You are on page 1of 12

J. Non-Newtonian Fluid Mech.

156 (2009) 165–176

Contents lists available at ScienceDirect

Journal of Non-Newtonian Fluid Mechanics


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

An anisotropic rotary diffusion model for fiber


orientation in short- and long-fiber thermoplastics
Jay H. Phelps, Charles L. Tucker III ∗
Department of Mechanical Science and Engineering, University of Illinois at Urbana-Champaign, 1206 West Green Street,
Urbana, IL 61801, United States

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

Article history: The Folgar–Tucker model, which is widely-used to predict fiber orientation in injection-molded compos-
Received 9 May 2008 ites, accounts for fiber–fiber interactions using isotropic rotary diffusion. However, this model does not
Received in revised form 14 August 2008 match all aspects of experimental fiber orientation data, especially for composites with long discontin-
Accepted 14 August 2008
uous fibers. This paper develops a fiber orientation model that incorporates anisotropic rotary diffusion.
From kinetic theory we derive the evolution equation for the second-order orientation tensor, correcting
Keywords:
some errors in earlier treatments. The diffusivity is assumed to depend on a second-order space tensor,
Fiber orientation
which is taken to be a function of the orientation state and the rate of deformation. Model parameters
Rotary diffusion
Long-fiber thermoplastics
are selected by matching the experimental steady-state orientation in simple shear flow, and by requiring
Injection molding stable steady states and physically realizable solutions. Also, concentrated fiber suspensions align more
slowly with respect to strain than models based on Jeffery’s equation, and we incorporate this behavior
in an objective way. The final model is suitable for use in mold filling and other flow simulations, and it
gives improved predictions of fiber orientation for injection molded long-fiber composites.
© 2008 Elsevier B.V. All rights reserved.

1. Introduction PDF for tens of thousands of nodes in an injection molding sim-


ulation is prohibitively expensive. Recasting the model in terms of
There is considerable value in being able to predict the fiber the second-order moment tensor of the orientation distribution [3]
orientation state in a flowing fiber suspension. Example uses are makes the calculation affordable for engineering purposes.
modeling the processing of fiber-reinforced plastics, and models of In the years since its introduction, there have been few major
paper making. An accurate treatment of fiber orientation is also an modifications to the Folgar–Tucker model. Experimental evidence
essential element in any rheological model of a fiber suspension. has shown that the kinetics of orientation change in concentrated
For non-Brownian fibers, the model proposed by Folgar and suspensions is much slower than Jeffery-based models predict
Tucker [1] has been widely used. This phenomenological model [4,5], and an objective model to account for this has recently been
uses Jeffery’s single-fiber equation to represent flow effects [2], proposed [6]. That model primarily modifies the Jeffery terms in
and adds an isotropic rotary diffusion term, with a diffusivity pro- the model. Another type of modification, which is the subject of this
portional to the scalar rate of deformation, to represent fiber–fiber paper, concerns the rotary diffusion term. Rotary diffusion models
interactions. Adding the rotary diffusion term gives steady orien- are interesting for both practical and theoretical reasons.
tation states that are not perfectly aligned, which is consistent On the practical side, the Folgar–Tucker model has proved useful
with experimental results. Rotary diffusion is proportional to the for modeling fiber orientation in short-fiber/thermoplastic com-
interaction coefficient CI , a scalar parameter that is adjusted to fit posites (SFTs) that are processed by injection molding, e.g. [7,8].
experimental data. However, for long-fiber thermoplastics (LFTs) the Folgar–Tucker
The Folgar–Tucker model was originally formulated as a model is less accurate. LFT pellets are usually prepared by a pultru-
Fokker–Planck type equation for the probability density function sion process, and the pellets typically have 10–13 mm long fibers
(PDF) of fiber orientation. However, numerically calculating the aligned along their length. This is a contrast to SFTs, where the fiber
length is typically 0.2–0.4 mm. When LFTs are injection molded
they develop fiber orientation patterns that are qualitatively similar
∗ Corresponding author. Tel.: +1 217 244 3822. to SFT orientation patterns, but quantitatively they achieve lower
E-mail addresses: jphelps@illinois.edu (J.H. Phelps), ctucker@illinois.edu flow-direction alignment than SFTs. Accurately predicting the fiber
(C.L. Tucker III). orientation in LFT composites requires a better model for fiber ori-

0377-0257/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.jnnfm.2008.08.002
166 J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176

entation, and it appears that modifying the rotary diffusion term is a general and concise way to describe orientation. This tensor is
could provide the requisite model behavior. given by
On the theoretical side, one modification to the rotary diffu- 
sion term was proposed by Ranganathan and Advani [9]. They A= pp dp, (1)
suggested that diffusivity should be dependent on the orientation

state, being higher when the fibers are randomly oriented and lower where dp denotes an integral over all possible fiber orientations,
when the fibers are aligned. They proposed a phenomenological and pp is the tensor product of the fiber orientation vector p with
model in which the interaction coefficient is inversely proportional itself.
to the average interfiber spacing. Because their rotary diffusion is The primary impetus for using A instead of to describe ori-
isotropic, their model gives the same steady-state distributions of entation comes from the numerical simulation of fiber orientation
fiber orientation as the Folgar–Tucker model; only the transient in complex flows. It is computationally more efficient to track the
approach to steady state is affected. Also, their model reaches an components of A over multiple time steps and many spatial nodes,
upper limit in volume fraction and aspect ratio when the average than it is to track the distribution function over the same dis-
interfiber spacing becomes negative. LFT materials fall well beyond cretization.
this limit, so Ranganathan and Advani’s model cannot be applied
to LFT materials. 2.1. Folgar–Tucker fiber orientation model
A different modification of rotary diffusion was proposed by
Fan et al. [10] and Phan-Thien et al. [11] who performed direct Turning attention to modeling the evolution of orientation, we
numerical simulations of multiple interacting fibers in a concen- first review the concept of orientation space. Since p is a unit vec-
trated suspension undergoing simple shear flow. They could not tor, the space of all possible orientations (the orientation space) is
fit their calculated steady-state orientation distribution using the the surface of a unit sphere. The probability density function is
Folgar–Tucker model, so they developed a rotary diffusion model in defined on this space. The probability density is conserved (as
which the scalar interaction coefficient CI was replaced by a second- the total probability is constant), and the probability density func-
order tensor C. This makes the rotary diffusion anisotropic. Fan and tion must satisfy a continuity equation at any time t. This continuity
Phan-Thien developed a moment-tensor equation for their model, equation is given by
and found values of the C tensor to fit their steady orientation states.
d
Significantly, they did not explore the dynamic properties of their = −∇ s · ( ṗ), (2)
dt
moment-tensor equation. We build on their approach in this work.
Koch [12] developed a detailed, mechanistic model of long-range where s represents the gradient operator on the surface of the unit
hydrodynamic interactions in semi-concentrated fiber suspen- sphere, or the gradient operator in orientation space. The probabil-
sions, and from this produced a model for the rotary diffusion. ity flux term ṗ may be decomposed into a hydrodynamic term
h
Like the Folgar–Tucker model, the diffusivity scales with the rate of ṗ and a diffusive flux vector q, giving
deformation. However, in Koch’s model the diffusivity varies with d h
the orientation state, and is anisotropic. No moment-tensor form = −∇ s · ( ṗ + q). (3)
dt
of this model was developed, so the properties of this model have
Typically, the hydrodynamic contribution is modeled by Jeffery’s
been largely unexplored.
equation for particle motion in a dilute suspension [2], or
In this paper we explore anisotropic rotary diffusion models
for fiber orientation. The treatment is phenomenological. In Sec- h
ṗ = W · p + (D · p − D : ppp), (4)
tion 2, after reviewing some background material, we compare
predictions of the Folgar–Tucker model to fiber orientation mea- where W = (1/2)(L − LT ) is the vorticity tensor, D = (1/2)(L − LT ) is the
surements in molded LFT samples, showing what types of model rate-of-deformation tensor, and  is the particle shape parameter
improvements are needed for LFTs. We then examine the Fan and (for long, slender fibers  → 1). L represents the velocity gradi-
Phan-Thien model, showing that it fails a simple test for correct ent tensor of the dilute suspension, with components Lij = ∂vi /∂xj
diffusive behavior, and the Koch model, showing that it has lim- where vi represents the component of the velocity in the xi direc-
ited usefulness for modeling LFTs. With these items as motivation, tion.
in Section 3 we derive the correct moment-tensor form for an In order to model the diffusion term appearing in Eq. (3), Folgar
anisotropic rotary diffusion (ARD) model that is characterized by a and Tucker [1] suggest the phenomenological relationship:
second-order tensor. The dynamic behavior of this model proves to q = −CI ˙ ∇ s , (5)
be very sensitive to the choice of model parameters, so we present
a systematic procedure for choosing robust parameter sets. Finally, where CI is the fiber–fiber interaction coefficient (a fitting param-
in Section 4 we show that the new ARD model gives improved fiber eter), and ˙ = (2D : D)1/2 is the scalar magnitude of D. In this
orientation predictions for LFT materials. The paper closes with a model, Folgar and Tucker scale q with , ˙ assuming that the rate
short discussion. of fiber–fiber collisions is proportional to ,˙ and the orientation
perturbation per collision is independent of .
˙
Advani and Tucker [3] utilize Eqs. (1) and (3)–(5) to develop a
2. Background and motivation time evolution equation for A. This is the standard Folgar–Tucker
model, and is given by
The orientation of a single, straight fiber can be characterized h d
Ȧ = Ȧ + Ȧ , (6)
by a unit vector p directed along the fiber axis. For collections of
fibers, the complete description of orientation is a probability den- where the hydrodynamic contribution is
sity function (p) where the probability of a fiber being oriented h
between p and p + dp equals (p) dp. Experimental measurements Ȧ = (W · A − A · W) + (D · A + A · D − 2A : D), (7)
of orientation are samples drawn from the distribution function . and the diffusive contribution is
A useful way to report the data is to compute the average values d
of certain functions of p. The second-order orientation tensor A [3] Ȧ = 2CI (I
˙ − 3A). (8)
J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176 167

