Computational Materials Science 108 (2015) 128–135

Contents lists available at ScienceDirect

Computational Materials Science

journal homepage: www.elsevier.com/locate/commatsci

Mechanical modeling based on numerical homogenization of an Al2O3/Al

composite manufactured via binder jet printing
K. Myers a,⇑, M. Juhasz b, P. Cortes a,c, B. Conner b
a
Materials Science and Engineering, Youngstown State University, Youngstown, OH 44555, USA
b
Mechanical and Industrial Engineering, Youngstown State University, Youngstown, OH 44555, USA
c
Civil/Envtl and Chemical Engineering, Youngstown State University, Youngstown, OH 44555, USA

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

Article history: The present research work takes advantage of a recently published numerical homogenization imple-
Received 6 January 2015 mentation in MATLAB to find the elasticity tensor of a ceramic–metallic composite (CMC) system to be
Received in revised form 17 June 2015 compared to an experimental data. Numerical homogenization is an efficient way to determine effective
Accepted 20 June 2015
macroscopic properties of a composite material. This technique represents an effective means to model a
Available online 3 July 2015
two-phase composite. In this work, an extension of a previously published numeric homogenization code
was investigated in order to model the compressive elastic modulus of the ceramic–metallic composites.
Keywords:
The extension to the numerical code makes use of physical micrograph images to accurately describe the
Numerical homogenization
Interpenetrating phase composites
phase distribution of the composite. Multiple micrographs were taken from each sample, to subsequently
Binder jet printing better represent the actual microstructure of the composite as a whole. The composites were created
Matlab using a binder jet 3D printing technology, where a ceramic precursor material was initially assembled,
followed by a molten metal infiltration process. It was found that the studied numerical homogenization
yielded an elastic modulus approximately 11.5% lower than the experimental data, suggesting a reliable
modeling technique for predicting the elastic tensor of CMCs.
Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction seems to be a feasible alternative. Binder jet process is a

The microstructure and mechanical properties of ceramic–
metallic composites (CMCs) have been extensively studied over
the last 20 years. CMCs can be manufactured through a large vari-
ety of techniques, including reactive metal penetration (RMP) [1],
directed metal oxidation (DIMOX) [2], and pressure assisted infil-
tration [3]. The composition of CMCs can be made based on a wide
variety of ceramic and metal materials. The most commonly stud-
ied CMC system is based on a Al2O3/Al composite which was orig-
inally patented by Breslin [1], who used a RMP process to produce
a co-continuous ceramic–metal composite (C4). In the RMP pro-
cess, a silica ceramic preform is reactively infiltrated with molten
aluminum to create the final composite. A reduction reaction
between the molten aluminum and silica creates the final
Al2O3/Al composite. Typically the ceramic precursor materials are
made by a slip or squeeze casting techniques. These manufacturing
techniques can be limited by the complexity of the part design. In
order to overcome the geometrical constraints associated with the
manufacturing of complex parts, the use of additive manufacturing
⇑ Corresponding author at: One University Plaza, Youngstown, OH 44555, USA.
E-mail address: Kmmyers01@student.ysu.edu (K. Myers).
layer-by-layer building approach which was originally established
at MIT [4]. In this process, a binder is physically deposited onto an
evenly spread powder bed following the design of a predetermined
CAD model. A fresh layer of powder is then laid on top of the pre-
vious layer and the process is repeated until the design is accom-
plished (see Fig. 1).
After the printing process, the binder is cured in order to harden
and glue the piece together. Subsequently, the binder is burned off
during the sintering cycle. The burning off of the binder creates a
natural porosity in the final part, which prevents the final part
from achieving highly dense parts. Yoo et al. [5] reported that they
were able to achieve a final density of 62.5%, without post process-
ing, for binder jet printed alumina. Although the porosity is not
always a desired feature on bulk ceramics, it may be beneficial
for the creation of infiltrated CMCs. In fact, most of the research
groups have used additive manufacturing to assemble a ceramic
backbone for subsequent metal infiltration, to create a multiphase
composite [6–17]. However, no previous studies on AM for produc-
ing multiphase composites have reported the use of the aforemen-
tioned RMP infiltration technique. This represents an opportunity
to explore the combination of AM with RMP to yield unique inter-
penetrated composites. Hence, the present paper will investigate

http://dx.doi.org/10.1016/j.commatsci.2015.06.031
0927-0256/Ó 2015 Elsevier B.V. All rights reserved.
 K. Myers et al. / Computational Materials Science 108 (2015) 128–135 129

