An ecient multiscale model of damping properties for lled

elastomers with complex microstructures

John A. Moore
a
, Ruizhe Ma
a
, August G. Domel
a
, Wing Kam Liu,
a,1,
a
Department of Mechanical Engineering, Northwestern University, 2145 Sheridan Road, Evanston, IL
60208-3111, USA
Abstract
This work proposes an ecient framework for prediction of lled elastomer damping properties
based on imaged microstructures. The eciency of this method stems from a hierarchical multi-
scale modeling scheme, in which the constitutive response of subcell regions, smaller than a repre-
sentative volume element (RVE), are determined using micromechanics; the resulting constitutive
parameters then act as inputs to nite element simulations of the RVE. It is shown that the mi-
cromechanics models of Halpin-Tsai and Mori-Tanaka are insucient for modeling subcells with
many ller clusters, and thus these models are augmented by an additional interaction term, based
on stress concentration factors. The multiscale framework is veried in two dimensions and extended
to predictions for three dimensional systems, which include the response of matrix-ller interphase
properties. The proposed multiscale framework shows a signicant improvement in computational
speed over direct numerical simulations using nite element analysis, and thus the proposed frame-
work allows for detailed parametric studies of microstructural properties to aid in the design of
lled elastomeric systems.
Keywords: A. Particle-reinforcement, B. Mechanical properties, C. Micro-mechanics, C.
Computational modeling, Imaged-based modeling

Corresponding author
Email address: [email protected] (Wing Kam Liu,)
1
Adjunct Professor under the Distinguished Scientists Program Committee at King Abdulaziz University (KAU),
Jeddah, Saudi Arabia
Preprint submitted to Composites Part B: Engineering November 3, 2013
1. Introduction
Filled elastomers have found many applications in a variety of elds spanning the automotive[1]
[2], civil engineering [3] [4] and even food packing industries [5]. In many applications polymers
provide desirable damping or energy dissipation properties but can lack the strength needed to
carry structural loads or resist wear. The addition of llers add strength to polymer systems but
can reduce damping properties. Wang [2] gives a specic example where it is shown that the wet
grip of tires is improved by increasing damping property values at room temperature. The addition
of strengthening llers, however, acts against these wet grip improvements by decreasing room
temperature damping values. Thus, as damping decreases, performance decreases as well, creating
a design dilemma, trading performance with strength.
A powerful tool to aid in this design decision would be one that could accurately predict the
damping properties of an existing material while simultaneously allowing for parametrized studies
of new virtual materials. This model would allow a designer to determine if a material meets both
the strength and damping goals before the material is produced and to gage what modications are
needed before production. Such a model is the goal of this work. This model will be based on the
premise that knowledge of the three dimensional (3D) morphology of a lled system (for example
from sectioned scanning electron microscopy) is key to accurate computational models. The major
challenge to be addressed is that direct numerical simulations (DNS) using computational tools
such as molecular dynamics or the nite element method (FEM) can be prohibitively slow if not
intractable when modeling a 3D microstructure on the scale of a representative volume element
(RVE). Thus, a multiscale approach is proposed here using micromechanics to homogenize the
morphology of sub-RVE scale subcells. A coarse nite element model is then used to simulate the
combined response of these subcells, resulting in a reduced computation size as compared to DNS
of the entire microstructure. While a homogenized model cannot account for all of the information
included in direct simulations, the advantage in computational speed will be substantial, making it
a powerful tool for materials design of lled elastomer systems.
Section 2 will outline previous micromechanical models and image based modeling. The material
properties used in each simulation will be given in Section 3. Section 4 will describe both the nite
element implementation as well as the multiscale modeling scheme. The sub-RVE mechanisms
aecting damping will be explored in Section 5. Section 6 will pose an extension to existing
micromechanics models to capture these sub-RVE mechanics. The implementation of such a model
2
in two dimensions (2D) and results are shown in Section 7 and compared to DNS to verify the
model. Once veried, the model is used to predict the response of a 3D microstructure in Section
8. The eects of an interphase region are then explored for a 3D system in Section 9. Finally,
conclusions are drawn in Section 10 and extensions are posed for future modeling and materials
design.
2. Background
Two micromechanics models are popular for representing the eective stiness of a lled com-
posite and will form the basis for the proposed work. These are the Mori-Tanaka [6] and Halpin-Tsai
[7] models. Both take into account the interaction between ller particles. To illustrate the nature
of this interaction, the framework of each of these models will be briey outlined. The formula-
tion of each theory can be understood by rst considering a series of inclusions with stresses and
strains characterized by Eshelbys solution [8](as described in [9]). The composite properties which
encompass both the eect of inclusions and the surrounding matrix are then given by fourth order
stiness tensor

L

L. The value of

L

L can be determined by postulating a matrix strain and stiness

properties

L

L for a region directly surrounding each inclusion. The assumptions regarding these
strain and stiness values are often what distinguish one micromechanical model from another. For
instance, if

L

L is dened as the matrix properties LLL

0
, and as the average strain of the total body,
then a dilute approximating is found, which is simply a sum of several traditional Eshelby solutions
for each inclusion. If

L

L is dened as the matrix properties LLL

0
and as the average strain in the
matrix alone,
0
, then the Mori-Tanaka stiness tensor is found. The Mori-Tanaka stiness tensor
is given as

L
MT
= LLL
0
+
N

r=1
c
r
(LLL
r
LLL
0
)TTT
r
[c
0
III +
N

n=1
c
n
TTT
n

1
, (1)
where c
o
is the volume fraction of matrix material, c
r
is the volume fraction of phase r, III is a fourth
order identity tensor, N is the number of distinct inclusion phases, and
TTT
r
= III +SSS
r
LLL
1
0
(LLL
r
LLL
0
)_
1
. (2)
Where the Eshelby tensor SSS
r
is a function of LLL
0
and
0
. Finally, if

L

L is chosen to be the average

composite properties

L

L, and is chosen as the average total strain then a self consistent approx-
imation is formed; in that, the self consistent stiness tensor

L

L
SC
is expressed as a function of
3
itself. Many approaches [7] solve for specic cases of

L

L
SC
. The Halpin-Tsai equations simplify the
self consistent approaches of Hermans [10] and Hilll [11] resulting in the following expression for
composite Youngs modulus [7],

E
yy
E
m
=
1 + 2Ac
f

1 c
f

(3)
=
(E
f
]E
m
) 1)
(E
f
]E
m
) + 2A
(4)
where A is the aspect ratio of the inclusion particle in two dimensions (i.e. L]t where L is the
particle length and t the particle thickness, L being in the direction of loading y), c
f
is the ller
volume fraction, E
m
is the Youngs modulus of the polymer matrix, E
f
the Youngs modulus
of the ller particles and E
yy
the longitudinal Youngs modulus of the composite system. In
the determination of Equations 3 and 4 a number of assumptions were made. Of note were the
assumption that the composite modulus is independent of matrix Poissons ratio and that all ller
particles are uniaxially aligned and identical [10]. Also, the term 2Aused in Equation 3 and 4 was
determined through comparison with available nite element results [10].
It can be shown from the above formulation that the Mori-Tanaka model and the self consistent
theory (on which Halpin-Tsai model is based) both account for the interaction of ller inclusions.
Mori-Tanaka assumes that the strain in a given inclusion can be determined using the average matrix
strain,
0
which diers from the average total strain. This dierence is attributed to the presence
of the other inclusions. Thus, a single inclusions response is a function of all other inclusions.
Self-consistent theory makes a similar assumption, that the inclusions stress is a function of the
average bulk properties