In this model, Ȧ represents the material derivative of A, I is the model, and it is given by
identity tensor, and A is the fourth-order orientation tensor, defined
d 
3
as h
= −∇ s · [ ṗ − CI ˙ ∇ s + wi (ei ei · p − ei ei : ppp)], (15)
 dt
i=1
A= pppp dp. (9)
h
where ṗ is Jeffery’s equation for particle motion. The term with the
In practice, A is approximated as a closure function of compo- summation represents an orientation-dependent probability flux.
nents of A. The present work utilizes the ORE orthotropic closure. The scalar coefficients wi are calculated as
The eigenvalue-based orthotropic closure family was introduced by 5(1 − )
Cintra and Tucker [13], and the coefficients for the ORE version are wi = − [(i D : ei ei − ei ei : A : D) + CI (1
˙ − 3i )], (16)
4
given by VerWeyst [14].
d where i and ei are once again the eigenvalues and eigenvectors
The Ȧ term approximately accounts for the fiber–fiber interac- of the orientation tensor A. This kinetic theory RSC model corre-
tions in a concentrated suspension. This interaction term is a rotary sponds to Eq. (12) in the same manner that Eq. (3) corresponds to
diffusion term. It acts isotropically, providing a flux that pulls fibers the Folgar–Tucker moment-tensor model of Eqs. (6)–(8).
toward a random state. While the RSC-equivalent kinetic theory model does contain the
moment tensors A and A, these tensors may be evaluated at each
2.2. The RSC fiber orientation model time step from their definitions, so no closure approximation is
needed to solve Eq. (15) [6]. For the remainder of this work, we
Recent experience has shown that the rate of orientation devel- return to the moment-tensor form of the RSC model.
opment in SFT materials is much slower than this model predicts
[4,5]. Wang et al. [6] introduce the reduced-strain closure (RSC) 2.3. Application to LFT injection molding
model, which slows the orientation kinetics in an objective fash-
ion, in order to achieve better agreement between experiment and The Folgar–Tucker model family has been implemented in
prediction. almost all commercial software programs for injection molding.
To develop the RSC model, Wang et al. [6] write the sym- For research purposes we have made a similar implementation in a
metric orientation tensor as a function of its eigenvalues i and research software package known as ORIENT. This software and its
3
eigenvectors ei , A =  e e . This expression may then be sub- implementation have a structure similar to commercial injection
i=1 i i i
stituted into the Folgar–Tucker model, and the material derivatives molding codes. A detailed description of the theory and numerical
of both the eigenvalues ˙ i and the eigenvectors ėi may be iso- methods behind this program is found in Bay and Tucker [7]. A brief
lated. The authors then make the phenomenological assumption summary follows here.
that the eigenvalue kinetics for the RSC model are slowed by a ORIENT is a finite difference program that calculates the veloc-
factor of , compared to the eigenvalue kinetics derived from the ity field and second-order orientation tensor A during mold filling
Folgar–Tucker model, while the expressions for the eigenvector for either an end-gated strip or a center-gated disk. ORIENT also
kinetics remain unchanged. This is given by takes into account non-isothermal conditions and solves for the
temperature field within the mold. ORIENT uses the Hele–Shaw
˙ RSC
i
= ˙ FT
i
, (10) approximation [15] for solving for the velocity field in mold-filling
operations, and accounts for the dependence of viscosity on both
and
strain rate and temperature.
RSC
ėi = ėi .
FT
(11) In solving Eq. (6) or (12), ORIENT needs a boundary condition
for orientation at the mold inlet. Typically, experimental data taken
RSC near the inlet is used as a boundary condition for each orientation
Using Eqs. (10) and (11), the tensor material derivative Ȧ may be
re-constructed. The result is the RSC model, given by tensor component. ORIENT accepts values of , CI , and  as input.
The outputs of ORIENT include vx (flow direction velocity), T, , ˙
RSC
Ȧ = (W · A − A · W) + {D · A + A · D and A as functions of x (flow direction coordinate) and z (thick-
ness direction coordinate), and pressure as a function of x, all at
− 2[A + (1 − )(L − M : A)] : D} + 2CI (I
˙ − 3A). (12) each time step during filling. Typically, we only use the data for
the last time step, when the cavity is full. There is no post-filling
The fourth-order tensors L and M are analytical functions of the
analysis.
eigenvalues and eigenvectors of A, and are given by
To obtain experimental LFT data against which to compare

3 ORIENT calculations, a series of glass-reinforced (40 wt%) speci-
L= i (ei ei ei ei ), (13) mens with polypropylene matrices were molded using a range
i=1
of injection speeds. Two geometries were considered: a 90 mm
long × 80 mm wide × 3 mm thick end-gated strip, also called an
and ISO plaque [16], and a 180 mm diameter, 3 mm thick center-gated
disk. After molding, the number-average fiber length (L̄n ) was

