Flows in Polymers, Reinforced Polymers and Composites - A Multi-Scale Approach-Spr

SPRINGER BRIEFS IN
APPLIED SCIENCES AND TECHNOLOGY
Christophe Binetruy
Francisco Chinesta
Roland Keunings
Flows in Polymers,
Reinforced
Polymers and
Composites
A Multi-Scale
Approach
123
SpringerBriefs in Applied Sciences
and Technology
More information about this series at http://www.springer.com/series/8884
Christophe Binetruy Francisco Chinesta
•
Roland Keunings
Flows in Polymers,
Reinforced Polymers
and Composites
A Multi-Scale Approach
123
Christophe Binetruy Roland Keunings
Ecole Centrale de Nantes Université catholique de Louvain
Nantes Louvain-la-Neuve
France Belgium
Francisco Chinesta
UMR CNRS
Ecole Centrale de Nantes
Nantes
France
ISSN 2191-530X ISSN 2191-5318 (electronic)

SpringerBriefs in Applied Sciences and Technology
ISBN 978-3-319-16756-5 ISBN 978-3-319-16757-2 (eBook)
DOI 10.1007/978-3-319-16757-2
Library of Congress Control Number: 2015934680
Springer Cham Heidelberg New York Dordrecht London

© The Author(s) 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper
Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
To Anne-Claire, Ofelia and Françoise
Preface
Over the last decades, an increasing number of functional and structural parts, made
so far with metals, has been progressively reengineered by replacing metallic
materials by polymers, reinforced polymers, and composites. The motivation for
this substitution may be the weight reduction, the simpler, cheaper or faster forming
process, or the ability to exploit additional functionalities.
The fillers usually employed cover a broad range involving many scales: (i) the
nanometer scale (e.g., carbon nanotubes, graphene, fullerene, nanodiamonds);
(ii) the micrometer to the millimeter scale (particles and short fibers); (iii) the
centimeter scale of fibers used in SMC and BMC composite processes; and finally
(iv) the macroscopic scale where fibrous reinforcements are made of continuous
fibers arranged in bundles.
When load-bearing capacities are especially looked for, continuous fiber rein-
forced polymers are selected. In that case, the impregnation of the reinforcement
with a low viscosity polymer involves the flow of a Newtonian or non-Newtonian
fluid in the complex multi-scale microstructure related to the fiber and tow
arrangement. Reinforced polymers are selected instead of high performance poly-
mers of equivalent properties since the latter are generally more expensive.
When looking for functional properties, the use of nano-charges opens a wide
spectrum of possibilities but also raises new challenges, such as dispersion of
charges into the polymer matrix and occurrence of aggregation and disaggregation
mechanisms. Suspensions of practical interest involve many scales and many
concentration regimes, the latter ranging from dilute to highly concentrated.
In the present monograph, we survey modern developments related to the
multi-scale modeling and simulation of polymers (Chap. 1, by Roland Keunings),
reinforced polymers that involve a flowing microstructure (Chap. 2, by Francisco
Chinesta), and continuous fiber-reinforced composites, wherein the fluid flows inside
a nearly stationary multi-scale microstructure (Chap. 3, by Christophe Binetruy).
January 2015 Christophe Binetruy

Francisco Chinesta
Roland Keunings
vii
Acknowledgments
The authors acknowledge the support of the ESI and FAURECIA Chairs, Ecole
Centrale de Nantes (France), as well as of the Institut Universitaire de France.
Moreover, they acknowledge the contribution of several colleagues from Ecole
Centrale de Nantes (France), Université catholique de Louvain (Belgium),
Université de Bretagne Sud (France), ENSAM Angers (France) and University of
Zaragoza (Spain), in particular: Emmanuelle Abisset-Chavanne, Amine Ammar,
Gilles Ausias, Sebastien Comas, Elias Cueto, Julien Ferec, Elena Lopez, Rabih
Mezher, and Marta Perez.
ix
Contents
1 Multi-scale Modeling and Simulation of Polymer Flow . . . . . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Rheology and Structure . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 A Hierarchy of Modeling Approaches . . . . . . . . . . . . . . . 2
1.1.3 The Role of Computational Rheology . . . . . . . . . . . . . . . 4
1.2 Macroscopic Modeling and Simulations . . . . . . . . . . . . . . . . . . . 5
1.2.1 Governing Equations and Numerical Challenges . . . . . . . . 6
1.2.2 A Spectrum of Numerical Methods and Applications . . . . . 8
1.3 Multi-scale Kinetic Theory of Viscoelastic Flow . . . . . . . . . . . . . 9
1.3.1 Probability Distribution of Molecular Configurations . . . . . 9
1.3.2 Fokker-Planck Equation and Kramers Relation . . . . . . . . . 10
1.3.3 Exploiting the Multi-scale Model: Three
Basic Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.4 Illustrative Examples of Kinetic Theory Models . . . . . . . . 13
1.4 Micro-Macro Simulations Using the Stochastic Approach . . . . . . 17
1.4.1 CONNFFESSIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.2 Numerical Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.3 Brownian Configuration Fields . . . . . . . . . . . . . . . . . . . . 20
1.4.4 Lagrangian Particle Methods. . . . . . . . . . . . . . . . . . . . . . 21
1.4.5 Variance Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5 Micro-Macro Simulations Using the Fokker-Planck Approach . . . 23
1.5.1 Proper Generalized Decomposition . . . . . . . . . . . . . . . . . 24
1.5.2 PGD Simulations of the Multi-bead-Spring
FENE Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5.3 Fokker-Planck Simulations of Complex Flows . . . . . . . . . 27
1.6 The Impact of Closure Approximations . . . . . . . . . . . . . . . . . . . 28
1.7 Illustrative Examples of Complex Flow Simulations . . . . . . . . . . 31
1.7.1 Early Macroscopic Simulations . . . . . . . . . . . . . . . . . . . . 31
1.7.2 Macroscopic Simulations with Advanced Models . . . . . . . 33
1.7.3 Impact of Closure Approximations in Complex Flows . . . . 36
xi
xii Contents
1.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.8.1 From Numbers to Rheology . . . . . . . . . . . . . . . . . . . . . . 37
1.8.2 The Infamous HWNP and Mesh Convergence . . . . . . . . . 38
1.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2 Complex Flows of Micro/Nano Structured Fluids:

Reinforced Polymer Composites . . . . . . . . . . . . . . . . . . . ....... 43
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 43
2.2 Dilute and Semi-dilute Suspensions. . . . . . . . . . . . . . . ....... 46
2.2.1 Multi-scale Description of Non-brownian
Rod Suspensions . . . . . . . . . . . . . . . . . . . . . . . ....... 46
2.2.2 Multi-scale Description of Brownian Rods
Suspensions . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 52
2.2.3 Semi-concentrated Regime . . . . . . . . . . . . . . . . ....... 58
2.3 Processing Flow Simulation . . . . . . . . . . . . . . . . . . . . ....... 59
2.3.1 Fixed Mesh Description of the Fluid Domain
Evolution and Flow Front Tracking by Using
a Volume of Fluid—VoF—Technique . . . . . . . . . . . . . . . 60
2.3.2 Updated Lagrangian Meshless Simulation. . . . . . . . . . . . . 65
2.4 Concentrated Suspensions Involving Rod Clusters. . . . . . . . . . . . 69
2.4.1 Rigid Clusters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.4.2 Deformable Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5 Advanced Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.5.1 On the Solution of the Fokker-Plank Equation . . . . . . . . . 80
2.5.2 Descriptions Based on Higher-Order Kinematics . . . . . . . . 84
2.5.3 Accounting for Rod Bending . . . . . . . . . . . . . . . . . . . . . 90
2.5.4 Delaying Orientation Mechanisms . . . . . . . . . . . . . . . . . . 96
2.5.5 Collective Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
2.5.6 Orientation Induced by an Electric Field . . . . . . . . . . . . . 101
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3 Flows of Simple Fluids in Complex Microstructures:

Composite Processing of Structural Polymer Composites . . . . . . . . 109
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.2 Description of Fibrous Microstructures in Composites . . . . . . . . . 113
3.2.1 Monodisperse Porosity Medium . . . . . . . . . . . . . . . . . . . 115
3.2.2 Bidisperse Porosity Medium . . . . . . . . . . . . . . . . . . . . . . 119
3.3 Governing Equations for Flows in Porous Media . . . . . . . . . . . . 121
3.3.1 Pore-Scale Isothermal Flow Description . . . . . . . . . . . . . . 122
3.3.2 Macroscopic Scale: Upscaling . . . . . . . . . . . . . . . . . . . . 122
Contents xiii
3.4 Flow of Simple Fluids in Non-deformable

Fibrous Microstructures . . . . . . . . . . . . . . . . . . . . . ......... 128
3.4.1 Negligible Inertia and Single-Scale
Porosity Medium . . . . . . . . . . . . . . . . . . . . ......... 128
3.4.2 Negligible Inertia and Double-Scale
Porosity Medium . . . . . . . . . . . . . . . . . . . . ......... 132
3.4.3 Fluid Inertia . . . . . . . . . . . . . . . . . . . . . . . . ......... 134
3.5 Flow of Complex Fluids in Non-deformable Fibrous
Microstructures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.5.1 Slightly Compressible Fluids . . . . . . . . . . . . . . . . . . . . . 135
3.5.2 Reactive Fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.6 Numerical Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Chapter 1
Multi-scale Modeling and Simulation
of Polymer Flow
Abstract The flow of polymeric fluids in complex geometries relevant to

processing applications can be simulated numerically using a wide variety of theo-
retical models. Simple mathematical models have a purely macroscopic nature and
focus only on the non-linear relationship between shear viscosity and shear rate.
More advanced models address the viscoelastic character of polymeric fluids, either
in a macroscopic or multi-scale framework. These advanced models and the related
numerical approaches are the subject of this first chapter, wherein we build upon and
update our previous reviews of the field [29, 31].
1.1 Introduction
1.1.1 Rheology and Structure
Polymer solutions and melts, like other rheologically-complex fluids, exhibit a variety
of non-Newtonian flow properties [3, 13, 49, 52]. A first departure from Newtonian
behaviour is that the shear viscosity of most polymeric fluids is not a constant mate-
rial property (at a given temperature), but rather a non-linear function of the shear
rate. This property can have dramatic effects in shear-dominated flows, such as flows
in cylindrical pipes. A wide variety of macroscopic phenomenological constitutive
equations are available that offer a very good fit, with few material parameters, of
the non-Newtonian shear viscosity of polymeric fluids [3]. In combination with the
conservation principles of continuum mechanics (for mass, linear momentum and
energy), these simple constitutive models yield generalized Navier-Stokes equations
that can be solved numerically with rather obvious extensions of well-known tech-
niques for Newtonian fluids [52]. It is much less obvious to take account of the many
other non-Newtonian properties of polymeric liquids related in general to memory
effects. Indeed, polymeric liquids are viscoelastic materials, meaning that the stress
endured by a fluid element depends upon the history of the deformation experienced
© The Author(s) 2015 1

C. Binetruy et al., Flows in Polymers, Reinforced Polymers
and Composites, SpringerBriefs in Applied Sciences and Technology,
DOI 10.1007/978-3-319-16757-2_1
2 1 Multi-scale Modeling and Simulation of Polymer Flow
Fig. 1.1 Two-scale view at polymer flow: macroscopic fluid trajectory and velocity field v (left) and
microscopic molecular configurations (here, of linear entangled polymers) within a fluid element
(right)
by that element. Viscoelastic properties include non-linear shear viscosities, but also
the presence of normal stresses in viscometric flows, high resistance to elongational
deformation, and various memory effects (such as stress relaxation) related to the
elasticity of the material. They are responsible for numerous flow phenomena of
scientific and industrial relevance [5] that need to be predicted, understood, and pos-
sibly controlled by means of a combination of suitable physical models and numerical
techniques.
The rheological properties of viscoelastic liquids are governed by the flow-induced
evolution of their internal microstructure [18, 19, 34]. In flowing polymers, the
relevant microstructure is the conformation of the macromolecules, namely their
orientation and degree of stretch. Each macroscopic fluid element contains a large
number of polymers with a statistical distribution of conformations. While the flow
alters the polymer conformations along the fluid trajectories, the macroscopic stress
carried by each material element is itself governed by the distribution of polymer
conformations within that element (Fig. 1.1).
Furthermore, the frozen-in microstructure which develops in processing flows
dictates the physical properties of the final product. One thus faces a challenging
non-linear coupling between flow-induced evolution of molecular configurations,
macroscopic rheological response, flow parameters (such as geometry and bound-
ary conditions), and final product properties. Theoretical modeling and methods of
computational rheology are needed to elucidate this coupling. Clearly, the purely
macroscopic approach of continuum mechanics will not be able to fully address the
problem at hand. Rather, a suitable multi-scale approach is called for.
1.1.2 A Hierarchy of Modeling Approaches
Modern research in computational materials science exploits, sometimes in a coupled

fashion, a hierarchy of theoretical models ranging from quantum mechanics to con-
tinuum mechanics. The difficulty with polymers is of course the gigantic number of
microstructural degrees of freedom and the broad range of time and length scales sep-
arating the relevant atomistic and macroscopic processes [41]. A modeling approach
based on quantum mechanics and related ab initio computational techniques must
1.1 Introduction 3
clearly be ruled out at the outset. Indeed, atomistic modeling is probably the most
detailed level of description that could realistically be applied in rheological studies,
by means of techniques of non-equilibrium molecular dynamics (e.g. [51]).
Atomistic flow simulations have been conducted to study the behaviour of
polymers near walls and geometrical singularities such as re-entrant corners (e.g.
[7, 10, 33]). In view of the significant computer resources involved in such calcu-
lations, the atomistic approach is currently restricted to very coarse models for the
polymer and to flow geometries of molecular dimensions. While it is doubtful that it
may be feasible soon (and even necessary, for that matter!) to solve flow problems of
macroscopic size using the atomistic approach, the potential of atomistic simulations
is great, however, in helping us resolve difficult issues such as wall slip. Consider-
ation of macroscopic flows found in processing applications calls for less detailed
mesoscopic models, such as those of kinetic theory.
The next level of description of a polymeric fluid is that of kinetic theory [4, 14].
Here, one ignores atomistic processes altogether and focuses rather on the evolution
of a more or less coarse-grained model of the polymer conformations. For example,
a linear polymer chain in a viscous solvent can be described by several models of
decreasing complexity: the freely jointed, bead-rod Kramers chain made of a number
of beads (of order 100) connected linearly by rigid segments, the bead-spring chain
made of a smaller number of beads (of order 10) connected by entropic springs, or
the single dumbbell, namely two beads connected by a spring. Clearly, these models
of kinetic theory are not meant to describe the chemical structure of the polymer.
They do, however, display in a more or less detailed fashion the important features
needed to describe the evolution of polymer conformations in a macroscopic flow.
Kinetic theory models for polymer solutions or melts are most naturally exploited
numerically by means of stochastic simulation or Brownian dynamics methods [44],
but a powerful alternative method known as the Proper Generalized Decomposition
(PGD) has been developed recently [9], as we shall discuss below. Use of a model
of kinetic theory in the numerical simulation of complex flows, in combination with
the macroscopic conservation laws, constitutes the so-called micro-macro approach
to computational rheology. This emerging methodology is the main theme of this
chapter.
Finally, besides atomistic modeling and kinetic theory lies the macroscopic
approach of continuum mechanics [3, 52]. Here, details of the fluid microstructure
are not taken into account, at least explicitly. Rather, the stress experienced by the
macroscopic fluid elements is related to the deformation history through a suitable
constitutive equation. Combined with the conservation laws, the constitutive model
yields a set of partial differential (or integro-differential) equations that can be solved
by means of a suitable grid-based numerical method, such as the finite element tech-
nique. The vast majority of research publications in computational rheology has
followed the macroscopic approach [29, 45, 52]. Very much like atomistic modeling
and kinetic theory, continuum mechanics offers a hierarchy of constitutive models
of diverse predictive ability. It should be stressed that most constitutive equations
used today in numerical work have been derived from a molecular model of kinetic
theory. Such molecular-based constitutive equations yield quantitative information
on the distribution of polymer conformations within a macroscopic fluid element
in the form of averaged quantities such as the second moment of the distribution
of conformations. Unfortunately, closure approximations of a purely mathematical
nature are often needed in the derivation of a constitutive model from kinetic theory,
and their impact can be significant indeed [28]. In view of the closure problem, the
link with the parent molecular model is thus somewhat polluted, and interpretation
of the macroscopic results in molecular terms becomes delicate.
In this context, micro-macro methods [31] that couple the coarse-grained molec-
ular scale of kinetic theory to the macroscopic scale of continuum mechanics have
an important role to play. In a micro-macro simulation, the conservation equations
are solved together with a model of kinetic theory. This approach is much more
demanding in computer resources than more conventional continuum simulations
that integrate a constitutive equation to evaluate the viscoelastic contribution to the
stress tensor. On the other hand, micro-macro techniques allow the direct use of
kinetic theory models and thus avoid potentially harmful closure approximations.
1.1.3 The Role of Computational Rheology
It is useful to close these introductory remarks with a discussion of the potential role
of computational rheology [29]. The viscoelastic character of a given flow is often
measured by the dimensionless Weissenberg number We, defined as the product of
a characteristic relaxation time of the fluid and a characteristic deformation rate of
the flow. While We vanishes for Newtonian fluids, it is of order 1 or 10 in the poly-
mer flows of interest here. The challenge for computational rheologists is to develop
numerical schemes for obtaining accurate numerical solutions to the governing equa-
tions at values of We of practical interest, using a physically-realistic mathematical
model. Let us consider for a while the (ideal) situation in which reliable computa-
tional rheology software would be readily available and usable by non-specialists on
standard computer hardware, and let us imagine some of the numerous opportunities.
For instance, the theoretical rheologist would have a tool at his or her disposal for
the critical evaluation in complex flows of a new theory, whether it be a constitutive
equation, a molecular model, or an improved description of boundary conditions.
Complex flows are characterized by transient kinematics in the Lagrangian sense
(i.e. following the fluid elements) and by a combination of shear and elongational
deformations; they indeed constitute a tough testing ground for any theoretical model.
The theoretical rheologist could also use tools of computational rheology to bridge the
gaps between the several levels of description used to predict the polymer dynamics
[41]. One example of such a difficult exercise is the evaluation in complex flows of
1.1 Introduction 5
suitable closure approximations that are invariably needed to derive a macroscopic

constitutive equation from a molecular model. Also, non-Newtonian fluid mechanics
studies could be performed to understand better the mechanisms responsible for
observed macroscopic flow phenomena (such as vortex growth and purely-elastic
instabilities).
The experimentalist would also benefit. For example, he or she could perform
useful computational rheometry work wherein the experiment is actually simu-
lated in order to better interpret the data, i.e. to translate better what is measured
(such as forces, torques, deformations) into well-defined rheological information.
Experimental problems (such as flow inhomogeneities and secondary motions) would
thus be identified, possibly eliminated through improved design of the rheometri-
cal equipment, or explicitly taken into account in the data reduction phase. Also,
the combination of flow experiments and numerical simulations could be used to
characterize rheological behaviour in complex flows, yielding in a more or less auto-
mated way the optimal rheological model and material parameters for the fluid under
investigation.
Finally, and most importantly in industrial practice, the polymer engineer could
perform elaborate Computer Aided Design studies in which the link between the
molecular architecture of the raw material and the final properties of the product
would be established, at least qualitatively. Production problems (such as extru-
sion instabilities [32]) would be predicted and partially overcome through improved
design. One could also think of using an on-line computational rheology model in
concert with appropriate control algorithms to provide for intelligent, physics-based
process control techniques.
These are but a few opportunities that a mature field of computational rheology
could generate. Not surprisingly, we remain far from this ideal situation. Neverthe-
less, significant collective progress has been made over the last three decades to the
extent that the above virtual picture may not be as distant as would appear.
1.2 Macroscopic Modeling and Simulations
Since its pioneering days (Circa 1975), computational rheology has adopted the
macroscopic approach. Over the last two decades, computer processing capacity has
reached a level such as to make feasible the complementary micro-macro approach,
which involves the coupled solution of the conservation laws and a microscopic
model of kinetic theory. In this section, we focus on the macroscopic approach.
Detailed reviews include [27, 30, 45].
1.2.1 Governing Equations and Numerical Challenges
Let us consider for the sake of illustration the case of incompressible isothermal flow
in a confined geometry, in the absence of body forces. The Cauchy stress tensor σ is
thus given by
σ = − pI + τ p + ηs γ̇ , (1.1)
where p is the pressure, I is the unit tensor, τ p is the polymer contribution to the
stress, and ηs γ̇ is a purely viscous component to the stress which involves the rate of
strain tensor γ̇ = ∇v + ∇v T and a constant viscosity coefficient ηs . Finally, ∇v is
the gradient of the macroscopic velocity field v. The Newtonian stress can be inter-
preted physically as the solvent contribution to the stress in polymer solutions, or as
the stress response associated with fast relaxation modes. In other cases, the New-
tonian component is added to correct the constitutive model itself (e.g., to eliminate
excessive shear-thinning).
In a complex flow, the velocity field is a priori unknown. The velocity and stress
fields are coupled through the conservation laws. Conservation of mass and linear
momentum is expressed as
Dv
∇ · v = 0, ρ = ∇ · σ, (1.2)
Dt
where ρ is the fluid density and the operator D/Dt denotes the Lagrangian or material
derivative ∂/∂t + v · ∇. In macroscopic simulations, the set of governing Eqs. (1.1)
and (1.2) is closed with a suitable constitutive equation for the polymer stress.
Over the years, both differential and (separable) single-integral models have been
used in numerical work, ranging from the upper-convected Maxwell (UCM) and
Lodge models, to more realistic equations such as the FENE-CR, Giesekus, Phan-
Thien-Tanner, Doi-Edwards, K-BKZ, and various other models derived from tube
theory [41]. Although they differ greatly in complexity as well as in predictive abil-
ity, macroscopic constitutive equations used in computational rheology display a
(deceptively) simple generic form. Indeed, differential models read
Dτ p
= f(τ p , ∇v), (1.3)
Dt
meaning that the Lagrangian derivative of the polymer stress τ p is given as a model-
dependent function f of the local polymer stress and velocity gradient evaluated
along the fluid trajectories. For the particular case of steady-state flow (in the Eulerian
sense), the left-hand side of (1.3) reduces to the convective term v·∇τ p . In a complex
flow, the fluid trajectories along which (1.3) applies are of course unknown a prior, and
one must solve the complete set of Eqs. (1.1)–(1.3). If direct coupling between modes
1.2 Macroscopic Modeling and Simulations 7
is neglected, a spectrum of relaxation times is readily taken into account by defining

the polymer stress as a sum of partial contributions and writing Eq. (1.3) for each
partial stress. We shall see later that macroscopic constitutive equations derived from
kinetic theory of solutions or melts give the polymer stress as an algebraic function of
a number of microstructural tensors, which themselves follow an evolution equation
similar to (1.3). In computational work, it has thus become natural to select the
microstructural tensors as primary variables, instead of the polymer stress. The basic
numerical issues remain identical, however.
Separable single-integral constitutive models have been less popular in computa-
tional studies [30]. They also have a deceptively simple generic form
t
τ p (t) = m(t − t )S(t, t )dt , (1.4)
−∞
where m is the memory function of linear viscoelasticity, and S is a model-dependent

non-linear strain measure relative to the present time t. The operator ·dt is a time
integral taken along the a priori unknown fluid trajectories parameterized by the past
time t . A spectrum of relaxation times is readily introduced through the memory
function.
The generic constitutive equations (1.3) and (1.4) clearly express the mem-
ory of polymeric liquids: the polymer stress carried by a fluid element at present
(Lagrangian) time t is a function of the deformation history experienced at past
times t by the element flowing along its trajectory.
Appropriate boundary conditions supplement the governing equations. One spec-
ifies components of either the velocity or the contact force at the boundary ∂Ω of the
flow domain Ω; the pressure is specified at one point of the flow domain if no normal
contact forces have been specified anywhere at the boundary. In view of the fluid’s
memory, additional boundary conditions must be specified when the flow domain Ω
contains an inlet boundary. Specification of the flow pre-history is often achieved by
assuming fully-developed flow upstream of the inlet boundary. For transient flows,
initial conditions are also needed for the velocity field.
Although they may look rather innocuous, the above macroscopic governing equa-
tions, supplemented with suitable boundary and initial conditions, actually present
formidable mathematical and numerical challenges which certainly were not antici-
pated in the mid 1970s by the pioneers of the field. Through a combination of numeri-
cal experiments and mathematical investigations, our collective understanding of the
basic reasons for these difficulties progressively developed during the 1980s. See
[27, 45] for a detailed analysis. In a nutshell, we now know the following: the gov-
erning equations are of mixed mathematical type (elliptic-hyperbolic), with possible
local changes of type; stress boundary layers develop in many flow fields where
the corresponding Newtonian fluid mechanical problem is smooth; stress singular-
ities (e.g. at re- entrant corners) are much stronger than in the Newtonian case; the
non-linear qualitative behaviour of the exact solutions can be very rich (e.g. mul-
tiplicity of solutions, bifurcations). To the present day, these features remain quite
difficult to handle numerically in an accurate and general way. Whether they reflect
the actual physics of polymeric liquids (in which case we have to live with them!) or
result from inadequate modeling (governing equations and/or boundary conditions)
remains essentially an open issue. We shall see later that similar challenges appear in
the context of micro-macro simulations [31]. Despite all these difficulties, the field
of computational rheology has made definite progress, and it has been gratifying
to observe in most cases that, the more realistic is the rheological model used in a
simulation, the easier it is to obtain numerically-accurate solutions (of course, once
a suitable technique has been developed to handle the complexity of the model).
1.2.2 A Spectrum of Numerical Methods and Applications
Standard Computational Fluid Dynamics (CFD) technology offers a wide spectrum