L

L resulting again in a single inclusions response being a function of all

other inclusions. This shows that both models allow for a single inclusions stress to be altered by
the presence of other inclusions and thus, an interaction between inclusions is modeled. The author
will dene this as a weak interaction, in that, the explicit increase in stress due to a pair or group
of inclusions in not modeled. The extensions proposed in Section 6 will attempt to capture a strong
interaction between inclusions by modeling stress risers based on the inuence of adjacent llers.
The application of these models to viscoelasticity is achieved by using the elastic-viscoelastic
correspondence principle. The details of which are given in the work of Brinson et al. [12]. In
short, an elastic modulus, such as Youngs modulus, can be viewed as complex and a function of
4
frequency, .
E

() = E

() + iE

(), (5)
where E

and E

are the storage and loss modulus respectively and i =

1. Where E

represents
the elastic response in-phase with the strain loading and E

represents the response 90 degrees

out-of-phase with the loading due to time dependent viscoelastic eects. The ratio of the loss to
storage modulus will be referred to here as tan (where is the phase lag of the material) and is
proportional to the energy dissipation or damping of the system [2].
tan =
E

(6)
This value can be directly related to properties of importance to many engineering applications and
thus this work will focus on determining tan for a lled elastomer. Using Equation 3 and Equation
4 and assuming a linear elastic ller model, Halpin-Tsai can be written in a complex form, where
E

yy
and E

m
are functions of frequency, as

yy
E

m
=
1 + 2Ac
p

1 c
f

(7)

=
(E
f
]E

m
) 1)
(E
f
]E

m
) + 2A
(8)
from which tan can be determined. Likewise, if a complex Poissons ratio is given as

and only
two phases are considered, then E

and

of the matrix and ller can be determined by inserting

the complex forms of SSS

r
, LLL

r
and LLL

0
into Equation 1, resulting in a complex Mori-Tanaka stiness
tensor [12], given by

L
,MT
= LLL

m
+ c
f
(LLL

f
LLL

m
)TTT

f
[c
m
III + c
f
TTT

1
, (9)
where TTT

f
is the complex form of Equation 2 evaluated using the material properties of the ller.
Brinson et al. [12] have studied

L

using Mori-Tanaka and have shown good agreement with nite

element simulations of a periodic computational unit cells with a single inclusion. Brinsons work,
however, was not intended to model larger cells with many interacting inclusions. Putz et al. [13]
used the complex form of the Halpin-Tsai expression to compare to experimental data for carbon
nanotube lled composites, but no work was done to remedy discrepancies between the model and
experiments.
5
A number of works have included a region of modied properties between the ller and matrix,
known as the interphase region (or IP region), into micromechanical models. Ji [14] proposed a three
phase model for composite stiness. This model accounted for a linear gradient in matrix properties
near llers and agreed well with experimental data for polystyrene/cloisite 10A nanocomposites [1];
however, these results where determined by tting a number of parameters in the model to agree
with experimental ndings [1]. The model was also limited to random ller microstructures and was
not applied to viscoelastic responses. Three phase micromechanics models, including Jis model,
along with self consistent models, such as [15] which include a particle coating phase, are intended
for randomly dispersed ller particles and do not account for a strong interaction between closely
grouped particles. Gusev and Lurie [16] used a four phase model of spherical ller particles to
predict tan for ordered and random microstructures. The four phase model and detailed nite
element simulations agreed well when only the two interphase regions were considered viscoelastic
(the matrix and llers were considered elastic), and a single frequency of 1 rad/sec was studied.
Gusev also observed that the results from ordered and random microstructures showed very similar
trends. From this observation they concluded that the eective properties of composites with
low aspect ratio llers lack a strong sensitivity to micro-structural variations, thus accounting for
the eectiveness of the four phase model. The present work will show that for arbitrary shaped
llers with viscoelastic matrix properties, considered over a range of frequencies, microstructure
plays a more prominent role, and thus it is expected, that the four phase model of Gusev may
prove less eective. H. Liu and Brinson [17] have also carried out detailed nite element analysis
to inform models of IP eects. Brinson numerically computed a strain concentration tensor for
various inclusion geometries, such as nanotubes and nanoplates. This analysis requires a specic
nite element analysis for all ller geometries and thus may become computationally expensive for
arbitrarily complex microstructures. A similar approach was employed by Song et al. [18]. This
model introduced a scalar strain amplication factor pre-multiplying the complex matrix modulus
of a lled system. The amplication factor was expressed in an semi-empirical form accounting for a
number of microstructural parameters but on some empirical factors in order to match experimental
data.
The model presented here will be capable of including IP eects but will rst focus on computing
mechanics of a sub-RVE cell containing many interacting ller particles. Then the IP properties will
be added for a 3D demonstration. The model will rely on extracting data from a microstructural
6
image. This image is broken down into binary voxels which represent either matrix or ller and
are illustrated in Figure 1. A recent work by Mishnaevsky [19] used voxel based images such as
a b c d
A
B
C
D
Filler
Matrix
Figure 1: voxel representation of microstructure
in Figure 1 combined with micromechanics. Mishnaevskys approach was to extract the volume
fraction of the voxel columns (i.e. a-d Figure 1) and determine the stiness of each using an
isostress Reuss model. The combined (composite) response of all columns was then calculated
using a Voigt (isostrain) model. Through this approach

E can be quickly calculated for arbitrarily
complex microstructures. This work was not extended to viscoelastic properties. The simulation
results where able to match experimental trends when modulus was compared to weight content of
nanoclays but failed to match the slope and magnitude of such a curve.
The work of Deng et al. [20] used this voxel representation to determine the interphase prop-
erties of a lled polymer system by comparing nite element analysis of a 2D, 300 x 300 voxel,
reconstructed image to experimental results. This work was extended to 3D in the study by Bren-
eman et al. [21]. For both of these works interphase thickness was limited to integer multiples of
the mesh size. The proposed model will speed up the voxel based calculations of Deng et al. and
Breneman et al. and also allow for more exibility in interphase thickness modeling.
The work shown here uses both Mori-Tanaka and Halpin-Tsai models extended to capture strong
interactions between particles (and IP in 3D) and will also capture arbitrarily complex structures
through the voxel representations.
7
3. Material Properties
In this work, elastomeric polymers with sti llers will be addressed. Examples of poly-
mer/matrix materials are styrene-butadienes [2] or ethylene-propylenes [18], while examples of
llers are carbon black or silica [2]. A viscoelastic polymer with frequency dependent properties
will be used for the matrix. Values of matrix properties E

m
and E

m
used in this work are based
on the soft material tan
m
studied by Brinson and Lin [12]. The values of E