3

M= ei ei ei ei . (14) 2.866 mm, and the weight-average fiber length (L̄w ) was 5.167 mm,
for a slow-filled ISO plaque specimen [17]. Experimental orien-
I=1
tation measurements were performed on polished cross-sections
The  parameter, which slows the orientation kinetics, must be cho- using the image analysis methods developed by Hine et al. [18]. Fur-
sen to fit experimental data. Typical values of  for SFT materials ther details of these moldings and orientation measurements can
range from 0.05 to 0.2 [6]. be found in Nguyen et al. [17]. We summarize some key findings
Although the RSC model was developed at the level of a here.
moment-tensor equation, there is an equivalent model at the level Let the A11 orientation direction correspond to the flow direction
of kinetic theory. Wang et al. [6] also derive such a kinetic theory during mold filling, while the A22 and A33 components correspond
168 J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176

steady-state orientation in simple shear flow. CI is usually chosen


to match A11 in this region. Typical CI parameter values for SFT com-
posites are 0.006–0.01; the best-fit value for a given data set will
depend on which closure approximation is used in the model.
For the LFT composites studied here,  and CI were set to 1/30
and 0.03, respectively, in order to fit the experimental values of A11 .
Compared to SFTs, this is an unusually large value of CI . In Fig. 1(a),
note that the A11 component is predicted with a reasonable level of
accuracy, while in Fig. 1(b) the A22 and A33 components are under-
and over-predicted, respectively. The same trends were observed
for other filling speeds, for both ISO plaque and disk geometries.
More importantly, Nguyen et al. [17] showed that micromechani-
cal predictions of material properties based on the predicted fiber
orientation distributions do not adequately capture the mechanical
anisotropy of the solid parts, while predictions using the measured
fiber orientation distributions do.
With a single, scalar CI parameter available to model the
fiber–fiber interactions, the Folgar–Tucker model does a good job
of predicting only one of the orientation components in our LFT
moldings. We hypothesize that a model which can anisotropically
model fiber–fiber interactions will be better able to predict all of
the orientation components Aij . This should be especially true for
LFT composites. In order to anisotropically model these interactions
one must introduce a global tensorial representation of the inter-
action (rotary diffusion) term. Previous attempts at such a model
will now be reviewed.

2.4. Previous ARD models

2.4.1. The Koch model


Koch [12] considered the influence of long-range hydrodynamic
fiber–fiber interactions on orientation in semi-dilute fiber sus-
pensions. Koch developed a mechanistic model for orientational
Fig. 1. Second-order orientation tensor components A11 (a), A22 , and A33 (b) for a diffusivity, and derived an expression for a spatial tensor C such
3 mm thick, slow-filled, long-glass-reinforced ISO plaque 90 mm long × 80 mm wide. that C˙ represents rotary diffusion. Using the notation of our work,
The ORIENT predictions are performed with both the standard Folgar–Tucker model this tensor is given by
(CI = 0.03) and the RSC variant ( = 1/30 and CI = 0.03).
nL3
C= [ˇ1 (D : A : D)I + ˇ2 D : A : D], (17)
˙ 2 ln2 ˛
to the cross-flow and thickness directions, respectively. Fig. 1(a)
and (b) show both the predicted and experimental values of A11 , where n is the number of fibers per unit volume, L is the fiber
A22 , and A33 versus the non-dimensional thickness coordinate z/h length, and ˛ = L/d is the fiber aspect ratio. While C is a spatial ten-
for a 3 mm thick, slow-filled, glass-reinforced ISO plaque at a posi- sor, rotary diffusion only operates in orientation space, and Koch
tion approximately halfway down the length of the channel. In points out that “only the components of the tensor tangent to
these plots, we observe a region near the center of the mold- the unit sphere have physical significance.” The model assumes
ing where the A11 component has a low value. This is referred a semi-dilute suspension, and so is expected to be valid for nL2 d
to as the core. The regions to either side of the core, with A11 less than about 3. In terms of fiber volume fraction vf , this require-
values ranging from approximately 0.65 to 0.7, are referred to as ment is vf < 3/4˛. Koch obtains values of ˇ1 = 3.16 × 10−3 and
the shell. Some moldings also show thin skin layers near the sur- ˇ2 = 1.13 × 10−1 by fitting the model to theoretical orientation dis-
face, where the fibers are less aligned in the flow direction than tributions in extensional flows.
in the shell. In general, LFT composites exhibit wider cores and In Eq. (17), A is the sixth-order orientation tensor, defined as
lower flow-direction (A11 ) alignment in the shell than their SFT 
counterparts. A= pppppp dp. (18)
Fig. 1(a) and (b) also highlight the need for the RSC model. The
standard Folgar–Tucker model predicts a very narrow core. How- In our work, where we seek an equation for Ȧ, the sixth-order ten-
ever, by slowing the orientation kinetics with the RSC model, the sor must be approximated from the second-order tensor using a
core remains wide, consistent with the experimental data. While closure approximation. Our work utilizes the invariant-based INV6
the need for this model is illustrated here with LFT composites, the closure approximation of Jack and Smith [19].
same treatment is also needed with SFTs [4]. In order to explore the effectiveness of this model, the orienta-
The value of CI is typically chosen to fit experimental data d
tion tensor C may be incorporated into a model for Ȧ , the diffusive
by considering the steady-state simple shear flow solution to the d
Folgar–Tucker or RSC model. Away from the mold inlet, the shell- contribution of Ȧ. We develop a general expression for Ȧ in Section
region flow closely approximates a simple shear flow, and fibers 3.1 of this paper. For Koch’s model, this is
in the shell have experienced very large shear strains during mold d
filling. Thus, we expect the orientation in the shell to match the Ȧ = [2C
˙ − 2(tr C)A − 5(C · A + A · C) + 10A : C]. (19)
J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176 169

2.4.2. The Fan and Phan-Thien model


d
Phan-Thien et al. [11] propose replacing the Ȧ term in Eq. (8)
with
d
Ȧ = [2C
˙ − 2 tr(C)A − 3(C · A + A · C) + 6A : C], (20)

where C is a spatial tensor describing fiber–fiber interactions. The


hydrodynamic term from Eq. (7) is unaffected. This model does sat-
isfy the symmetry requirement on Ȧ, and Phan-Thien et al. [11]
corrected an earlier problem from Fan et al. [10] so that the model
maintains tr A = 1. However, our own preliminary calculations with
Eq. (20) produced unrealistic dynamic behavior, and we discovered
that this model suffers from a major flaw.
When fibers are isotropically distributed, the second-order ori-
entation tensor A takes on the exact value of
1
A= I, (21)
Fig. 2. Dynamic solution for three cases of the Koch model in simple shear flow. 3
L = 1000 ␮m and d = 17 ␮m.
and the components of the fourth-order orientation tensor A are
given exactly by
Combining Eqs. (6), (7), (17) and (19) gives a complete description 1
Aijkl = (ı ı + ıik ıjl + ıil ıjk ), (22)
of orientation kinetics utilizing Koch’s anisotropic rotary diffusion 15 ij kl
model.
where ıij is the Kronecker delta. Introducing these tensors into Eq.
Fig. 2 shows the orientation evolution of the A11 and A33 tensor
(20) gives
components in simple shear flow (v1 = x ˙ 3 , where we use ˙ = 1)
for the Koch model, for three examples of fiber volume fraction d 4 4
Ȧ = C− (tr C)I. (23)
in simple shear flow. In each case, L = 1000 ␮m and d = 17 ␮m, giv- 5 15
ing ˛ = 58.8, which is representative of LFT fiber dimensions after d
injection molding [17]. vf = 0.19 corresponds to the volume fraction Thus, for an isotropic distribution of fibers, Ȧ = / 0 for a general
typically seen in LFT composites. C. In other words, when the fibers are randomly distributed, the
One can compare the simple shear flow results at steady state diffusion term of this model would pull the orientation away
to the experimental data shown in Fig. 1. In Fig. 1, the shell region from isotropy. Diffusion provides a flux that is proportional to a
(which is dominated by simple shear flow) of the LFTs exhibits gradient—in this case a gradient of . For an isotropic orientation
experimental A11 and A33 components of approximately 0.65 and state the gradient ∂ /∂p is zero, so diffusive flux must be zero, and
d
0.01, respectively. In Fig. 2, however, the Koch model for vf = 0.19 Ȧ must be zero. The model of Phan-Thien et al. [11] does not have
gives A11 and A33 equal to approximately 0.43 and 0.23, respec- this property, so it must be incorrect.
tively, at steady state. This result is overly diffusive, as the A11 and Further examining the derivation of Phan-Thien et al. [11], we
A33 components do not change significantly from their initial con- conclude that their projection of C onto the orientation space was
dition (A = (1/3)I). However, corresponds to nL2 d = 14.23, which is incorrect. They correctly note that the rotary diffusion tensor C
well outside of the semi-dilute regime for which Koch’s model was should be projected onto a surface normal to the orientation vec-
developed. If we reduce vf to 0.05, the Koch model does match the tor p. In projecting the tensor C, they left-multiply it by the tensor
target of A11 = 0.65 at steady state, but it gives A33 = 0.10, which is (I − pp). However, the correct projection involves both a left- and
almost identical to the Folgar–Tucker model. This suggests that the a right-multiplication by (I − pp), and the right-multiplication is
ˇ1 term in Eq. (17), which is an isotropic diffusivity whose magni- omitted from their derivation.
tude varies with orientation state, is dominating the steady-state Rotary diffusion can neither translate a fiber nor change a fiber’s
behavior compared to the anisotropic ˇ2 term. length. Thus, diffusion is defined only on the unit spherical surface
Fig. 2 also gives the Koch model orientation evolution for a fiber traced by all possible fiber orientations, and not in the direction
volume fraction of vf = 0.01 (nL2 d = 0.749). In this case, the A11 com- that the fiber is pointing. Rotary diffusion for fiber orientation is
ponent rises above unity, while the A33 component falls below zero, properly defined two-dimensionally in the orientation space. The
results that are non-physical. The C tensor given by the Koch model derivation of Phan-Thien et al. [11] does not draw a clear distinction
will always have positive eigenvalues, which should guarantee a between spatial quantities (such as p and C) and surface quanti-
physically plausible solution for y(p,t). However, we have imple- ties in the orientation space (such as s , q, or ). We correct this
mented Koch’s model in an equation for Ȧ, which requires closure problem below by developing a model that incorporates a proper
approximations for both the fourth-order and sixth-order orienta- two-dimensional understanding of rotary diffusion.
tion tensors. The non-physical results at low vf are almost certainly
caused by the closure approximations for A or A. These errors might 3. Theory
be alleviated by using other closures, though the ORE closure for A
and the INV6 closure for A are among the best available closures to 3.1. Model development
date.
Because of the need to calculate the sixth-order tensor, the In order to properly introduce anisotropic rotary diffusion,
Koch model is computationally much more expensive than the we consider a two-dimensional rotary diffusivity surface tensor,
Folgar–Tucker model, yet in the best case it provides no better denoted Dr and defined in surface spherical coordinates. The diffu-
match to experimental data. These observations indicate that the sive flux vector q for the ARD model is then
Koch model provides no advantage for predicting fiber orientation
in LFT composites. ˙ r · ∇s ,
q = −D (24)
170 J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176

