation of viscoplastic [87] S.R. Niezgoda, D.T. Fullwood, S.R. Kalidindi, Delineation of the space of 2-
behavior in heterogeneous polycrystals, Textures Microstruct. 11 (1989) point correlations in a composite material system, Acta Mater 56 (2008)
57e71. 5285e5292.
[59] H. Garmestani, S. Lin, B.L. Adams, S. Ahzi, Statistical continuum theory for [88] S. Torquato, Statistical description of microstructures, Annu. Rev. Mater. Res.
large plastic deformation of polycrystalline materials, J. Mech. Phys. Solids 49 32 (2002) 77e111.
(2001) 589e607. [89] A. Çeçen, T. Fast, S.R. Kalidindi, Versatile algorithms for the computation of 2-
[60] H. Garmestani, S. Lin, B.L. Adams, Statistical continuum theory for inelastic point spatial correlations in quantifying material structure, Integr. Mater.
behavior of a two-phase medium, Int. J. Plast. 14 (1998) 719e731. Manuf. Innov. 5 (2016) 1e15.
[61] J.R. Willis, Variational and related methods for the overall properties of [90] H.J. Bunge, W.T. Roberts, Orientation distribution, elastic and plastic
composites, Adv. Appl. Mech. 21 (1981) 1e78. anisotropy in stabilized steel sheet, J. Appl. Crystallogr. 2 (1969) 116e128.
[62] J. McCoy, Macroscopic response of continua with random microstructures, [91] I.M. Gel’fand, R.A. Minlos, Z.Y. Shaprio, G. Cummins, T. Boddington,
Mech. Today 6 (1981) 1e40. H.K. Farahat, J.E. Mansfield, Representations of the rotation and Lorentz
[63] J. Korringa, Theory of elastic constants of heterogeneous media, J. Math. Phys. groups and their applications, Phys. Today 17 (1964) 48.
509 (2001) 509e513. [92] M.A. Groeber, M.A. Jackson, DREAM.3D: a digital representation environ-
[64] D.T. Fullwood, B.L. Adams, S.R. Kalidindi, A strong contrast homogenization ment for the analysis of microstructure in 3D, Integr. Mater. Manuf. Innov. 3
formulation for multi-phase anisotropic materials, J. Mech. Phys. Solids 56 (2014) 5.
(2008) 2287e2297. [93] D.M. Turner, S.R. Kalidindi, Statistical construction of 3-D microstructures
[65] S. Torquato, G. Stell, Microstructure of two-phase random media. I. The n- from 2-D exemplars collected on oblique sections, Acta Mater 102 (2016)
point probability functions, J. Chem. Phys. 77 (1982) 2071. 136e148.
[66] M.W. Priddy, N.H. Paulson, S.R. Kalidindi, D.L. McDowell, Strategies for rapid [94] B. Smith, Microstructure-sensitive Plasticity and Fatigue of Three Titanium
parametric assessment of microstructure-sensitive fatigue for HCP systems, Alloy Microstructures, Masters thesis, Georgia Institute of Technology, 2013.
Prep (2017). [95] G. Lütjering, Influence of processing on microstructure and mechanical
[67] D.B. Brough, D. Wheeler, J.A. Warren, S.R. Kalidindi, Microstructure-based properties of (aþb) titanium alloys, Mater. Sci. Eng. A 243 (1998) 32e45.
knowledge systems for capturing process-structure evolution linkages, Curr. [96] C.-h. Goh, J.M. Wallace, R.W. Neu, D.L. McDowell, Polycrystal plasticity
Opin. Solid State Mater. Sci. (2015), http://dx.doi.org/10.1016/ simulations of fretting fatigue, Int. J. Fatigue 23 (2001) 423e435.
j.cossms.2016.05.002. [97] J.R. Mayeur, D.L. McDowell, A three-dimensional crystal plasticity model for
[68] A. Gupta, A. Cecen, S. Goyal, A.K. Singh, S.R. Kalidindi, Structure-property duplex Ti-6Al-4V, Int. J. Plast. 23 (2007) 1457e1485.
linkages using a data science approach: application to a non-metallic in- [98] M. Zhang, J. Zhang, D.L. McDowell, Microstructure-based crystal plasticity
clusion/steel composite system, Acta Mater 91 (2015) 239e254. modeling of cyclic deformation of Ti-6Al-4V, Int. J. Plast. 23 (2007)
[69] Y.C. Yabansu, D.K. Patel, S.R. Kalidindi, Calibrated localization relationships 1328e1348.