m
and E

m
and tan
m
are shown as a functions of frequency in Figure 2. The polymer matrix is considered nearly in-
10
6
10
4
10
2
10
0
10
2
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
x 10
8
Y
o
u
n
g
s

s

M
o
d
u
l
u
s
,

P
a
Frequency, 1/second


E
E"
10
6
10
4
10
2
10
0
10
2
0
0.5
1
1.5
2
2.5
t
a
n

Frequency, 1/second
Figure 2: Material properties of matrix based on data from [12]
compressible with a Poissons ratio of
m
= 0.48. The density of both phases is considered small
(
m
=
f
= 1x10
9
kg/m
3
) to avoid any inertial eects in nite element analysis which are absent in
micromechanical modeling and negligible in dynamic mechanical analysis (DMA) testing used to
characterize viscoelastic properties. The ller is considered to be sti as compared to the matrix
and of a brittle nature. Thus, the ller is considered to have E
f
= 1x10
9
Pa. A ller Poissons
ratio of
f
= 0.4 is used in this work based on the value used in [20]. The values of E
f
and
f
are
considered frequency independent.
8
4. Finite Element Model
Finite element modeling will be used for two purposes in this work. First, as a verication
tool for both subcell and cell models, and second, as an RVE level simulation tool of the cell (see
Figure 3). The properties of an RVE are dened in the Micromechanics text [9]. For this work an
Verification
Multiscale Model
Cell (DNS)
(a)
Cell (Homogenized)
(c)
Subcell (DNS)
(b)
Figure 3: a) 2D voxel model of cell, colors represent volume fraction of subcell region b) 10x10 subcell used for DNS
verication and to determine physical descriptors as input to the homogenized cell model c) homogenized cell model
colors represent dierent material properties
RVE cell (Figure 3a) is broken down into subcells (Figure 3b); each subcells average constitutive
properties are then calculated using micromechanics. A coarse nite element mesh of the cell is then
created with each element representing a subcell (Figure 3c). The result is a hierarchical multiscale
modeling procedure. This procedure recognizes that the response of each subcell will dier from
other subcells and from the response of the RVE as a whole (if this was not true then the subcell
would be considered the RVE). Thus, the cell nite element model allows for strain inhomogeneities
within an RVE by allowing each element to have dierent material properties based on each subcells
microstructure.
The FEM models used to verify subcell calculations and the FEM models used to simulate
the RVE cell will employ the same boundary conditions and nite element formulations; only the
meshes will dier between the subcell and cell. Both the subcell and cell models will use the
9
following assumptions. The boundary conditions shown in Figure 4 will be used. The right edge of
x
y

A
T
I
E
D

T
O

A

Figure 4: Example boundary conditions and mesh, circles represent roller supports and
0
is a uniform applied strain
the mesh in Figure 4 is tied to node A in the x-direction such that the right edge remains straight.
This constraint forces periodic stress boundary conditions on the model which account for Poissons
eect. In 2D, a plane strain assumption will be used to model the subcell/cell as a piece of a much
larger bulk of material. A displacement resulting in an applied strain of
0
is applied in the y-
direction as shown in Figure 4. The FEM analysis will be linear; thus, the magnitude of
0
will
have a linear eect. While an applied strain of
0
yy
= 1.0 + 0.0i is assumed, this value is somewhat
arbitrary because responses from other applied small strains can be linearly scaled from this result,
and the composite values of E

, E

and tan are independent of the real value of

0
yy
.
To account for the nearly incompressible nature of the matrix, hybrid four node quadrilateral
elements are used in 2D to avoid numerical locking [22] (see Abaqus manual for more details [23]).
Hybrid eight node bricks are used in 3D. Finally, the matrix and llers are considered perfectly
bonded. The interface nodes between each phase are concurrent, and no contact or multiple point
constraint are used to connect the phases. Three example meshes are shown in Figure 5. Figure
5a is a subcell mesh based on voxel representations, 5b is a subcell mesh for a single particle and
10
5c is the coarse homogenized mesh in 2D. Again the reader is reminded that 5a and 5b are for
verication only, where as 5c is used as part of the multiscale model.
Verification Subcell Meshes
Mutiscale Mesh
Random Particle
Subcell
(a)
Homogenized
Cell
(c)
Single Particle
Subcell
(b)
Figure 5: a) uniform subcell mesh with random ller particles b) ne mesh of single particle c) coarse mesh of
homogenized cell
5. Subcell Mechanisms
The model extension presented in Section 6 is based on the mechanics of a subcell. The goal is
that such a model can represent both single ller particle and multiple ller particles (as in Figure
5a-b). The power of the model will be that it can represent a subcell much larger than a single
ller particle, which leads to fewer homogenized subcells and thus a fast cell model. It is also a goal
that the model extension be rooted in the subcell mechanics such that phenomenological tting
parameters are avoided as much as possible.
To motivate such a model, two types of subcell structures were studied by DNS using the FEM.
The two types of structures are: a single ller particle and random particles. A subcell containing
a single particle can be viewed as the smallest useful subcell. A subcell containing random particles
can be viewed as a much largest subcell where clusters appear to be single points. Alternatively,
a subcell containing a single particle can be viewed as the most ordered and periodic structure
while a subcell containing random particles can be viewed as the most disordered and random.
11
Using the FEM model shown in Figure 5b, the peak value of tan for a number of ller volume
fractions are shown in Figure 6. The average response of 100 random subcells (see Figure 5a)
0 0.05 0.1 0.15 0.2 0.25 0.3
0
0.5
1
1.5
2
2.5
3
Volume Fraction
P
e
a
k

t
a
n

single particle
average of 100 random subcells
Figure 6: peak tan values for single particles subcells and the average peak tan values of 100 random particles
subcells (100 subcells per volume fraction)
for varying volume fractions are also plotted. It can be seen that the viscoelastic response diers
greatly between the random cluster and single particle subcells. The magnitude of the real stress
response for an random subcell and a single particle subcell are shown in Figure 7. This gure
highlights that there is a strong interaction between the two particles near the right-hand-side of
Figure 7b, represented by regions of high stress. The interaction is much weaker in Figure 7b
2
. A
model of the strong interaction in Figure 7b will be discussed in the next section. Based on Figure
7, along with the study of the response of many other subcells with random clusters, it is concluded
that the interaction between particles is a large contributor to the discrepancy between the random
2
the high stress region spanning to the edge in 7a is an interaction with the particle in the neighboring subcell,
due to periodic boundary conditions
12
Figure 7: Real normal stress in the y for a single particle and random particles both at 10% volume fraction. Both
contours are at a frequency of 1x10
2
(1/second). The light grey areas in the random structure are particles where
the stress contour has been removed for comparison of matrix stress.
cluster and single particle responses shown in Figure 6. Thus, the forthcoming model extension will
account for the level of particle interaction as a descriptor which has a strong eect on tan.
Turning now to the micromechanics models, two descriptors are shown to aect subcell response:
volume fraction and aspect ratio of the ller particles. Figure 8 shows tan predictions using both
the Mori-Tanaka and Halpin-Tsai models, in 2D, for a particle with aspect ratio of 1.0 and varying
volume fractions. It can be seen from Figure 8 that volume fraction has a strong eect on both
the peak value of tan and the frequency of the peak. The frequency of the peak moves to a lower
frequency range as volume fraction increases. Mori-Tanaka predictions of peak tan are slightly
higher at low volume fractions and lower at high volume fractions. Overall, it is shown that both
micromechanics models predict similar values for tan at a given volume fraction with a constant
aspect ratio. The aspect ratio is then varied for each model and the eect is shown in Figure 9. It
can be seen that each model shows dierent responses at high aspect ratios. Halpin-Tsai predicts a
double peaked tan at high aspect ratio while Mori-Tanaka predicts only a slight asymmetry due to
aspect ratio. Halpin-Tsai is more sensitive to aspect ratio than Mori-Tanaka. The type of clusters
considered in this work are relatively compact (A= 1 to 3); thus, the eect of aspect ratio on tan
will be smaller than that of volume fraction and particle interaction but will still be considered. To
conclude, the three main physical descriptors of the subcell, which will inform the micromechanics
model extension, are:
13
10
-6
10
-4
10
-2
10
0
10
2
0
0.5
1
1.5
2
2.5
Frequency, 1/second