where s is the surface gradient operator defined in the orientation The second integral may then be re-arranged and expanded such
space. Here, we retain the assumption of Folgar and Tucker that that
q ∝ .˙ 
Ȧdij = −˙ (∇ s ) · [(tT Bt) · (∇ s pi pj )] dp
Although diffusion is properly recognized as two-dimensional
on the surface of a unit sphere, the orientation tensor whose evo-  
= −˙ ∇ s · [ (tT Bt) · ∇ s pi pj ]dp+˙ ∇ s · [(tT Bt)·∇ s pi pj ] dp
lution we are attempting to model is a space tensor, expressed in
global Cartesian spatial coordinates. Some method is needed to pass

= ˙ ∇ s · [(tT Bt) · ∇ s pi pj ] dp.
diffusion information between these two coordinate systems. To do (32)
this, we define Dr as the projection of a three-dimensional space
tensor B onto the local surface coordinates of the unit sphere. This The periodicity argument once again applies to the first integral in
projection of B is given by the second line, reducing it to zero.
The surface gradient operator can be written in surface spherical
Dr = tT Bt, (25)
coordinates as
where, using spherical coordinates for the orientation space, t is ∂ 1 ∂
the 3 × 2 matrix, e.g. [20]: ∇s = e + e , (33)
∂  sin  ∂
 
cos  cos −sin where e and e are unit vectors in the local  and directions
t= cos  sin cos . (26)
respectively. With t, B (via its dependence on p), and the compo-
−sin  0 nents of p all trigonometric function of  and , Eq. (33) may be
The tensor B could presumably change with orientation p, so we substituted into Eq. (32). The components of C which enter into
write it as a function of a symmetric global tensor C and a tensorial the resulting equation may be factored outside the integral, leav-
representation of orientation, pp. Hand [21] gives the representa- ing only trigonometric functions of  and inside the integral. The
tion theorem for the ways that B can depend on C and pp. This is remaining trigonometric functions inside the integral can be re-
arranged so that, upon integrating over all p, only components of
B = [a1 + a2 tr C + a3 tr C2 + p · (a4 C + a5 C2 ) · p]I the A tensor remain. These multiply components of the C tensor
that had previously been factored out of the integral.
+ [a6 + a7 tr C + a8 p · C · p]C + a9 C2 . (27)
From that point, an extensive algebraic exercise yields each term
Each ai could, most generally, be a scalar function of the joint of the rotary diffusion portion of the fiber orientation tensor evo-
invariants of C and pp. However, to keep the analysis tractable, we lution model. Reconstruction of the tensor representation gives
impose the constraint that B be no more than quadratic in C, which
d
constrains each ai parameter to be a constant. ˙ 1 + a2 tr C + a3 tr C2 )(2I − 6A) + a4 [2(A
Ȧ = (a ˙ : C)I
Using Eqs. (3), (24), (25) and (27), we can write the rotary dif-
+ 2(C · A + A · C) − 10A : C] + a5 [2(A
˙ : C2 )I + 2(C2 · A + A · C2 )
fusion portion of the distribution function’s continuity equation
as − 10A : C2 ] + (a
˙ 6 + a7 tr C)[2C − 2(tr C)A − 5(C · A + A · C)
 d
d + 10A : C] + a8 {2C
˙ tr (A · C) + 2(C2 · A + A · C2 )
= ˙ ∇ s · [(tT Bt) · ∇ s ]. (28)
dt
− 2(tr C)(A : C) − 4A : C2 − 7[C · (A : C) + (A : C) · C]
Multiplying each side of this equation by the tensor pp, integrating + 14C : A : C} + a9 [2C
˙ 2
− 2(tr C2 )A − 5(C2 · A + A · C2 )
over all orientations p, and factoring the time derivative out of the
left-hand side yields: + 10A : C2 ]. (34)

 d  h
d d For the entire orientation tensor evolution equation, Ȧ = Ȧ +
pp dp = Ȧ = ˙ {∇ s · [(tT Bt) · ∇ s ]}pp dp. (29) d h
dt Ȧ , the hydrodynamic term Ȧ remains unchanged from the
Folgar–Tucker model. Note that setting a1 = CI and a2–9 = 0 recov-
Here, note that the quantity inside the braces on the right-hand
ers the Folgar–Tucker model, and that setting a6 = 1 and a1–5,7–9 = 0
side is a scalar, the surface divergence of a surface flux vector q.
recovers a corrected form of the Phan-Thien model, Eq. (20). More-
This scalar multiplies the space tensor pp. Although pp is spatial, its
over, this model satisfies the symmetry requirements on Ȧ (with
components pi pj are scalar functions of the spherical coordinates  d
d the assumption that C is symmetric) and also gives Ȧ = 0 for a
and . Thus, Ȧ may be examined component-by-component, using random distribution of fibers (A = (1/3)I), for all choices of a1–9 and
the scalar equation: all choices of C.
 Equation (34) is our most general form of the ARD model. How-
Ȧdij = ˙ pi pj ∇ s · [(tT Bt) · ∇ s ] dp. (30) ever, owing to our earlier decision to constrain each ai parameter to
be a scalar constant in order to preserve tractability in the deriva-
Expanding the surface divergence via the product rule gives tion of the model, this is not the most general model. By treating
 the ai parameters as joint invariants of the C and pp tensors, a host
Ȧdij = ˙ ∇ s · [pi pj (tT Bt) · ∇ s ] dp of models may be developed using this derivation template. How-
ever, we refrain from developing any more terms for this model,
 and instead focus our attention on simplifying the model at hand
− ˙ (∇ s pi pj ) · [(tT Bt) · ∇ s ] dp. (31) to make it useful for applications.
Preliminary numerical experiments indicated that not all of the
Note that (and its gradient), the components of p, and the surface ai terms from Eq. (34) were useful in predicting fiber orientation,
diffusion tensor Dr = tT Bt are all periodic over the orientation space, and we chose to retain only the a6 term. While this is an arbitrary
so the first integral evaluates to zero. With B (and by extension Dr ) choice, we can argue for it heuristically as follows. First, note that
symmetric, we can re-order the dot product in the second integral. the a1–3 terms retain the same form as the isotropic Folgar–Tucker
J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176 171