CAD File Spread Powder Deposit Binder Drop Layer Final Part

Repeat

Fig. 1. Schematic for the binder jet printing process.

and model the elastic properties of an interpenetrating ceramic– RVE is the smallest volume of a structure which contains the
metal Al2O3/Al composite manufactured by binder jet printing required information under the chosen numeric scheme to repre-
and infiltrated via RMP, using a numerical homogenization tech- sent the macroscopic properties of the structure. In heterogeneous
nique. Here, the predicted elastic properties will be compared to materials, forming a RVE for all phases within the continuum is the
the experimental data to evaluate the reliability of the aforemen- process of homogenization. A main benefit of homogenization is
tioned technique. that the physical and the mechanical properties can be determined
analytically without need to test the material. This is especially
important in composites materials, since statistically relevant test-
1.1. Modeling of composite materials
ing of the numerous variables applicable to composites would be
There are many models and methods that have been studied to onerous and expensive.
While there exist many techniques for homogenization of a
simulate the physical properties of CMCs and interpenetrating
phase composites (IPCs); such as the Voigt and Reuss bounds multiphase composite, the method of cells, or unit cell method,
and Hashin spheres are two prominent homogenization methods.
[18], and the rule of mixtures [19]. More unique methods of creat-
ing unit cells have been studied by Wegner and Gibson [20] and Hashin and Shtrikman [28] used a variational approach and multi-
layer spheres to model particles in a multiphase material. These
Feng et al. [21]. However, all these previously mentioned methods
do not account for random geometries. Alternatively, a modeling spheres then form the RVE for this multiphase material. The unit
cell method, depending on which definition is taken, is synony-
based on finite element analysis, which divide an interpenetrating
phase composite (IPC) into cubic ‘‘voxels’’, seems to initially repre- mous of the RVE approach, and in practice, the two concepts are
interchangeable. Both of these methods discretize a material into
sent a suitable modeling approach, since the phase distribution of
the materials can be assigned with the corresponding properties. a periodic repeating structure, and therefore the quality of the
result of homogenization is directly tied to the quality of the
However, such ‘‘voxels’’ approach does not take into account the
randomness of the microstructure [22]. Agarwal et al. [23] used RVE. The authors find that the unit cell approach of Andreassen
and Andreasen [25] to be advantageous due to the ability to incor-
an effective mesh free method to describe IPCs. Kaminsky and
Kleiber [24] modeled the randomness of an IPC using a two-step porate micrographs into the homogenization. This allows for the
creation of multiple, but not necessarily unique unit cells, which
homogenization method. Although some of the methods can
describe an IPC’s material properties accurately, they require a will discretize the domain.
According to the theory of homogenization, the macroscopic
lot of computing time. On the other hand, Andreassen and
Andreasen [25] have offered a MATLAB code based on a homoge- elasticity tensor for two distinct periodic phases is given by Eq.
(1) [27];
nization technique which is computing friendly. Hence, the present
study uses this MATLAB code, with additional modifications to Z  
1 
accurately model the elastic tensor of a 3D printed Al2O3/Al C Hijkl ¼ C pqrs 0ðijÞ
epq  eðijÞ
pq ers0ðklÞ  ersðklÞ dV ð1Þ
jVj V
composite.
where V is the volume of the unit cell, C Hijkl is the homogenized elas-
1.2. Homogenization background ticity tensor, and C pqrs , which is a function of position, is the local
elasticity tensor. Here, C pqrs can be obtained by
Materials like composites, solid foams, bone, and ceramic
matrix composites consist of multiple, distinct phases. Each of C pqrs ðxÞ ¼ C mat:1
pqrs b
mat:1
ðxÞ þ C mat:2
pqrs b
mat:2
ðxÞ ð2Þ
these phases has their own individual physical and mechanical
where b(x) is an indicator function to determine the phase
properties, and it is the sum of these phases, that form a heteroge-
(mat.1 = material 1 and mat.2 = material 2) for a given position,
neous material. Most heterogeneous materials, CMC’s included, 0ðijÞ ðijÞ
exhibit a random arrangement of phases throughout its continuum and epq is the macroscopic strain field. Here, epq represents the
structure. Prediction of the mechanical properties of these materi- local strain fields and it is given by
als can be performed through the micro-mechanics theory. This !
ðijÞ ðijÞ
typically makes use of a representative volume element (RVE), or ðijÞ 1 @up @uq
epq ¼ þ ð3Þ
a statistical volume element (SVE). A RVE is the smallest volume 2 @xq @xp
of a structure which still represents the macroscopic properties
of the structure, while a SVE is smallest volume of a structure that Therefore, Eq. (1) can be found by solving the elasticity equation,
statistically represents the macroscopic properties of the structure Z Z
[26,27]. It should be noted that a RVE and SVE are very similar in C ijpq eij ðv Þepq ðukl ÞdV ¼ C ijpq eij ðv Þe0ðklÞ
pq dV ð4Þ
V V
theory, where both intend to capture the macroscopic properties,
but only the mechanism differs between them. Additionally, the where v is a virtual displacement within the unit cell, which is cal-
authors suggest an addendum to the definition of the RVE; the culated by weighted residuals or finite element techniques.