t
a
n

Matrix
5% Filler
25% Filler
45% Filler
65% Filler
85% Filler
Mori-
Tanaka
Halpin-
Tsai
Figure 8: tan predictions for 2D systems with aspect ratios of 1.0 at varying volume fractions base on Halpin-
Tsai(red) and Mori-Tanaka(black) models
volume fraction of ller particles
percentage of ller particles interacting
aspect ratio of ller particles.
6. Micromechanics Model
To begin the development of a micromechanical model extension, it is rst observed that results
from the complex versions of Halpin-Tsai and Mori-Tanaka models agree quite well with DNS of a
10
-6
10
-4
10
-2
10
0
10
2
0
0.5
1
1.5
2
2.5
3
Frequency, 1/second

t
a
n

Halpin-Tsai Model
10
-6
10
-4
10
-2
10
0
10
2
0
0.5
1
1.5
2
2.5
3
Frequency, 1/second

t
a
n

Mori-Tanaka Model
Matrix
AR=1
AR=5
AR=10
AR=15
AR=20
AR=25
AR=30
Figure 9: tan predictions for 2D systems with volume fractions of 0.20 at varying aspect ratio base on Halpin-
Tsai(red) and Mori-Tanaka(black) models
14
single particle with volume fractions less than less than 30%, as shown in Figure 10. Thus, the model
0 0.05 0.1 0.15 0.2 0.25 0.3
0
0.5
1
1.5
2
2.5
3
Volume Fraction
P
e
a
k

t
a
n

DNS (single particle)
DNS (random particles)
Model (single particle HalpinTsai)
Model (random particle HalpinTsai)
Model (single particle MoriTanaka)
Model (random particle MoriTanaka)
Figure 10: comparison of DNS peak tan results to the predictions made using the model extension and the two
base models, Halpin-Tsai(red) and Mori-Tanaka(black)
extension will extend either Mori-Tanaka or Halpin-Tsai in such a way that if a subcell encloses
only a single cluster, then the model extension degrades to the base model (either Mori-Tanaka
or Halpin-Tsai). The second note on the forthcoming model extension is that it will not address
the exact mechanisms at every point in the subcell but will rather model the average response. To
that means, the model will carry less information than detailed simulations but also with much less
computationally expensive.
The extension of Mori-Tanaka or Halpin-Tsai starts with the average stress predicted by these
models,
b
. This will be called the stress of the base model (given by the superscript b). It is then
postulated that a region of the matrix has a higher stress than
b
, which is due to the interaction
of ller particles. The average stress in this interaction region will be denoted as,
int
. The
volume fraction of this elevated stress region is denoted as w
1
. The base and interaction stresses
15
are combined using the following rule-of-mixtures type breakdown of stress
= (1 w
1
)
b
+ w
1

int
. (10)
The next step is to formulate an expressions for
int
. To determine the interaction stress, it is
postulated that there is a region of the matrix with an elevated stress given by

int
= K

t

b
, (11)
where K

t
is the complex version of the traditional tensile stress concentration factor. Equation
11 assumes that each ller is surrounded by an innite domain of material and that this material
has the properties of
b
(as shown in Figure 11). Figure 11 shows two possible approximations

(a)
(b)
Figure 11: Two assumptions for the far eld properties used to determine inclusion response a) an Eshelby type
assumption b) a self-consistent type assumption which will be used in this work
for the matrix properties far from the inclusion of interest. Figure 11a assumes that these far eld
properties are that of the matrix, and thus the inclusion stress will be a function of E
m
. Figure 11b
assumes that the far eld properties are those of the composite, and thus the inclusion stress will be
a function of E
b
. The later is a self consistent type assumption. The assumption in Figure 11b will
be used for this work, as it takes into account surrounding microstructure. From this approximation,
K

t
is dened as the ratio of the peak matrix stress ,which occurs at the ller matrix interface for
16
dilute microstructure, to the stress far from the inclusions,
b
. The stress concentration factor is
expressed then as
K

t
=

peak
yy

b
yy
, (12)
where the y-direction is the loading direction. It is then noted that the peak stress occurs at the
largest y coordinate of the inclusion at the interface between the ller and matrix (point a seen in
Figure 12). At this point the force continuity condition holds; this condition [22] is expressed as,
Figure 12: Real normal stress in the y for a single particle in a dilute system (volume fraction of 1%). Contour is
for a frequency of 1x10
2
(1/second)

m
a+
nnn
a+
=
f
a
nnn
a
, (13)
where nnn is a unit normal vector, a+ and a represent points just above (+) and below() point a as
shown in Figure 13. It is also noted that in an Eshelby problem the inclusion stress is considered
constant, an assumption which will be adopted here. The validity of this assumption is illustrated
in both Figure 7a and Figure 12. Thus,

f
a
=
f
. (14)
Since nnn
a+
and nnn
a
are equal, substituting Equation 14 into Equation 13 and using only the
yy
components gives

m
yy(a+)
=
f
yy
. (15)
Since a+ is the location of peak yy stress,
m
yy(a+)
=
peak
yy
, and thus the complex stress concentration
factor, resulting from inserting Equation 15 into Equation 12, is given by
K

t
=

f
yy

b
yy
, (16)
17

Filler
Matrix

Figure 13: point a+ and a across matrix/ller interface at which continuity is considered
where both
f
yy
and
b
yy
are complex values. The real and imaginary parts of K

t
are plotted against
frequency as shown in Figure 14. This Figure shows that reasonable values of real K
t
between 1.0
and 2.0 are estimated. Note that the stress concentration factor for a rigid inclusion is 1.478 [24].
However for imaginary values of K
t
values between zero and negative numbers are estimated, both
of which are not physically consistent with the problem at-hand. The eect of the imaginary stress
concentration factor is thus neglected and K

t
becomes
K

t
= R{

f
yy

b
yy
) + 1i, (17)
where R() represents the real part of a function (). Therefore, the average stress can be partitioned
into real and imaginary parts as
R(
yy
) = (1 w
1
)R(
b
yy
) + w
1
R(K

t

b
yy
), (18)
I(
yy
) = I(
b
yy
), (19)
where I() represents the imaginary part of a function (). The stress real stress concentration
factor shown in Figure 14 and the assumption that imaginary stress concentration factor is unity,
agrees reasonably well with nite element calculations of K

t
for a low volume fraction, especially
at high frequencies. Overall, this model for K

t
captures the general trend seen from nite element
results, that the real part of K

t
exceeds the imaginary part across the frequency spectrum. The
18
10
6
10
4
10
2
10
0
10
2
0.5
0
0.5
1
1.5
2
Frequency, 1/second