model. Although the C tensor can affect these terms as a contribu-


tion to their leading scalar factor, these terms remain expressions
for isotropic rotary diffusion. From an application standpoint, these
terms are equivalent to a Folgar–Tucker model with an adjustable
CI parameter and can only reach the same steady orientation states
as Folgar–Tucker. Thus, we do not further consider the a1–3 terms.
Next, the remaining a5 and a7–9 terms are quadratic with respect
to C, as opposed to the a4 and a6 terms, which are linear with respect
to C. Solving for the components of C which best fit experimental
orientation data is decidedly more difficult with these quadratic
terms than with the linear terms. Thus, we set the a5 and a7–9 terms
aside as a matter of convenience.
Finally, by adjusting the remaining, linear, a4 and a6 parameters
relative to one another, a host of models that exhibit anisotropic
rotary diffusion and that are linear in the C tensor may be devel-
oped. However, the kernel of the a4 term, given by Eq. (27), is
(p·C·p)I. Thus, the a4 term allows the local rate of diffusion to
Fig. 3. Transient orientation kinetics in simple shear flow using the ARD model with
depend on orientation. However, at any single orientation p, rotary bi = (2.0257, −26.9707, 43.0677, 1.1676 × 10−5 , −7.1671).
diffusion remains locally isotropic. The a6 term, though, exhibits
true local anisotropy. By allowing C itself to be a function of
orientation, we may then also introduce orientation dependent Together, Eqs. (36) and (37) form what is referred to in the remain-
changes in the degree of isotropy. This choice is equivalent to der of this work as the ARD fiber orientation model. This model’s
setting B = a6 C in Eq. (27). We thus discard the a4 term, and focus equivalent kinetic theory model is
attention on the a6 term.
d h
Thus far, we have imposed no restrictions on the values that the = −∇ s · [ ṗ − (t
˙ T Ct) · ∇ s ], (38)
components of C may take. For the overall model to be objective, C dt
must be a function of other objective tensors that appear in Eq. (34). where t is defined in Eq. (26).
We could create a quite general objective model by requiring C = Just as the RSC model of Eq. (12) gives an objective means to
C(D, A, A, A). The Koch model, Eq. (17), is an example of this. How- slow the orientation kinetics of the Folgar–Tucker model of Eq. (6),
ever, again for simplicity, we further assume that C = C(D, A). The one can also develop an ARD-RSC model which slows the kinetics
representation theorem once again comes from Hand [21], where of Eq. (37). Applying the RSC treatment embodied in Eqs. (10) and
(11) to Eq. (37) gives
b4 b5 b6
C = b1 I + b2 A + b3 A2 + D + 2 D2 + (D · A + A · D) ARD,RSC
˙ ˙ ˙ Ȧ = (W · A − A · W) + {D · A + A · D − 2[A
b7 b8 + (1 − )(L − M : A)] : D} + {2[C
˙ − (1 − )M : C]
+ (D · A2 + A2 · D) + 2 (D2 · A + A · D2 )
˙ ˙
− 2(trC)A − 5(C · A + A · C)
b9 2 2
+ (D · A + A2 · D2 ). (35) + 10[A + (1 − )(L − M : A)] : C}. (39)
˙ 2
d Note that setting  equal to 1 reduces the ARD-RSC model to the ARD
This model ensures the objectivity of C, and also makes Ȧ propor- model. The ARD-RSC model is the model whose properties we will
tional to ,
˙ consistent with the Folgar–Tucker model. Once again, explore. An effective technique for choosing the b1–5 parameters
each scalar bi could most generally be a function of the nine joint will now be introduced.
invariants of A and D. However, to keep the model tractable, we
constrain the bi coefficients to be scalar constants and set b6–9 = 0, 3.2. Model parameter selection
using:
b4 b5 Properly assigning values to the bi fitting parameters is a
C = b1 I + b2 A + b3 A2 + D + 2 D2 . (36) critical step in implementing the ARD model. The consequences
˙ ˙
of poor parameter selection are shown in Fig. 3. Here, the
As with the parameters of the general ARD model, our scalar orientation approaches a reasonable-looking steady state and
constant constraint on the parameters of the definition of the C then falls into an oscillatory pattern, indicative of a dynamic
tensor simplifies the model for our applications. instability in the model. Such behavior is clearly undesir-
The second choice to set b6–9 equal to zero simplifies the pro- able.
cess of fitting the bi parameters to experimental data. The choice to The process of parameter selection first considers the ability
eliminate the terms with dependencies on products of A and D was of the model to produce the desired final orientation state in
pragmatic, and no theoretical consideration was given to including a molded part. In the injection molding process, the flow very
the other terms. The selection of the remaining b1–5 parameters near the upper and lower mold walls closely approximates a sim-
will be covered in Secttion 3.2. ple shear flow (SSF). Moreover, the strains accumulated there by
Returning to our ARD fiber orientation model in which we have the end of the filling are very large. Thus, the fiber orientation
eliminated all but the a6 term, one may now set a6 equal to unity state close to the walls should be the steady state in simple shear
without loss of generality, since C remains adjustable via the b1–5 flow.
parameters. This simplifies our general ARD model to Phan-Thien et al. [11] choose the components of C in order to fit
a desired steady-state orientation in simple shear flow. However,
ȦARD = (W · A − A · W) + (D · A − A · D − 2A : D)
our C tensor in Eq. (36) is a function of A and D in order to retain
+ [2C
˙ − 2(tr C)A − 5(C · A + A · C) + 10A : C]. (37) objectivity in our model for Ȧ. Thus, we adopt a similar technique
172 J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176

Fig. 4. (a) Parameter search for ARD-RSC model in simple shear flow. (♦) A solution that both is stable and has positive steady-state eigenvalues for C, () a stable solution
with negative eigenvalues in C, and () an unstable steady state. Here, the target SSF steady-state orientation matrix is given in Eq. (41). (b-d) Stable solutions with positive
steady-state eigenvalues of C are successively analyzed for PEF, UEF, and BEF flows. Physically realizable solutions that give positive eigenvalues of C solutions are marked
with a plus (+) symbol, and advance to the next flow under consideration. Physically realizable solutions that do not meet the eigenvalue of C requirement are marked
with a mid-dot (·) symbol, and non-physical solutions are denoted with a times (×) symbol. (d) The valid solution located farthest from the boundary of invalid solutions is
additionally marked with a oplus (⊕) symbol.