130 K. Myers et al. / Computational Materials Science 108 (2015) 128–135
1.3. MATLAB code
As mentioned before, the MATLAB code of Andreassen and
Andreasen [25] was extended upon in this research, to model
CMCs with random microstructures. Their code was created to
use a homogenization technique to find the elastic tensor of a
two phase composite system. The present work has implemented
an extension which consists of using photographs of the
microstructure of the composite as the homogenization platform.
This code is provided in Appendix A. The microstructure pho-
tographs are converted into a binary code, and incorporated into
the MATLAB program, which in turn calculates the phase distribu-
tion of the composite from the inputted image. The modified code
allows input from multiple images in order to achieve statistical
relevance. Here, the code first prompts the user to enter the
amount of samples (images) that will be analyzed, as well as the
elastic modulus and Poisson’s ratio for each component of the
composite to calculate their individual Lamé parameters (Eqs.
(5)–(7), respectively). The experimental data was run with a sam-
ple that had the dimensions of 8.58  13.37  22.38 mm
(width  thickness  height). The sample’s dimensions and orien-
tation to the compression load put the sample under plane stress
conditions, therefore, the Lamé parameters for plane stress (k, l,
and ^k), will be calculated in the modified code for the simulation.
These Lamé parameters relate the elastic and shear modulus of a
homogeneous and isotropic material.
mE
k¼
ð1 þ mÞð1  2mÞ
E respectively. The fifth input is the angle, u, which is considered
l¼
2ð1 þ mÞ
^k ¼ 2lk
k þ 2l
Following the incorporation of all desired images into the code, the
images are cropped by using the built-in MATLAB function ‘‘im-
crop’’, and subsequently converted into a grayscale version, through
the MATLAB function ‘‘graythresh’’. The ‘‘graythresh’’ function
determines the interface phase boundaries in order to distinguish
between them. The interface value combined with the grayscale
image, produces a purely black and white ‘‘binary’’ version of the
image using the command ‘‘im2bw’’. A dialog box then follows
and prompts the user to adjust the image contrast, if necessary, to
best capture the microstructure while mitigating possible image
defects. The resulting image represents a matrix of 1’s and 0’s which
is used in the homogenization function. The homogenization func-
tion requires a square matrix for calculation purposes. An example
of an original inputted image used in this research work and its bin-
ary image can be seen in Fig. 2.
Fig. 2. Original microstructure image of Al2O3/Al composite (left) and its binary image (right).
The phase distribution of the sample is then calculated by using
the area of the square cropped image. The code counts the number
of 0’s in the matrix and divide them by the total number of pixels
in the image. This provides the phase distribution of both phases in
the system. In order to collect the phase distribution of multiple
samples, a new variable is created for each image for the black
and white phases. It is worth mentioning that the MATLAB code
created by Andreassen and Andreasen [25] provides a random
matrix for the integers 1 and 2 for a 200  200 matrix. This random
matrix is to represent the microstructure of a two-phase composite
system. The extension added in this work was done to represent
the actual microstructure of a known composite system. To apply
the modified code to the known phase distribution of the
Al2O3/Al composite investigated here, the binary matrix was chan-
ged into 1’s and 2’s to fit the original existing code. Also, the binary
matrix needs to have a specific numerical value which was per-
formed by using the command ‘‘double’’, which gives the matrix
an appropriate data structure.
The homogenization code established by Andreassen and
Andreasen [25], also requires six inputs to run. The first two inputs
are the length and width of the unit cell. Since the image is
squared, the generic value of 1 is used for both. The third and
fourth inputs are the individual Lamé parameters (^ k and l) for each
component, which as mentioned before, were calculated by the
Young’s modulus and Poisson’s ratio of the two investigated mate-
rials. For this research work, the Lamé constants were calculated
using values of 69 GPa and 0.3 for aluminum’s Young’s modulus
and Poisson’s ratio respectively. For alumina, the values of
ð5Þ
375 GPa and 0.22 for the Young’s modulus and Poisson’s ratio
ð6Þ the angle between the horizontal axis and the left wall of the unit
cell. Andreassen and Andreasen [25] state that the angle u can be
ð7Þ between 45° and 135° to represent a repeatable parallelogram of
the unit cell. For the present research work, the angle was set to
90°. The final input is a matrix representing the microstructure
of the two phase system. In this study, the matrix was the square
binary image of the Al2O3/Al system. Following the entrance of
the previous six inputs, the homogenization technique runs auto-
matically. When the program is finished, it outputs the elastic ten-
sor and the phase distribution percent of the composite, which
were averaged across all of the inputted samples.
2. Experimental
Fused silica (CE Minerals) was sieved through a 90 lm screen to
be used in the binder jet printing process. The 3D printing was
accomplished using an X1-Lab (Ex One) machine. The parts were
designed as simple rectangular boxes in SolidWorks CAD software,
and the dimension were made large enough to fit a strain gauge for
 K. Myers et al. / Computational Materials Science 108 (2015) 128–135 131