K
t



Real
Imaginary
Figure 14: complex tensile stress concentration factor as a function of frequency for 20% volume fraction
ller stress, needed in Equation 16, is again determined using the self consistent assumption

f
=
inc,MT
(SSS, E
b
,
b
, E
f
,
f
), (20)
where thedenotes a function,
inc,MT
is the inclusion stress predicted by Mori-Tanaka, and E
b
and
b
the Youngs modulus and Poissons ratio of the composite predicted either by Mori-Tanaka
or Halpin-Tsai. Note that Mori-Tanaka is based on Eshelbys solution, which gives direct access to
inclusion (i.e. ller) stress and strain. Inclusion response in not readily provided by the Halpin-Tsai
model and thus Mori-Tanaka is used to determine ller stress regardless of the base model. This
is somewhat of an inconsistency for 2D analysis but is accepted with the purpose of including the
Halpin-Tsai model in this study, because it is the preferred model in many situations and by many
modelers. For 3D, the base model will always be Mori-Tanaka, remedying this inconsistency. Also
a general approximation is that v
b
= v
m
+ 0i, which is an assumption that is not strictly true but
generally results in negligible error.
19
7. Implementation
7.1. Subcell Model
To implement the model extension shown in Section 6, an average aspect ratio is determined
for each subcell as
A=
1
V
subcell
Nc

c=1
A
c
V
c
, (21)
where c is an index for each ller cluster, N
c
the number of ller clusters and V
subcell
and V
c
are
the volume of the subcell and each ller cluster respectively. The aspect ratio for each cluster A
c
is determined, in 2D, as
A
c
=
l
major
l
minor
, (22)
where l
major
is the length of the major axis of a ller cluster, and l
minor
is the length of the minor
axis. The major axis is considered to always be in the loading direction. These lengths are measured
as the dierence between the maximum and minimum coordinates of a cluster in each direction.
Also for implementation, the form of w
1
must be determined. Again, w
1
represents the volume
fraction of matrix with an average stress of K

t

b
yy
. This is in general dicult to determine exactly,
but it can be estimated as a function of the three main physical descriptors discussed above
w
1
w
1
(c
f
, , A), (23)
where is the percentage of ller particles interacting. In this work, only a simple form of w
1
will
be posed, but further extensions will be described in Section 10. One of the simplest forms is
w
1
= . (24)
The power of this assumption is that when only one ller particle is present, = 0 and the extended
model degrades to the base model. This makes physical sense, in that, for a single inclusion the
region where stress is K

t

b
yy
is only a single point (with a volume fraction of zero). Also w
1
in
Equation 24 increases as the number of particles interacting increases and so does the physical
volume with an elevated stress.
However, one detail of the assumption in Equation 24 needs to be further addressed. This detail
is that can take on a range of values as
R | 0 1. (25)
20
Since w
1
cannot exceed the volume fraction of the matrix, the average stress must be recast as
=

(1 w
1
)
b
+ w
1

int
, if w
1
(1 c
f
)
(1 c
f
)
int
+ c
f

f
, if w
1
> (1 c
f
)
(26)
where
f
is the ller/inclusion stress.
To determine , a 10 x 10 voxel subcell is considered. First, the number of clusters is determined
using the MATLAB function, BWLABEL, each cluster getting a label as shown in Figure 15. The
12
1 12
3 3
3 3 13
4 9
6 7 14
5 11 15
5
8 10
2
Figure 15: cluster labels for voxel subcell image
algorithm for searches through a list of lled voxels; if a lled voxel is within an integer distance,
dr, of another lled voxel, then the current voxel is considered to be interacting as shown by the
grey voxels in Figure 16. Then is determined as
=
number of lled voxels interacting
number of lled voxels
. (27)
A cluster interacting with itself, such as the merlon structure in Figure 16d, is considered interacting
since a high stress region will exist between merlons. Assuming a constant aspect ratio, subcells
of random and single particles are considered, where is calculated for each subcell. The results
are shown in Figure 10. It can be seen that the model captures both the single inclusion response
as well as the drop in tan due to interactions in random microstructures. It is also noted that
21
(a)
c
f
= 0.1; = 0.2
(b)
(c)
c
f
= 0.2; = 0.3
(d)
c
f
= 0.11; = 0.55
c
f
= 0.1; = 0.2
dr
Figure 16: a) an example of 10x10 subcells with 10% volume fraction, the grey particles are considered interacting,
dr = 2 voxels and the percentage of interacting particles, , is given b) the real normal stress at a frequency of 1x10
2
(1/second) in the y direction for the microstructure in a), the interaction stress between grey voxels is shown c) an
example of 10x10 subcells with 20% volume fraction d) self-interaction of a cluster with a merlon structure
dr was set to a value of 2 voxels for the results shown in Figure 10. The value of dr is considered
somewhat of a tting parameter, but, based on subcell DNS, a value of 2 or 3 seems reasonable.
Further nite element or molecular dynamics studies (such as those in [25] ) could determine more
exact values for dr.
7.2. RVE Cell Model
Using the micromechanics model extension outlined in the previous sections, a 2D RVE mi-
crostructure is considered as shown in Figure 3a. The microstructure was constructed using a 2D
kinetic Monte Carlo algorithm explained in Appendix A. This structure is meant to be somewhere
between a purely random structure and and an array of single inclusion subcells. The response
of both base models with the addition of the proposed model extension are shown in Figure 17.
It can be seen from Figure 17 that both the extended Halpin-Tsai and Mori-Tanaka models agree
well with DNS results. It can also be seen that the extended models are an improvement over the
22
prediction made by the original Halpin-Tsai model.
10
6
10
4
10
2
10
0
10
2
0
0.5
1
1.5
2
2.5
3
Frequency, 1/second
t
a
n