of fitting the bi parameters to a desired steady-state orientation For simple shear flow, the target orientation and velocity
tensor. gradient tensors give Ȧ23 = Ȧ12 = 0 at all times. Also, with the
Once the model has the desired steady-state behavior, we con- requirement that tr A = 1 (and, by extension, tr Ȧ = 0), one of the
sider its dynamic properties. These include stability of the steady equations for the three diagonal components of Ȧ is redundant.
state, positive eigenvalues in the C tensor, and physically plausible This leaves three independent equations, and five unknown model
transient solutions. Each of these criteria will now be discussed in parameters. Thus, one can choose any two of the bi parameters
detail. independently, and then solve directly for the remaining three. We
choose b3 and b5 as the independent parameters, and solve for b1 ,
3.2.1. Matching steady-state orientation data b2 , and b4 to match the target steady-state orientation in simple
In a 1–3 simple shear flow, the velocity gradient tensor L is given shear flow. The necessary linear equations are derived in Appendix
by A. This method is guaranteed to yield a set of parameters for either
  the ARD or ARD-RSC model in which the target orientation tensor
0 0 1
is a steady state of the chosen model in simple shear flow. How-
LSSF = 0 0 0 . (40)
ever, depending on the values b3 and b5 , the model may or may not
0 0 0
behave in other desirable ways, for either the simple shear flow or
For the LFT samples examined in our experiments, the orientation for other flows.
data near the walls is well represented by
  3.2.2. Stability
0.65 0 0.03
Fig. 3 showed the oscillatory pattern into which an ARD model
ASSF = 0 0.34 0 . (41)
solution may settle when the fitting parameters are not carefully
0.03 0 0.01
chosen. Note that, at around time t = 8, the model appears to be
With target values for the components of A and a given flow (i.e., settling into a nearly steady solution, before beginning to oscillate.
a given L tensor), the only unknown quantities in the ARD or ARD- This is consistent with a saddle point instability.
RSC models are the bi parameters (and  in the case of the ARD-RSC In order to eliminate instabilities such as those observed in Fig. 3,
model) and the components of Ȧ. If the targeted orientation tensor we perform a linear stability analysis of the model at the desired
A is a steady-state solution, the value of the components of Ȧ are steady state. In a 1–3 simple shear flow, the independent com-
all then zero, leaving a set of algebraic equations that is linear in bi . ponents of the fiber orientation tensor are A11 , A33 , and A31 . We
J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176 173

numerically calculate the derivatives of the independent compo-


nents of Ȧ with respect to the independent components of A at the
steady state, giving
⎡ ⎤
∂Ȧ11 ∂Ȧ11 ∂Ȧ11
⎢ ∂A11 ∂A33 ∂A31 ⎥
⎢ ⎥
∂Ȧ ⎢ ∂Ȧ33 ∂Ȧ33 ∂Ȧ33 ⎥
=⎢ ⎥. (42)
∂A ⎢ ∂A11 ∂A33 ∂A31 ⎥
⎢ ⎥
⎣ ⎦
∂Ȧ31 ∂Ȧ31 ∂Ȧ31
∂A11 ∂A33 ∂A31
For a stable steady state, the eigenvalues of this tensor must all be
negative.

3.2.3. Eigenvalues of diffusivity tensor


Next, one should consider that negative diffusivity in the 2D ori-
entation space has no physical meaning. Thus, the two eigenvalues
Fig. 5. Dynamic solution to the ARD-RSC model in simple shear flow with  = 1/30,
of Dr must be positive. Given the definition of Dr as a projection of and the bi parameter set given in Eq. (46).
B, ensuring that B (and by extension C) has positive eigenvalues will
ensure that Dr will also have positive eigenvalues. This check on C is
made at every time step during a dynamic orientation calculation
for each flow of interest. can be inserted into the model, and the validity of the model for
the other flows may be assessed. As previously “good” solutions
3.2.4. Validity in other flows fail to meet these increasing criteria, they are discarded from fur-
Finally, the model’s behavior under other flow conditions must ther consideration. Eventually we find a family of parameter sets
be assessed. Although there is usually no experimental data from that satisfies all of the criteria above.
other flows against which to compare orientation results, we want We choose to systematically select b3 and b5 , and linearly solve
the solutions generated to be physically valid. The criteria for for b1 , b2 , and b4 for each (b3 , b5 ) pair. The ability of each solution
validity are: 0 ≤ A11 , A22 , A33 ≤ 1 and −0.5 ≤ A23 , A31 , A12 ≤ 0.5. We to meet the stability and positive eigenvalue of C requirements in
consider transient orientation behavior in planar elongational flow simple shear is then recorded. It is convenient to select a grid of
(PEF), uniaxial elongational flow (UEF), and biaxial elongational (b3 , b5 ) values, and display the ability of each solution to meet the
flow (BEF). Their respective velocity gradient tensors are SSF criteria in a plot in (b3 , b5 ) space. The ability of the parameter
  sets to meet the criteria of physically plausible solutions and pos-
1 0 0 itive eigenvalues of C in the extensional flows can be displayed in
LPEF = 0 −1 0 , (43)
subsequent plots.
0 0 0 Fig. 4(a) gives a criteria plot for the SSF solution given the target A
  of Eq. (41). Here we have utilized the ARD-RSC model with  = 1/30,
1 0 0
L UEF
= 0 −0.5 0 , (44) as our initial mold-filling simulations with the RSC model indicated
0 0 −0.5 a significant slowing of the orientation kinetics in LFT materials.
There are many solutions that meet the stability and C eigenvalue
and criteria for simple shear. The different elongational flows (Fig. 4(b)
 
0.5 0 0 and (d)) gradually eliminate some of these b3 –b5 combinations, but
BEF
L = 0 0.5 0 . (45) do leave multiple solutions that meet all of the previously described
0 0 −1 criteria for all of the flows. To select a single parameter set from
the remaining valid solutions (Fig. 4(d)), we implement a boundary
3.3. Searching for satisfactory model parameters erosion technique common in image processing to isolate the point
furthest from the boundary of the region of valid solutions. This
Summarizing all of the criteria, a valid set of parameters bi must point is also indicated in Fig. 4(d).
give a solution to either the ARD or ARD-RSC model in which Fig. 5 gives the orientation kinetics in simple shear flow for the
chosen parameter set, which is
• the steady, simple shear flow solution is equal to the target ori-
entation tensor, bi =(1.924 × 10−4 , 5.839 × 10−3 , 4.0 × 10−2 , 1.168 × 10−5 , 0).
• the steady, simple shear flow solution is linearly stable,
• the C tensor has positive eigenvalues at all times in all flows, and (46)
• the transient solution is physically valid in simple shear, as well
as planar, uniaxial, and biaxial elongation flows. The initial condition in Fig. 5 is A = (1/3)I, a random arrangement
of fibers. The evolution of the A23 and A12 components are not
Stability is not explicitly verified in the extensional flows. How- reported, as they remain at their initial zero values. The orientation
ever, we have not observed any instabilities in any of the extensional kinetics of all of the valid bi solution sets are quite similar, so Fig. 5
flows for any parameter set. is representative of this entire family of solutions for the ARD-RSC
If the choice of the two independent bi parameters is done sys- model with  = 1/30 and the target steady state of Eq. (41). Suit-
tematically, one can utilize the stability criteria and the positive able parameter sets for other target orientation states can be found
eigenvalue of C criteria to develop a family of solutions which sat- using the same procedure. This ARD-RSC model gives the type of
isfies the first three described solution criteria for simple shear flow. behavior we desire for incorporation into a mold-filling simulation
Then, using these same parameters, a new velocity gradient tensor for LFT materials.
174 J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176