of numerical algorithms based on finite element, finite volume, finite difference,
boundary integral, spectral methods, and combinations thereof. A similar diversity
exists in modern computational rheology, whose fair discussion requires a rather
thick monograph indeed [45]. In what follows, we wish to point to methods and
applications that define in our opinion the current state of the art. The reader will
not find here a personal list of best-buys. Indeed, it has been our experience over the
years that a particular numerical scheme (very much like a particular constitutive
model) enjoys a range of application where it behaves at its best, and perhaps better
than other competing techniques, while it may be quite limited or even totally useless
in other situations. Practitioners of computational rheology also know that the only
way to really evaluate and compare different numerical methods is to implement and
test them all for oneself.
This being said, it is fair to state that most of the published work deals with finite
element methods for solving 2d steady-state flows (in the Eulerian sense) using
a differential constitutive equation. Extensions of these methods have been made
available for the temporal stability analysis of complex flows, and the computation
of transient problems. Over the last two decades, finite volume methods have also
been actively developed, sometimes in combination with finite elements, to solve
2d and 3d time-dependent problems. For an extensive review, see [45]. Recent work
representative of the current state of the art include [12, 54].
Methods for integral models have also progressed considerably, most notably with
the introduction of Lagrangian finite element schemes (wherein the mesh deforms
with the fluid) and the Deformation field method (which uses a fixed Eulerian grid).
These techniques have opened the way for 2d and 3d transient simulations with
integral models. They are detailed in [30]. For examples of recent work, see [48, 59].
In view of the progress made in numerical technology, macroscopic simulations
have been exploited over the last decade in two important tasks, namely the evaluation
of constitutive equations for solutions and melts in benchmark complex flows, usually
through a detailed comparison with experimental observations, and computational
1.2 Macroscopic Modeling and Simulations 9
rheometry, or use of numerical simulation to aid the experimentalist in reducing his

or her data. It is in such studies that computational rheology has already proven
useful, in spite of its relatively young age.
We now turn to the complementary micro-macro approach.
1.3 Multi-scale Kinetic Theory of Viscoelastic Flow
Micro-macro methods of computational rheology [31] couple the coarse-grained,

mesoscopic molecular scale of kinetic theory to the macroscopic scale of contin-
uum mechanics. They thus offer a two-scale description of the fluid behaviour.
In a micro-macro simulation, the conservation equations are solved together with a
model of kinetic theory. As noted previously, this approach is much more demanding
in computer resources than more conventional continuum simulations that integrate
a constitutive equation to evaluate the viscoelastic contribution to the stress ten-
sor. On the other hand, micro-macro techniques allow the direct use of kinetic theory
models and thus avoid potentially dangerous closure approximations. Since the early
1990s, the field has developed considerably following the introduction of the CON-
NFFESSIT method [35, 43]. Being relatively new, micro-macro techniques have to
date been implemented only for models of kinetic theory with few configurational
degrees of freedom, such as non-linear dumbbell models of dilute polymer solutions
and single-segment tube models of linear entangled polymers. Nevertheless, their
potential range of applications is very wide indeed. Before describing the numerical
implementation of the micro-macro approach, it is useful to consider the generic
multi-scale formalism on which it is based.
1.3.1 Probability Distribution of Molecular Configurations
For the sake of illustration, let us consider a dilute solution of linear flexible poly-
mers in a Newtonian solvent. The configuration of an individual polymer chain is
influenced by a number of mechanisms, including Brownian, elastic, and drag forces.
Indeed, the solvent molecules undergo thermal agitation and continuously bombard
the polymer chain; this results in a diffusion process driven by a stochastic, Brown-
ian force. Moreover, an elastic restoring force of entropic nature arises when the
chain uncoils. The solvent also exerts viscous drag along the chain which affects the
polymer configuration when different parts of the chain feel a different velocity, i.e.
when a velocity gradient is imposed on the fluid.
In the framework of kinetic theory, the molecular configuration of an individual
polymer chain can be defined in a rather coarse fashion by the vector X connecting
the two chain ends. This simple approach provides a measure of both molecular
orientation and stretch. The set of possible values for X is called the configuration
space C, which has here a dimension NC equal to 3.
Within each macroscopic fluid element, located at position x in the flow domain
Ω, there is a large collection of polymer chains characterized at time t by a probability
distribution (pdf) of configurations ψ(X, x, t). The distribution function ψ is such
that ψ(X, x, t)dX gives the probability of finding a polymer with configuration
between X and X + dX at time t and position x.
Under equilibrium conditions, namely when the macroscopic velocity vector field
v vanishes identically, configurations are distributed according to the equilibrium
probability function ψeq (X) that results, in the present illustrative model, from a bal-
ance between Brownian and elastic forces. Furthermore, the polymer or viscoelastic
contribution to the stress τ p , reduces to an isotropic tensor of no rheological impor-
tance.
Under non-equilibrium conditions, namely when the polymer solution undergoes
a macroscopic flow with velocity field v(x, t), the polymer chains also feel a velocity
gradient ∇v(x, t), and viscous drag exerted by the solvent affects their configuration.
The distribution function is no longer equal to the equilibrium value ψeq (X), and it
generally varies both in space and time, namely ψ = ψ(X, x, t). Furthermore, the
flow-induced distribution of configurations induces a viscoelastic stress field τ p (x, t)
resulting from anisotropic orientation and stretch of the polymer chains.
Thus, while the kinematics alter the distribution of molecular configurations along
the flow trajectories, the stress experienced by each macroscopic fluid element is itself
governed by the distribution of configurations within that element. Furthermore,
velocity and stress fields are coupled through the conservation principles. Clearly,
predicting viscoelastic flow using a molecular model of kinetic theory is a challenging
non-linear problem. We now briefly describe its generic mathematical formulation.
1.3.2 Fokker-Planck Equation and Kramers Relation
Kinetic theory yields two basic building blocks: the diffusion or Fokker-Planck equa-
tion that governs the evolution of the distribution function, and the Kramers relation
that links the viscoelastic stress to the distribution function [4, 14, 44].
The Fokker-Planck equation has the general form
Dψ ∂ 1 ∂ ∂
=− · {A ψ} + : {D ψ}. (1.5)
Dt ∂X 2 ∂X ∂X
Here again, the symbol D/Dt denotes the Lagrangian or material derivative ∂/∂t +v·
∇, where ∇ is the del operator with respect to position x in physical space. Molecules
are thus assumed to be convected by the macroscopic velocity field v. The vector
X defines the coarse-grained configuration and has dimension NC . The factor A is
1.3 Multi-scale Kinetic Theory of Viscoelastic Flow 11
an NC -dimensional vector that defines the drift or deterministic component of the

molecular model. The macroscopic velocity gradient ∇v enters in the formulation
of A, and is assumed constant over the molecular length scale (first-gradient theory).
Finally, D is a symmetric, positive definite NC × NC matrix that embodies the
diffusive or stochastic component of the molecular model. In general, both A and
D depend upon (X, x, t). Suitable boundary conditions must be specified for ψ, in
both configuration and physical spaces.
It is useful to note that the Fokker-Planck equation is a convection-diffusion
equation defined in configuration and physical spaces. It combines in fact a continuity
equation expressing conservation of probability,
Dψ ∂
=− · {J}, (1.6)
Dt ∂X
and a constitutive equation for the flux of probability J,
1 ∂
J=Aψ − · {D ψ}. (1.7)
2 ∂X
A particular kinetic theory model is thus defined by a specific constitutive equation
for the flux of probability, via A and D. Examples pertaining to polymer solutions
and melts will be given later.
The second building block of a kinetic theory model is Kramers’ relation, namely
an expression relating stress to molecular configurations. It takes the form of a partic-
ular average computed with the distribution function over all possible configurations:

τp = g(X) ψ dX = g(X). (1.8)
C
Here, g is a model-dependent tensorial function of configuration. Thus, for a given

velocity field, and starting from a specified initial condition for ψ at time t0 , integra-
tion of the Fokker-Planck equation (1.5) yields the distribution function at all time
t ≥ t0 , everywhere in the flow domain Ω. Use of the stress-configuration relation
(1.8) then provides the viscoelastic stress field.
In a complex flow, the velocity field is a priori unknown; velocity and stress fields
are coupled through the conservation laws. For isothermal and incompressible flows,
conservation of mass and linear momentum is then expressed as
Dv
∇ · v = 0, ρ = ∇ · {− pI + τ p + ηs γ̇ }. (1.9)
Dt
Appropriate boundary conditions supplement these conservation equations, as dis-
cussed in Sect. 1.2.1.
1.3.3 Exploiting the Multi-scale Model: Three Basic

Approaches
The set of coupled Eqs. (1.5), (1.8) and (1.9), supplemented with suitable initial
and boundary conditions in both physical and configuration spaces, is the generic
multi-scale formulation of viscoelastic flow that we consider in the present chapter.
To date, three basic approaches have been adopted for exploiting this model:
The continuum approach, wherein a constitutive equation of continuum mechanics
that relates the viscoelastic stress to the deformation history is derived from, and
replaces altogether, the kinetic theory model (1.5) and (1.8). The derivation process
usually involves closure approximations. The resulting constitutive model takes the
form of a differential, integral, or integro-differential equation. It yields molecular
information in terms of averaged quantities, such as the second moment XX of the
distribution;
The Fokker-Planck approach, wherein one solves the generic problem (1.5), (1.8)
and (1.9) as such, in both configuration and physical spaces. The distribution func-
tion is thus computed explicitly as a solution of the Fokker-Planck equation. The
viscoelastic stress is merely a by-product obtained from (1.8);
The stochastic approach, which draws on the mathematical equivalence between
the Fokker-Planck equation (1.5) and the following Itô stochastic differential equa-
tion [44]:
dX = A dt + B · dW, (1.10)
where D = B · BT and W is a Wiener stochastic process of dimension NC . In a

complex flow, the stochastic differential equation (1.10) applies along individual
flow trajectories; the time derivative is thus a material derivative. Instead of solving
the deterministic Fokker-Planck equation (1.5), one solves the associated stochastic
differential equation (1.10) for a large ensemble of realizations of the stochastic
process X by means of a suitable numerical technique [44]. The distribution function
is not computed explicitly, and the viscoelastic stress (1.8) is readily obtained as an
ensemble average.
The continuum approach outlined above has been adopted throughout the devel-
opment of computational rheology, as discussed briefly in Sect. 1.2. In particular,
the first successful finite element simulations of die swell and laminar jet breakup in
viscoelastic liquids were obtained in the early 1980s with the Oldroyd-B constitutive
equation [11, 26]. The latter is mathematically equivalent to the simplest kinetic
theory model of a dilute solution of flexible linear polymers in a Newtonian sol-
vent, known as the Hookean dumbbell model [4]. Three decades later, macroscopic
numerical techniques based upon the continuum approach remain under active devel-
opment.
In the sequel, we focus on the micro-macro methods needed to implement the

Fokker-Planck and stochastic approaches. Before describing micro-macro techniques
in some detail, we consider briefly the FENE dumbbell and Doi-Edwards models.
These two models are no longer representative of the state of the art in kinetic theory.
They do however display enough complexity for our purpose.
1.3.4 Illustrative Examples of Kinetic Theory Models
1.3.4.1 The FENE Dumbbell Model
The simplest non-linear kinetic theory model of a dilute polymer solution is known as
the Finitely Extensible Non-linear Elastic (FENE) dumbbell model [4]. The polymer
solution is viewed as a flowing suspension of dumbbells that do not interact with each
other and are convected by the Newtonian solvent. Each dumbbell consists of two
identical Brownian beads connected by an entropic spring. Here, the configuration
X is the three-dimensional vector Q connecting the two beads (Fig. 1.2).
A FENE dumbbell cannot extend beyond the polymer contour length Q 0 . Thus,
the configuration space C is the set of three-dimensional vectors of magnitude less
than Q 0 . The Fokker-Planck equation (1.5) for the distribution function ψ(Q, x, t)
reads

Dψ ∂ 2 2kT ∂ ∂
=− · κ · Q − Fc (Q) ψ + · ψ, (1.11)
Dt ∂Q ζ ζ ∂Q ∂Q
where κ = ∇v T is the transpose of the velocity gradient, ζ is a friction coefficient,

Fc (Q) is the entropic connector force, T is the absolute temperature, and k is the
Boltzmann constant. The distribution function vanishes at the boundary of the con-
figuration space C, namely for all Q of magnitude Q 0 . We consider entropic springs
that follow Warner’s force law [4],
H
Fc (Q) = Q, (1.12)
1 − Q2 /Q 20
Fig. 1.2 Left dumbbell model (red) of a linear polymer (blue). Right non-interacting dumbbells
within a macroscopic fluid element
where H is a spring constant. The stress-configuration relation (1.8) gives
τ p = n QFc (Q), (1.13)
where n is the dumbbell number density and the angular brackets · denote
the configuration space average C · ψ dQ. Finally, the Itô stochastic differential
equation (1.10) that is equivalent to (1.11) reads

2 4kT
dQ = [κ · Q − Fc (Q)] dt + dW, (1.14)
ζ ζ
where W is the three-dimensional Wiener process, namely a Gaussian stochastic

process with vanishing mean and covariance W(t1 ) W(t2 ) = min(t1 , t2 ) I.
The limit of infinitely extensible dumbbells (Q 0 → +∞) is the Hookean dumb-
bell model, which is equivalent to the Oldroyd-B constitutive equation. The FENE
dumbbell model does not, however, have a mathematically equivalent, closed-form
constitutive equation. A closure approximation is thus needed to exploit the model
in continuum simulations. The simplest closure, due to Peterlin, replaces the FENE
spring force (1.12) by the pre-averaged FENE-P approximation
H
Fc (Q) ≈ Q. (1.15)
1 − Q2 /Q 20
It is then possible to obtain from (1.11) and (1.13) the FENE-P constitutive equation,
namely a differential equation for the configuration tensor A = QQ:
DA 4kT 4H/ζ
− κ · A − A · κT = I− A, (1.16)
Dt ζ 1 − tr(A)/Q 20
supplemented by an algebraic expression relating stress to average configuration,
H
τp = n A. (1.17)
1 − tr(A) /Q 20
Note that the FENE-P constitutive equation is of the form,

DA
= f(A, ∇v), (1.18)
Dt
which is precisely the generic formulation of differential constitutive models given
in Sect. 1.2.1. The Peterlin approximation can be very poor indeed [28, 50], and
much better closure approximations are available [36, 37]. At any rate, closure-
approximated dumbbell models (such as FENE-P) are very useful in the develop-
ment and evaluation of micro-macro methods, since the micro-macro results can be
compared to those obtained with the continuum approach.
1.3.4.2 The Doi-Edwards Model with Independent Alignment
The simplest tube model of entangled linear polymers is the classical Doi-Edwards
model with independent alignment. Reptation is the single relaxation mechanism
described by the model, and the dynamics of individual tube segments are decou-
pled. Constraint release (thermal or convective), tube stretch, and contour length
fluctuations are ignored [14]. The configuration X is defined by the unit orientation
vector u of a tube segment and the normalized contour label s ∈ [0, 1] of the primitive
chain, where s = 0 and 1 correspond to the chain ends (Fig. 1.3).
The distribution function ψ(u, s, x, t) is such that ψ(u, s, x, t)duds is the joint
probability that at time t and position x a tube segment has an orientation in the
interval [u, u+du] and contains the chain segment labelled in the interval [s, s +ds].
Thus, the configuration space C is B(0, 1) × [0, 1], where B(0, 1) is the surface of
the unit sphere centered at the origin. The Fokker-Planck equation (1.5) for the
Doi-Edwards model reads
Dψ ∂ 1 ∂ 2ψ
=− · [(I − uu) · κ · u ψ] + 2 , (1.19)
Dt ∂u π τd ∂s 2
where τd is the disengagement time, namely the characteristic time for a chain to
escape from its tube by reptation. The boundary conditions in configuration space
specify that ψ is isotropic at s = 0 and 1, namely
1
ψ(u, 0, x, t) = ψ(u, 1, x, t) = δ(|u| − 1), (1.20)
4π
Fig. 1.3 a Test chain within

entangled linear polymers.
b Test chain (red) within
virtual tube formed by
entangled chains (green).
c Axis of the tube or
primitive chain (red),
parameterized by s and with
local orientation u. When
tube stretch is ignored, the
tube length L is constant and
normalized to 1
where δ is the Dirac delta function. The Kramers relation (1.8) reads
τ p = G uu, (1.21)
where G is an elastic modulus, and the angular brackets · denote the configuration
1
space average 0 B(0,1) · ψ duds. Finally, the Itô stochastic differential equation
(1.10) that is equivalent to (1.19) is given by

2
du = (I − uu) · κ · u dt, ds = dW, (1.22)
π 2 τd
where W is the one-dimensional Wiener process. The evolution equation for u is

deterministic; it expresses the assumption that tube segments orient with the macro-
scopic flow. The equation for s is that of a purely-diffusive process. The coupling
between u and s arises through the boundary conditions (1.20): when the process
s reaches either 0 or 1, it is reflected and u is reset to a randomly oriented unit
vector [44].
In view of its (relative) simplicity, the basic Doi-Edwards model can be cast into
an equivalent integral constitutive model. Consider a fluid particle whose position
at present time t is given by x; its position at some past time t is x . The particle
motion is described by the displacement function x = x (x, t, t ), while the defor-
mation gradient F(t, t ) is defined as ∂x /∂x. The Doi-Edwards model is formulated
in a Lagrangian framework and has the generic form of separable single-integral
constitutive equations given in Sect. 1.2.1. It thus gives the viscoelastic stress τ p at
a moving fluid particle, according to
t
τ p (t) = G m(t, t )Q(t, t ) dt , (1.23)
−∞
where the integral is computed along the past trajectory of the fluid particle, para-
meterized by the past time t . Here, m is a memory function given by
∞
8 (2k + 1)2 (t − t )
m(t, t ) = exp − , (1.24)
π 2 τd τd
k=0
and Q is a strain measure defined as
(F(t, t ) · u(t ))(F(t, t ) · u(t ))

Q(t, t ) = uut = , (1.25)
|F(t, t ) · u(t )|2 t
where · t denotes the average computed with the isotropic distribution (1.20).
Thus, simulations based on the basic Doi-Edwards tube model can be performed
either with a micro-macro numerical method, or with a continuum technique for
integral constitutive equations such as the method of deformation fields [46]. Again,
this is very useful for the validation of numerical developments. More sophisticated
tube models that take account of additional physics such as stretch and constraint
release generally do not have an equivalent closed-form constitutive equation.
We now turn to available numerical techniques for exploiting the generic multi-
scale formulation of visoelastic flows. We start with the stochastic approach.
1.4 Micro-Macro Simulations Using the Stochastic

Approach
1.4.1 CONNFFESSIT
The CONNFFESSIT method [35, 43] is the first implementation of the stochas-
tic approach defined in Sect. 1.3.3. The acronym stands for “Calculation Of Non-
Newtonian Flow: Finite Elements and Stochastic SImulation Technique”. Although
it can deal with Eulerian steady flows, CONNFFESSIT is most naturally exploited
as a time-marching scheme. A steady-state flow, if it exists, is then obtained as the
long-time limit of the calculations. At the start of a simulation, a large number of
model molecules (such as dumbbells, bead-spring chains, or any other coarse-grained
model) is distributed uniformly over the entire flow domain. Their initial configu-
ration is drawn from the equilibrium distribution. As the simulation proceeds, the
molecules are convected along flow trajectories. At each time step, solution of the
conservation laws (1.9) is decoupled from the integration of the stochastic differen-
tial equation (1.10) for each molecule. More precisely, a typical time step is defined
by the following algorithm:
S1 Solve the conservation equation (1.9) by means of a standard finite element
method to obtain updated values for the velocity and pressure fields, treating the
current approximation to the viscoelastic stress field as a known body force term
in the momentum equation;
S2 Using the updated velocity field, compute the path of each model molecule
convected by the macroscopic flow;
S3 For each molecule, integrate the stochastic differential equation (1.10) along the
molecule’s path to update its configuration;
S4 Update the viscoelastic stress field by means of the average (1.8) computed over
the local ensemble of molecules currently located within each finite element.
A global ensemble of N g molecules (e.g. FENE-P and FENE dumbbells) is distrib-
uted over the elements of the mesh. Consider a typical time step tn → tn+1 = tn +Δt.
S1 is performed using a classical penalty finite element technique, with bi-linear
quadrilateral elements for the velocity. Each quadrilateral element is split into two
triangles over which the viscoelastic stress has a constant value assigned to the cen-
tral Gauss integration point. The time-stepping scheme is implicit in the velocity
and explicit in the viscoelastic stress. The body force term ∇ · τ p in the momen-
tum equation is evaluated with stress values known at time tn . Thus, S1 yields the
approximated velocity field va at tn+1 . S2 is the classical problem of particle tracking
j
in a given finite element velocity field. The updated location rn+1 of the centre of
mass of the jth molecule ( j = 1, 2, ..., N g ) at time tn+1 is obtained by integrating
dr j
= va (r j , t), (1.26)
dt
j
using the explicit, first-order Euler scheme and the initial condition rn known from
the previous time step. The simplest numerical technique for implementing S3 is the
Euler-Maruyama scheme,
j j j j j
Xn+1 = Xn + A(Xn , tn )Δt + B(Xn , tn ) · ΔWn . (1.27)
j
Here, Xi denotes the configuration of the jth molecule at time ti , and the random
j
vector ΔWn has independent Gaussian components with zero mean and variance
Δt. Note that each model molecule feels its own Brownian random force, namely
j j
ΔWn1 and ΔWn2 are uncorrelated for j1 = j2 . The Euler-Maruyama scheme is of
weak order 1. When used with FENE dumbbells, it can lead to difficulties if Δt is
too large. Indeed, the length of an individual dumbbell can become larger than the
upper bound Q 0 , which is unphysical. Use of the predictor-corrector scheme of weak
order 2 solves this problem [44]. Finally, S4 amounts to approximating the stress-
configuration relation (1.8) by means of a local ensemble average computed in each
element (or possibly sub-element) of the mesh. The updated viscoelastic stress in
element k is thus given by
1
τ kp,n+1 = k
g(Xln+1 ). (1.28)
Nloc l
k molecules that are currently