DNS
HT w/ extension
MT w/ extension
HT w/o extension
Matrix
Figure 17: results using the model extension are compared to DNS results for a 2D system. The response from the
original Halpin-Tsai model and pure matrix are shown for comparison
8. Three Dimensional Modeling
It it known that models of viscoelastic systems in 2D and 3D generally yield dierent results.
In some circumstances this discrepancy is due to dierent percolation limits in 2D and 3D. While
more frequently this discrepancy can be attributed to the fact that area fraction in 2D and volume
fraction in 3D are not equivalent. In plane strain analysis, as was presented previously, area fraction
and volume fraction were identical, but this is rarely the case for real microstructures. For example,
a circular inclusion with a xed radius in 2D may have an area fraction of 20%; a spherical inclusion
with the same radius in 3D will have a volume fraction of 6.7%. Thus the viscoelastic response
of the 2D system with an area fraction of 20% will resemble the response of a 3D system with a
volume fraction of 6.7%. The mapping between 2D area fraction and 3D volume fraction is straight
forward based on simple geometry for a circle and sphere respectively; however, a mapping of this
type for an arbitrary and complex microstructure would be very dicult to determine analytically.
23
Due to this diculty, the 2D model, which was useful for comparing to tractable DNS results, is
less useful for making predictions that can be compared to 3D experiential data; thus, a 3D model
is explored here. To cast the model implemented in Section 7 into a three dimensional form, a
3D microstructure was generated. The 3D microstructure was generated by randomly placing ller
particles in a 100 x 100 x 100 voxel matrix as is shown in Figure 18. While the previous 2D work
Figure 18: 3D random microstructure, black voxels are ller, white voxels matrix, inset shows mesh detail
extends to 3D without any modication to the proposed theory, a few implementation aspects need
claried in the context of 3D analysis. The rst is the base model itself. Mori-Tanaka is based on
the Eshelby tensor, which is formulated in 3D. Halpin-Tsai on the other hand relies solely on a 2D
aspect ratio. Thus, only the Mori-Tanaka model will be used in 3D. The form of the Eshelby tensor
used considers each inclusion to be a prolate spheroid with two equal minor axes and one longer
major axis and is given in [9]. The major axis was considered to be aligned with the load, as was
the case in 2D. To characterize this shape, the aspect ratio was determined as:
A=
l
major
mean|l
minor1
, l
minor2
|
, (28)
where l
major
is the length of the major axis of a ller cluster. The lengths l
minor1,2
represent the
length of each minor axis respectively. The aspect ratio of every cluster in a subcell was calculated
and volume-averaged as in Equation 22 in 2D.
24
The computational cost savings of the model extension become evident in 3D analysis. For
illustration purposes the author has run a 3D viscoelastic calculations for a 80 x 80 x 80 mesh which
considers 50 frequencies. Using 64 processors on the Quest computing cluster [26] this simulation
took approximately 15 hours. For the proposed model with a 100 x 100 x 100 microstructure
the simulation took approximately 5 minutes using 2 Intel Core i7-950, 3.07GHz processors.
Approximately half of this time was spent calculating homogenized properties using the extended
micromechanics model and the other half in the nite element calculation. The speedup is attributed
solely to the fact that the homogenization reduces the size of the nite element calculation. For the
100 x 100 x 100 microstructure only a 10 x 10 x 10 nite element mesh was needed using the proposed
homogenization procedure, resulting in substantial speedup. The proposed model extension has
substantially increased the speed of calculations of viscoelastic properties for 3D microstructures,
making the simulation of large RVE scale microstructure possible in 3D and smaller domain sized
calculations tractable without advanced computer resources.
9. Three Dimensional Modeling with Interphases
It has been shown in [2] [27] [25] [28] [29] [30] [31] [32] that the properties of many polymers
dier from that of the bulk matrix in regions near llers. These regions are referred to as interphase
regions. The eect of llers on matrix properties is predominant in nanocomposites where the
surface area is substantial [27]. The volume fraction and average thickness of the interphase region
has been investigated by the author using course grain molecular dynamics simulations for polymer
chains conned between non-attractive spherical nanoparticles [25]. It was found that the interphase
volume fraction can be expressed as a function of ller volume fraction as
c
IP
= (1 c
f
)|1 exp(3.07c
f
)|. (29)
It was also determined that the interphase thickness d
IP
can be directly estimated from c
IP
. For
relatively dilute ller clusters, the interphase thickness was expressed as
c
IP
L
3
sim
= M
NP
4|(R
F
+ d
IP
)
3
R
3
f
|]3 (30)
where L
sim
is the length of the side of a simulation box, M
NP
is the number of ller nanoparticles
and R
f
is the bare radius of the nanoparticles. Using Equation 29 the volume fraction of the
interphase region can be incorporated into an expression for average composite stress. The volume
25
fraction of the interphase will from here on be referred to as
2
. The 3D average composite stress
with an IP region is then expressed as
=

(1 w
1
)
b
+ (w
1
w
2
)
int
(LLL
b
) + w
2

int
(LLL
IP
), if w
1
w
2
(1 w
2
)
b
+ (w
1
)
int
(LLL
IP
) + (w
2
w
1
)
IP
(LLL
IP
), if w
2
> w
1
.
(31)
For simplicity it has been assumed that K
t
is the same as used for analysis without IP and that
neither w
1
or w
2
exceeds the volume fraction of the matrix (1c
f
). The expressions in Equation 31
assume that the IP and interaction regions are concentric. From Equation 31, the response of 3D
the system can be determined if the properties of the interphase region are known. The constitutive
expression for the IP region will be expressed as a volume-average over the IP region as
LLL
IP
=
1

IP

IP

LLL
IP
(xxx)d, (32)
where
IP
is the IP volume, and

LLL
IP
is a local quantity of interphase material properties as a
function of distance xxx from the ller surface.
Two simple models for the interphase region are examined here. Figure 19 shows the variation
in peak E

and E

using a linear and exponential transition between the ller and matrix properties
over the IP region. The expressions for E

with a linear and exponential transition over the width

0 0.2 0.4 0.6 0.8 1
4
5
6
7
8
9
10
x 10
8
normalized IP distance (x/d
IP
)

M
a
x
i
m
u
m

I
P

S
t
o
r
a
g
e

M
o
d
u
l
u
s
,

P
a



linear
exponential
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
x 10
8
normalized IP distance (x/d
IP
)

M
a
x
i
m
u
m

I
P

L
o
s
s

M
o
d
u
l
u
s
,

P
a

Figure 19: Linear and exponential transition of IP properties over the IP thickness
of the interphase are given receptively in
E

IP
= E
f
+ (E

m
E
f
) x, (33)
E

IP
= E
f
_
E

m
E
f
_
x
, (34)
26
where x is a normalized distance, x]d
IP
. Equation 32 was evaluated using using both Equation
33 and 34 for