later testing. During the printing, the silica powder was spread at a
0.1 mm/s speed. The binder was a water based solution proprietary
to Ex One. After printing the silica part, it was cured at 190 °C for
4 h and subsequently sintered at 1500 °C for 4 h. The sintered
Fig. 4. The average elastic tensor found for Sample 1 using 10 images.
parts were then infiltrated by submerging them in a molten alu-
minum bath. The final composites were then ground and polished
for optical microscopy. The images used in the homogenization Subsequently, the elastic tensor was then used to calculate the
technique were taken with a Zeiss Axiphot microscope at elastic modulus Ex assuming plane stress conditions by:
200 magnification. The experimental modulus was found using 2 3 2 32 3 2 32 3
a universal compression testing machine with a loading rate of rx C 11 C 12 C 16 ex Ex Ex myx 0 ex
6 7 6 76 7 1 6 76 7
0.1 mm/min. 4 ry 5 ¼ 4 C 21 C 22 C 26 54 ey 5 ¼ 4 Ey mxy Ey 0 54 ey 5
1  myx mxy
rxy C 61 C 62 C 66 exy 0 0 Gxy exy
ð8Þ
3. Results and discussion
where mxy and myx were calculated by:
Two different samples were used for the homogenization tech- C 21
nique. Both of which were created by printing a simple cube shape
mxy ¼ ð9Þ
C 22
using the fused silica powder. The first sample was sintered at C 12
1500 °C and then metal-infiltrated. At the sintering temperature myx ¼ ð10Þ
C 11
of 1500 °C, the sample showed no drooping or warping of the
printed shape. The second sample, which was sintered at And subsequently Ex was calculated by:
1650 °C, showed drooping and warping from the original dimen- Ex ¼ C 11 ð1  myx mxy Þ ð11Þ
sions due to the high sintering temperature. Therefore, experimen-
tal quasi-static testing was not preformed due to the parts warping The calculated values for mxy, myx, and Ex for the 10 images of Sample
and only the homogenization analysis was done. The modeling 1 were found to be 0.287, 0.286, and 147.8 GPa respectively. In con-
results of this second sample were compared to literature data. It trast, the sample yielded an experimental elastic modulus of
should be noted that lower sintering temperatures results in sam- 167 GPa. This represents a theoretical–experimental deviation of
ples with higher volume fraction of the metal phase, which is the 11.5%. Here, it was found that Sample 1 has a phase distribution
result of the lower sintering density of the printed silica. of 49.7% Al2O3 and 50.3% Al. For comparison 20 images were ana-
Ten images from both samples were taken with an optical lyzed for Sample 1. The calculated values for mxy, myx, and Ex for 20
microscope with 200 magnification. Three images that were images of Sample 1 were found to be 0.289, 0287, and 144.6 GPa
taken from Sample 1 can be seen in Fig. 3, where the white phase respectively, with a deviation of 13.4%. From the simulated results,
in the images is the aluminum, and the dark phase is the alumina. it can be observed that the addition of 10 more images in the anal-
The 10 images for Sample 1 were inputted into the MATLAB ysis process has a negligible difference (see Fig. 5) when comparing
code for analysis. The elastic tensor for each image was calculated 20 images to 10 images.
and then averaged together. The elastic tensor that was calculated It is possible to achieve a deviation in the elastic tensor that
by the homogenization technique using the 10 images can be seen is computed from an individual image due to the contrast
in Fig. 4. changing, random cropping of the image, and the program