Here, the counter l runs over the indices of the Nloc
present in the kth element. The time step tn → tn+1 is thus completed.
Applications of the CONNFFESSIT approach to dumbbell models of dilute poly-
mer solutions, tube models of entangled polymers, colloidal dispersions, and liq-
uid crystalline polymers, have been reported for steady-state and transient two-
dimensional flows and are reviewed in [31].
1.4.2 Numerical Challenges
The original CONNFFESSIT scheme has had a significant impact on the development
of micro-macro techniques. It does however present three basic difficulties. First,
tracking the motion of the molecules with a simple explicit scheme can be inaccurate
in complex flow geometries, especially in the vicinity of impervious walls where
1.4 Micro-Macro Simulations Using the Stochastic Approach 19
numerical errors can result in molecules leaving the flow domain. Tracking all N g
individual molecules is also very expensive. At every time step, one must know
the index of the finite element in which each molecule is currently located. Given
the large number of molecules and elements, typically N g = O(105 − 107 ) and
Nelt = O(103 − 104 ) in two-dimensional simulations, one cannot use a brute force
search of complexity O(Nelt × N g ) at each time step, and more elaborate methods
are needed such as use of element neighbour lists ordered on the basis of the local
direction of flow. The second difficulty is related to the statistical accuracy of the
viscoelastic stress.For a given velocity field, the ensemble average (1.28) carries
a statistical error Θ/Nloc k , where Θ = V ar [g(X)] is the variance of g(X). Thus,
each element of the mesh must, during the entire simulation, contain enough model
molecules for the local ensemble average (1.28) to be accurate. Ideally, one should
k = O(102 −103 ) molecules in each element and at all times. This is not easy
have Nloc
(and maybe even impossible) to ensure, especially in flow regions where large spatial
gradients develop in the velocity and stress fields, and where one would precisely
wish to use very small elements. A means of reducing the statistical error without
increasing the size Nlock of the ensemble is to apply variance reduction methods.
The third difficulty is that CONNFFESSIT stress fields typically show spurious
spatial oscillations which affect, via the momentum equation, the numerical accuracy
of the velocity field [31]. These basic issues motivated the development of improved
techniques which we review next.
A breakthrough came with the method of Brownian configuration fields, that uses
correlated local ensembles [22]. In this approach, the same local ensemble of model
molecules is defined initially within each material element. As time evolves, the
configurations of the kth molecule in all local ensembles is determined using the
same sequence of random numbers, i.e. the sequence only depends on the index k.
Thus, strong spatial correlations develop in the stress fluctuations within neighbour-
ing material elements (which feel a very similar flow history), and evaluation of the
divergence of the stress in the momentum equation leads to partial cancellation of the
fluctuations. The method of Brownian configuration fields is an Eulerian implemen-
tation of correlated local ensembles. Alternatively, the so-called Lagrangian particle
methods [17, 21, 55] provide a Lagrangian implementation of the same idea. In com-
parison with CONNFFESSIT, these techniques produce stress and velocity fields
that are significantly smoother in space. Moreover, the cost of generating sequences
of random numbers is reduced drastically. Of course, the use of correlated local
ensembles introduces artificial spatial correlations. It is thus forbidden altogether in
problems dominated by physical fluctuations (such as flows on a molecular length
scale).
1.4.3 Brownian Configuration Fields
The method of Brownian configuration fields [22, 39] uses correlated ensembles of
model molecules and completely avoids the tracking problem. Instead of computing
the configuration of discrete molecules along flow trajectories, this method deter-
mines the evolution of a finite number of Eulerian configurations fields. Each field
feels a random Wiener process that is uniform in space.
Nf
Let {Xk (x, t)}k=1 denote an ensemble of N f configuration fields defined over the
entire flow domain Ω. In a typical simulation, N f is of order 103 . The field Xk (x, t)
represents the configuration of the kth model molecule in the local ensemble at
position x and time t. At time t = 0, Xk is set to a spatially uniform value Xk0 , drawn
from the equilibrium distribution for example; all kth molecules thus have the same
initial configuration. The subsequent evolution of the field Xk is governed by
dXk (x, t) = {−v(x, t) · ∇Xk (x, t) + A(Xk (x, t))} dt + B(Xk (x, t)) · dWk (t).
(1.29)
This is indeed the Eulerian formulation of the stochastic differential equation (1.10),
with a familiar convection term v · ∇Xk . Note however that the stochastic factor
dWk only depends on time. As a result, the spatial gradient of the configuration
field is a well defined function of the spatial coordinates. One can thus discretize the
field Xk by means of a standard finite element approximation. Also, in the numerical
integration of (1.29), the same sequence of random numbers is used to determine,
everywhere in the flow domain, the configuration of the kth model molecule in the
local ensemble.
At any time t, an ensemble average computed over all configuration fields yields
the approximated viscoelastic stress,
Nf
1
τ pa (x, t) = g(Xk (x, t)). (1.30)
Nf
k=1
Note that the CONNFFESSIT local ensembles (1.28) have a size Nloc that varies
both with time and position in the finite element mesh. In contrast, the ensemble
average (1.30) involves a number N f of realizations that is fixed and independent of
mesh refinement. This provides a much more effective control of the statistical error.
In their early implementation for Hookean dumbbells [22], the authors solve the N f
evolution Eq. (1.29) for the configuration fields by means of a discontinuous Galerkin
(DG) finite element method, with explicit Euler-Maruyama time stepping. In the
DG approach, the interpolation functions are discontinuous across neighbouring
elements, which induces only weak coupling between elements. Thus, at each time
step, all configuration and stress calculations can be performed at the element level.
Spatial discretization of the conservation laws is achieved by means of the Discrete
Elastic-Viscous Split Stress (DEVSS) formulation.
In [53], the method of configuration fields has been applied to to the basic Doi-
Nf
Edwards model described in Sect. 1.3.4.2. Let us define a set {uk (x, t), sk (t)}k=1 of
N f configuration fields uk and associated random walkers sk . In view of (1.22), the
field uk evolves according to
∂
uk (x, t) + v(x, t) · ∇uk (x, t) = [I − uk (x, t)uk (x, t)] · κ(x, t) · uk (x, t), (1.31)
∂t
while the random walker sk performs Brownian motion,

2
dsk = dWk (t), (1.32)
π τd
2
with a reflecting boundary condition when it reaches 0 or 1. Note that sk is not a func-
tion of position x. Whenever sk is reflected, the associated field uk is instantaneously
reset to a spatially uniform, random field drawn from the isotropic distribution on the
surface of the unit sphere; its subsequent evolution is again governed by (1.31) until
the next reflection of sk . Here again, the method of configuration fields produces
much smoother results than CONNFFESSIT. Indeed, in a typical CONNFFESSIT
simulation of the Doi-Edwards model, individual tube segments are convected by the
flow and have their own associated random walkers. Thus, not only are the tube seg-
mental orientations reset at uncorrelated times (when the associated random walker
is reflected), they are also reset to uncorrelated random values. This induces severe
spatial oscillations in the numerical results for velocity and stress.
1.4.4 Lagrangian Particle Methods
The Lagrangian particle method (LPM) introduced in [21] is another further develop-
ment of CONNFFESSIT. Here again, the conservation laws are solved at each time
step by means of a standard finite element technique, with the viscoelastic stress
known from the previous step. In LPM, the stress is computed at a number Npart of
Lagrangian particles that are convected by the flow. Each Lagrangian particle carries
a number Nd of model molecules. Tracking is thus performed for Npart particles
instead of Npart × Nd molecules, using a very accurate fourth-order Runge-Kutta
scheme within the parent element. Along the path of each particle, LPM solves the
stochastic differential equation (1.10) for each of the Nd molecules. Note that LPM
can be used either with uncorrelated or correlated local ensembles of model mole-
cules. In the first case, Npart × Nd independent Wiener processes drive the stochastic
evolution of molecular configurations, as in a CONNFFESSIT simulation. In the sec-
ond case, the same initial ensemble of molecules is used in each Lagrangian particle,
and the same Nd independent Wiener processes are generated to compute the con-
figuration of corresponding molecules in each particle. In fact, LPM with correlated
ensembles of molecules can be viewed as a (Lagrangian) method of characteristics

for solving the evolution equation (1.29) that governs a particular Brownian configu-
ration field; the number N f of fields then corresponds to the number Nd of molecules
carried by each Lagrangian particle. The updated viscoelastic stress is computed at
each Lagrangian particle by means of a local ensemble average involving the con-
figurations of Nd molecules. At the end of a typical time step for the solution of the
conservation laws, we have at our disposal values of the viscoelastic stress at discrete
Lagrangian particles with a known position. These Lagrangian results then feed the
Eulerian discretized momentum balance as follows: within each finite element, LPM
computes the linear least-squares polynomial that best passes through the available
Lagrangian stress data. Clearly, this procedure requires that at least three Lagrangian
particles be present in each element at all times. LPM results obtained for non-linear
dumbbells [21] are in excellent agreement with those provided by the method of
configuration fields.
Use of LPM with highly graded finite element meshes requires a large number
of Lagrangian particles. This motivated the development of the adaptive Lagrangian
particle method (ALPM) in [17]. At each time step, ALPM ensures that all ele-
ments of the mesh have a number of Lagrangian particles in the user-specified inter-
elt,min elt,max
val [Npart , Npart ]. This necessitates an elaborate adaptive procedure wherein
Lagrangian particles are either created or destroyed. Once a new particle is created,
the configuration of the local Nd molecules must be properly initialized in order
to allow for the subsequent solution of the stochastic differential equation (1.10)
along the particle’s path. Initialization is performed at the element level, by means
of a least-squares approximation based on the current configuration of molecules in
neighbouring particles. ALPM is superior to LPM both in terms of cost and numerical
accuracy. Unfortunately, it is much more intricate to implement, and can only be used
with correlated local ensembles in view of the initialization step for newly-created
particles.
A further development of LPM is the backward-tracking Lagrangian particle
method (BLPM) [55]. Instead of dropping Lagrangian particles into the flow and
following them through the flow domain, the particle positions at which the vis-
coelastic stress is evaluated are fixed and specified a priori. The stress is computed
by tracking the particles backwards in time over a single time step Δt, evaluating
suitable initial values of configurations at that point, and then integrating the stochas-
tic differential equation forwards along the obtained trajectories. Let r B denote an
arbitrary fixed position in the mesh. In general, the particle trajectory leading to that
position changes from one time step to the next, so that different Lagrangian particles
arrive at r B as time evolves. The position r I at tn−1 of the Lagrangian particle which
is located at r B at current time tn is obtained by integrating (1.26) backwards in time,
tn
rI = rB − v dt. (1.33)
tn−1
Then, starting from the previously computed values of the Nd molecular configu-
rations X j at r I and time tn−1 , one obtains the updated configurations at r B and tn
by integrating the stochastic differential equation (1.10) along the path joining r I
and r B . The initialization process requires that an Eulerian field be computed from
the Lagrangian data for each realization X j and at each time step. As with ALPM,
correlated ensembles of molecules must be used for the initialization process to be
meaningful. In [55], BLPM has been implemented for two-dimensional transient
flows of non-linear dumbbell models. In each quadrilateral element of the mesh, the
fixed positions r B are defined at the nine nodal positions.
For all problems which can be solved with correlated ensembles of model mole-
cules, BLPM is the best available Lagrangian particle method, both in terms of
numerical behaviour and cost. BLPM is also very efficient in continuum simulations
with a constitutive equation of differential type [56, 58]. More generally, Lagrangian
particle methods take account in a most natural way of the purely convective nature
of both the deterministic and stochastic differential equations (1.3) and (1.10).
1.4.5 Variance Reduction
The stochastic formulation of viscoelastic flows raises several mathematical issues of

fundamental interest. In particular, an important topic is that of the statistical variance
of the numerical solution. In the continuous (or exact) formulation of the coupled
problem, the velocity and viscoelastic stress fields are deterministic, while only
the molecular configurations are random. In the discrete problem, all variables are
random and the numerical accuracy of a simulation is dictated by the variance of the
discrete variables. In practice, a large variance in the numerical results would imply
that independent micro-macro simulations yield vastly different solutions for the
same flow problem. We have already discussed a significant beneficial effect of using
correlated ensemble of model molecules: at each time step, the approximate velocity
and viscoelastic stress fields are much smoother functions of x than those obtained
with uncorrelated ensembles. But what is the effect on the variance? It has been
shown in [21], by means of numerical experiments, that use of correlated ensembles
reduces the variance of the velocity but increases the variance of the viscoelastic
stress. This counter-intuitive experimental finding must be due to the non-linear
coupling between the conservation laws and the stochastic differential equations.
Indeed, for a given deterministic velocity field, the variance of the approximate
stress is not affected by the spatial dependence of the Wiener processes.
1.5 Micro-Macro Simulations Using the Fokker-Planck

Approach
We have seen in the previous section that control of the statistical noise is a major
issue in stochastic micro-macro simulations. This problem does not arise at all in the
Fokker-Planck approach, since the viscoelastic stress τ p and the primary unknown
fields (v, p, ψ) governed by Eqs. (1.5), (1.8) and (1.9) are deterministic quantities.
The difficulty, however, is that the Fokker-Planck equation (1.5) must be solved for ψ
in both physical and configuration spaces. This necessitates a suitable discretization
procedure for all relevant variables, namely position x, configuration X, and time
t (in transient flow). The dimensionality of the problem can be daunting, and con-
sideration of molecular models with many configurational degrees of freedom does
not appear feasible. This probably explains why relatively few studies based on the
Fokker-Planck approach have appeared in the literature. The recent introduction of
the powerful technique now know as Proper Generalized Decomposition (PGD) has
opened the way for robust and efficient simulations with the Fokker-Planck approach,
as discussed in the sequel.
1.5.1 Proper Generalized Decomposition
Problems defined in high-dimensional spaces abound. For example, we have seen

that coarse-grained models of kinetic theory result in a Fokker-Planck equation for
the distribution function that must be solved in both configuration space, physical
space and temporal domain. Until recently (circa 2005), the direct numerical solution
of the Fokker-Planck equation has been limited to models having but few (2 or 3)
configurational degrees of freedom (see e.g. the review [31]).
In [1, 2], a new technique has been proposed that is able to circumvent, or at least
alleviate, the curse of dimensionality. This method is based on the use of separated
representations. It basically consists in constructing by successive enrichment an
approximation of the solution in the form of a finite sum of N functional products
involving d functions of each coordinate. In contrast with the shape functions of clas-
sical methods, these individual functions are unknown a priori. They are obtained by
introducing the approximate separated representation into the weak formulation of
the original problem and solving the resulting non-linear equations. The complexity
of the method grows linearly with the dimension d of the space wherein the problem
is defined, in vast contrast with the exponential growth of classical mesh-based tech-
niques. The method was later coined Proper Generalized Decomposition (or PGD),
as in many cases the number N of terms in the separated representation needed to
obtain an accurate solution is found to be close to that of the optimal decomposi-
tion obtained by applying a posteriori the Proper Orthogonal Decomposition to the
problem solution. The PGD is detailed in a recent book [9].
Let us briefly present the main ideas behind the PGD by considering an abstract
problem defined in a space of dimension d for the unknown field u(x1 , . . . , xd ).
Here, the coordinates xi denote any usual coordinate (scalar or vectorial) related to
physical space, time, or conformation space, for example, but they could also include
problem parameters such as boundary conditions or material parameters. We seek a
solution for (x1 , . . . , xd ) ∈ Ω1 × · · · × Ωd .
1.5 Micro-Macro Simulations Using the Fokker-Planck Approach 25
The PGD yields an approximate solution in the separated form:

N
u(x1 , . . . , xd ) ≈ Fi1 (x1 ) × · · · × Fid (xd ). (1.34)
i=1
The PGD approximation is thus a sum of N functional products involving each

j
a number d of functions Fi (x j ) that are unknown a priori. It is constructed by
successive enrichment, whereby each functional product is determined in sequence.
j
At a particular enrichment step n + 1, the functions Fi (x j ) are known for i ≤ n from
the previous steps, and one must compute the new product involving the d unknown
j
functions Fn+1 (x j ). This is achieved by invoking the weak form of the problem
under consideration. The resulting discrete system is non-linear, which implies that
iterations are needed at each enrichment step. A low-dimensional problem can thus
j
be defined in Ω j for each of the d functions Fn+1 (x j ).
If M nodes are used to discretize each coordinate, the total number of PGD
unknowns is N × M × d instead of the M d degrees of freedom involved in standard
mesh-based discretizations. Moreover, all numerical experiments carried out to date
with the PGD show that the number of terms N required to obtain an accurate solution
is not a function of the problem dimension d, but it rather depends on the regularity
of the exact solution. The PGD thus avoids the exponential complexity with respect
to the problem dimension.
In many applications studied to date, N is found to be as small as a few tens, and
in all cases the approximation converges towards the solution associated with the
complete tensor product of the approximation bases considered in each Ω j . Thus,
we can be confident about the generality of the separated representation (1.34),
but its optimality depends on the solution regularity. When an exact solution of a
particular problem can be represented with enough accuracy by a reduced number
of functional products, the PGD approximation is optimal. If the solution is a non-
separable function for the particular coordinate system used, the PGD solver proceeds
to enrich the approximation until including all the elements of the functional space,
i.e. the M d functions involved in the full tensor product of the approximation bases
in each Ω j . For more details, see [9].
1.5.2 PGD Simulations of the Multi-bead-Spring FENE Model
In [1], the PGD has been used to compute solutions of the Fokker-Planck equation,
for given flow kinematics, in configuration spaces of dimension 20 using the Multi-
Bead-Spring FENE model of dilute polymer solutions.
The MBS FENE chain consists of d + 1 beads connected by d non-linear entropic
springs [4]. The molecular configuration is thus given by the concatenation of d
end-to-end connecting vectors Q j for each spring, i.e. X = (Q1 , . . . , Qd ). For
homogeneous flows, the pdf is of the form Ψ (Q1 , . . . , Qd , t) and its evolution is
Fig. 1.4 PGD simulation of steady-state simple shear flow of the Multi-Bead-Spring FENE model
(10 springs with 2D orientation). First mode for end spring (left) and central spring (right)
governed by a Fokker-Planck equation of the general form (1.5). In order to demon-

strate the applicability of the PGD, the authors considered steady-state, simple shear
flow calculations for chains with 10 springs having 2D orientation [1]. The distribu-
tion function, thus defined in a space of dimension 20 (!), is sought in the separated
form
N
Ψ (Q1 , . . . , Q10 ) ≈ F j1 (Q1 ) × · · · × F j10 (Q10 ). (1.35)
j=1
A mesh consisting of 104 nodes was used for approximating each function F jk (Qk ). A
√
simple shear flow was applied (W e = 2) and 3 terms in the separated representation
were found sufficient to accurately describe the steady-state distribution function. We
depict in Fig. 1.4 the functions defining the first mode for the end and central springs,
i.e. functions F1k (Qk ), k = 1 and 5. Notice that the results are sharper at the center of
the chain, indicating that the central springs are more stretched than the ones located
near the chain ends.
This particular simulation implied 10 × 104 = 105 degrees of freedom (10 spring
connectors whose approximation functions were defined using a mesh of 104 nodes).
A standard finite element solution would have required of the order of (104 )10 = 1040
degrees of freedom for computing an equivalent solution.
1.5 Micro-Macro Simulations Using the Fokker-Planck Approach 27
1.5.3 Fokker-Planck Simulations of Complex Flows
As in the stochastic approach, solution of the conservation and Fokker-Planck equa-

tions are decoupled at each time step [31]. The new velocity field is inserted in the
Fokker-Planck equation, which is solved in configuration and physical spaces to
update the distribution function ψ. To this end, use of a fully implicit method would
be prohibitively expensive. It is imperative to somehow decouple the computation in
physical space Ω from that in configuration space. In [38], the authors achieve this
by splitting each time step into two consecutive sub-steps,
ψ̃ − ψn ∂ 1 ∂ ∂
=− · { A ψn } + : { D ψn }, (1.36)
Δt ∂X 2 ∂X ∂X
ψn+1 − ψ̃
+ v n · ∇ψn+1 = 0. (1.37)
Δt
Equation (1.36) for the intermediate value ψ̃ is defined at each collocation point of
the spectral element mesh covering the flow domain Ω. It is solved by means of a
Galerkin spectral method. The first sub-step is thus an ensemble of explicit, local
updates in configuration space. The second sub-step (1.37) is an implicit, global
update of the distribution function in physical space. This has been implemented for
various low-dimensional kinetic theory models (NC = 2 or 3).
Being relatively new, the PGD remains to be fully exploited for simulating com-
plex polymer flows. It has however been applied in the context of complex flows of
fiber suspensions [42, 47], which, as we shall see in the next chapter, involve micro-
macro governing equations that have a very similar generic form as those discussed
here. In [42], the authors considered a suspension of rod-like rigid particles which
can flocculate to create aggregates. These aggregates are continuously broken by the
flow. Thus, aggregation and disaggregation mechanisms coexist and two populations
of particles can be identified: the one related to free rods (pendant population) and
the one associated with the aggregated rods (active population). The kinetic theory
description of such systems contains two coupled Fokker-Planck equations for the
orientation distribution Ψ (p, x, t) and Φ(p, x, t) of the active and pendant rigid rods,
respectively. Here, the configuration of rods, i.e. their orientation, is given by the unit
vector p.
The authors [42] considered the flow in a converging channel. The steady-state
flow kinematics (assumed undisturbed by the presence of the suspended particles)
were computed by solving the Stokes equations. The coupled Fokker-Planck equa-
tions were integrated along individual flow streamlines, with separated representa-
tions for both orientation distribution functions:
N
Ψ (p, s) PiΨ (p) · SiΨ (s)
≈ . (1.38)
Φ(p, s) PiΦ (p) · SiΦ (s)
i=1
Here, s denotes the curvilinear coordinate along a chosen streamline.

Fig. 1.5 Fokker-Planck simulation of a complex flow of a two-population fiber suspension [42].
Orientation distribution of active (Ψ ) and pendant (Φ) populations in a converging flow. The orien-
tation distribution is represented on the unit sphere at various positions along individual streamlines
Figure 1.5 shows the resulting orientation distribution of both populations at

particular points along individual flow streamlines. The orientation distribution is
directly depicted on the unit surface, and the colour scheme indicates the intensity
of the orientation in each direction.
1.6 The Impact of Closure Approximations
In [28], we have shown for the simplest non-linear kinetic theory of dilute polymeric
solutions (FENE dumbbells), that the Peterlin closure approximation used to derive a
macroscopic constitutive equation (FENE-P) has a significant impact on the statistical
and rheological properties of the model. To do so, we have compared results obtained
for given flow kinematics with both the FENE model and its FENE-P approximation.
Figure 1.6 shows typical results for start-up of elongational flow, in terms of the
temporal evolution of the mean square polymer extension and the polymer contribu-
tion to the stress. Results for the FENE model are obtained either by solving the FENE
stochastic differential equation (line) or the Fokker-Planck equation (circles) given
1.6 The Impact of Closure Approximations 29
Fig. 1.6 Start-up of elongational flow for FENE and FENE-P fluids [28]
Fig. 1.7 Inception of elongational flow for FENE and FENE-P fluids [28]
in Sect. 1.3.4.1. In this particular example, the final steady-state values reached with
FENE and FENE-P models are identical, while the transients differ significantly. This
result, obtained for specified homogeneous kinematics, is relevant to more complex
flow situations. Indeed, in a complex flow field, the polymer molecules experience a
time-dependent velocity gradient (at least in the Lagrangian sense) as they are con-
vected along the flow trajectories. We thus expect drastic differences in simulations
of complex flows between the FENE kinetic theory and its approximate FENE-P
version, as we shall see in Sect. 1.7.3.
Differences between FENE and FENE-P results are even more drastic when one
looks at their statistical properties. We show in Fig. 1.7 the distribution of dumbbell
length obtained in start-up of elongational flow. A total of 103 FENE or FENE-P
dumbbells were used in these stochastic simulations, whose results are shown here
in the form of histograms. The initial condition is that of equilibrium. For FENE
dumbbells, the distribution evolves in time towards a highly-extended state localized
near the maximum dumbbell length Q 0 (green dot). The FENE-P dumbbells behave
in a very different manner. Indeed, many individual dumbbells do cross the limit
Q 0 as they are deformed by the flow field. Moreover, the steady-state distribution is
Fig. 1.8 Start-up of elongation followed by relaxation [37]. FENE model (thick curve), FENE-P
closure (dotted curve), two FENE-L closures (other curves)
drastically different from that obtained with the FENE dumbbells. In fact, it can be
shown that the configuration distribution function for FENE-P dumbbells is always
Gaussian, and thus never localized, whatever the flow kinematics. Thus, nothing pre-
vents individual FENE-P dumbbells from deforming beyond the maximum allowed
length Q 0 of the FENE theory. It is only the average Q2 that is bounded for FENE-P
dumbbells. Clearly, the Peterlin closure approximation has somewhat polluted the
very meaning of the parameter Q 0 of the underlying molecular theory. These results
have motivated the development of more sophisticated and much more accurate clo-
sure approximations known as FENE-L models [36, 37].
The above results do not tell the whole story. In [37, 50], we have considered
inception of elongational flow followed by relaxation. Figure 1.8 compares the tem-
poral evolutions of the polymer stress and mean square molecular extension obtained
with the FENE model and its various FENE-P and FENE-L approximations. In this
numerical experiment, cessation of flow occurs at dimensionless time of 1.5.
Clearly, the two FENE-L closure models are more accurate than FENE-P. But
more importantly, it is useful to plot, for the same numerical experiment, the polymer
stress as a function of molecular extension for various values of time, as indicated
by the dots in Fig. 1.8. The resulting plot, shown in Fig. 1.9, is quite surprising in
that the FENE model shows a hysteresis. This feature is highly significant as it
rules out the existence of a stress-optic law for FENE dumbbells wherein the stress-
optic coefficient would be a function of the second moment of the configuration
distribution function alone [50]. While the two FENE-L closure models also predict
a hysteresis in qualitative agreement with the parent FENE theory, the Peterlin closure
approximation (red curve) simply obliterates hysteric behaviour altogether.
Thus, we see that closure approximations can have a dramatic qualitative impact
indeed. A similar conclusion has been reached in the study of chaotic dynamics in
sheared liquid crystalline polymers [20].
1.7 Illustrative Examples of Complex Flow Simulations 31
Fig. 1.9 Hysteresis in start-up of elongation followed by relaxation [37]
1.7 Illustrative Examples of Complex Flow Simulations
In this section, we present a selection of complex polymer flow simulations obtained