LLL
IP
and used in a 3D simulation of tan for a random microstructure. The results
are shown in Figure 20. These results are also compared to the tan of the IP region alone, as
10
6
10
4
10
2
10
0
10
2
0
0.5
1
1.5
Frequency, 1/second

t
a
n

composite w/o IP
composite w/ linear IP
composite w/ exponetial IP
linear IP
exponential IP
Figure 20: Eect of an IP region on tan for a 3D random microstructure. Properties of the IP region are shown for
reference
indicated by dashed lines in Figure 20. It is shown in Figure 20, that the presence of an IP region
reduces tan and that the linear IP region reduces tan more than the exponential region.
The goal of the 3D model is to make predictions which can be translated to materials design.
Thus, a brief study of the IP thickness is shown here. This is by no means the only variable which
can be explored but showcases the type of parametric studies possible in a short timespan with
a fast model. For cases where the IP properties vary exponentially (i.e. Equation 34), the IP
distance was calculated for each subcell. The addition of ller reduces tan and the presence of an
exponential IP region further decreases tan response. As was shown in [2], it is often desirable
to have tan as high as possible in certain frequency ranges to maximize performance of systems
such as tire materials. Thus, it would also be desirable to know how much d
IP
should be reduced
to see a substantial increase in tan. Figure 21 shows the response of peak tan to decreases in IP
thickness. For this study, d
IP
is determined for each subcell using Equation 30, then reduced by
varying amounts. Exponentially varying IP properties were also used for the study in Figure 21.
The average IP thickness was d
I
P = 0.13 voxels, which showcases that the model can account for a
non-integer voxel IP thicknesses. It can be seen that the eect of IP distance on peak tan becomes
nonlinear at approximately 80% of the predicted d
IP
. However, it is shown that to increase peak
27
0 20 40 60 80 100
0.4
0.6
0.8
1
1.2
1.4
% of predicted IP thickness
p
e
a
k

t
a
n

Figure 21: peak tan as a function of decreasing IP thickness; IP properties are exponentially varying across the IP
region
tan to within 20 % of the value without an interphase region would require an 80% reduction in
d
IP
(i.e. 20% of d
IP
). A detailed consideration of the processing and functionalization of the ller
particles would reveal if this reduction is possible.
Another study was also performed showing the eect of shifting the frequency of tan in the IP
region. All tan values were uniformly shifted a given number of decades. The results are shown
in Figure 24. It can be seen that a shifted IP with no reduction in tan amplitude results in little
change in composite properties. In the 2D work of Deng et al. [20], a larger eect of shifting
IP properties was shown. They studied a minimum IP volume fraction of 24.4% where as in the
current study the average IP volume fraction is 10.8% which could account for the reduced eect.
From these two studies, the following conclusions can be made. First, the eect of IP thickness is
nonlinear; the eects of reducing IP thickness are amplied for large percentage reductions. Second,
the IP thickness would need to be reduced 80% to achieve 80% of the composite properties seen
without an IP region. Third, if the IP region is less lossy than the matrix, then the maximum value
of tan for a given volume fraction is shown in Figure 20 as the tan with no IP. Further increase
28
10
6
10
4
10
2
10
0
10
2
0
0.5
1
1.5
Frequency, 1/second

t
a
n

Interphase shift, in decades
0
3
2
1
+1
+2
+3
Figure 22: Eect of shifting the frequencies of IP properties on 3D tan responses. Numbers in the legend represent
the number of decades IP tan was shifted, with (+) being a shift to higher frequencies
in tan would require changing variables other than IP thickness. Fourth, a linear gradient in
the IP region reduces tan more than an exponential gradient, and fth, amplitude changes in IP
tan have much more of an eect on composite response than frequency shifts, for low IP volume
fractions of 10% or less. These results would hopefully give materials designers some guidance
into how to modify the interphase properties, or alternatively inform designers as to the possibility
of increasing tan by means of IP property changes, perhaps a lower volume fraction would be
required to meet design goals.
It should be noted that the IP properties used here are very simple, and much more complex
forms exist. For example, the author has chosen the IP properties to vary from that of the ller
to that of the matrix. However, it is known that there will be a jump in properties between ller
and matrix. It was not the goal of this work to characterize this jump, and thus the author uses
the simple form shown in Figure 19 as an example. It also can be shown that the IP properties are
a function of not only the distance from a ller but also the eects of ller-ller interactions such
that

LLL
IP
=

LLL
IP
(xxx, ). This dependence could easily be included in the proposed model extension
29
(as is available) but was not considered here. Another use of the proposed model is to leave
LLL
IP
as an unknown and compare it to experimental data. From this comparison the value of LLL
IP
,
which showed good agreement between experimental results and the model, could be considered
the appropriate value of interphase properties. Thus, the proposed model could be used to back
out interphase properties of a polymer, (as was done in [20]) which may be dicult to access exper-
imentally, much like researchers in MicroElectroMechanical Systems (MEMS) back out microscale
materials properties based on the response of micro-beams. If the determined properties vary for
dierent region of the microstructure using

LLL
IP
=

LLL
IP
(xxx), then a dependence of IP properties on a
ller-ller interaction is evident, and the form

LLL
IP
=

LLL
IP
(xxx, ) should be considered. It is possible
that the appropriate form of

LLL
IP
(xxx, ) could be backed out as well. The determination of this
dependence would increase knowledge of the polymer system, which showcases the usefulness of the
proposed model extension not only for design but also for discovery.
10. Summary and Conclusions
A model extension to the traditional Mori-Tanaka and Halpin-Tsai equations was proposed
here. This extension captured eects of stress risers near interacting ller particles to allow for
larger subcells when homogenizing an RVE scale cell. The model extension agreed well with DNS
calculations in 2D for subcells with a medium number of ller clusters. 3D predictions showed
the eect of interphase regions and showcased the substantial increase in computational speed over
direction simulations.
The form of the interaction volume expression was based on a very simple formulation. Future
work may express this term based on more detailed microstructural metics such as, two point
correlation function or surface area of the ller clusters. This work was meant to characterize tan,
which can be determine from a one dimensional (1D) loading. Thus, the expressions for average
stress were reduces to 1D as well, which greatly simplied the determination of stress concentration
factor. If three dimensional properties are of interest, a more sophisticated method would be needed
to determine K
t
for 3D load states. Recent work by the author and his collaborators in generalized
continuum mechanics [33] [34] could oer a method to determine 3D K
t
, where extra degrees of
freedom account for subcell scale stress states.
Overall, the hope is that the model extension proposed here will allow materials designers to
explore complex elastomeric systems in more detail, resulting in better, safer and more ecient
30
materials and discoveries of complex micro or nanoscale mechanism, which were previous out of the
scope of current modeling procedure.
Acknowledgement
The support of this research by National Science Foundation (NSF) is gratefully acknowledged.
We would also like to thank Brendan Abberton and Wylie Stroberg for their contribution of the
kinetic Monte Carlo code.
Appendix A. Kinetic Monte Carlo
The 2D microstructure was generated using a Monte Carlo simulation based on the 2D kinetic
Ising model. The details of a 2D Ising model are given in [35]. The model initializes a voxelized
system with a given ller volume fraction. The Monte Carlo method will exchange matrix and ller
voxels in such a way that volume fraction is conserved. The model calculates the energy change, E,
due to exchanging a voxel. If the E < 0, then the change is accepted. If E > 0, then the exchange
is accepted based on the probability P = exp(E]k
b
T), where k
b
is Boltzmanns constant and T
the system temperature. The energy is calculated as E = J
<i,j>
S
i
S
j
where < i, j > is the sum
over the nearest neighbor voxels i, j and S
i,j
are spins mapped from ller concentration, C
f
. This
mapping is S
i
= 2C
f
i
1. The concentration is determined by C(T) = C
c
|1|sinh(2J]k
b
T)|