5. Conclusion

Adding anisotropic rotary diffusion (ARD) to a Jeffery-type fiber


orientation model offers greatly improved ability to match exper-
imental data, compared to the isotropic Folgar–Tucker model.
In this paper we have derived the moment-tensor equation for
an anisotropic rotary diffusion that is described by a second-
order space tensor, correcting an error in the equation given by
Phan-Thien et al. [11]. Combining this equation with a closure
approximation for the sixth-order orientation tensor allowed us
to explore the model of Koch [12]. That model was found to offer
no improvement over the Folgar–Tucker form, at least for modeling
long-fiber thermoplastic composites. Moreover, at low fiber volume
fractions our implementation of Koch’s model gave non-physical
results, probably induced by the closure approximation.
Proceeding with a phenomenological model, we allowed the
rotary diffusion tensor to depend on fiber orientation state and on
the rate-of-deformation tensor. The resulting model has five scalar
parameters: three are chosen to match a target steady-state orien-
tation in simple shear flow, and the other two are chosen to ensure a
stable steady state, positive eigenvalues in the diffusion tensor, and
physically valid results in various elongational flows. We show that
the new model provides improved predictions of fiber orientation
in injection molded long-fiber thermoplastic composites.
While long-fiber thermoplastics provided the practical moti-
vation for this work, our anisotropic rotary diffusion model may
have other applications as well. Fan et al. [10] and Phan-Thien et
al. [11] found that their direct numerical simulation results for
semi-concentrated fiber suspensions could not be matched by the
Folgar–Tucker model. Our ARD model can easily be fitted to this
type of result using the parameter-selection procedure shown here.
For short-fiber reinforced thermoplastics, the Folgar–Tucker model
can match experimental rheology data for the shear viscosity, but
Fig. 6. Second-order orientation tensor components A11 (a), A22 , and A33 (b) for a in some cases cannot simultaneously match the normal stress data
3 mm thick, slow-filled, glass-reinforced ISO plaque 90 mm long × 80 mm wide. The [6]. Because the ARD model offers the ability to tailor all compo-
ORIENT predictions use the RSC variant of the Folgar–Tucker model with  = 1/30 and nents of the orientation tensor at steady state, it should provide a
CI = 0.03, while the ORIENT-ARD predictions use the ARD-RSC model with  = 1/30
better match to rheological data for fiber-reinforced materials.
and the bi parameter set given in Eq. (46).
Our model for the rotary diffusion tensor, Eq. (36), is purely
phenomenological. It would be very helpful to have a mechanis-
4. Application to LFT injection molding tic model that would give the rotary diffusion as a function of local
orientation, fiber orientation state, and type of deformation. When
The ARD model family was implemented in ORIENT to predict such models are developed, the framework of this paper will allow
the development of orientation during mold filling in LFT compos- those models to be implemented in moment-tensor form, so that
ites. Using the ARD-RSC model with  = 1/30 and the bi parameter their properties can be explored and their usefulness determined.
set given in Eq. (46), we compared the predicted orientation tensor
components to the experimental components of the same spec- Acknowledgements
imen and the same location that was examined in Fig. 1. This
comparison is shown in Fig. 6. The work presented here was supported by the Department of
In this comparison, the A11 data is reasonably well predicted Energy through the Pacific Northwest National Laboratory. Com-
by both the RSC and ARD-RSC models, as seen in Fig. 6(a). How- ments and suggestions from Drs. Mark T. Smith, Ba Nghiep Nguyen,
ever, using the new ARD-RSC fiber orientation model (Fig. 6(b)), and James D. Holbery have been most helpful. The experimental ori-
the A22 and A33 components are much better predicted than before entation measurements made by Vlastimil Kunc of the Oak Ridge
(Fig. 1(b)). There is still a poor fit near z/h = 0 in both the A11 and A22 National Laboratory were integral to this work, and we are grateful
comparisons, which seems due to asymmetry in the experimental for his input.
data (ORIENT assumes symmetry about the midplane z = 0). This
asymmetry could come from mold wall temperature differences,
Appendix A. Derivation of linear equations for bi
resulting in asymmetric cooling of the polymer melt, or from the
parameters
fact the gate in this cavity is not located symmetrically about that
cavity midplane.
Consider the ARD-RSC model, Eq. (39), and the definition of the
Finally, using the orientation distribution predicted by the ARD-
global, spatial rotary diffusion tensor C in Eq. (36). Combining these
RSC model to predict elastic moduli in the flow and crossflow
two equations gives
directions gives results that compare favorably with the measured
moduli [17,22]. This is an improvement compared to the agree- ARD,RSC
Ȧ = (W · A − A · W) + (D · A + A · D − 2[A + (1 − )(L
ment observed when the RSC model was used to predict the fiber
orientation distribution. − M : A)] : D) + b1 (2[I
˙ − (1 − )M : I] − 2(tr I)A
J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176 175

− 5(I · A + A · I) + 10[A + (1 − )(L − M : A)] : I) (51) by choosing the rows corresponding to Ȧ11 , Ȧ33 , and Ȧ31 . We
write this as
+ b2 (2[A
˙ − (1 − )M : A] − 2(tr A)A − 5(A · A · A · A) ⎧ ⎫
⎪ b1 ⎪
+ 10[A + (1 − )(L − M : A)] : A) + b3 (2[A
˙ 2

⎨ b2 ⎪

− (1 − )M : A2 ] − 2(tr A2 )A − 5(A2 · A + A · A2 ) [B̃] b3 = {H̃}, (52)

⎪ ⎪
⎩ b4 ⎪

+ 10[A + (1 − )(L − M : A)] : A2 ) b5
+ b4 (2[D − (1 − )M : D] − 2(tr D)A − 5(D · A + A · D) where [B̃] is a 3 × 5 matrix, and {H̃} is a 3 × 1 column vector.
b5 Now, five bi parameters must satisfy three equations. We choose
+ 10[A + (1 − )(L − M : A)] : D) + (2[D2 b3 and b5 as independent variables, and modify Eq. (52) such that

 
b1
− (1 − )M : D2 ] − 2(tr D2 )A − 5(D2 · A + A · D2 )
[B̂] b2 = {Ĥ}, (53)
+ 10[A + (1 − )(L − M : A)] : D2 ), (47) b4

which is linear in bi . Noting that M : I = I, L : I = A, A : I = A, M : where [B̂] is a 3 × 3 matrix, and {Ĥ} is the 3 × 1 column vector with
A = A, L : A = A2 , tr A = 1, tr I = 3, and tr D = 0, the model can be sim- components:
plified, leaving Ĥi = H̃i − B̃i3 b3 − B̃i5 b5 , (54)
ARD,RSC
Ȧ = (W · A − A · W) + (D · A + A · D − 2[A where i ranges from 1 to 3.
Following through with the algebra, the components of the [B̂]
+ (1 − )(L − M : A)] : D) + 2b1 (I
˙ − 3A) matrix and the {Ĥ} vector are
+ 10b2 (A
˙ : A − A2 ) + 2b3 (A
˙ 2
− (tr A2 )A B̂11 = 2(1
˙ − 3A11 ), (55)
2 3
+ 5A : A − 5A ) + b4 (2[D − (1 − )M : D] B̂21 = 2(1
˙ − 3A33 ), (56)
− 5(D · A + A · D) + 10[A + (1 − )(L − M : A)] : D) B̂31 = −6A
˙ 31 , (57)
b5 B̂12 = 10(A
˙ 11ij Aij − A1i Ai1 ), (58)
+ (2[D2 − (1 − )M : D2 ] − 2(tr D2 )A − 5(D2 · A

B̂22 = 10(A
˙ 33ij Aij − A3i Ai3 ), (59)
+ A · D2 ) + 10[A + (1 − )(L − M : A)] : D2 ). (48)
B̂32 = 10(A
˙ 31ij Aij − A3i Ai1 ), (60)
We now utilize a notation in which the components of a symmetric
ARD,RSC
tensor are written as a column vector. For Ȧ (now denoted Ȧ B̂13 = 2[D11 − (1 − )M11ij Dij ] − 5(D1i Ai1 + A1i Di1 )
for brevity), this is
+ 10[A11ij Dij + (1 − )(L11ij Dij − M11ij Aijkl Dkl )], (61)
⎧ ⎫

⎪ Ȧ11 ⎪
⎪ Ȧ ⎪
⎪ ⎪


⎨ 22 ⎪ ⎬
Ȧ33 B̂23 = 2[D33 − (1 − )M33ij Dij ] − 5(D3i Ai3 + A3i Di3 )
{Ȧ} = . (49)

⎪ Ȧ23 ⎪
⎪ + 10[A33ij Dij + (1 − )(L33ij Dij − M33ij Aijkl Dkl )], (62)