over the years in our group. The discussion is by necessity extremely brief, and the
reader is invited to consult the references for a deeper exposition.
1.7.1 Early Macroscopic Simulations
We first show macroscopic simulation results obtained in the early 1980s. While the
numerical techniques developed in these pioneering days are now for the most part
outdated, these results were then the first successful attempts at predicting significant
viscoelastic effects in complex flows. Also shown are corresponding experimental
observations [5].
In Fig. 1.10, we consider die swell of an Oldroyd-B fluid, computed with a mixed
finite element method and a suitable algorithm to handle the a priori unknown free
surface [11]. Significant die swell is predicted at high We. A zoom of the mesh close to
the die exit is shown, as well as the predicted streamlines. While the mesh would now
be considered rather coarse, these calculations were then taxing available computer
resources to their limit.
Prediction of the vortex growth observed in steady-state flow of dilute solutions
through an abrupt contraction has been another early challenge in this community.
Figure 1.11 shows the finite element results [24] obtained with the Phan Thien-Tanner
Fig. 1.10 Die swell: early numerical results for the Oldroyd-B fluid [11]
Fig. 1.11 Flow through a contraction: early numerical results for the Phan Thien-Tanner fluid [24]
Fig. 1.12 Jet breakup: early numerical results for the Oldroyd-B fluid [6, 26]
fluid, a constitutive equation derived from the kinetic theory of temporary polymeric
networks. The predicted viscoelastic vortex is indeed significantly bigger than its
Newtonian counterpart.
Finally, we show in Fig. 1.12 the simulation of viscoelastic jet breakup obtained
with the Oldroyd-B fluid [26]. This is a moving boundary problem driven by sur-
face tension. We show snapshots of the axisymmetric computational domain, which
extends over half the wave length of the disturbance, as well as a comparison
between experiments and numerical predictions for the temporal evolution of the
jet radius disturbance [6]. Quantitative agreement with experimental data for a dilute
polymer solution is obtained, revealing the significant stabilizing effect of elonga-
tional stresses. Access to a supercomputer (a Cray X-MP) was mandatory to run
these demanding simulations.
1.7.2 Macroscopic Simulations with Advanced Models
Here, we discuss recent simulation results [58] obtained with constitutive equations
and a numerical technique that are representative of the current state of the art.
The rheological models are the so-called decoupled or coupled Double-Convection-
Reptation models with Chain Stretch (DCR-CS) [40, 57]. These have been derived
via a tube theory that goes much beyond the Doi-Edwards model discussed in
Fig. 1.13 The decoupled double-convection-reptation model with chain stretch [40]
Fig. 1.14 Zoom of 4:1:4 constriction geometry with rounded corners and medium-size mesh [58]
Sect. 1.3.4.2. Both models account, but in a different manner, for reptation, convective
constraint release, and Rouse relaxation. The equations for the decoupled DCR-CS
model are shown in Fig. 1.13. Here, the orientation tensor S and tube stretch λ are
pre-averaged separately and each is governed by its own evolution equation. The
model has 4 material parameters, shown in red.
In [58], we considered the start-up flow through an axisymmetric 4:1:4 constric-
tion with rounded corners, depicted in Fig. 1.14 along with the mesh. Steady-state
regimes were obtained as a long-time limit of transient BLPM simulations starting
from the equilibrium state.
Fig. 1.15 Boundary layer for the second invariant of the rate of strain tensor [58]
Fig. 1.16 DCR-CS models: predicted vortex growth with increasing Weissenberg number [58]
Figure 1.15 displays the contours of the second invariant of the rate-of-deformation
tensor at the highest Weissenberg number for which we have performed calculations,
W e = 50. A boundary layer develops near the constriction wall. This is quite a
challenge for any numerical technique, and is typical of viscoelastic flows in con-
fined complex geometries.
Steady-state streamlines are shown in Fig. 1.16. With both coupled and decoupled
models, a significant vortex growth is predicted.
High-We results obtained with the coupled DCR-CS model are given in Fig. 1.17.
As shown in Fig. 1.18, large levels of tube stretch, close to the specified maximum
value λmax = 2.5, are predicted by the coupled DCR-CS model at high-We. These
correlate with significant upstream vortex activity.
Fig. 1.17 Results obtained with the coupled DCR-CS model at high Weissenberg number [58]
Fig. 1.18 Tube stretch predicted with the coupled DCR-CS model [58]
1.7.3 Impact of Closure Approximations in Complex Flows
We end this section with an illustrative example of the use of the micro-macro sto-
chastic approach for simulating a complex flow. In [55], we used BLPM to compute
the transient flow of a FENE fluid in a 4:1 contraction, starting from rest. The same
flow was also simulated with the FENE-P model, with identical flow and material
parameter values, in order to assess the impact of the closure approximation.
Figure 1.19 shows the temporal evolution of the streamlines for both FENE and
FENE-P fluids. A steady-state is reached at a dimensionless time of about 10. It is
clear from these results that the Peterlin closure approximation has a dramatic impact
on the predicted macroscopic flow features: the FENE model exhibits a corner and
a lip vortex, while the FENE-P approximation shows a big corner vortex only.
1.8 Discussion
Following [29], we end this chapter with a few personal remarks on computational
rheology.
1.8 Discussion 37
Fig. 1.19 Start-up flow through a contraction computed with BLPM: FENE versus FENE-P [55]
1.8.1 From Numbers to Rheology
From a user’s viewpoint, any serious exercise in computational rheology should in

our opinion follow a three-step path: (1) get the numbers, (2) check their numerical
accuracy, and (3) check their physical accuracy.
At the outset, step (1) was found to be by no means a trivial matter: obtaining
numerical solutions of the discrete, non-linear algebraic equations at significant val-
ues of the Weissenberg number We has long been difficult or even impossible (this
is known as the High Weissenberg Number Problem or HWNP, which is discussed
in detail in [25, 27]).
Though overlooked at times, the importance of step (2) is simply paramount.
One must indeed convince oneself and others that the numerical results obtained
constitute sufficiently accurate approximations of the exact solutions of the governing
equations. Since the mathematical analysis of numerical methods for viscoelastic
fluids is quite difficult and of limited applicability to complex flows, step (2) is usually
performed by means of careful (and often expensive) mesh refinement experiments.
Finally, step (3) questions the very validity of the physical model that is imple-
mented in the computer simulation (constitutive equation or molecular model, val-
ues of the material parameters, initial and boundary conditions). It is performed by
comparing simulation results (such as flow kinematics, overall pressure drop, and
birefringence) to available experimental observations. Since the pioneering days,
overall progress along this three-step path has been sometimes chaotic but overall
quite steady. Difficulties do remain, however, some of which we now wish to discuss
briefly.
1.8.2 The Infamous HWNP and Mesh Convergence
It is fair to state that the HWNP is now partially resolved, in that high-We numerical
solutions have been reported over the years for a variety of flow problems. It could be
argued, however, that viscoelastic flow computations are not yet the routine and safe
procedure that more classical sub-fields of computational mechanics (e.g. numerical
linear elasticity) have generated, and that their success is not guaranteed even in the
very restricted sense of “getting the numbers”.
Many, if not most, viscoelastic flow problems of interest are characterized by stress
singularities or boundary layers of essentially unknown mathematical structure. In
this context, an outside observer should not necessarily be impressed by the high-We
solutions reported in the literature. Indeed, some have been generated with numerical
schemes that either explicitly or implicitly smooth out the difficulties. A similar issue
arises in classical CFD of Newtonian fluids, in the quest for numerical solutions
at high Reynolds numbers. Getting high-Re or high-We (smooth) solutions using
risky, “false-diffusion” numerical schemes often essentially means that the numerics
have computed an approximate solution to a different problem than the one under
investigation.
One could argue that mesh refinement experiments (step (2) alluded to above)
would readily settle the matter of numerical accuracy. When feasible, as far as
computer resources are concerned, these are unfortunately not always successful
in that the range of Weissenberg numbers that can be covered in the simulations may
decrease as the grid is further refined [25, 27]. This problem often occurs in the
presence of a stress singularity, thus preventing the safe interpretation of the rheo-
logical results obtained in the vicinity of the singular point. One should also keep
in mind that a mesh refinement experiment, while essential in assessing the accu-
racy of the numerical results, does not constitute a rigourous mathematical proof of
convergence.
In a seminal paper [15], the authors have shown that a large class of differential
constitutive models can be transformed into an equation for the matrix logarithm
of the conformation tensor. Under this transformation, the extensional components
of the deformation field act additively, rather than multiplicatively. As discussed in
[15], this transformation is motivated by numerical evidence that the HWNP may
be caused by the failure of polynomial-based approximations to properly represent
exponential profiles developed by the conformation tensor. The log-conformation
approach significantly alleviates the HWNP [16, 23].
Finally, the above discussion does not imply, of course, that all results reported
in the literature for high values of We are necessarily inaccurate. We simply wish
to stress that the significant advances made over the last two decades in expanding
the field to more detailed levels of description of polymeric liquids should not hide
the absolute necessity of careful numerical validation of present and future methods,
whether they be macroscopic, micro-macro, or atomistic. The role of benchmark
flow problems is crucial in this regard [12].
1.9 Conclusions 39
1.9 Conclusions
The multi-scale modeling of viscoelastic flows is an exciting development in compu-

tational rheology, which nicely complements the efforts made by the community on
more conventional continuum simulations. Micro-macro methods allow the analysis
of complex flow using a kinetic theory model that does not have an equivalent closed-
form constitutive equation. The difficult theoretical problem of obtaining accurate
closure approximations is thus bypassed altogether, and the computational rheol-
ogist gains access to a potentially much wider modeling space. Micro-macro tech-
niques are of course more demanding in computer resources than continuum methods,
but they are generally easier to implement efficiently on parallel computers. Until
recently, available stochastic and Fokker-Planck micro-macro techniques have been
implemented for models of kinetic theory having but few configurational degrees of
freedom. Introduction of the Proper Generalized Decomposition has changed the sit-
uation drastically, as this new technique opens the way for micro-macro simulations
with more advanced kinetic theory models.
This short survey of modeling and numerical approaches in computational rheol-
ogy applied to polymeric liquids demonstrates that the field has expanded consider-
ably in scope since its inception thirty five years ago. In the early days, researchers
were struggling to compute the flow of memory liquids using constitutive equa-
tions of very limited (though not altogether vanishing) value for rheological studies,
such as the second-order fluid and the upper-convected Maxwell model. Computing
with these “simple” constitutive equations is very difficult indeed, as the community
soon discovered with surprise and dismay. The problems met then by all players in
the field launched a series of numerical and mathematical investigations, constantly
nourished by challenging experimental observations such as those on Boger fluids,
which significantly increased our understanding of the basic issues. In particular,
that more complex and realistic models are actually easier to exploit in numerical
studies, once the proper numerical methodology has been developed, is a welcome
and healthy fact that has long been reported and continuously been witnessed since.
Computational rheology has clearly gone a long way since the first attempts (back in
the mid 1970s) to predict the flow of a memory fluid in a complex geometry. Today,
the most sophisticated constitutive equations and molecular models developed by
theoretical rheologists can in principle be implemented numerically. Furthermore,
our colleagues active in experimental work make an increasing use of simulation
software to gain further insight into their observations on real materials. Finally,
very detailed molecular dynamics simulations are becoming feasible that should
help us to elucidate difficult issues such as wall rheology. These facts are a clear
signature of progress. Difficult issues do remain, however, which point to the ever-
present need for the careful evaluation of numerical techniques in benchmark flow
problems. The numerical challenges met by computational rheologists are intimately
linked to mathematical and physical considerations. This makes computational rhe-
ology fascinating, difficult, and altogether fun. On the modeling side, computational
rheology will continue to be exploited for the evaluation and use in complex flows of
theoretical models for the polymer dynamics. It will help identify the actual math-
ematical structure of boundary layers and stress singularities observed in many cir-
cumstances. It will also help the theoretical rheologist assess the relative importance
of the particular physical mechanisms put forward in his or her theories. One could
also envisage using a modeling approach which would couple, in an automatic and
adaptive fashion, different levels of description of the same fluid in different regions
of the flow. For example, detailed atomistic models could be used to describe the
solid wall and its immediate fluid vicinity, while a series of kinetic theory and contin-
uum mechanics models of decreasing levels of detail could be used at progressively
larger distances from the wall. The proper matching between the different modeling
domains would be done automatically. Most probably, very different numerical tech-
niques, best suited to the local level of description, would be used in each domain.
The corresponding numerical scheme would thus be adaptive both in numerical and
modeling terms. These indeed are exciting and challenging perspectives for bright
students now entering the field.
References
1. A. Ammar, B. Mokdad, F. Chinesta, R. Keunings, A new family of solvers for some classes
of multidimensional partial differential equations encountered in kinetic theory modeling of
complex fluids. J. Non-Newtonian Fluid Mech. 139, 153–176 (2006)
complex fluids. Part II. J. Non-Newtonian Fluid Mech. 144, 98–121 (2007)
3. R.B. Bird, R.C. Armstrong, O. Hassager, Dynamics of Polymeric Liquids, Volume 1: Fluid
Mechanics (Wiley, New York, 1987)
4. R.B. Bird, C.F. Curtiss, R.C. Armstrong, O. Hassager, Dynamics of Polymeric Liquids, Volume
2: Kinetic Theory (Wiley, Mew York, 1987)
5. D.V. Boger, K. Walters, Rheological Phenomena in Focus (Elsevier, New York, 1993)
6. D.W. Bousfield, R. Keunings, G. Marrucci, M.M. Denn, Nonlinear analysis of the surface
tension driven breakup of viscoelastic filaments. J. Non-Newtonian Fluid Mech. 21, 79–97
(1986)
7. B. Busic, J. Koplik, J.R. Banavar, Molecular dynamics simulation of liquid bridge extensional
flows. J. Non-Newtonian Fluid Mech. 109, 51–89 (2003)
8. F. Chinesta, A. Ammar, A. Leygue, R. Keunings, An overview of the Proper Generalized
Decomposition with applications in computational rheology. J. Non-Newtonian Fluid Mech.
166, 578–592 (2011)
9. F. Chinesta, R. Keunings, A. Leygue, The Proper Generalized Decomposition for Advanced
Numerical Simulations. A Primer (Springerbriefs) (Springer, New York, 2014)
10. M. Cieplak, J. Koplik, J.R. Banavar, Boundary conditions at a fluid-solid interface. Phys. Rev.
Lett. 86, 803–806 (2001)
11. M. Crochet, R. Keunings, Finite element analysis of die swell of a highly elastic fluid. J.
Non-Newtonian Fluid Mech. 10, 339–356 (1982)
12. F.A. Cruz, R.J. Poole, A.M. Afonso, F.T. Pinho, P.J. Oliveira, M.A. Alves, A new viscoelastic
benchmark flow: stationary bifurcation in a cross-slot. J. Non-Newtonian Fluid Mech. 214,
57–68 (2014)
13. J. Dealy, R.G. Larson, Structure and Rheology of Molten Polymers (Hanser Publications,
Munich, 2006)
References 41
14. M. Doi, S.F. Edwards, The Theory of Polymer Dynamics (Clarendon Press, Oxford, 1987)
15. R. Fattal, R. Kupferman, Constitutive laws for the matrix-logarithm of the conformation tensor.
J. Non-Newtonian Fluid Mech. 123, 281–285 (2004)
16. R. Fattal, R. Kupferman, Time-dependent simulation of viscoelastic flows at high Weissenberg
number using the log-conformation representation. J. Non-Newtonian Fluid Mech. 126, 23–37
(2005)
17. X. Gallez, P. Halin, G. Lielens, R. Keunings, V. Legat, The adaptive Lagrangian particle method
for macroscopic and micro-macro computations of time-dependent viscoelastic flows. Comput.
Meth. Appl. Mech. Engng. 180, 345–364 (1999)
18. W.W. Graessley, Polymeric Liquids and Networks: Structure and Properties (Garland Science,
New York, 2003)
19. W.W. Graessley, Polymeric Liquids and Networks: Dynamics and Rheology (Garland Science,
New York, 2008)
20. M. Grosso, R. Keunings, S. Crescitelli, P.L. Maffettone, Prediction of chaotic dynamics in
sheared liquid crystalline polymers. Phys. Rev. Letters 86, 3184–3187 (2001)
21. P. Halin, G. Lielens, R. Keunings, V. Legat, The Lagrangian particle method for macroscopic
and micro-macro viscoelastic flow computations. J. Non-Newtonian Fluid Mech. 79, 387–403
(1998)
22. M.A. Hulsen, A.P.G. van Heel, B.H.A.A. van den Brule, Simulation of viscoelastic flows using
Brownian configuration fields. J. Non-Newtonian Fluid Mech. 70, 79–101 (1997)
23. M.A. Hulsen, R. Fattal, R. Kupferman, Flow of viscoelastic fluids past a cylinder at high
Weissenberg number: stabilized simulations using matrix logarithms. J. Non-Newtonian Fluid
Mech. 127, 27–39 (2005)
24. R. Keunings, M.J. Crochet, Numerical simulation of the flow of a viscoelastic fluid through an
abrupt contraction. J. Non-Newtonian Fluid Mech. 14, 279–299 (1984)
25. R. Keunings, On the high Weissenberg number problem. J. Non-Newtonian Fluid Mech. 20,
209–226 (1986)
26. R. Keunings, An algorithm for the simulation of transient viscoelastic flows with free surfaces.
J. Comput. Phys. 62, 199–220 (1986)
27. R. Keunings, Simulation of viscoelastic fluid flow, in Fundamentals of Computer Modeling for
Polymer Processing, ed. by C.L. Tucker III (Carl Hanser Verlag, Munich, 1989), pp. 402–470
28. R. Keunings, On the Peterlin approximation for finitely extensible dumbbells. J. Non-
Newtonian Fluid Mech. 68, 85–100 (1997)
29. R. Keunings, A Survey of Computational Rheology. Plenary Lecture, ed. by D. M. Binding et al.
Proceedings of 13th International Congress on Rheology, vol. 1 (British Society of Rheology,
Glasgow, 2000), pp. 7–14
30. R. Keunings, Finite Element Methods for Integral Viscoelastic Fluids, ed. by D.M. Binding,
K. Walters. Rheology Reviews (British Society of Rheology, Glasgow, 2003), pp.167–195
31. R. Keunings, Micro-macro Methods for the Multi-scale Simulation of Viscoelastic Flow Using
Molecular Models of Kinetic Theory, ed. by D.M. Binding, K. Walters. Rheology Reviews
(British Society of Rheology, Glasgow, 2004), pp. 67–98
32. R. Koopmans, J. Den Doelder, J. Molenaar, Polymer Melt Fracture (CRC Press, Boca Raton,
2010)
33. J. Koplik, J.R. Banavar, Re-entrant corner flows of Newtonian and non-Newtonian fluids. J.
Rheol. 41, 787–805 (1997)
34. R.G. Larson, The Structure and Rheology of Complex Fluids (Oxford University Press, New
York, 1999)
35. M. Laso, H.C. Öttinger, Calculation of viscoelastic flow using molecular-models: the CON-
NFFESSIT approach. J. Non-Newtonian Fluid Mech. 47, 1–20 (1993)
36. G. Lielens, P. Halin, I. Jaumain, R. Keunings, V. Legat, New closure approximations for the
kinetic theory of finitely extensible dumbbells. J. Non-Newtonian Fluid Mech. 76, 249–279
(1998)
37. G. Lielens, R. Keunings, V. Legat, The FENE-L and FENE-LS closure approximations to the
kinetic theory of finitely extensible dumbbells. J. Non-Newtonian Fluid Mech. 87, 179–196
(1999)
38. A. Lozinski, C. Chauvière, J. Fang, R.G. Owens, Fokker-Planck simulations of fast flows of
melts and concentrated polymer solutions in complex geometries. J. Rheol. 47, 535–561 (2003)
39. C. Mangoubi, M.A. Hulsen, R. Kupferman, Numerical stability of the method of Brownian
configuration fields. J. Non-Newtonian Fluid Mech. 157, 188–196 (2009)
40. G. Marrucci, G. Ianniruberto, Flow-induced orientation and stretching of entangled polymers.
Philos. Trans. R. Soc. A 361, 677–688 (2003)
41. Y. Masubuchi, Simulating the flow of entangled polymers. Annu. Rev. Chem. Biomol. Eng. 5,
11–33 (2014)
42. B. Mokdad, A. Ammar, M. Normandin, F. Chinesta, J.R. Clermont, A fully deterministic micro-
macro simulation of complex flows involving reversible network fluid models. Math. Comput.
Simul. 80, 1936–1961 (2010)
43. H.C. Öttinger, M. Laso, Smart polymers in finite-element calculations, in Theoretical and
Applied Rheology, ed. by P. Moldenaers, R. Keunings, Proceedings of XIth International
Congress on Rheology, vol. 1 (Elsevier, Amsterdam, 1992), pp. 286–288
44. H.C. Öttinger, Stochastic Processes in Polymeric Fluids (Springer, Berlin, 1996)
45. R.G. Owens, T.N. Phillips, Computational Rheology (Imperial College Press, London, 2002)
46. E.A.J.F. Peters, M.A. Hulsen, B.H.A.A. van den Brule, Instationary Eulerian viscoelastic flow
simulations using time separable Rivlin-Sawyers constitutive equations. J. Non-Newtonian
Fluid Mech. 89, 209–228 (2000)
47. E. Pruliere, A. Ammar, N. El Kissi, F. Chinesta, Recirculating flows involving short fiber
suspensions: Numerical difficulties and efficient advanced micro-macro solvers. Arch. Comput.
Methods Eng, State Art Rev 16, 1–30 (2009)
48. H.K. Rasmussen, Catastrophic failure of polymer melts during extension. J. Non-Newtonian
Fluid Mech. 198, 136–140 (2013)
49. M. Rubinstein, R.H. Colby, Polymer Physics (Oxford University Press, Oxford, 2003)
50. R. Sizaire, G. Lielens, I. Jaumain, R. Keunings, V. Legat, On the hysteretic behaviour of dilute
polymer solutions in relaxation following extensional flow. J. Non-Newtonian Fluid Mech. 82,
233–253 (1999)
51. P.S. Stephanou, C. Baig, V.G. Mavrantzas, Projection of atomistic simulation data for the
dynamics of entangled polymers onto the tube theory: calculation of the segment survival
probability function and comparison with modern tube models. Soft Matter 7, 380–395 (2011)
52. R.I. Tanner. Engineering Rheology (Oxford University Press, Oxford, 2000)
53. A.P.G. van Heel, M.A. Hulsen, B.H.A.A. van den Brule, Simulation of the Doi-Edwards model
in complex flow. J. Rheol. 43, 1239–1260 (1999)
54. M.M. Villone, F. Greco, M.A. Hulsen, P.L. Maffettone, Simulations of an elastic particle in
Newtonian and viscoelastic fluids subjected to confined shear flow. J. Non-Newtonian Fluid
Mech. 210, 47–55 (2014)
55. P. Wapperom, R. Keunings, V. Legat, The backward-tracking lagrangian particle method for
transient viscoelastic flows. J. Non-Newtonian Fluid Mech. 91, 273–295 (2000)
56. P. Wapperom, R. Keunings, Numerical simulation of branched polymer melts in transient
complex flows using pom-pom models. J. Non-Newtonian Fluid Mech. 97, 267–281 (2001)
57. P. Wapperom, R. Keunings, G. Ianniruberto, Prediction of rheometrical and complex flows
of entangled linear polymers using the DCR model with chain stretch. J. Rheol. 47, 247–265
(2003)
58. P. Wapperom, R. Keunings, Impact of decoupling approximation between stretch and orienta-
tion in rheometrical and complex flow of entangled polymers. J. Non-Newtonian Fluid Mech.
122, 33–43 (2004)
59. P. Wapperom, A. Leygue, R. Keunings, Numerical simulation of large amplitude oscillatory
shear of a high-density polyethylene melt using the MSF model. J. Non-Newtonian Fluid Mech.
130, 63–76 (2005)
Chapter 2
Complex Flows of Micro/Nano Structured
Fluids: Reinforced Polymer Composites
Abstract The motion of an ellipsoidal particle immersed in a Newtonian fluid was

studied in the pioneering work of Jeffery in 1922. Suspensions of industrial interest
usually involve particles with a variety of shapes. Moreover, suspensions composed
of rods (a limit case of an ellipsoid) aggregate, leading to clusters with particular
shapes that exhibit, when immersed in a flow, an almost rigid motion. In this chapter,
we revisit the modeling and simulation of suspensions involving rods throughout the
different scales of description (microscopic, mesoscopic and macroscopic) and the
different concentration regimes (dilute, semi-dilute, semi-concentrated and concen-
trated), involving gradually richer physics.
2.1 Introduction
Suspensions involving particles can be described at the microscopic scale by tracking