4|
18
where J is approximated as J = k
n
T]0.4407 and C
c
a critical concentration.
The aforementioned Monte Carlo simulation will not necessarily represent the physics of a lled
elastomer system; however, it will give a microstructure which is less ordered then periodic single
particles but more morphologically diverse than randomly distributed ller voxels. Ideally, this
microstructure will be determined based on imaging as was performed by Deng et al. [20], but this
type of imaging was beyond this scope of this work.
References
[1] V. Mittal, Optimization of polymer nanocomposite properties, Wiley. com, 2009.
[2] M.-J. Wang, Eect of polymer-ller and ller-ller interactions on dynamic properties of lled
vulcanizates, Rubber Chemistry and Technology 71 (3) (1998) 520589.
31
[3] I. D. Aiken, J. M. Kelly, F. F. Tajirian, Mechanics of low shape factor elastomeric seismic
isolation bearings, Vol. 89, Earthquake Engineering Research Center, College of Engineering,
University of California, 1989.
[4] C. W. Roeder, J. F. Stanton, Elastomeric bearings: state-of-the-art, Journal of Structural
Engineering 109 (12) (1983) 28532871.
[5] T. V. Duncan, Applications of nanotechnology in food packaging and food safety: Barrier
materials, antimicrobials and sensors, Journal of colloid and interface science 363 (1) (2011)
124.
[6] M. T, T. K, Average stress in matrix and average elastic energy of materials with mistting
inclusions, acta metallurgica 21 (1973) 571.
[7] A. JE, J. Halpin, Primer on composite materials: analysis, Technomic, 1969.
[8] J. D. Eshelby, The determination of the elastic eld of an ellipsoidal inclusion, and related
problems, Proceedings of the Royal Society of London. Series A. Mathematical and Physical
Sciences 241 (1226) (1957) 376396.
[9] J. Qu, M. Cherkaoui, Fundamentals of Micromechanics of Solids, John Wiley & Sons, Inc,
2006.
[10] C. L. Tucker III, E. Liang, Stiness predictions for unidirectional short-ber composites: review
and evaluation, Composites science and technology 59 (5) (1999) 655671.
[11] R. Hill, Theory of mechanical properties of bre-strengthened materials: I. elastic behaviour,
Journal of the Mechanics and Physics of Solids 12 (4) (1964) 199212.
[12] L. Brinson, W. Lin, Comparison of micromechanics methods for eective properties of multi-
phase viscoelastic composites, Composite Structures 41 (3) (1998) 353367.
[13] K. Putz, R. Krishnamoorti, P. F. Green, The role of interfacial interactions in the dynamic
mechanical response of functionalized swntps nanocomposites, Polymer 48 (12) (2007) 3540
3545.
[14] X. L. Ji, J. K. Jing, W. Jiang, B. Z. Jiang, Tensile modulus of polymer nanocomposites,
Polymer Engineering & Science 42 (5) (2002) 983993.
32
[15] G. Li, Y. Zhao, S.-S. Pang, Four-phase sphere modeling of eective bulk modulus of concrete,
Cement and Concrete Research 29 (6) (1999) 839845.
[16] A. A. Gusev, S. A. Lurie, Loss amplication eect in multiphase materials with viscoelastic
interfaces, Macromolecules 42 (14) (2009) 53725377.
[17] H. Liu, L. C. Brinson, A hybrid numerical-analytical method for modeling the viscoelastic
properties of polymer nanocomposites, Journal of applied mechanics 73 (5) (2006) 758768.
[18] Y. Song, Q. Zheng, Application of two phase model to linear viscoelasticity of reinforced
rubbers, Polymer 52 (3) (2011) 593596.
[19] L. Mishnaevsky Jr, Micromechanical analysis of nanocomposites using 3D voxel based material
model, Composites Science and Technology 72 (10) (2012) 11671177.
[20] H. Deng, Y. Liu, D. Gai, D. A. Dikin, K. W. Putz, W. Chen, L. Catherine Brinson, C. Burkhart,
M. Poldne, B. Jiang, et al., Utilizing real and statistically reconstructed microstructures for
the viscoelastic modeling of polymer nanocomposites, Composites Science and Technology
72 (14) (2012) 17251732.
[21] C. M. Breneman, L. C. Brinson, L. S. Schadler, B. Natarajan, M. Krein, K. Wu,
L. Morkowchuk, Y. Li, H. Deng, H. Xu, Stalking the materials genome: A data-driven ap-
proach to the virtual design of nanostructured polymers, Advanced Functional Materials.
[22] T. Belytschko, B. Moran, W. K. Liu, K. Elkhodary, Nonlinear nite element analysis for
continua and structures, 2nd Edition, Wiley, in press.
[23] D. Simulia, Abaqus 6.11 theory manual, Providence, RI, USA: DS SIMULIA Corp.
[24] P. WD, Petersons stress concentration factors, New York: Wiley, 1997.
[25] Y. Li, M. Kroger, Dynamic structure of unentangled polymer chains in the vicinity of non-
attractive nanoparticles, Soft Matter.
[26] Quest, Northwestern University, http://www.it.northwestern.edu/research/adv-
research/hpc/quest/.
33
[27] T. Ramanathan, A. Abdala, S. Stankovich, D. Dikin, M. Herrera-Alonso, R. Piner, D. Adam-
son, H. Schniepp, X. Chen, R. Ruo, et al., Functionalized graphene sheets for polymer
nanocomposites, Nature Nanotechnology 3 (6) (2008) 327331.
[28] K. I. Winey, R. A. Vaia, Polymer nanocomposites, MRS bulletin 32 (04) (2007) 314322.
[29] P. Akcora, S. K. Kumar, V. Garca Sakai, Y. Li, B. C. Benicewicz, L. S. Schadler, Segmental
dynamics in pmma-grafted nanoparticle composites, Macromolecules 43 (19) (2010) 82758281.
[30] G. N. Toepperwein, N. C. Karayiannis, R. A. Riggleman, M. Krger, J. J. de Pablo, Inuence
of nanorod inclusions on structure and primitive path network of polymer nanocomposites at
equilibrium and under deformation, Macromolecules 44 (4) (2011) 10341045.
[31] J. Torres, P. Nealey, J. De Pablo, Molecular simulation of ultrathin polymeric lms near the
glass transition, Physical Review Letters 85 (15) (2000) 3221.
[32] Y. Li, M. Kroger, W. K. Liu, Nanoparticle eect on the dynamics of polymer chains and their
entanglement network, Physical Review Letters 109 (11) (2012) 118001.
[33] K. I. Elkhodary, M. Steven Greene, S. Tang, T. Belytschko, W. K. Liu, Archetype-blending
continuum theory, Computer Methods in Applied Mechanics and Engineering.
[34] M. S. Greene, Mechanics and physics of solids, uncertainty, and the archetype-genome exem-
plar, Ph.D. thesis, Northwestern University (2012).
[35] L. D. Landau, E. M. Lifshitz, Statistical Physics, Part 1, 3rd Edition, Vol. 5, Butterworth-
Heinemann, 1980.
34