Fig. 3. Representative example of the images used in the homogenization technique of the Al2O3/Al composite.
132 K. Myers et al. / Computational Materials Science 108 (2015) 128–135

45 cropping the data again to form a square matrix. To evaluate

40 this deviation, the homogenization technique was ran 10 times,
35 using the same 10 images, cropping a random portion of the
image with choosing the most appropriate phase contrast. The
30
Stress (MPa)

resulting average had the elastic modulus at 145.7 GPa with a

25 Experimental Results standard deviation of 0.75 GPa. This investigation shows that
20 10 Pictures there is also a very minimal deviation from using the same
15 20 Pictures images on multiple occasions.
10 All of the images used in the homogenization process were
5
taken from the same plane of the cubic sample. The microstruc-
ture of the other two orthogonal planes show similar random
0
0 0.00005 0.0001 0.00015 0.0002 0.00025 0.0003 heterogeneous phase distribution, suggesting that the results
Strain in the other planes would yield similar elastic moduli in the
simulation. The microstructures of all three planes can be seen
Fig. 5. Stress–strain curve for the experimental and homogenization results. in Fig. 6.

B
C
A

A B C

Fig. 6. Microstructure of all three planes of Sample 1.

Fig. 7. Images that were used in the homogenization technique of the Al2O3/Al (Sample 2).