the motion of each one of the particles involved in the system. This approach is
based on three main elements: (i) the knowledge of the equation governing the
particle motion in the fluid flow; (ii) the introduction of the particle effects on the
flow kinematics if coupled simulations are envisaged; and (iii) the availability of
computational resources for tracking efficiently millions of particles.
In dilute suspensions, the motion of ellipsoidal particles can be accurately
described by using Jeffery’s equation [1]. When the concentration becomes large
enough, interactions cannot be neglected any longer and the calculation becomes
more complex from the computational point of view. At this scale, currently avail-
able simulations remain quite far from the scenarios of industrial interest.
For circumventing the difficulties related to simulations at the microscopic scale,
these being more computational than conceptual, coarser models were introduced.
The interested reader can refer to [2] and the references therein for a review on
multi-scale approaches in the context of computational rheology.
Mesoscopic kinetic theory models result from coarsening microscopic
descriptions. In kinetic theory models the individuality of the particles is lost in
favour of a statistical description that substitutes the entities by a series of confor-
mation coordinates [3, 4]. For example, when considering a suspension of rods, the
DOI 10.1007/978-3-319-16757-2_2
44 2 Complex Flows of Micro/Nano Structured Fluids . . .
mesoscopic description consists in giving the fraction of rods that at position x and
time t are oriented along direction p. This information is contained in the probabil-
ity distribution function—pdf—whose conservation balance results in the so-called
Fokker-Planck equation governing its evolution. Fokker-Planck equations involve
the flow induced conformation evolution. In the case of a suspension of rods, the
flow induced conformation (orientation) evolution is given, as indicated above, by
Jeffery’s equation. Since the pdf depends on the physical coordinates (space and time)
and a series of conformational coordinates, the associated Fokker-Planck equation is
multidimensional. Standard mesh-based discretization techniques fail when address-
ing multidimensional models. This issue is known as the curse of dimensionality and
it explains the few number of existing works addressing the solution of kinetic theory
models within the Fokker-Planck framework.
For circumventing the curse of dimensionality at the mesoscopic scale, several
techniques based on the use of particles were proposed and widely employed. Here
the particles are not real particles, but rather should be viewed as computational
particles that allow one to describe the main suspension features (rheology, properties
related to the particles conformation, etc.). Despite the fact of considering a discrete
description, the level of detail in the description and the richness of the physics is
exactly the same that the one associated with the use of Fokker-Planck descriptions,
and obviously the solutions computed by using both descriptions are in the limit of
convergence exactly the same.
The use of the continuous description based on the solution of the Fokker-Planck
equation remains challenging because of the high dimensionality that it involves.
On the other hand, when employing its discrete counterpart, the main difficulty is
related to the extremely large number of particles to be considered. This number
depends on the model output of interest. When only the moments of the distribution
are concerned, a moderate number of particles is enough. However, when one is
interested in the pdf itself, the number of computational particles could become
extremely large.
Solution procedures based on the use of particles at the mesoscopic scale have
been extensively employed by many authors [5–13]. On the other hand, there are
few works focusing on the solution of Fokker-Planck equations by using standard
discretization techniques [7, 14]. We proposed some years ago a new solution tech-
nique called Proper Generalized Decomposition based on the use of separated rep-
resentations in order to ensure that the complexity scales linearly with the model
dimensionality [15, 16]. This technique consists in expressing the unknown field
as a finite sum of functional products, i.e. expressing a generic multidimensional
function u(x1 , . . . , xd ) as:

i=N
u(x1 , . . . , xd ) ≈ Fi1 (x1 ) · · · Fid (xd ). (2.1)
i=1
The interested reader can refer to [17–19] and the references therein for a deep
analysis of this technique and its applications in computational rheology.
2.1 Introduction 45
At the macroscopic scale, the pdf is substituted by some of its moments. Here
the level of detail and the involved physics are sacrificed in favour of computa-
tional efficiency. The equations governing the time evolution of these moments usu-
ally involve closure approximations whose impact on the results is unpredictable
[20, 21]. Alternatively, macroscopic equations are carefully postulated, within a top-
down approach, in order to guarantee the model objectivity and its thermodynamical
admissibility.
In the case of dilute suspensions of short fibers, the three scales have been
extensively considered to model the associated systems without major difficulties.
However, difficulties appear as soon as the concentration increases. In the semi-dilute
and semi-concentrated regimes, fiber-fiber interactions occur, but in general they can
be accurately modeled by introducing a sort of randomizing diffusion term [22].
There is a wide literature concerning dilute and semi-dilute suspensions, addressing
modeling [23–27], flows [28–32] and rheology [33, 34]. These models describe quite
well the experimental observations. When the concentration increases, rod interac-
tions cannot be neglected anymore and appropriate models addressing these intense
interactions must be considered, as for example the one proposed in [35]. Recent
experiments suggest that short fibers in concentrated suspensions align more slowly
as a function of strain than models based on Jeffery’s equation predict [36]. For
addressing this issue Wang el al. [36] proposed the use of a strain reduction fac-
tor, however this solution violates objectivity. Later, the same authors proposed an
objective model by decoupling the time evolution of both the eigenvalues and the
eigenvectors of the second-order orientation tensor [37]. In [38] an anisotropic rotary
diffusion is proposed for accounting for the fiber-fiber interactions and the model
parameters were selected by matching the experimental steady-state orientation in
simple shear flow and by requiring stable steady states and physically realizable
solutions.
The worst scenario concerns concentrated suspensions involving entangled
clusters exhibiting aggregation/disaggregation mechanisms. A first approach in that
sense was proposed in [39]. The first natural question is how to describe such systems.
At the macroscopic scale, one could try to fit some power-law constitutive equation,
however, this description does not hold for the microstructure. At the microscopic
scale, direct numerical simulations describing complex fiber-fiber interactions can
be carried out in small enough representative volumes [40–42]. A natural candidate
to be a reasonable compromise between (fine) micro and (fast) macro descriptions
consists in considering again a kinetic theory description.
The main issue with such an approach lies in the fact that it must include two
scales, the one involving the aggregates and the one related to the rods compos-
ing the aggregates. What are the appropriate conformational coordinates? How to
determine the time evolution of these conformational coordinates? How to represent
simultaneously both scales, the one related to the aggregates and the other related to
the fibers? How to derive the interaction mechanisms?
In [43], the authors propose a first attempt to describe such clusters from a micro-
mechanical model. Later in [44], the authors compared the model predictions with
direct numerical simulation in the case of rigid and deformable clusters. An enriched
description taking into account the polidispersity of fibers constituting the cluster
within a multi-scale framework was addressed in [45] in the case of rigid clusters.
In this chapter, we start by revisiting in Sect. 2.2 the multi-scale description
of dilute and semi-dilute suspensions of rigid rods, whose numerical solution is
addressed in Sect. 2.3. More concentrated regimes involving rod clustering are con-
sidered in Sect. 2.4, and finally in Sect. 2.5 we consider some advanced topics related
to work in progress.
Remark 1 In the present chapter we consider the following tensor products, where
Einstein’s summation convention is assumed:
• if a and b are first-order tensors, the single contraction “·” reads (a · b) = a j b j ;
• if a and b are first-order tensors, the dyadic product “⊗” reads (a ⊗ b) jk = a j bk ;
• if a and b are first-order tensors, the cross product “×” reads (a×b) j = ε jmn am bn ,
where ε jmn are the components of the Levi-Civita tensor ε (also known as permu-
tation tensor);
• if a and b are respectively second and first-order tensors, the single contraction
“·” reads (a · b) j = a jm bm ;
• if a and b are second-order tensors, the single contraction “·” reads (a · b) jk =
a jm bmk ;
• if a and b are second-order tensors, the double contraction “:” reads (a : b) =
a jk bk j .
• if a and b are third-order tensors then the triple contraction “∵” reads (a ∵ b) =
a jkm bmk j ;
• if a and b are fourth-order tensors then the fourth contraction “::” reads (a :: b) =
a jkmn bnmk j .
2.2 Dilute and Semi-dilute Suspensions
2.2.1 Multi-scale Description of Non-brownian Rod

Suspensions
We consider a suspending medium consisting of a Newtonian fluid of viscosity

η in which there are suspended rigid and non-Brownian rods. We assume as first
approximation that their presence and orientation do not affect the flow kinematics
that is defined by the velocity field v(x, t), with x ∈ Ω ∈ Rd , d = 2, 3. The multi-
scale modeling involves nine main conceptual bricks, the first three related to the
microscopic description (conformation, kinematics and dynamics), the next three
related to the mesoscopic scale (conformation, kinematics and dynamics) and the
last three to the macroscopic description (conformation, kinematics and dynamics).
We describe all of them in what follows.
2.2 Dilute and Semi-dilute Suspensions 47
2.2.1.1 Particle Conformation
The conformation of each rod R of length 2L can be described from its orientation
given by the unit vector p located at the rod center of gravity G and aligned along
the rod axis. Inertial effects are neglected in the sequel.
2.2.1.2 Particle Conformation Evolution
The orientation evolution of an ellipsoidal particle is described by Jeffery’s equation

[1]. This equation can be derived in a very simple way illustrated in Fig. 2.1. We
consider a system consisting of a rod and two beads located at both rod ends where
we assume that hydrodynamic forces apply. We assume that the forces that apply on
each bead F depend on the difference of velocities between the fluid and the bead,
the first one given by v0 + ∇v · pL and the second one by vG + ṗL. Thus, the force
F(pL) reads:
F(pL) = ξ(v0 + ∇v · pL − vG − ṗL), (2.2)
where ξ is the friction coefficient, v0 the fluid velocity at the rod center of gravity,
and vG the velocity of the center of gravity.
If F applies on the bead at pL, then the force on the opposite bead at −pL reads
F(−pL) = ξ(v0 − ∇v · pL − vG + ṗL). (2.3)
Fig. 2.1 Hydrodynamic forces applied on a rod immersed in a Newtonian fluid

By adding Eqs. (2.2) and (2.3), and neglecting inertial effects, we obtain the force
balance
F(pL) + F(−pL) = 2ξ(v0 − vG ) = 0, (2.4)
which implies v0 = vG , that is, the rod center of gravity is moving with the fluid
velocity. For simplicity of notation, we shall write F = F(pL) and F(−pL) = −F.
As the resulting torque must also vanish, the only possibility is that the force F
acts along p, that is F = λp, with λ ∈ R. Thus, we can write
λp = ξ L(∇v · p − ṗ). (2.5)
Premultiplying Eq. (2.5) by p and taking into account that p · p = 1 and conse-
quently p · ṗ = 0, we have
λ = ξ L (∇v : (p ⊗ p)) , (2.6)
which gives
ξ L (∇v : (p ⊗ p)) p = ξ L(∇v · p − ṗ). (2.7)
We have thus obtained Jeffery’s equation
ṗ = ∇v · p − (∇v : (p ⊗ p)) p. (2.8)
Remark 2 As the factor ξ L appears in both sides of Eq. (2.7), the rod kinematics
does not contain size effects.
2.2.1.3 Particle Contribution to the Stress
The forces applied at the rod ends pL and −pL are respectively λp and −λp, i.e.
directed along the rod and in equilibrium by construction.
With λ given by Eq. (2.6), we have
F(pL) = ξ L(∇v : (p ⊗ p))p. (2.9)
By applying Kramers’ formula, the corresponding contribution to the stress is

given by
τ p = pL ⊗ F = ξ L 2 (∇v : (p ⊗ p))p ⊗ p, (2.10)
which can be rewritten as
τ p = ξ L 2 ∇v : (p ⊗ p ⊗ p ⊗ p). (2.11)
2.2.1.4 Population Description
There are two natural descriptions of a population of N rods:

• The first one consists in specifying each rod orientation by considering the unit
vector aligned along its axis, that is, by considering pi , i = 1, . . . , N . As dis-
cussed in the next section, the main drawback of this approach lies in the necessity
of tracking the evolution of each “computational” rod by solving the correspond-
ing Jeffery equation, and even if conceptually there is no major difficulty, the
computing cost could be excessive in most practical applications.
• The second approach is the introduction of the PFD ψ(x, t, p) that gives the
fraction of rods that a position x and time t are oriented along direction p.
Despite the fact that both mesoscopic models involve the same physics and
richness of description, the main advantage of the second one is the manipulation of
a scalar continuous function instead of the discrete description involved in the first
approach. The price to be paid when using the description based on the use of the
pdf is its inherent multidimensionality, because in that framework the pdf depends
on the standard space and time coordinates and also on the conformation coordinates
that the microstructural description involves, i.e. p in the present case.
2.2.1.5 Description of the Population Evolution
• When the population is described from the individuals describing it, whose con-
formation is given by vectors pi , i = 1, . . . , N , the evolution of each one is given
by Jeffery’s equation:
ṗi = ∇v · pi + (∇v : (pi ⊗ pi ))pi , ∀i = 1, . . . , N . (2.12)
• The alternative description consists in using the pdf ψ(x, t, p) that satisfies the
normalisation condition:

ψ(x, t, p) dp = 1, ∀x, ∀t. (2.13)
S
with S the surface of the unit sphere, where vector p is defined.
Conservation of probability yields
∂ψ
+ ∇x · (ẋ ψ) + ∇ p · (ṗ ψ) = 0, (2.14)
∂t
where, for inertialess rods, ẋ = v(x, t), and the rod rotary velocity is given by
Jeffery’s equation:
ṗ = ∇v · p − (∇v : (p ⊗ p)) p. (2.15)
The price to pay is the increase of the model dimensionality, as the orientation
distribution is defined in a 6-dimensional domain, i.e. ψ : (x, t, p) → R+ where
x ∈ Ω ⊂ Rd , d = 3; t ∈ I ⊂ R+ , p ∈ S .
2.2.1.6 Contribution of the Particle Population to the Stress
Again, we consider the two alternative descriptions:

• When the population is described in a discrete manner by means of the pi vectors,
the contribution of rods to the suspension stress is calculated by adding their
individual effects, that is:
N
(x,t) N
(x,t)
τ (x, t) = τi = ξ L 2 ∇v : (pi ⊗ pi ⊗ pi ⊗ pi ), (2.16)
i=1 i=1
where N (x, t) refers to the number of computational rods located at time t in the
neighborhood of x.
• When the population is described from the pdf, the sum in Eq. (2.16) is replaced
by an integral in the conformation space S :

τ (x, t) = τ p ψ(x, t, p) dp
S
= 2ηN p ∇v : (p ⊗ p ⊗ p ⊗ p) ψ(x, t, p) dp. (2.17)
S
Here, the particle number N p accounts for the particle concentration and the viscos-
ity is used instead of the friction coefficient to be consistent with the usual notation.
In terms of the fourth-order orientation tensor

A= p ⊗ p ⊗ p ⊗ p ψ(x, t, p) dp, (2.18)
S
we obtain:
τ = 2ηN p (A : ∇v), (2.19)
which in view of the symmetry of A can be rewritten as
τ = 2ηN p (A : D). (2.20)
2.2.1.7 Macroscopic Description
As just discussed, discrete descriptions are computationally expensive because of

the large number of rods that must be considered in order to derive accurate-enough
model outputs. On the other hand, Fokker-Planck descriptions are rarely considered
in view of the curse of dimensionality that introduction of conformation coordinates
implies. Thus, standard mesh-based discretization techniques, as finite differences,
finite elements or finite volumes, fail when addressing models defined in high-
dimensional spaces.
For these reasons, mesoscopic models were coarsened to derive macroscopic mod-
els defined in standard physical domains, involving space and time. At the macro-
scopic scale, the orientation distribution function is substituted by its moments for
describing the microstructure [46]. Usually, macroscopic descriptions of rod suspen-
sions are based on the use of the first two non-zero moments, i.e. the second and the
fourth-order moments, a and A, defined by

a= p ⊗ p ψ(x, t, p) dp, (2.21)
S
and
A= p ⊗ p ⊗ p ⊗ p ψ(x, t, p) dp, (2.22)
S
respectively.
2.2.1.8 Microstructural Macroscopic Evolution
The microstructural evolution described at the macroscopic scale considers the time
evolution of the pdf moments. The time evolution of the second-order orientation
tensor is given by:

ȧ = (ṗ ⊗ p + p ⊗ ṗ) ψ dp
S
= (∇v · p − (∇v : (p ⊗ p)) p) ⊗ p ψ dp
S

+ p ⊗ (∇v · p − (∇v : (p ⊗ p)) p) ψ dp
S
= ∇v · a + a · (∇v)T − 2 A : ∇v. (2.23)
This equation involves the fourth-order moment A. The time derivative of the
fourth-order moment, using the same rationale, involves the sixth-order moment A ,
and so on.
Thus, an approximate closure relation is needed in order to express the fourth-order
moment A as a function of the lower-order moment a. Different closure relations have
been introduced and widely used [47–50]. With the quadratic closure relation (that is
only exact when all rods are locally aligned in the same direction), the fourth-order
moment is approximated as follows:
A ≈ a ⊗ a. (2.24)
This gives
ȧ ≈ ∇v · a + a · (∇v)T − 2 (∇v : a)a, (2.25)
and invoking again symmetry considerations,
ȧ ≈ ∇v · a + a · (∇v)T − 2 (D : a)a. (2.26)
2.2.1.9 Moment-Based Stress
We obtained previously the expression of the rod population contribution to the stress:
τ = 2ηN p (A : ∇v), (2.27)
which involves the fourth-order moment A. There is no closure issues when A is

calculated from the pdf ψ by using (2.22). When one proceeds at the macroscopic
scale, however, wherein the pdf is not available, a closure relation must be considered
for either
• writing A from the knowledge of a, itself being calculated by integrating (2.23)
with an appropriate closure relation (e.g. Eq. (2.25) when considering the quadratic
closure),
or
• calculating A by solving the equation that governs its time evolution in which, as
just commented, the sixth-order moment appears requiring again an appropriate
closure.
The first route is the simplest one and the most used in practice. It leads to
τ = 2ηN p (Acr (a) : ∇v), (2.28)
where the superscript “cr ” refers to the use of an appropriate closure relation.
With the quadratic closure, the stress reads:
τ = 2ηN p (a : ∇v)a ≡ 2ηN p (a : D)a. (2.29)
2.2.2 Multi-scale Description of Brownian Rods Suspensions
Until now Brownian effects were neglected as well as hydrodynamical interactions

between rods. The Brownian effects are due to the fluid’s molecular bombardment
acting on the beads. These effects were widely analyzed in [10, 51, 52] when focusing
on a microscopic description.
2.2.2.1 Microscopic Description
In this case the road beads are subjected to the hydrodynamical forces and the ones
coming from such bombardment. The first one was introduced previously
F H = ξ(∇v · pL − ṗL), (2.30)
where the superscript “H ” refers to its hydrodynamic nature. Now the Brownian
force F B is assumed to apply during a short time interval δt following a certain
statistical distribution concerning its magnitude and its orientation. The first one is
assumed described by a Gaussian distribution of zero mean and a certain standard
deviation and the one related to the orientation by a uniform distribution on the unit
circle (2D) or on the unit sphere (3D).
Brownian forces applying in the rod direction are assumed equilibrated. However,
the components of those forces perpendicular to the rod axis contribute to the rod
rotation, and then they affect the rod rotary velocity. In what follows, for the sake of
clarity, we restrict our analysis to the 2D case. Because the rod inertia is neglected,
the resultant torque vanishes, implying:
F H · t + F B · t = 0, (2.31)
where t is the unit vector tangent to the unit circle. By introducing in this balance
the expression of the hydrodynamic force, we obtain
FB · t
t T · ∇v · p − ṗ = − , (2.32)
ξL
where the fact that ṗ = ṗ t was taken into account. The previous equation can be
rewritten in the form:
FB · t
ṗ = t T · ∇v · p t + t, (2.33)
ξL
that, using the vectorial equivalence

T
t · ∇v · p t = ∇v · p − pT · ∇v · p p, (2.34)
results in
FB · t
ṗ = ∇v · p − pT · ∇v · p p + t
ξL
F B − (F B · p)p
= ∇v · p − pT · ∇v · p p + , (2.35)
ξL
where we can notice that the rotary velocity is given by the Jeffery expression ṗ J
complemented with a term describing the Brownian effects ṗ B :

ṗ J = ∇v · p − pT · ∇v · p p, (2.36)
and
F B − (F B · p)p
ṗ B = , (2.37)
ξL
from which
ṗ = ṗ J + ṗ B . (2.38)
Finally, we should discuss the effects of such Brownian contribution to the extra-
stress tensor.
Consider a rod aligned along the x-axis, such that pT = (1, 0) and the fluid at
rest. This rod is subjected to a continuous bombardment from the solvent molecules
(remember that the regime is dilute enough for neglecting the rod-rod interactions).
The components of forces aligned with the rod axis do not participate to the rod
rotation and by averaging on a time Δt δt the contributions of the components of
those forces along the rod direction vanish. On the contrary, the ones perpendicular
to the rod will participate to the stress. To derive the expression of this Brownian
contribution we consider that due to a Brownian force the rod rotates a small angle
δθ > 0, with the rod orientation being defined by pδθ . Considering the Brownian
force applying at that position [51, 52], i.e. F B · tδΘ , we obtain the contribution
to the virial stress given by:

sin(δθ ) · cos(δθ ) − cos2 (δθ )
− pδθ ⊗ F B
tδθ = F B
, (2.39)
sin2 (δθ ) − sin(δθ ) · cos(δθ )
where the negative sign accounts for the fact that the hydrodynamic force considered
in the virial stress applies in the opposite direction of the Brownian force.
We can notice in that expression two facts: (i) the trace is zero, and (ii) the
contribution is non-symmetric. However, we can imagine that in other rod, the same
Brownian force applies in the opposite direction, leading to an angle −δθ , from
which

sin(δθ ) · cos(δθ ) cos2 (δθ )
p−δθ ⊗ F B t−δθ = F B , (2.40)
− sin (δθ )
2 − sin(δθ ) · cos(δθ )
and then, after averaging, we obtain a Brownian contribution to the extra-stress due
to rods aligned on the x-direction (i.e. ϕ = 0)

β 0 1 0
τϕ=0
B
≈ =β = βU, (2.41)
0 −β 0 −1
that becomes almost symmetric and traceless. Here we considered a population

average, that, invoking ergodicity, is equivalent to averaging in Δt.
Now, for rods aligned in any other direction ϕ, it suffices to apply a rotation of
angle ϕ to tensor τϕ=0
B :
τϕB = βRϕT · U · Rϕ , (2.42)
with
cos(ϕ) sin(ϕ)
Rϕ = , (2.43)
− sin(ϕ) cos(ϕ)
that leads to:

1
cos2 (ϕ) sin(ϕ) · cos(ϕ) 0
τϕB = 2β − 2 . (2.44)
sin(ϕ) · cos(ϕ) sin2 (ϕ) 0 1
2
This can be written as

I
τpB = 2β p ⊗ p − , (2.45)
2
where I is the identity tensor.

For a population of rods pi , i = 1, . . . , N , the contribution of Brownian effects
is finally:

i=N
I
τ B = 2β pi ⊗ pi − . (2.46)
2
i=1
2.2.2.2 Mesoscopic Description
At the mesoscopic scale, we postulate that Brownian effects try to randomize the
rod orientation distribution, i.e. a mechanism that can be modeled by assuming a
diffusion term in the Fokker-Planck equation:
∂ψ
+ ∇x · (vψ) + ∇ p · (ṗψ) = ∇ p · Dr ∇ p ψ , (2.47)
∂t
with the flow-induced orientation term given by Jeffery’s equation:

ṗ = ∇v · p − pT · ∇v · p p, (2.48)
where Dr is the rotary diffusion. We can notice that in absence of flow, i.e. v(x, t) = 0
the Fokker-Planck equation reduces to
∂ψ
= ∇ p · Dr ∇ p ψ , (2.49)
∂t
which ensures a steady-state isotropic orientation distribution: ψ(x, t → ∞, p) =

2π in 2D and ψ(x, t → ∞, p) = 4π in 3D. The higher is the rotational diffusion
1 1
the faster the isotropic orientation distribution in reached.