⎪ ⎪


⎩ Ȧ31 ⎪

Ȧ12
B̂33 = 2[D31 − (1 − )M31ij Dij ] − 5(D3i Ai1 + A3i Di1 )
Defining H as
+ 10[A31ij Dij + (1 − )(L31ij Dij − M31ij Aijkl Dkl )], (63)
H = Ȧ − (W · A − A · W) − (D · A + A · D − 2[A

+ (1 − )(L − M : A)] : D), (50)


Ĥ1 = Ȧ11 − (W1i Ai1 − A1i Wi1 ) − {D1i Ai1 + A1i Di1 − 2[A11ij Dij
we may combine Eqs. (48) and (50), adopt the notation of Eq. (49),
+ (1 − )(L11ij Dij − M11ij Aijkl Dkl )]} − 2b3 (A
˙ 1i Ai1 − A11 Aij Aji
and write the ARD-RSC model as
⎧ ⎫ + 5A11ij Aik Akj − 5A1i Aij Aj1 ) −
b5
{2[D1i Di1 − (1 − )M11ij Dik Dkj ]
⎪ b1 ⎪

⎨ b2 ⎪
⎬ ˙
[B] b3 = {H}. (51) − 2A11 Dij Dji − 5(D1i Dij Aj1 + A1i Dij Dj1 ) + 10[A11ij Dik Dkj

⎪ ⎪
⎩ b4 ⎪
⎭ + (1 − )(L11ij Dik Dkj − M11ij Aijkl Dkm Dlm )]}, (64)
b5

Here [B] is a 6 × 5 matrix whose components Bij are the coefficients


of the parameters bi in Eq. (48), and {H} is a 6 × 1 column vector. Ĥ2 = Ȧ33 − (W3i Ai3 − A3i Wi3 ) − {D3i Ai3 + A3i Di3 − 2[A33ij Dij
The components Bij and Hi depend on Ȧ, A, D, W, and  (since A, L, + (1 − )(L33ij Dij − M33ij Aijkl Dkl )]} − 2b3 (A
˙ 3i Ai3 − A33 Aij Aji
and M are functions of A).
For a steady-state orientation, Ȧ = 0. Then, given a simple shear b5
+ 5A33ij Aik Akj − 5A3i Aij Aj3 ) − {2[D3i Di3 − (1 − )M33ij Dik Dkj ]
flow and a target steady-state value of A, all of the Bij and Hi com- ˙
ponents are determined. However, due to the property that tr A = 1
− 2A33 Dij Dji − 5(D3i Dij Aj3 + A3i Dij Dj3 ) + 10[A33ij Dik Dkj
and to the symmetries of the simple shear flow (SSF), there are only
three independent scalar equations in Eq. (51). Thus, we modify Eq. + (1 − )(L33ij Dik Dkj − M33ij Aijkl Dkm Dlm )]}, (65)
176 J.H. Phelps, C.L. Tucker III / J. Non-Newtonian Fluid Mech. 156 (2009) 165–176

Ĥ3 = Ȧ31 − (W3i Ai1 − A3i Wi1 ) − {D3i Ai1 + A3i Di1 − 2[A31ij Dij [8] R.S. Bay, C.L. Tucker III, Fiber orientation in simple injection moldings. Part II.
Experimental results, Polym. Compos. 13 (4) (1992) 332–342.
+ (1 − )(L31ij Dij − M31ij Aijkl Dkl )]} − 2b3 (A
˙ 3i Ai1 − A31 Aij Aji
[9] S. Ranganathan, S.G. Advani, Fiber–fiber interaction in homogeneous flows of
nondilute suspensions, J. Rheol. 35 (1991) 1499–1522.
b5 [10] X. Fan, N. Phan-Thien, R. Zheng, A direct simulation of fibre suspensions, J.
+ 5A31ij Aik Akj − 5A3i Aij Aj1 ) − {2[D3i Di1 − (1 − )M31ij Dik Dkj ] Non-Newtonian Fluid Mech. 74 (1998) 113–135.
˙ [11] N. Phan-Thien, X. Fan, R.I. Tanner, R. Zheng, Folgar–Tucker constant for a fibre
suspension in a Newtonian fluid, J. Non-Newtonian Fluid Mech. 103 (2002)
− 2A31 Dij Dji − 5(D3i Dij Aj1 + A3i Dij Dj1 ) + 10[A31ij Dik Dkj 251–260.
[12] D.L. Koch, A model for orientational diffusion in fiber suspensions, Phys. Fluids
+ (1 − )L31ij Dik Dkj − M31ij Aijkl Dkm Dlm )]}, (66) 7 (1995) 2086–2088.
[13] J.S. Cintra Jr., C.L. Tucker III, Orthotropic closure approximations for flow-
induced fiber orientation, J. Rheol. 39 (1995) 1095–1122.
Eq. (53) may now be solved to find b1 , b2 , and b4 for any choice [14] B.E. VerWeyst, Numerical predictions of flow-induced fiber orientation in 3-D
of b3 and b5 . geometries, Ph.D. Thesis, University of Illinois at Urbana-Champaign, Urbana,
IL, 1998.
[15] C.A. Hieber, S.F. Shen, A finite-element/finite-difference simulation of the
References injection-molding filling process, J. Non-Newtonian Fluid Mech. 7 (1980)
1–32.
[1] F. Folgar, C.L. Tucker III, Orientation behavior of fibers in concentrated suspen- [16] International Organization for Standardization, Plastics—injection moulding
sions, J. Reinf. Plast. Compos. 3 (1984) 98–119. of test specimens for thermoplastic materials. Part 5. Preparation of standard
[2] G.B. Jeffery, The motion of ellipsoidal particles immersed in a viscous fluid, Proc. specimens for investigation anisotropy, ISO 294-5:2001 (2001).
R. Soc. A 102 (1922) 161–179. [17] B.N. Nguyen, Pacific Northwest National Laboratory, Personal communication
[3] S.G. Advani, C.L. Tucker III, The use of tensors to describe and predict fiber (2008).
orientation in short fiber composites, J. Rheol. 31 (1987) 751–784. [18] P.J. Hine, N. Davidson, R.A. Duckett, A.R. Clarke, I.M. Ward, Hydrostatically
[4] H.M. Huynh, Improved fiber orientation predictions for injection-molded com- extruded glass–fiber reinforced polyoxymethylene. I. The development of fiber
posites, Master’s Thesis, University of Illinois at Urbana-Champaign, 2001. and matrix orientation, Polym. Compos. 17 (1996) 720–729.
[5] M. Sepehr, G. Ausias, P.J. Carreau, Rheological properties of short fiber filled [19] D.A. Jack, D.E. Smith, An invariant based fitted closure of the sixth-order orienta-
polypropylene in transient shear flow, J. Non-Newtonian Fluid Mech. 123 tion tensor for modeling short-fiber suspensions, J. Rheol. 49 (2005) 1091–1115.
(2004) 19–32. [20] R. Aris, Vectors, Tensors, and the Basic Equations of Fluid Mechanics, Prentice-
[6] J. Wang, J.F. O’Gara, C.L. Tucker III, An objective model for slowing orien- Hall, Englewood Cliffs, NJ, 1962.
tation kinetics in concentrated fiber suspensions: theory and rheological [21] G.L. Hand, A theory of anisotropic fluids, J. Fluid Mech. 13 (1962) 33–64.
evidence, J. Rheol. 52 (5) (2008) 1179–1200. See also US Patent 7,266, [22] B.N. Nguyen, V. Kunc, B. Frame, J.H. Phelps, C.L. Tucker III, S.K. Bapana-
469 (2007). palli, J.D. Holbery, M.T. Smith, Fiber length and orientation in long-fiber
[7] R.S. Bay, C.L. Tucker III, Fiber orientation in simple injection moldings. Part I. injection-molded thermoplastics. Part I. Modeling of microstructure and elactic
Theory and numerical methods, Polym. Compos. 13 (4) (1992) 317–331. properties, J. Compos. Mater. 42 (2008) 1003–1029.

You might also like