Porosity
Porosity

Fig. 8. Micrograph of an Al2O3/Al composite highlighting the porosity (left) and the resulting binary image used in the homogenization technique (right).
 K. Myers et al. / Computational Materials Science 108 (2015) 128–135
Fig. 7 shows some of the images taken from Sample 2 for its
homogenization analysis. Indeed, due to the higher sintering tem-
perature (1650 °C), which resulted in a denser structure, the code
calculated a higher volume fraction of ceramic phase than in the
first sample. Here, the calculated average phase distribution for
the Sample 2 was 61.1% Al2O3 and 38.9% Al.
For Sample 2, ten images were used in the homogenization
technique, which yielded an elastic modulus Ex equal to
180.8 GPa. Breslin et al. [29] reports their C4 Al2O3/Al to have an
elastic modulus of 215 GPa (15.9% difference). However, their
investigated C4 material was approximately 65% Al2O3 and 35%
Al. In contrast, the averaged phase distribution for the simulated
specimen (Sample 2) has less volume fraction of the ceramic phase
(61.8% Al2O3). Hence, the slightly higher amount of the Al phase in
Sample 2, caused the predicted elastic modulus to be below than
the reported by Breslin et al.
It should be noted that both samples used in the homogenization
technique have a small degree of porosity through the sample. The
presence of porosity can result in an error during the modeling pro-
cess, since the porosity is black in the images being used, and the
program recognizes these pixels as part of the darker phase (in this
case, as the ceramic Al2O3). Therefore, the modeled homogenized
modulus could result in an increased error of the elastic modulus
due to a misidentification of phases. Fig. 8 shows how the MATLAB
program recognizes porosity as an actual phase. Here, the porosity
is highlighted in a red box in both images. In the actual micrograph
of the sample (left), the porosity can be easily seen by its pure black
color, which is darker than the Al2O3 phase. Whereas in the binary
image (right), it is shown that the porosity is recognized as the dark
(ceramic) phase in the image. Increasing the magnification of the
images used will allow the program to detect smaller differences
in the microstructure of composite materials.
133
4. Conclusions
A homogenization technique was implemented on a 3D printed
and infiltrated ceramic–metallic composite using a MATLAB code
created by Andreassen and Andreasen [25]. The MATLAB code
was expanded by turning a micrograph of a two phase system into
a binary matrix, which was subsequently used in a homogeniza-
tion technique. Several micrographs of a Al2O3/Al composite were
used to have a representative average phase distribution of the
sample. Two different Al2O3/Al composite with different
microstructures were modeled. The homogenization technique of
a high temperature sintered structure predicted a modulus of
180.8 GPa, which was 15.9% lower than the reported in the litera-
ture. In contrast, the modeling homogenization of a sample sin-
tered at a lower temperature (1500 °C), resulted in an elastic
modulus of 147.8 GPa, which is a value of approximatelly 13.4%
lower than that obtained experimentally. It was also found that
using 20 images instead of 10, there is little deviation in the results
for the modeling process. The homogenization of 10 images, ana-
lyzed only once, in the MATLAB program was preformed in under
1.5 h, proving that this technique is a rapid and reliable way to find
an elastic modulus of a two phase system.
The initial simulated results found in this work show that the
described MATLAB implementation of a homogenization technique
to be promising. It quickly and accurately describe the mechanical
properties of two phase composite system. It is suggested for
future work, that the unit cell be optimized to lower the percent
error between the simulated and experimental data even more.
This can be achieved by incorporating a statistical volume element
(SVE) with the use of a ‘‘local’’ RVE approach. Indeed, extensions of
the currently described work into three dimensions, porosity, and
also three or more phases is a likely source of future work.
Appendix A
(continued on next page)
134 K. Myers et al. / Computational Materials Science 108 (2015) 128–135