Thus, at the mesoscopic level, introduction of Brownian effects seems quite sim-
ple. The question is what are the microscopic and macroscopic counterparts of this
diffusion term?
The Fokker-Planck equation can be rewritten in the form:
∂ψ
+ ∇x · (vψ) + ∇ p · (ṗψ) − ∇ p · Dr ∇ p ψ (2.50)
∂t
∂ψ
= + ∇x · (vψ) + ∇ p · p̃ψ˙ = 0,
∂t
where the effective rotary velocity p̃˙ is given by

∇pψ
p̃˙ = ∇v · p − pT · ∇v · p p − Dr , (2.51)
ψ
which contains the flow-induced Jeffery contribution ṗ J plus the Brownian one ṗ B ,
i.e. p̃˙ = ṗ J + ṗ B , with
∇pψ
ṗ B = −Dr . (2.52)
ψ
The Fokker-Planck multidimensionality issue was usually circumvented by using

stochastic strategies for solving the Ito equation related to its Fokker-Planck counter-
part. It is important to mention that any Fokker-Planck equation has an Ito counterpart,
however the reciprocal is not true [53].
At the mesoscopic scale, the Brownian contribution to the extra-stress tensor
results from the generalization of Eq. (2.46):

I I
τ B = 2γ Dr p⊗p− ψ(p) dp = 2γ Dr a − . (2.53)
S 2 2
When moving towards the macroscopic scale, the Brownian contribution to the extra-
stress is defined by Eq. (2.53). However, at the macroscopic scale the microstructure
is defined by the different moments of the orientation distribution. In what follows, we
derive the contribution of Brownian effects to the equation governing the evolution
of the second-order moment.
We start from the definition of the second-order moment

a= p ⊗ p ψ dp, (2.54)
S
whose time derivative involves now the effective rotational velocity p̃˙

ȧ = p̃˙ ⊗ p + p ⊗ p̃˙ ψ dp, (2.55)
S
or

ȧ = ṗ J + ṗ B ⊗ p + p ⊗ ṗ J + ṗ B ψ dp
S
= ṗ J ⊗ p + p ⊗ ṗ J ψ dp
S

+ ṗ B ⊗ p + p ⊗ ṗ B ψ dp = ȧ J + ȧ B , (2.56)
S
where, as proved before, the flow-induced microstructure evolution ȧ J is given by
ȧ J = ∇v · a + a · (∇v)T − 2A : D. (2.57)
We now calculate the expression of the remaining contribution ȧ B :

ȧ =
B
ṗ B ⊗ p + p ⊗ ṗ B ψ dp, (2.58)
S
with ṗ B given by
∇pψ
ṗ B = −Dr . (2.59)
ψ
For the sake of clarity, we consider again the 2D case for which
∂ψ
∂θ
ṗ B = −Dr t, (2.60)
ψ
where t is the unit tangent vector to the unit circle. In this case, Eq. (2.58) reduces to

∂ψ
ȧ B = −Dr (t ⊗ p + p ⊗ t) dθ. (2.61)
S ∂θ
Now, integrating Eq. (2.61) by parts and taking into account
dp
= t, (2.62)
dθ
and
dt
= −p, (2.63)
dθ
we obtain:
ȧ = −2Dr
B
(p ⊗ p − t ⊗ t) ψ(θ ) dθ. (2.64)
S
It is easy to prove that
t ⊗ t + p ⊗ p = I → t ⊗ t = I − p ⊗ p, (2.65)
that allows us to write Eq. (2.64) in the form

ȧ B = −2Dr (2p ⊗ p − I) ψ(θ ) dθ = −2Dr (2a − I) , (2.66)
S
or
I
ȧ = −4Dr a −
B
. (2.67)
2
We can notice that, in absence of flow, ȧ J = 0, and then ȧ = ȧ B ,

I
ȧ = −4Dr a − , (2.68)
2
thus ensuring an isotropic steady state, i.e. a(t → ∞) = I

2 in the 2D case.
Thus, the macroscopic orientation equation reads:

I
ȧ = ∇v · a + a · (∇v) − 2 · A : D − 2d Dr a −
T
, (2.69)
d
where d is the space dimension: d = 2, 3.
2.2.3 Semi-concentrated Regime
Semi-dilute and semi-concentrated regimes have been widely addressed, most of

time by using phenomenological approaches. The most common approach consists
in considering that rod-rod interactions tend to randomize the orientation distribution.
Thus, a second diffusion coefficient is introduced for accounting for rod interactions.
However, in the present case, that diffusion coefficient should scale with the flow
intensity in order to ensure that in absence of flow the microstructure does not evolve
artificially because of such a diffusion term. In general, the interaction diffusion
coefficient D I is assumed of the general form
D I = C I · f (D eq ), (2.70)
where
√ D eq is related to the second invariant of the rate of strain tensor, i.e. D eq =
2D : D. The simplest choice consists in considering the dependence f (D eq ) =
eq
D [22].
Obviously, there are finer approaches based on the direct simulation where the
rod-rod interactions are taken into account explicitly [40, 54]. A nice mesoscopic
based macroscopic approach was proposed in [35] in which interactions are explicitly
described from the introduction of some interaction tensors that can be obtained from
the orientation moments [55].
2.3 Processing Flow Simulation
The flow model for short fibers suspensions is defined by the following equations:
• The balance of linear momentum, without inertia and mass terms,
∇ · σ = 0, (2.71)
where σ is the stress tensor.

• The incompressibility condition,
∇ · v = 0, (2.72)
where v is the velocity field.

• The constitutive equation, with a quadratic closure relation for the fourth-order
orientation tensor,
σ = − pI + 2ηD + 2ηN p (a : D) a. (2.73)
The flow model is defined in the volume Ω f (t) occupied by the fluid at time t.
On its boundary, Γ f (t) ≡ ∂Ω f (t), either the velocity or the traction is imposed:
v (x ∈ Γ1 ) = v g , (2.74)
and
σ · n (x ∈ Γ2 ) = Fg , (2.75)
with Γ1 ∪ Γ2 = Γ f (t), Γ1 ∩ Γ2 = ∅, and where n(x) is the outward unit normal to

the domain boundary.
The inflow boundary will be denoted by Γ − :
Γ − = {x ∈ Γ1 , v(x) · n(x) < 0}. (2.76)

At the flow front Γ f f , Γ f f ⊂ Γ2 , a null traction is usually prescribed, i.e. Fg (x ∈

Γ f f ) = 0.
For calculating the orientation tensor involved in the constitutive equation, three
routes are possible:
• Considering a population of computational particles pi whose position xi (t) results
from the integration of
dxi (t)
= v(xi , t), (2.77)
dt
and whose orientation pi (t) is obtained from the integration of the Jeffery equation
or its Brownian counterpart.
• Solving the Fokker-Planck equation and then evaluating the integral defining the
orientation tensor a.
• Integrating the equation governing the evolution ȧ of the second-order orientation
tensor a. Taking into account its hyperbolic character, this can be performed accu-
rately by using the method of characteristics or a stabilized Eulerian integration.
2.3.1 Fixed Mesh Description of the Fluid Domain

Evolution and Flow Front Tracking by Using
a Volume of Fluid—VoF—Technique
As just indicated, the flow model is defined in the part Ω f (t) of the whole domain
Ω (of boundary Γ ) occupied by the fluid at each time t. In order to update the fluid
domain, we introduce the fluid presence function, I (x, t). This function takes a unit
value in the fluid region and vanishes in the empty domain:
1 i f x ∈ Ω f (t)
I (x, t) = , (2.78)
0 i f x ∈ Ωe (t)
with the empty domain Ωe (t) = Ω − Ω f (t).

The evolution of this function is given by the following scalar and linear advection
equation
∂I
+ v · ∇ I = 0, (2.79)
∂t
which is defined in the whole domain. The fluid presence function must verify a
boundary condition on Γ − (inflow boundary), which for simulating filling processes
reads:
I (x ∈ Γ − , t) = 1, (2.80)
as well as an initial condition. If we assume the domain empty at the initial time
t = 0, the initial condition is
I (x, t = 0) = 0. (2.81)
2.3 Processing Flow Simulation 61
The discretization of the equations of motion can be carried out by means

of a standard mixed finite element technique, using for example an enriched P1
(P1 + “bubble”) C 0 approximation of the velocity field and a linear P1 −C 0 approx-
imation of the pressure field. This functional approximation verifies the inf-sup con-
dition, also known as the LBB (Ladyzhenskaya-Babuska-Brezzi) condition. In order
to extend the variational formulation of the flow equations defined in Ω f (t) to the
whole domain Ω, we enforce a pseudo-behavior in the empty volume, defined by
v = 0 and p = 0 [56]. Combining both the flow and the pseudo-behavior variational
formulations, we obtain:
Find v ∈ (H 1 (Ω))3 and p ∈ L 2 (Ω) verifying the essential boundary conditions
v(x ∈ Γ − ) = v g and v(x ∈ Γ − Γ − ) = 0, such that

f (I ) σ : D∗ dΩ + αv (1 − f (I )) v · v∗ dΩ = 0, (2.82)
Ω Ω

f (I ) ∇ · v p ∗ dΩ + α p (1 − f (I )) p p ∗ dΩ = 0, (2.83)
Ω Ω
∀v∗ ∈ (H01 (Ω))3 and ∀ p ∗ ∈ L 2 (Ω), with f (I = 1) = 1 and f (I = 0) = 0 and

where the expression of the stress tensor is given by the constitutive equation (2.73).
In this variational formulation, H 1 (Ω) and L 2 (Ω) denote the usual Sobolev and
Lebesgue functional spaces and H01 (Ω) is the functional space of velocities vanishing
on the domain boundary Γ . The choice of the functions f (I ), αv and α p is a key
point for obtaining a numerical scheme without numerical dissipation and with a
low diffusion of the flow front (required to locate accurately moving boundaries).
The necessity of introducing functions αv and α p follows from the dimensional
consistency requirements of Eqs. (2.82) and (2.83).
In [56] a linear combination of both variational formulations is considered, i.e.
the fluid fraction was taken as weight function, i.e. f (I ) = I , and both parameters
αv and α p were considered constant.
To justify the required values of both parameters, we considered in [29] an element
Ω e which just starts its filling process. In this case the fluid fraction in this element
I e takes an intermediate value 0 < I e < 1. If parameters αv and α p are too high, the
over-imposition of the zero velocity condition in the empty region or in the elements
which start their filling process, tends to perturb the natural flow front movement. On
the other hand, if those parameters are too small, the equation of motion as well as the
fluid incompressibility will be over-imposed in the element Ω e in spite of the little
amount of fluid existing in this element which just starts its filling process. Thus, this
pseudo-incompressibility, derived from the small values of both parameters, retards
the filling process. In the same way, the resulting flow front thickness increases, and
the localization of the flow front remains uncertain. In order to improve the flow
front location (reducing its numerical thickness, i.e. the number of elements in the
flow direction with a fluid fraction 0 < I < 1 ), it was proposed in [56] to use small
values of both parameters combined with a local mesh adaptation in the flow front
neighbourhood.
Another choice able to reduce the flow front thickness was proposed in [29]:
1 i f I e ≥ Ith
f (I e ) = , (2.84)
0 i f I e < Ith
where Ith represents a threshold value, close to one. The first advantage of this
choice is that no combinations of the fluid flow and the pseudo-behavior variational
formulations are employed in the partially filled elements. Thus, only in the case that
I e ≥ Ith the flow model will be enforced in the element Ω e . However, as proved
below, a judicious choice of parameters αv and α p will be essential to enforce the
mass conservation reducing the flow front thickness to one element.
To illustrate the parameters choice, we consider a one-dimensional domain. We
denote by Ω e a generic element where a linear approximation of the velocity field is
considered. Thus, the velocity at each point of Ω e can be expressed from their nodal
values ve and ve+1 , where the coordinates of both nodes are x e and x e+1 (x e+1 > x e )
respectively. We assume the inlet located at x = 0, where the injection velocity is
enforced to a value vi . Now, let I e < Ith be the fluid fraction existing in the element
Ω e at time t. We will assume that the element located upstream of Ω e , Ω e−1 , is fully
filled, i.e. I e−1 = 1 and that the downstream element remains empty, i.e. I e+1 = 0.
At time t, with I e−1 = 1, the flow equations are imposed in the element Ω e−1 .
Thus, the velocity at node x e tends to be the injection velocity as a direct consequence
of the fluid incompressibility and the one-dimensional flow regime. On the other
hand, the node x e also belongs to the element Ω e where due to the fact that I e < Ith
a null velocity is enforced, implying the annulation of both nodal velocities. An
evident conflict appears because each element containing the node x e enforces a
different value of the velocity at this node. Due to the continuity assumed in the
velocity interpolation the resulting velocity at this node will be an intermediate value
between the injection velocity (required by the flow equations) and a null velocity
(enforced by the pseudo-behavior). Therefore, the extended variational formulation
to the whole domain, as has been presented until now, affects the natural flow front
progression. To avoid these undesired effects one possibility consists in an adequate
choice of parameters αv and α p . If we consider
1 i f Ie = 0
α p (I e ) = αv (I e ) = α(I e ) = , (2.85)
ε i f Ie > 0
where ε is a constant small enough, the nodal velocity at node x e is dominated by

the flow equations enforced in the element Ω e−1 (the contribution of the pseudo-
behaviour imposed in the element Ω e , and consequently affected by the constant ε
is negligible). Thus the velocity at node x e approaches the exact value (the injection
velocity vi ). The velocity at node x e+1 will be null due to the imposition of the
pseudo-behavior (null velocity) in all the elements containing this node. Thus, the
filling process of the element Ω e will be accurately achieved and the flow front area
is reduced to a single element improving significantly the flow front location.
Finally, we describe a first-order discontinuous finite element technique (higher-

order discretization also exist) for solving the advection problems governing the
evolution of fields I (x, t) and a(x, t).
In order to alleviate the notation we consider the linear and scalar advection
equation Eq. (2.79) governing the evolution of I (x, t).
Taking into account the fluid incompressibility, ∇ · v = 0, Eq. (2.79) can be
rewritten in the form
∂I
+ ∇ · (v I ) = 0, (2.86)
∂t
whose integral conservation form in element Ω e reads

∂I
dΩ + ∇ · (v I )dΩ = 0. (2.87)
Ωe ∂t Ωe
Using the divergence theorem, the previous equation becomes

∂I
dΩ + v · n I d S = 0, (2.88)
Ωe ∂t ∂Ω e
where ∂Ω e denotes the boundary of Ω e , and n is the outward unit normal to the
element boundary. The element boundary can be divided in two parts: ∂ + Ω e through
which the flow leaves the element (i.e. the part of the boundary verifying the relation
v · n > 0); and the inflow boundary ∂ − Ω e , through which the fluid is coming to
element Ω e (verifying v · n < 0). Thus, the mass balance yields

∂I
dΩ + v · n IdS + v · n I d S = 0. (2.89)
Ωe ∂t ∂−Ωe ∂+Ωe
We now consider the simplest approximation choice, consisting of a constant value

of function I (x, t) into each element. In this case function I (x, t) is not defined on
the element boundary, and the flow coming to element Ω e is approximated taking
− −
into account its value in the upstream neigbouring elements Ω ei , I ei . In the same
way, the outflow rate will depend on the fluid fraction existing in the element Ω e , I e .
Thus, the previous equation can be written as
∂Ie e − − +
|Ω | = q ei I ei − q e I e , (2.90)
∂t
i
+
where |Ω e | is the volume of Ω e , q e represents the flow rate leaving the element
−
Ω e , and q ei denotes the flow coming to element Ω e from the upstream neigbouring
−
element Ω ei .
Considering the time derivative, the Taylor’s expansion of I e (t + Δt) reads
∂Ie 1 ∂2 I e
I e (t + Δt) = I e (t) + Δt + (Δt)2 + · · · (2.91)
∂t t 2 ∂t 2 t
With a first-order explicit scheme, the update reads:
∂Ie
I e (t + Δt) = I e (t) + Δt
∂t t
ei− − +
i q I ei − q e I e
= I e (t) + Δt. (2.92)
|Ω e |
Figure 2.2 depicts different stages of the filling process of a complex 3D cavity.
It represents the field I (x, t) at 6 different filling times. In addition to the evolution
of the fluid domain Ω f (t), the velocity field and the orientation state was calculated
at each time step. The velocity field resulted from solving the weak form (2.82)
and (2.83), whereas the orientation state was calculated by integrating the advection
Eq. (2.69)

I
ȧ = ∇v · a + a · (∇v)T − 2 · A : D − 6Dr a − , (2.93)
3
which, being purely advective, was integrated by applying exactly the same scheme
as just described in the case of the fluid fraction I (x, t).
Fig. 2.2 Mould filling

simulation: evolution of the
presence of fluid field
2.3.2 Updated Lagrangian Meshless Simulation
Even if there are many variants of techniques operating on a fixed mesh, as the one
just presented, all of them give similar results despite theirs apparent differences.
Another possibility consists in solving the flow problem in the fluid domain Ω f (t)
that must be updated at each time step. One could imagine that at time t a mesh F t is
associated to the fluid domain Ω f (t). This mesh consists of a number of nodes Nn ,
with coordinates xi (t), and a number of elements Ne . By solving the flow problem
in Ω f (t) we have access to the velocity field at each node of the mesh, from which it
can be interpolated everywhere in Ω f (t). Now, one could imagine updating the nodal
positions with the material velocity, i.e. xi (t + Δt) ≈ xi (t) + v(xi (t), t) · Δt. It is
important to mention that during this Lagrangian update, purely advective equations,
as the one governing the evolution of the orientation tensor a, can be integrated
accurately by using the method of characteristics.
Now, a tentative mesh F˜ t+Δt could be defined, for example from the Delaunay
triangulation related to the updated nodal positions xi (t + Δt). However, it is well
known that such procedure entails the distortion of the elements involved in the
updated meshes and thus it compromises the solution accuracy. For this reason, a
remeshing step is compulsory. Nodes are redistributed within the fluid domain to
ensure that the associated mesh F t+Δt does not contain elements that are too dis-
torted.
Thus, standard updated Lagrangian approaches involve frequent remeshing steps.
In addition to their computational complexity (mainly in 3D), they require a projec-
tion of all the fields depending on the flow history (e.g. the orientation state) from the
trial mesh F˜ t+Δt to the updated one F t+Δt , with the associated numerical diffusion.
To alleviate these drawbacks, meshless methods were introduced some years ago
and intensively used in many engineering applications. Meshless methods can be
defined as those discretization techniques able to construct field approximations
whose quality does not depend on the geometrical quality of a subjacent mesh.
That is, when using meshless approximations, one could expect to define the fields
approximation at each time t from the nodal positions resulting of moving them with
the material velocity all along the whole simulation. They constitute an appealing
choice for simulating flows involving internal variables (e.g. orientation), because
advective fields are integrated very accurately by using the method of characteristics
and the description of free or moving boundaries is a trivial task [57].
The interested reader can refer to [58, 59] and the references therein. The main
difficulty of most meshless methods concerns the enforcement of essential boundary
conditions because most of them (the Natural Element Method being an exception)
do not verify the Kroenecker’s delta property when approximating a field u(x) by
u h (x), i.e. u h (xi ) = u(xi ), with xi the coordinates of a generic node n i .
The Natural Element Method—NEM—is a nice compromise because it allows
for a robust interpolation despite the quality of the subjacent mesh and it allows for
the enforcement of essential boundary condition like in the finite element method,
among many other appealing properties [59].
Fig. 2.3 Delaunay triangulation and Voronoi diagram
The NEM is based on the application of two classes of Natural Neighbour interpo-
lation (Sibson [60] and Belikov [61]) to the discretization of a variational formulation.
In what follows, we will refer to the first one.
Sibson interpolation relies on the concepts of Delaunay triangulations and Dirichlet
tesselations of a set of nodes to build the shape functions (see Fig. 2.3). A Delaunay
triangulation (tetrahedrization in three-dimensions) is the unique triangulation for a
given set of nodes that satisfies the empty circumcircle criterion, that is, given the
three nodes of a triangle (four nodes on a tetrahedron) and the circle (sphere) that
passes through them, no one of the other nodes lies inside this circle (sphere). A
Voronoi diagram or a Dirichlet tessellation is the dual structure of a Delaunay trian-
gulation. For a given node n i , the associated Voronoi cell is composed by all of the
points which are closer to the node n i than to any other node. Formally,
Ti = {x ∈ R3 : d(x, xi ) < d(x, x j ) ∀ j = i}, (2.94)
where Ti is the Voronoi cell and d(·, ·) represents the Euclidean distance. It is clear
from Fig. 2.3 that the Delaunay triangulation is defined over the convex hull of the
set of points.
In a similar way, the second-order Voronoi cell is defined as the locus of the points
that have the node n i as the closest node and the node n j as the second closest node:
Ti j = {x ∈ R3 : d(x, xi ) < d(x, x j ) < d(x, xk ), ∀k = i, j; i = j}. (2.95)
Thus, if a new point is added to a given cloud of points, the natural neighbour
coordinates of this point x with respect to one of his neighbours xi is the ratio of the
cell Ti that is transferred to Tx when adding x to the initial cloud of points, to the
total area of Tx . In other words, with κ(x) and κi (x) being the Lebesgue measures
of Tx and Txi respectively, the natural neighbour coordinates of x with respect to the
node n i is defined as
κi (x)
φi (x) = , (2.96)
κ(x)
which, using the notation indicated in Fig. 2.4, with A the polygon area, gives
Aab f e
φ1 (x) = . (2.97)
Aabcd
It is straightforward to prove that natural element shape functions define a partition

of unity. The resulting shape function is shown in Fig. 2.5.
A model variable u (u represents a scalar, a vector or a tensor) can be approximated
in 1D, 2D or 3D in the form:

Nn
uh (x) = φi (x)ui , (2.98)
i=1
with ui = u(xi ).
Fig. 2.4 Natural

neighbourhood coordinates
Fig. 2.5 Sibson’s shape

function
Fig. 2.6 NEM simulation of the flow leaving an extrusion die: orientation state (left) and velocity
field (right)
The main issue related to the use of these interpolants is the treatment of non-
convex domains, since, as indicated before, the Delaunay triangulation proceeds on
the convex hull of the set of nodes. In that case, the simplest alternatives are the
use of alpha-shapes [62] or the construction of the functional approximation on the
so-called Constrained Voronoi Diagram—CVD—[63].
NEM-based discretization techniques are particularly appropriate for simulating
very large thermomechanical transformations exhibiting possible free and moving
boundaries [57, 64], and in particular those involving flow dependent internal vari-
ables, e.g. molecules conformation [65] or rod orientation [30, 66]. Figure 2.6 depicts
the orientation and velocity fields in an extrusion simulation [31]. The orientation
field is represented by an ellipse at each nodal position, whose semi axes are pro-
portional to the orientation tensor eigenvalues and their orientation are given by the
corresponding eigenvectors.
In this simulation, the nodal position is updated at each time step from the material
velocity. Then, the fluid domain is extracted by using the alpha-shape constructor
[62]. Next, the flow Eqs. (2.71), (2.72) and (2.73) are solved in the just updated fluid
domain by using a Sibson approximation of the velocity field and considering the
pressure field constant in each Voronoi cell. This approximation does not verify the
LBB condition but generally works very well. It is much less expensive than using
approximation fulfilling the LBB stability condition [67].
The solution of the flow problem at time t +Δt in Ω f (t +Δt) needs specifying the
orientation state a(x ∈ Ω f (t + Δt), t + Δt) that appears in the constitutive equation
(2.73). Because of the meshless character of the discretization here employed, we
can consider the cloud of nodes throughout the whole simulation. As we consider
the orientation state attached to the nodes, and the orientation equation being purely
advective, one could proceed to its integration along the nodal pathlines by using the
method of characteristics, that is, proceeding from Eq. (2.69) in 2D:

I
ȧ = ∇v · a + a · (∇v)T − 2 · A : D − 4Dr a − . (2.99)
2
We consider for each node i = 1, . . . , Nn the simplest explicit first-order integration
a(xi (t + Δt), t + Δt) − a(xi (t), t)

Δt
I
= ∇v · a + a · (∇v)T − 2 · A : D − 4Dr a − , (2.100)
2
where
xi (t + Δt) = xi (t) + v(xi (t), t) · Δt, (2.101)
and in the right-hand side of Eq. (2.100), orientations and velocities are computed at
positions xi (t) and time t.
2.4 Concentrated Suspensions Involving Rod Clusters
Having considered rods immersed in a flow, we extend the approach to more com-
plex configurations. Suspensions of industrial interest composed of rods (a limit
case of an ellipsoid) aggregate, leading to clusters with particular shapes that
exhibit, when immersed in a flow, an almost rigid motion. These situations are
currently encountered when considering carbon nanotubes suspensions as discussed
in [33, 39, 68]. In what follows, we address the modeling of both rigid and deformable
clusters.
2.4.1 Rigid Clusters
We represent rigid clusters of general shape in both a discrete manner, assuming they
are all composed of N /2 rods (N being an even integer) involving N beads, and in a
continuous manner from the continuous pdf describing the configuration of those rod
beads. In a rigid cluster, there is no relative motion between the rods composing it.
2.4.1.1 Discrete Description
First, we consider a 3D rigid cluster consisting of N /2 rods R j of length 2L j . We

assume that each rod R j contains two beads at its ends on which hydrodynamic
forces apply. Thus, the cluster contains N beads Bi , i = 1, . . . , N . The location of
each bead Bi with respect to the cluster center of gravity G is represented by Si qi ,
where qi is the unit vector pointing from G to Bi . The cluster is sketched in Fig. 2.7.
Fig. 2.7 Rigid cluster

composed of rods
Brownian effects are neglected and then only flow-induced hydrodynamic forces
must be considered (hydrodynamic interactions are not considered in this work).
Forces Fi apply on each bead located at position Si qi (Fig. 2.7) and are proportional
to the difference of velocity between the one of the flow unperturbed by the presence
of the cluster at the bead location and the one of the bead itself:
Fi = ξ (V0 + ∇v · qi Si − VG − q̇i Si ), (2.102)
where ξ is the friction coefficient, v the flow velocity field, V0 the fluid velocity at
the cluster centre of gravity G and VG the velocity of the cluster center of gravity
itself.
By adding all the forces we obtain

i=N
i=N
i=N
0= Fi = N (V0 − VG ) + ∇v · Si qi − Si q̇i . (2.103)
i=1 i=1 i=1
Both sums in Eq. (2.103) vanish, the first one as a direct consequence of the
definition of the center of gravity, and the second because the cluster is assumed
rigid. Thus, Eq. (2.103) becomes
V0 = VG , (2.104)
implying that the cluster center of gravity is moving with the fluid velocity at that
position.
The torque created by forces applied on bead Bi is given by
Mi = Si qi × Fi . (2.105)
Neglecting inertial effects, the resulting torque for the whole cluster must vanish:
2.4 Concentrated Suspensions Involving Rod Clusters 71

i=N
Mi = 0. (2.106)
i=1
Taking Eqs. (2.105) and (2.102) into account, we have

i=N
i=N
Si2 qi × (∇v · qi ) = Si2 qi × q̇i . (2.107)
i=1 i=1
If we define the cluster angular velocity ω such that
q̇i = ω × qi , (2.108)
torque equilibrium reads

i=N
i=N
Si2 qi × (∇v · qi ) = Si2 qi × (ω × qi ). (2.109)
i=1 i=1
Now, by using the vector triple product relationship a × (b × c) = b (a · c) −

c (a · b), the right-hand side reads

i=N
i=N
Si2 qi × (ω × qi ) = Si2 (ω (qi · qi ) − qi (qi · ω)) , (2.110)
i=1 i=1
which, taking into account the normality of vectors qi and the fact that qi (qi · ω) =
(qi ⊗ qi ) · ω, becomes

i=N
i=N
Si2 qi × (ω × qi ) = Si2 (I − (qi ⊗ qi )) · ω, (2.111)
i=1 i=1
where I the unit matrix.