for elastic response of polycrystalline aggregates, Acta Mater 81 (2014) [99] J.R. Mayeur, D.L. McDowell, R.W. Neu, Crystal plasticity simulations of fret-
151e160. ting of Ti-6Al-4V in partial slip regime considering effects of texture, Com-
[70] A. Çeçen, T. Fast, E. Kumbur, S. Kalidindi, A data-driven approach to estab- put. Mater. Sci. 41 (2008) 356e365.
lishing microstructure-property relationships in porous transport layers of [100] F. Bridier, P. Villechaise, J. Mendez, Slip and fatigue crack formation processes
polymer electrolyte fuel cells, J. Power Sources 245 (2014) 144e153. in an a/b titanium alloy in relation to crystallographic texture on different
[71] T. Fast, S.R. Kalidindi, Formulation and calibration of higher-order elastic scales, Acta Mater 56 (2008) 3951e3962.
localization relationships using the MKS approach, Acta Mater 59 (2011) [101] B.D. Smith, D. Shih, D.L. McDowell, Cyclic plasticity experiments and poly-
4595e4605. crystal plasticity modeling of three distinct Ti alloy microstructures, Int. J.
[72] T. Fast, S.R. Niezgoda, S.R. Kalidindi, A new framework for computationally Plast. (2013), http://dx.doi.org/10.1016/j.ijplas.2013.10.004.
efficient structure-structure evolution linkages to facilitate high-fidelity [102] M.W. Priddy, Exploration of Forward and Inverse Protocols for Property
scale bridging in multi-scale materials models, Acta Mater 59 (2011) Optimization of Ti-6Al-4V, Ph.D. thesis, Georgia Institute of Technology,
699e707. 2016.
[73] S.R. Kalidindi, S.R. Niezgoda, A.A. Salem, Microstructure informatics using [103] J.R. Mayeur, Three Dimensional Modeling of Ti-al Alloys with Application to
higher-order statistics and efficient data-mining protocols, Jom 63 (2011) Attachment Fatigue, Ph.D. thesis, Georgia Institute of Technology, 2004.
34e41. [104] C.P. Przybyla, Microstructure-sensitive Extreme Value Probabilities of Fa-
[74] S.R. Niezgoda, Y.C. Yabansu, S.R. Kalidindi, Understanding and visualizing tigue in Advanced Engineering Alloys, Ph.D. thesis, Georgia Institute of
microstructure and microstructure variance as a stochastic process, Acta Technology, 2010.
Mater 59 (2011) 6387e6400. [105] Hibbett, Karlsson, Sorensen, ABAQUS/standard: User's Manual, 1 edition,
[75] G. Landi, S.R. Niezgoda, S.R. Kalidindi, Multi-scale modeling of elastic Hibbitt, Karlsson & Sorensen, 1998.
response of three-dimensional voxel-based microstructure datasets using [106] C.P. Przybyla, D.L. McDowell, Microstructure-sensitive extreme value prob-
novel DFT-based knowledge systems, Acta Mater 58 (2010) 2716e2725. abilities for high cycle fatigue of Ni-base superalloy IN100, Int. J. Plast. 26
[76] S.R. Kalidindi, S.R. Niezgoda, G. Landi, S. Vachhani, T. Fast, A novel framework (2010) 372e394.
for building materials knowledge systems, Comput. Mater. Contin. 17 (2010) [107] P.C. Kern, Improvements to the Computational Pipeline in Crystal Plasticity
103e125. Estimates of High Cycle Fatigue of Microstructures, Master's thesis, Georgia
[77] K. Pearson, On lines and planes of closest fit to systems of points in space, Institute of Technology, 2016.
London, Edinburgh, Dublin Philos. Mag. J. Sci. 2 (1901) 559e572. [108] X. Wu, G. Proust, M. Knezevic, S.R. Kalidindi, Elastic-plastic property closures
[78] D.T. Fullwood, S.R. Niezgoda, S.R. Kalidindi, Microstructure reconstructions for hexagonal close-packed polycrystalline metals using first-order bound-
from 2-point statistics using phase-recovery algorithms, Acta Mater 56 ing theories, Acta Mater 55 (2007) 2729e2737.