Appendix A (continued)
 K. Myers et al. / Computational Materials Science 108 (2015) 128–135
References
[1] M.C. Breslin, Process for Preparing Ceramic–Metal Composite Bodies, U.S.
Patent 5214011, May 25, 1993.
[2] L. Yang, D. Zhu, C. Xu, Ju. Zhang, Ji. Zhang, Metall. Mater. Trans. A 27A (1996)
2094–2099.
[3] D. Horvitz, I. Gotman, Acta Mater. 50 (2002) 1961–1971.
[4] H.J. Jee, E. Sachs, Adv. Eng. Softw. 31 (2) (2000) 97–106.
[5] J. Yoo, M.J. Cima, S. Khanuja, E.M. Sachs, Structural ceramic components by 3D
printing, in: Solid Freeform Fabrication Symposium, 1993, pp. 40–50.
[6] L.D. Wegner, L.J. Gibson, Int. J. Mech. Sci. 42 (2000) 943–963.
[7] C. San Marchi, M. Kouzeli, R. Rao, J.A. Lewis, D.C. Dunand, Scr. Mater. 49 (9)
(2003) 861–866.
[8] R. Melcher, S. Martins, N. Travitzky, P. Greil, Mater. Lett. 60 (4) (2006) 572–575.
[9] R. Melcher, N. Travitzky, C. Zollfrank, P. Greil, J. Mater. Sci. 46 (5) (2010) 1203–
1210.
[10] W. Zhang, N. Travitzky, P. Greil, J. Am. Ceram. Soc. 91 (9) (2008) 3117–3120.
[11] W. Zhang, R. Melcher, N. Travitzky, R.K. Bordia, P. Greil, Adv. Eng. Mater. (2009)
1039–1043.
[12] Z. Fu, L. Schlier, N. Travitzky, P. Greil, Mater. Sci. Eng., A 560 (2013) 851–856.
[13] C.R. Rambo, N. Travitzky, K. Zimmermann, P. Greil, Mater. Lett. 59 (8–9) (2005)
1028–1031.
135
[14] X. Yin, N. Travitzky, P. Greil, Int. J. Appl. Ceram. Technol. 4 (2) (2007) 184–190.
[15] S. He, X. Yin, L. Zhang, X. Li, L. Cheng, Trans. Nonferrous Met. Soc. China 19 (5)
(2009) 1215–1221.
[16] B. Nan, X. Yin, L. Zhang, L. Cheng, J. Am. Ceram. Soc. 94 (4) (2011) 969–972.
[17] X. Yin, N. Travitzky, P. Greil, J. Am. Ceram. Soc. 90 (7) (2007) 2128–2134.
[18] S. Torquato, Int. J. Solids Struct. 37 (2000) 411–422.
[19] M.K. Aghajanian, N.H. MacMillan, C.R. Kennedy, S.J. Luszcz, R. Roy, J. Mater. Sci.
24 (1989) 658–670.
[20] L.D. Wegner, L.J. Gibson, Int. J. Mech. Sci. 42 (2000) 943–964.
[21] X.Q. Feng, Y.W. Mai, Q.H. Qin, Comput. Mater. Sci. 28 (2003) 486–493.
[22] Z. Poniznik, V. Salit, M. Basista, D. Gross, Comput. Mater. Sci. 44 (2008) 813–
820.
[23] A. Agarwal, I.V. Singh, B.K. Mishra, J. Compos. Mater. 47 (2012) 1407–1423.
[24] M. Kaminski, M. Kleiber, Int. J. Numer. Methods Eng. 47 (2000) 1001–1027.
[25] E. Andreassen, C.S. Andreasen, Comput. Mater. Sci. 83 (2014) 488–495.
[26] S. Nemat-Nasser, M. Hori, Micromechanics: Overall Properties of
Heterogeneous Materials, Elsevier, Amsterdam; New York, 1999.
[27] S. Torquato, Random Heterogeneous Materials: Microstructure and
Macroscopic Properties, Springer, New York, 2002.
[28] Z. Hashin, S. Shtrikman, J. Mech. Phys. Solids 11 (1963) 127–140.
[29] M.C. Breslin, J. Rignalda, L. Xu, M. Fuller, J. Seeger, G.S. Daehn, T. Otani, H.L.
Frazier, Mater. Sci. Eng., A 195 (1995) 113–119.