Now, the left-hand side of Eq. (2.109) can be rewritten by using the third-order
Levi-Civita permutation tensor ε such that (u × v) = ε : (v ⊗ u). We obtain

i=N
i=N
Si2 qi × (∇v · qi ) = Si2 ε : (∇v · (qi ⊗ qi )) , (2.112)
i=1 i=1
which finally yields

i=N
i=N
Si2 ε : (∇v · (qi ⊗ qi )) = Si2 (I − (qi ⊗ qi )) · ω. (2.113)
i=1 i=1
Expressed in a more compact form, we have

i=N
ε : ∇v · Si2 (qi ⊗ qi )
i=1

i=N
i=N
= Si2 I− Si2 (qi ⊗ qi ) · ω. (2.114)
i=1 i=1
2.4.1.2 Continuous Description
The continuous description considers the pdf Υ (q, S) giving the fraction of rod beads
located at position qS, that can be expressed as
Υ (q, S) = ψ(q; S) Γ (S), (2.115)
where ψ(q; S) is the angular distribution of beads located at distance S from the
cluster center of gravity G, and Γ (S) is the fraction of beads located at that distance
S. The normality condition reads

ψ(q; S) dq = 1, ∀S. (2.116)
S
By defining the conformation tensor c S related to the population of beads located

at distance S with respect to the cluster center of gravity G as

cS = q ⊗ q ψ(q; S) dq, (2.117)
S
all sums in the previous expression (2.114) must be substituted by the corresponding
integrals in the length and orientation domains, L and S respectively, weighted by
the distribution function ψ(q; S) Γ (S).
In particular Eq. (2.114) becomes:

ε : ∇v · S2 (q ⊗ q)ψ(q; S) dq Γ (S) d S
L S

= S 2 Γ (S) d S I − S2 (q ⊗ q)ψ(q; S) dq Γ (S) d S · ω,
L L S
(2.118)
which gives, using definition (2.117),

ε : ∇v · S 2 c S Γ (S) d S
L

= S 2 Γ (S) d S I − S 2 c S Γ (S) d S · ω. (2.119)
L L
By introducing the mean square length β and tensor c̃ according to

β= S 2 Γ (S) d S, (2.120)
L
and
c̃ = S 2 c S Γ (S) d S, (2.121)
L
and due to the unit trace of c S , i.e. T r (c S ) = 1, we can infer that β is in fact the trace
of c̃, that is β = T r (c̃).
Thus, the particle kinematics finally results:
ε : (∇v · c̃) = (T r (c̃)I − c̃) · ω, (2.122)
or dividing by T r (c̃),
ω = (I − c)−1 · (ε : (∇v · c)) , (2.123)
where the conformation tensor c is given by
c̃
c= . (2.124)
T r (c̃)
Remark 3 In the 2D case, c · ω = 0 and then ω = (ε : (∇v · c)). Moreover, if all

beads are located at the same distance S, we recover the expression given in [44].
An extremely important consequence of this analysis is that rigid clusters com-

posed of rods having the same conformation tensor c rotate at the same angular
velocity.
As the conformation tensor c is symmetric and positive definite, it has real eigen-
values and eigenvectors. In 3D, the three mutually perpendicular eigenvectors will be
denoted by u1 , u2 and u3 , with the associated eigenvalues τ1 , τ2 and τ3 respectively.
Imagine a rigid cluster composed of three rods oriented in directions u1 , u2 and
√ √ √
u3 with respective lengths τ1 , τ2 and τ3 . The conformation tensor of such a
three-rod cluster coincides with c and then both tensors have the same rotary velocity.
Thus in [45] we proved the link between a rigid cluster compose of three mutually
orthogonal rods and the Jeffery triaxial-ellipsoid.
2.4.1.3 Multi-scale Description
In this chapter, we have only addressed the first brick of the proposed approach to
the multi-scale modeling of suspensions involving rigid clusters composed of rigid
rods. In way of perspectives for future work, we conclude with an overview of the
complete nine-step approach to be followed:
1. The cluster conformation has been successfully described by tensor c given in

Eq. (2.124). This choice was motivated by the fact that all clusters having the
same conformation tensor c have the same kinematics as just proved.
2. Cluster conformation evolution. Knowing the cluster kinematics ω, we can obtain
the expression of q̇ in order to define the conformation evolution from ċ.
3. Cluster contribution to the stress. The force acting on each rod bead involved in
the rigid cluster must be taken into account by invoking Kramers’ rule.
4. Population description. The population of rigid clusters at the mesoscopic scale
can be represented by means of either a discrete or a continuous description. In
the discrete framework, the population is described from different computational
individuals, each one characterized by its conformation tensor ci , i = 1, . . . , N .
Within the continuous framework, the suspension is characterized by the pdf Ψ ,
which now depends on the physical coordinates (space x and time t) and the
conformation coordinate c. Thus, the pdf reads Ψ (x, t, c). It gives the fraction of
clusters that at position x and time t have a conformation given by c.
5. Description of the population evolution. Within the discrete framework, the popu-
lation evolution is obtained by integrating the evolution equation for each cluster
conformation ċi . Within the continuous framework, the pdf Ψ (x, t, c) evolves
according to its associated Fokker-Planck equation, with the knowledge of ċ.
6. Contribution of cluster population to the stress requires adding the contribution
of all computational clusters, within the discrete framework, or to integrate in
conformation space when proceeding within the continuous framework.
7. The macroscopic description uses a coarser description based on the moments of
the pdf. Here, the simplest choice consists of the first moment C(x, t) defined as

C(x, t) = c Ψ (x, t, c) dc. (2.125)
C
8. Microstructural macroscopic evolution. In order to derive the time evolution of

the first moment C, we should consider its time derivative and propose adequate
closure relations.
9. The moment-based stress requires consideration of the stress expression obtained
in step (6) above. Here again, introduction of a suitable closure relation is required.
2.4.2 Deformable Clusters
In this section, we consider a more realistic scenario consisting of deformable clus-

ters. We consider two types of forces applied on each bead of a generic rod R (of
length 2L), one due to the fluid-rod friction once more modeled fas
FiH = ξ (v0i + ∇v · pi L − viG − ṗi L), (2.126)

where the superscript “H ” refers to its hydrodynamic nature. Here, v0i is the fluid
velocity (assumed unperturbed by the presence of the cluster) at the rod Ri centre
of gravity Gi , and viG the velocity of the rod Ri centre of gravity Gi .
The other force FiC is due to the rod entanglements. This last force is assumed
scaling with the difference between the rigid motion velocity (the one that the bead
would have if the cluster would be rigid) vB R and the real one v + ṗL.
G
As proved in the previous section, the bead velocity when assuming the cluster
rigid reads:
R
vB = V0 + W · r + W · pL , (2.127)
where tensor W derives from the rigid rotary velocity ω given by Eq. (2.123) from
W · p = ω × p, ∀p. (2.128)
V0 is the unperturbed fluid velocity at the cluster centre of gravity G and r is the
vector connecting the cluster centre of gravity G to the rod centre of gravity G , i.e.
r = G − G.
Thus, the contribution to the bead force due to collective effects (a sort of mean
field) when considering the generic rod Ri reads:
FiC = μ (V0 + W · ri + W · pi L − viG − ṗi L). (2.129)
This expresses that if the hypothetical rigid cluster is moving faster that the bead
then this sort of mean field pushes the bead.
When μ is large enough, forces FiC dominate the momentum balances enforcing
the cluster rotary velocity W to each rod composing it (rigid behaviour).
The resulting force at bead pi L results Fi (pi L) = FiH (pi L) + FiC (pi L). The
linear momentum balance, neglecting inertia effects, reads
Fi (pi L) + Fi (−pi L) = 0, (2.130)
that implies
ξ v0i − ξ viG + μV0 + μW · ri − μviG = 0, (2.131)
or
ξ v0i + μV0 μ
viG = + W · ri . (2.132)
ξ +μ ξ +μ
When μ = 0, this gives the standard non-interacting rod model (dilute suspension),
with v0i = viG , and when ξ = 0, we recover the solid motion viG = V0 + W · ri .
Now, taking into account the first gradient framework
v0i = V0 + ∇v · ri , (2.133)
Eq. (2.132) reduces to:
ξ μ
viG = V0 + ∇v · ri + W · ri . (2.134)
ξ +μ ξ +μ
This equation allows for the trajectory calculation of each rod centre of gravity.
Now, we consider the equation governing the evolution of the rods orientations, that
is ṗ.
Coming back to the force applied on each rod bead,
Fi = ξ (v0i + ∇v · pi L − viG − ṗi L)

+ μ (V0 + W · ri + W · pi L − viG − ṗi L), (2.135)
and taking (2.131) into account, we obtain
Fi (pi L) = L (ξ(∇v · pi − ṗi ) + μ(W · pi − ṗi )) . (2.136)
Now, if we define the equivalent traceless gradient G,
μ W + ξ ∇v
G= , (2.137)
ξ +μ
Eq. (2.136) can be rewritten as
Fi (pi L) = L(ξ + μ) (G · pi − ṗi ) , (2.138)
that allows for using all the rationale considered when developing the Jeffery model,
by replacing the friction coefficient by (ξ + μ) and the gradient of velocity by G.
Thus, the rod rotary velocity becomes:
ṗi = G · pi − (G : (pi ⊗ pi )) pi . (2.139)
The term G : (pi ⊗ pi ) deserves some additional comment. Since W and Ω are
both skew-symmetric, we obtain: W : (pi ⊗ pi ) = 0, and the same in the case of
ξ
considering Ω, implying G : (pi ⊗ pi ) = ξ +μ D : (pi ⊗ pi ) and then:
ξ
ṗi = G · pi − (D : (pi ⊗ pi )) pi (2.140)
ξ +μ
This can be rewritten as

μ ξ ξ
ṗi = W · pi + ∇v · pi − (D : (pi ⊗ pi )) pi
ξ +μ ξ +μ ξ +μ
ξ μ ξ μ
= ṗiJ + W · pi = ṗiJ + ṗ R , (2.141)
ξ +μ ξ +μ ξ +μ ξ +μ i
where ṗiJ represents the hydrodynamic contribution in absence of collective effects

(dilute regime described by the Jeffery equation), and ṗiR the one coming from the
rod entanglements that results in a rigid-like cluster kinematics.
We can notice that, when ξ μ, hydrodynamic effects are preponderant and the
rod kinematics are governed by the Jeffery equation, i.e. ṗi ≈ ṗiJ . In the opposite
case, μ ξ , the cluster is too rigid and the rods adopt the velocity dictated by the
rigid cluster kinematics ṗi ≈ ṗiR .
Macroscopic description: cluster inertia and shape
From Eq. (2.134), the position of each rod composing the cluster can be integrated.
It seems natural that knowing the location of the centre of gravity of each rod, the
cluster inertia and shape could be obtained.
If we define the cluster inertia tensor S from:

S= r ⊗ r ψ(r) dr, (2.142)
R
its time evolution is given by

Ṡ = (ṙ ⊗ r + r ⊗ ṙ) ψ(r) dr, (2.143)
R
with ṙ = vG − V0 , where the fact that the (rigid) cluster centre of gravity is moving
with the fluid velocity at that position was taken into account. Thus, we obtain
ξ μ
ṙ = ∇v · r + W · r. (2.144)
ξ +μ ξ +μ
Now, by introducing Eqs. (2.144) into (2.143) and making use of (2.142), we
obtain
ξ μ
Ṡ = ∇v · S + S · (∇v)T + (W · S − S · W) , (2.145)
ξ +μ ξ +μ
where we used the fact that WT = −W. The vorticity tensor appearing in both ∇v
and W ensures the objectivity of the cluster shape evolution.
In Eq. (2.145), the first term involving the velocity gradient ∇v is responsible of
the cluster deformation, whereas the second one involving W only rotates the cluster
like a rigid solid because of collective effects. The standard flow-induced rotation
(in absence of collective effects) is contained in the vorticity tensor. When μ = 0,
the cluster deforms and rotates according to the flow vorticity. When μ = 0, there
is an extra-rotation induced by collective effects. Finally, when ξ = 0 (μ = 0), the
cluster rotates without experiencing a deformation.
The trace of S gives information on the cluster size. Let us define the shape tensor
s from the normalized inertia tensor
Fig. 2.8 Cluster shapes
S
s= . (2.146)
T r (S)
The shape tensor being symmetric and positive definite, it can be diagonalized
and represented by an ellipsoid (ellipse in 2D) whose eigenvalues describe the cluster
orientation and its eigenvalues its shape. Figure 2.8 depicts different clusters and their
associated shape tensors.
Macroscopic description: orientation
Now, concerning the orientation, if we assume the orientation of the rods of a
deformable cluster given by the orientation distribution ψ(p), the time derivative
of its second order moment ȧ is given by

ȧ = (ṗ ⊗ p + p ⊗ ṗ) ψ(p) dp. (2.147)
S
By considering the expression of the microscopic velocity ṗ,
ξ μ
ṗ = ṗ J + ṗ R , (2.148)
ξ +μ ξ +μ
we obtain:
ξ μ
ȧ = ȧ J + ȧ R , (2.149)
ξ +μ ξ +μ
with ȧ J and ȧ R resulting once Eq. (2.148) is introduced into Eq. (2.147) by assuming
μ = 0 and ξ = 0, respectively:
ȧ J = ∇v · a + a · (∇v)T − 2 A : D
(2.150)
ȧ R = W · a + a · WT
Here, A is the fourth-order moment defined by:

A= p ⊗ p ⊗ p ⊗ p ψ(p) dp. (2.151)
S
The objectivity of the resulting evolution equation of ȧ was proved in [43]. Usually,
the fourth-order moment A is expressed from the second-order one by considering
any of the numerous closure relations proposed in the literature [47–50].
We just proved that the conformation of a deformable cluster is defined from
both the orientation tensor a and the shape tensor S. The evolution of both tensors
makes use of tensor W that depends itself on the conformation tensor c. Tensor c is
very close to tensor S; the first is defined from the vectors joining the cluster centre
of gravity with each rod bead, whereas the second one involves vectors joining the
cluster centre of gravity with the rods centre of gravity. The rod length being very
small, both tensors are very close, and consequently tensor W can be evaluated by
using S instead of c.
To validate this approach, we come back to the definition of the conformation
tensor c (2.124),
c̃
c= , (2.152)
T r (c̃)
where the discrete form of c̃ reads

N
c̃ = (Si qi ) ⊗ (Si qi ). (2.153)
i=1
If we consider the contribution of the two beads of a generic rod R j , we obtain
j j j j j j j j
c̃ j = (S1 q1 ) ⊗ (S1 q1 ) + (S2 q2 ) ⊗ (S2 q2 )
= (r j + Lp j ) ⊗ (r j + Lp j ) + (r j − Lp j ) ⊗ (r j − Lp j )
= 2r j ⊗ r j + 2L 2 p j ⊗ p j , (2.154)
which implies
N N

2 2

c̃ = c̃ j = 2r j ⊗ r j + 2L 2 p j ⊗ p j = 2 S + L 2 a , (2.155)
j=1 j=1
from which we obtain conformation tensor
S + L 2a
c= . (2.156)
T r (S) + L 2
Thus, as soon as L 2 T r (S), we can assume c ≈ S.

Cluster shape versus orientation
In the model just proposed, the cluster conformation is described from the orientation
and shape tensors and the trace of the inertia tensor. In principle, orientation and
shape are considered independent from one another, even if some correlation could
Fig. 2.9 Cluster shapes versus orientation: rods fully aligned
Fig. 2.10 Cluster shapes versus orientation: isotropic orientation
be expected from a physical point of view. Thus, we can define a spherical cluster with
all its rods aligned or a very elongated cluster with all its rods isotropically oriented.
Figures 2.9 and 2.10 depict different scenarios between these two limit cases.
2.5 Advanced Topics
2.5.1 On the Solution of the Fokker-Plank Equation
Since kinetic theory descriptions involve a probability distribution function depend-

ing on space, time and a number of conformational coordinates, the associated
Fokker-Planck equations suffer the so-called curse of dimensionality typical of prob-
lems defined in highly dimensional spaces.
2.5 Advanced Topics 81
Thus, mesh-based discretization techniques fail for discretizing the problem

because the number of degrees of freedom involved in a mesh or grid increases
exponentially with the space dimension.
In what follows, we consider some alternatives to standard mesh-based discretiza-
tion that are able to address the solution of Fokker-Planck equations associated with
kinetic theory descriptions.
We consider, without loss of generality, a generic Fokker-Planck equation where
diffusion terms are represented by fluxes qx and qa operating respectively in the
physical and conformational spaces:
∂Ψ
+ ∇x · (v Ψ ) + ∇a · (ȧ Ψ ) = −∇x · qx − ∇a · qa . (2.157)
∂t
2.5.1.1 Method of Particles for Solving Advection-Dominated Problems
This technique described in detail in [6, 8] consists in approximating the initial distri-
bution Ψ (x, t = 0, a) from M Dirac’s masses ai0 at each one of the Q positions x0j :
Q
M
j
Ψ (x, t = 0, a) = αi δ(a − ai0 ) δ(x − x0j ). (2.158)
j=1 i=1
This represents a sort of approximation based on Q · M computational particles Pi j

with initial positions and conformations given by

xi0j = x0j , i = 1, . . . , M ; j = 1, . . . , Q
, (2.159)
ai0j = ai0 , i = 1, . . . , M ; j = 1, . . . , Q
and whose position and conformation will be evaluated all along the flow simulation,
from which the distribution will be reconstructed.
When considering the purely advective balance equation
∂Ψ
+ ∇x · (v Ψ ) + ∇a · (ȧ Ψ ) = 0, (2.160)
∂t
the time evolution of position and conformation of each particle Pi j is calculated

by integrating τ =t
xi j (t) = xi0j + τ =0 v(xi j (τ )) dτ
τ =t . (2.161)
ai j (t) = ai0j + τ =0 ȧi j (ai j (τ ), xi j (τ )) dτ
As the position update only depends on the velocity field, that itself only depends
on the position, it can be stressed that particles Pi j , i = 1, . . . , M are following
the same trajectory in the physical space, having x0j as departure point.
Now, the orientation distribution at time t can be reconstructed from
Q
M
j
Ψ (x, t, a) = αi δ(a − ai j (t)) δ(x − xi j (t)). (2.162)
j=1 i=1
Obviously, smoother representations can be obtained by considering appropriate

regularizations of the Dirac’s distribution, as the one usually performed within the
SPH (Smooth Particles Hydrodynamics) framework [5, 6].
When considering diffusion terms, there are two main routes based on the use of
particles, one of stochastic nature, the other fully deterministic.
For illustrating the procedure when models involve diffusion terms, we consider
the Fokker-Planck equation
∂Ψ
+ ∇x · (v Ψ ) + ∇a · (ȧ Ψ ) = −∇x · qx − ∇a · qa , (2.163)
∂t
where qx and qa are two diffusive fluxes operating in the physical and conformational
spaces respectively, both modeled from a Fick-type law:
qx = −Dx · ∇x Ψ
. (2.164)
qa = −Da · ∇a Ψ
• Within the stochastic framework, diffusion terms can be modeled from appro-
priate random variables within a Lagrangian or a Eulerian description, the last
one known as Brownian Configurations Fields (BCF). Both approaches were con-
sidered in our former works on the solution of Fokker-Planck equations [8, 69].
Within the Lagrangian stochastic framework and starting from the initial cloud of
computational particles Pi j representing the initial distribution Ψ (x, t = 0, a),
the simplest particles updating reads
xi j (tn+1 ) = xi j (tn ) + v(xi j (tn )) Δt + Rx (Δt)

, (2.165)
ai j (tn+1 ) = ai j (tn ) + ȧi j (ai j (tn ), xi j (tn )) Δt + Ra (Δt)
where Δt is the time step and both random updates Rx and Ra depend on the
chosen time step (see [53] for more details as well as for advanced stochastic
integrations).
Obviously, because of the random effects operating in the physical space, the M
particles initially located at each position x0j , j = 1, . . . , Q, will follow different
trajectories in the physical space along the simulation. In order to obtain accurate
enough results, we must consider a rich enough representation, that is, a large
population of particles. For this purpose, we must consider large enough M and Q.
The large number of particles to be tracked seems a disadvantage of the approach at
first sight, but it must be noticed that the integration of each particle is completely
independent of all the others, making possible the use of HPC on massively parallel
computing platforms.
• The technique introduced for treating purely advective equations can be extended
for considering diffusion contributions as was described in [5] and that we revisit
in what follows, within a fully deterministic approach. Eq. (2.163) can be rewritten
as:
∂Ψ qx qa
+ ∇x · v + Ψ + ∇a · (ȧ + Ψ = 0, (2.166)
∂t Ψ Ψ
or
∂Ψ
+ ∇x · (ṽ Ψ ) + ∇a · ã˙ Ψ = 0, (2.167)
∂t
where the effective velocities ṽ and ã˙ are given by:
ṽ = v − 1
Ψ Dx · ∇x Ψ
. (2.168)
ã˙ = ȧ − 1
Ψ Da · ∇a Ψ
Now, the integration scheme (2.161) can be applied by replacing material and
˙
conformation velocities, v and ȧ, by their effective counterparts ṽ and ã:
τ =t
xi j (t) = xi0j + τ =0 ṽ(xi j (τ )) dτ
τ =t . (2.169)
ai j (t) = ai0j + τ =0 ã˙ i j (ai j (τ ), xi j (τ )) dτ
This fully deterministic particle description requires much less particles that its
stochastic counterpart, but as noticed in Eq. (2.168), the calculation of the effec-
tive material and conformational velocities requires the derivative of the pdf Ψ
with respect to both the physical and the conformational coordinates. To do so,
the distribution must be reconstructed all along the simulation (at each time step),
which constitutes a serious drawback for its implementation on massively parallel
computing platforms. Moreover, to make possible the calculation of the distribu-
tion derivatives, the Dirac distribution must be regularized in order to ensure its
derivability.
2.5.1.2 Separated Representations for Solving Difusion-Dominated

Problems
When the diffusion effects are dominant, the techniques presented in the previous
section become inefficient because they require an excessive number of particles to
produce accurate enough results, in particular for reconstructing the distribution. In
this case, standard mesh-based discretizations seem a better choice. However, as dis-
cussed before, mesh-based discretizations fail when addressing highly dimensional
models as it is the case when addressing the solution of the previous introduced
Fokker-Planck equation. Separated representations seem the most appealing choice.
Considering the Fokker-Planck equation
∂Ψ
+ ∇x · (v Ψ ) + ∇a · (ȧ Ψ )
∂t
= ∇x (Dx · ∇x Ψ ) + ∇a (Da · ∇a Ψ ) , (2.170)
there are many separated representation choices. The most natural one consists in
separating time, physical and conformational spaces, i.e.

N
Ψ (x, t, a) ≈ X i (x) · Ti (t) · Ai (a). (2.171)
i=1
Thus, when proceeding with the Proper Generalized Decomposition—PGD—

constructor [43], we must solve of the order of N 2D or 3D (depending on the
dimension of the physical space) boundary value problems—BVP—for calculating
functions X i (x), the same number of 1D initial value problems—IVP—for calcu-
lating functions Ti (t), and finally the same number of 2D or 5D (the number of
components of tensor a in the 2D or 3D case, where the symmetry and its unit trace
has been taking into account) BVP for calculating functions Ai (a). A discussion
of the difficulties related to the solution of the former multidimensional problems
involving the conformational coordinates was addressed in [44].
2.5.2 Descriptions Based on Higher-Order Kinematics
The models proposed until now consider a first-order kinematics, that is, a constant
velocity gradient at the particle scale. However, when the kinematics become rich
enough or the particle size becomes large with respect to the characteristic flow
dimension, first-order modeling must be enriched. In what follows, we address its
extension to second and third-gradient flow kinematics.
2.5.2.1 Second-Gradient Description with Concentrated Forces

at the Rod Beads
We now consider a rod R but with a higher-order description of the fluid velocity
field at the rod scale. Again, forces applied on each bead F depend on the difference
of velocities between the fluid and the bead, the first one now including the second-
order velocity gradient H according to v(pL) = v0 + ∇v · pL + (H : (p ⊗ p)) L 2
and the second one given by vG + ṗL. Thus, the force F(pL) reads:
F(pL) = ξ (v0 + ∇v · pL + (H : (p ⊗ p)) L 2 − vG − ṗL), (2.172)
1 ∂ 2 vi
where the third-order tensor H is defined by Hi jk = 2 ∂ x j ∂ xk .
Obviously, if F applies on the bead pL, then at the opposite bead −pL the resulting
force reads
F(−pL) = ξ (v0 − ∇v · pL + (H : (p ⊗ p)) L 2 − vG + ṗL). (2.173)
By adding Eqs. (2.172) and (2.173) and enforcing the force balance, neglecting
again inertial effects, we obtain
F(pL) + F(−pL) = 0, (2.174)
than implies v0 − vG = −(H : (p ⊗ p))L 2 , that is, the rod center of gravity has a
relative velocity with respect to the fluid at this position.
Remark 4 The fact of having obtained a non-zero relative velocity v0 −vG = 0 does
not imply the existence of a migration mechanism, as proved in [45, 70] and [71].
Since the resulting torque must also vanish, the only possibility is that force F
acts along p, that is F = λp, with λ ∈ R. Thus we can write
λp = ξ (∇v · pL − ṗL), (2.175)
which yields the same Jeffery equation that was obtained when considering the first-
order velocity gradient:
ṗ = ∇v · p − (∇v : (p ⊗ p)) p. (2.176)
The forces being again aligned in the rod direction, one could infer that the second
gradient does not suffice for activating bending mechanisms.
Until now, forces were assumed applied at the rod ends (beads). However, we can
distribute them all along the rod length as commonly considered when calculating
particles motion by using DPD (dissipative particle dynamics) methods.
In the next section, we consider forces applied all along the rod length as was
considered in [70]. We prove that as soon as a second-gradient description is retained,
bending mechanism appears naturally.
2.5.2.2 Second-Gradient Description of Rods with Distributed Forces
We consider now the system illustrated in Fig. 2.11 consisting of a rod R and the
hydrodynamical forces applied all along its length. With the same reasoning and
notation as above, the applied distributed force f(s) at position sp, s ∈ [−L , L]
reads
f(s) = ξ (v0 + ∇v · ps + (H : (p ⊗ p))s 2 − vG − ṗs). (2.177)

Fig. 2.11 Distributed hydrodynamic forces applied on a rod immersed in a Newtonian fluid, con-
sidering the second-order velocity gradient
The resultant force F must vanish, that is
+L
F= f(s) ds = 0, (2.178)
−L
that leads to the following expression for the sliding velocity (rod-fluid relative
velocity at the rod center of gravity):
L2
v0 − vG = − (H : (p ⊗ p)) . (2.179)
3
Thus, the distributed force reads:

L2
f(s) = ξ ∇v · ps + (H : (p ⊗ p)) s 2 − − ṗs , (2.180)
3
which leads to the moment m(s)

L2
m(s) = sp × ξ ∇v · ps + (H : (p ⊗ p)) s 2 − − ṗs , (2.181)
3
from which we can evaluate the resultant moment M
+L
M= m(s) ds = p × ξ α (∇v · p − ṗ) = 0, (2.182)
−L
+L
with α = −L s 2 ds = 23 L 3 . This again yields Jeffery’s equation
ṗ = ∇v · p − (∇v : (p ⊗ p)) p, (2.183)
i.e. the same equation as that obtained by assuming forces applied at the rod beads.
Introducing Jeffery’s equation (2.183) into the distributed force expression (2.180),
we obtain

L2
f(s) = ξ (H : (p ⊗ p)) s 2 − + (∇v : (p ⊗ p))ps . (2.184)
3
This force can be decomposed into two components, i.e. one, f (s), aligned with
the rod and the other, f ⊥ (s), perpendicular to it:

L2
f (s) = ξ (H ∵ (p ⊗ p ⊗ p))p s 2 − + (∇v : (p ⊗ p))ps , (2.185)
3
and

L2
f ⊥ (s) = ξ ((H : (p ⊗ p)) − (H ∵ (p ⊗ p ⊗ p)) p) s 2 − . (2.186)
3
Remark 5 We can notice from Eq. (2.186) that, when considering distributed forces
within a first-gradient description (H = 0), the perpendicular component of the
resulting distributed forces vanishes, i.e. f ⊥ (s) = 0 and bending mechanisms are
once again absent. However, bending seems possible as soon as second-gradient
descriptions implying H = 0 are retained.
The resultant axial force F reads
L
L2
F = f (s) ds = ξ (∇v : (p ⊗ p))p, (2.187)
2
0
expression that corresponds to the one obtained in the case of concentrated forces if
we consider the following relation between the distributed and concentrated friction
coefficients:
L
ξ = ξ. (2.188)
2
The main difference with respect to the situation in which forces were assumed
applied at the rod beads, occurs when considering the distributed force perpendicular
to the rod. In this case
L
F = f ⊥ (s) ds = 0,
⊥
(2.189)
0
in agreement with the results found when considering concentrated forces, but the
distributed force implies the existence of a bending moment M (s)k (acting in the
out-of-plane direction defined by the unit vector k)
L
M (s)k = (r − s)p × f ⊥ (r ) dr. (2.190)
s
When the rod is rigid there are no major consequences, but in the case of flexible
rods this force creates rod bending with the associated elastic effects.
Remark 6 In the case of elastic rods, the bending moment implies rod curvature.
Within the small strain and displacement hypotheses, the curvature is given by
d 2u M (s)
2
= , (2.191)
ds EI
where u(s) is the rod deflection with respect to its undeformed configuration, E the
elastic modulus and I the moment of area of the rod cross section with respect to
the out-of-plane direction. By integrating this equation twice, we can obtain the rod
bent configuration that, as expected, is symmetric with respect to the rod center of
gravity.
2.5.2.3 Third-Gradient Description
Now, we assume that the forces F that act on each bead depend again on the difference
of velocities between the fluid and the bead, but the former now including the third-
order velocity gradient J, i.e. that implies the fluid velocity at the bead position given
by v(pL) = v0 +∇v ·pL +(H : (p ⊗ p)) L 2 +(J ∵ (p ⊗ p ⊗ p)) L 3 and the second
one by vG + ṗL. Thus, the force F(pL) reads (see Fig. 2.12):
F(pL) = ξ (v0 + ∇v · pL + (H : (p ⊗ p))L 2

+ (J ∵ (p ⊗ p ⊗ p))L 3 − vG − ṗL), (2.192)
where the third-order velocity gradient J is the fourth-rank tensor with components
Ji jkm = 16 ∂ x j ∂∂ xvki∂ xm .
3
Obviously, if F acts on the bead pL, then for the opposite bead at −pL the resulting
force reads
F(−pL) = ξ (v0 − ∇v · pL + (H : (p ⊗ p))L 2

− (J ∵ (p ⊗ p ⊗ p))L 3 − vG + ṗL). (2.193)
By adding Eqs. (2.192) and (2.193) and enforcing the force balance
F(pL) + F(−pL) = 0, (2.194)

Fig. 2.12 Hydrodynamic

forces acting on a rod
immersed in a Newtonian
fluid, considering the
third-order velocity gradient

we obtain again v0 − vG = − H : (p ⊗ p) L 2 .
As the resulting torque must also vanish, the only possibility is that the force F
acts along p, that is F = λp, with λ ∈ R. Thus, we can write

λp = ξ ∇v · pL + (J ∵ (p ⊗ p ⊗ p))L 3 − ṗL , (2.195)
which multiplyied by p yields an expression of λ:

λ = ξ L ∇v : (p ⊗ p) + L 2 (J :: (p ⊗ p ⊗ p ⊗ p)) . (2.196)
We thus have

F = λp = ξ L ∇v : (p ⊗ p)p
+ L 2 (J :: (p ⊗ p ⊗ p ⊗ p))p , (2.197)
which yield the rotary velocity
ṗ = ∇v · p − (∇v : (p ⊗ p)) p + L 2 (J ∵ (p ⊗ p ⊗ p)
− (J :: (p ⊗ p ⊗ p ⊗ p))p), (2.198)
where we can identify Jeffery’s contribution ṗ J and a third-order correction ṗT h that
is affected by L 2 :
ṗ J = ∇v · p − (∇v : (p ⊗ p)) p, (2.199)

and
ṗT h = J ∵ (p ⊗ p ⊗ p) − (J :: (p ⊗ p ⊗ p ⊗ p))p, (2.200)
with the total rotary velocity given by:
ṗ = ṗ J + L 2 ṗT h . (2.201)
2.5.3 Accounting for Rod Bending
In this section, inspired from second-order fluids, we extend this higher-order frame-
work to higher-order rod suspensions.
2.5.3.1 Second-Gradient Fluid Description
Within the first-gradient formulation, the internal power for a Newtonian fluid Wint
reads
Wint = T : ∇v dx, (2.202)
Ω
where T = − pI + τ with τ = 2ηD. The associated balance of momentum reads:
ρ v̇ = ∇ · T. (2.203)
Fried and Gurtin [72] proposed a second-gradient formulation involving the vor-
ticity gradient. They proposed a non-standard form of the principle of virtual power
with the internal power given by

Wint = T : ∇v dx + G : ∇ω dx, (2.204)
Ω Ω
where ω is the vorticity vector

ω = ∇ × v, (2.205)
and G is the so-called hyper-stress. The associated generalized momentum balance

reads:
ρ v̇ = ∇ · T + ∇ × (∇ · G). (2.206)
The following constitutive equation was assumed in [73] for the fluid hyper-stress:

G = ηL f2 ∇ω + ι(∇ω)T , (2.207)
where the parameter ι ∈ [−1, 1] controls the asymmetry of the hyper-stress and
ensures a non-negative dissipation. In the previous equation L f2 > 0 is known as
the fluid gradient length.
In the case of an incompressible fluid, the introduction of both the stress and the
hyper-stress constitutive equations into the generalized momentum balance leads to:

ρ v̇ = −∇ p + ηΔ v − L f2 Δv . (2.208)
where ηL f2 represents the so-called hyper-viscosity [73].

See [73] for a discussion on the associated boundary conditions. The higher-
order velocity derivatives involved in Eq. (2.208) require the enforcement of a larger
number of boundary conditions compared to the standard first-gradient formulation.
2.5.3.2 Second-Gradient Description of a Dilute Suspension

of Flexible Rods
Inspired by the use of the vorticity ω in the above second-gradient fluid flow descrip-
tion, we consider now the rod beads being subjected to two actions, as depicted in
Fig. 2.13:
• A hydrodynamic force F that depends on the difference of velocities between the
fluid and the bead. The fluid velocity at the bead position v(pL) taking into account
second-gradient effects reads
v(pL) = v0 + ∇v · pL + H : (p ⊗ p)L 2 . (2.209)
The bead velocity is again given by vG + ṗL. Thus, the resulting hydrodynamic
force reads
Fig. 2.13 Hydrodynamic forces applied on a flexible rod immersed in a Newtonian fluid
F(pL) = ξ (v0 + ∇v · pL + H : (p ⊗ p)L 2 − vG − ṗL). (2.210)
• A hydrodynamic torque M that depends on the difference between the differen-

tial vorticity at the bead position and the bead rotary velocity ϕ̇. The differential
vorticity is the difference between the vorticity existing at the bead position minus
the one existing at the rod center of gravity. This torque could have its origin in
the distributed forces applied along the rod length as illustrated in Sect. 2.5.2.2.
Thus, the resulting torque reads
M(pL) = ξ R (∇ω · pL − ϕ̇(pL)), (2.211)
where ξ R is the rotary friction coefficient.

A forces balance yields
v0 − vG = −H : (p ⊗ p)L 2 , (2.212)
implying a second-order relative velocity of the rod center of gravity with respect
to the fluid velocity at this position. Again, for notational simplicity, we consider
F = F(pL) = −F(− pL).
As soon as we consider a first gradient of the vorticity-based bending mechanism
∇ω · pL, it is easy to prove [74] that the rod kinematics remains unchanged relative
to Jeffery’s model, and that this term only affects rod bending. Thus, we obtain:
• The standard expression of the forces applying on the rod beads:
F = ξ L (∇v : (p ⊗ p)) p, (2.213)
• Jeffery’s rod kinematics:
ṗ = ∇v · p − (∇v : (p ⊗ p) p, (2.214)
• The bending mechanism. Being ϕ̇ = ϕ̇(pL) = −ϕ̇(− pL) [74] and assuming a
linear elastic behavior of the rod, the relation between the applied torque M and
the bending angle ϕ at the rod bead is given by
L
ϕ= M, (2.215)
EI
or
EI
M= ϕ = κ ϕ, (2.216)
L
where E is the elastic modulus and I the moment of area of the rod cross section.
Introducing the expression of the moment M (2.211) into Eq. (2.216), we obtain
ξ R (∇ω · pL − ϕ̇) = κ ϕ, (2.217)
whose integration allows calculating ϕ(t) and from it the moment M.
Remark 7 It is noteworthy that:

• In first-gradient flows, rods orient according to Jeffery’s equation without experi-
encing bending effects because the vorticity gradient vanishes.
• In second-order flows, rods orient according to Jeffery’s equation experiencing a
bending induced by the second gradient.
• In rigid kinematics, a rod moves with the fluid without experiencing any relative
movement. Since the vorticity gradient vanishes, there are no bending effects,
ensuring the formulation objectivity.
The rod kinematics are then fully described by p and ϕ. Knowing p, we can
locate both rod beads. Then, knowing the bead bending angle ϕ and all the forces
and moments being applied at both ends, the deformed rod configuration is parabolic
and symmetric with respect to the center of gravity. Thus, from (p and ϕ), we can
predict the bent configuration.
The stress has two components, the standard one and the one related to the hyper-
stress. The first one is related to forces applied at both opposite beads acting in the
rod direction and includes the suspending medium contribution τ f = 2ηD:

N
τ = τ f + τ r = 2ηD + pi ⊗ Fi
i=1

N
= 2ηD + β∇v : pi ⊗ pi ⊗ pi ⊗ pi , (2.218)
i=1
with the total stress T given by
T = − p I + τ, (2.219)
which, as expected, is symmetric.

The second one is related to second-gradient fluid contribution and rod bending.
We consider a partition of the total hyper-stress G consisting of the fluid G f and the
rods G̃r contributions, with G = G f + G̃r .
We consider a standard form of the second-gradient fluid contribution G f ,

G f = η L f2 ∇ω + ι(∇ω)T , (2.220)
where L f2 is the fluid gradient length.

On the other hand, we compute the rod contribution to the hyper-stress Gr from

N
Gr = υ pi ⊗ Mi , (2.221)
i=1
from which we assume the more general form

T
G̃r = Gr + ς Gr . (2.222)
Making use of the orientation distribution function Ψ (x, t, p, ϕ), we can proceed
to the calculation of the stress within a continuous macroscopic description. Con-
sidering the stress expression (2.218) and the use of the quadratic closure relation
A ≈ a ⊗ a, we obtain:
τ = τ f + τ r = 2ηD + 2ηN p (D : A) ≈ 2ηD + 2ηN p (D : a) a, (2.223)
with T = − p I + τ .
On the other hand, the hyper-stress is given by:

Gr = υ p ⊗ M Ψ dp dϕ, (2.224)
S ×C
and, by using the expression of the moment,

G = υκ
r
p ⊗ ϕ Ψ dp dϕ
S ×C
= κ̃ p ⊗ ϕ Ψ dp dϕ. (2.225)
S ×C
In the previous expressions, the distribution function Ψ = Ψ (x, t, p, ϕ) (its

dependence on the different coordinates is omitted for the sake of clarity) gives
the fraction of rods that at position x and time t have a conformation given by (p, ϕ).
The domains S and C refer respectively to the domains in which coordinates p and
ϕ are defined.
Remark 8 One could expect that the moment (2.225) vanishes, however taking into
account that Ψ (p, ϕ) = Ψ (−p, −ϕ), the integral does not vanish.
In order to close the formulation, we consider the second-order tensor g (such that
Gr = κ̃ g):
g= p ⊗ ϕ Ψ dp dϕ, (2.226)
S ×C
and compute its time derivative taking into account the expressions of ṗ (2.214) and
ϕ̇ (2.217):

ġ = (ṗ ⊗ ϕ + p ⊗ ϕ̇) Ψ dp dϕ
S ×C

= (∇v · p − (∇v : (p ⊗ p)) p) ⊗ ϕ Ψ dp dϕ
S ×C

κ
+ p ⊗ ∇ω · pL − R ϕ Ψ dp dϕ
S ×C ξ

= ∇v · g − ∇v : p ⊗ p ⊗ p ⊗ ϕ Ψ dp dϕ
S ×C
κ
+ La · (∇ω) − R g. T
(2.227)
ξ
If we adopt the following closure relation:

p ⊗ p ⊗ p ⊗ ϕ Ψ dp dϕ ≈ a ⊗ g, (2.228)
S ×C
then we obtain a closed form for the evolution of g:

κ
ġ = ∇v · g − (∇v : a)g + La · (∇ω)T − g. (2.229)
ξR
Multiplying Eq. (2.229) by κ̃, we obtain the time evolution of the contribution of
rods to the hyper-stress:
κ r
Ġr = ∇v · Gr − (∇v : a)Gr + L κ̃a · (∇ω)T − G . (2.230)
ξR
If we consider the coefficient affecting the vorticity gradient in Eq. (2.230) and
take into account the relations κ̃ = υκ and κ = E I /L, we obtain
L κ̃ = E(υ I ) = ELr2 , (2.231)
where Lr2 represents the rod gradient length. The last term in equation (2.230)
involves the coefficient κ/ξ R with reciprocal time units, which, in absence of flow,
controls the relaxation to the undeformed reference configuration. Thus, the inverse of
this coefficient has the meaning of a relaxation time that we denote by T . Moreover,
taking into account the symmetry of tensor a, we have ∇v : a = D : a. Thus, Eq.
(2.230) can be rewritten as
1 r
Ġr = ∇v · Gr − (D : a) Gr + ELr2 a · (∇ω)T − G . (2.232)
T
It is important to notice that the above equation involves two closure relations,
the one related to the fourth-order orientation tensor A and the one expressed in Eq.
(2.228).
The macroscopic flow model can thus be summarized as follows:
• Generalized momentum balance:
ρ v̇ = ∇ · T + ∇ × (∇ · G), (2.233)
• Mass balance
∇ · v = 0, (2.234)
• Constitutive equation:
T = − p I + τ, (2.235)
with τ = τ f + τ r :
τ f = 2ηD, (2.236)
and
τ r = 2ηN p (∇v : a)a = 2ηN p (D : a)a, (2.237)
• Hyper-stress G = G f + G̃r :

G f = ηL f2 ∇ω + ι(∇ω)T , (2.238)
and T
G̃r = Gr + ς Gr , (2.239)
with
1 r
Ġr = ∇v · Gr − (D : a)Gr + ELr2 a · (∇ω)T − G . (2.240)
T
2.5.4 Delaying Orientation Mechanisms
While the models proposed until now allowed the introduction of rod bending mech-
anisms, the rotary velocity remained unchanged however. In this section, we propose
alternative models able to control the rod rotary velocity.
A simple proposals consists in considering rods subjected to three actions:
1. The hydrodynamic force, which reads as in previous sections
F(pL) = ξ (v0 + ∇v · pL + (H : (p ⊗ p))L 2 − vG − ṗL). (2.241)

2. A torque acting at each bead involving the first gradient of potential Υ . Thus, the
resulting torque at bead pL reads
M(pL) = ξ I (∇Υ · pL − ϕ̇(pL)). (2.242)
3. A torque M at the rod centre of gravity scaling with the difference between Υ
and the objective rotary velocity θ̇ − ω,
M = ξ R (Υ − (θ̇ − ω)). (2.243)
Here, ω is the vorticity: ω = ∇ × v, such that ω × p = Ω · p, with Ω the

antisymmetric component of the velocity gradient.
A force balance yields again
v0 − vG = −H : (p ⊗ p)L 2 . (2.244)
By enforcing the moment balance, we obtain
2Lp × F + M + M(pL) + M(−pL) = 0, (2.245)
which reads

2ξ L 2 p × ∇v · p − θ̇ +ξ R (Υ −(θ̇ −ω))−ξ I (ϕ̇(pL) + ϕ̇(−pL)) = 0. (2.246)
Taking into account p × Ω · p = p × ω × p = ω, we obtain the rod rotary velocity
2ξ L 2
θ̇ = ω + (p × D · p)
2ξ L 2 + ξ R
ξR 2ξ I
+ Υ − ϕ̇, (2.247)
2ξ L + ξ
2 R 2ξ L 2 + ξ R
with D being the symmetric component of the velocity gradient and ϕ̇= ϕ̇(pL) +2ϕ̇(−pL) .
The previous equation can be rewritten as:
2ξ L 2
ṗ = Ω · p + (D · p − (D : (p ⊗ p))p)
2ξ L 2 + ξ R
ξR 2ξ I
+ Υ × p − ϕ̇ × p, (2.248)
2ξ L 2 + ξ R 2ξ L 2 + ξ R
where it can be noticed that the objectivity is ensured, and for that, consideration of
the relative velocity (θ̇ − ω) in the expression of M was a key choice.
Remark 9 With ξ R = ξ I = 0, we recover the Jeffery kinematics. With ξ I = 0, we

remove elastic effects but the Jeffery kinematics are affected when considering ξ R
large enough. In the case of ξ R 2ξ L 2 (or ξ R → ∞), we obtain
ṗ ≈ Ω · p + Υ × p = (ω + Υ ) × p. (2.249)
When Υ = −ω, rods can translate without rotating at all. On the other hand, if
Υ = 0, we obtain rigid kinematics with ṗ = Ω · p governed by the flow vorticity.
The force acting on the bead reads

2ξ L 2 2ξ L 2
F(pL) = ξ L D · p 1 − + (D : (p ⊗ p))p
2ξ L 2 + ξ R 2ξ L 2 + ξ R
ξ Rξ L 2ξ I ξ L
− Υ ×p+ ϕ̇ × p. (2.250)
2ξ L + ξ
2 R 2ξ L 2 + ξ R
The force applied on the beads has a component acting along the rod direction
F (pL) = ξ L(D : (p ⊗ p))p, (2.251)
and another perpendicular to it and contained in the plane in which the rod orientation
is defined
ξ Rξ L
F⊥ (pL) = (D · p − (D : (p ⊗ p))p)
2ξ L 2 + ξ R
ξ Rξ L 2ξ I ξ L
− Υ × p + ϕ̇ × p. (2.252)
2ξ L 2 + ξ R 2ξ L 2 + ξ R
This implies that

F⊥ (−pL) = −F⊥ (pL). (2.253)
This modeling approach allows for modifying the Jeffery kinematics while keep-
ing a first-gradient description of the induced torques able to activate rod bending.
The choice of the interaction potential Υ is the key point of this modeling
approach. A simple choice consists in assuming that at position pL the microstruc-
ture is described by the second-order moment a(x, t) of the orientation distribution
function ψ(x, t, p). The simplest choice for the interaction potential Υ is the rotary
velocity of the eigenvectors of tensor a.
2.5.5 Collective Effects
We start by obtaining the Jeffery equation but now with the force F applied at the
rod beads (see Fig. 2.14) calculated from the Cauchy stress τ .
Fig. 2.14 Forces applied on

a rod immersed in a flowing
fluid
Thus, we write
F(pL) = τ · p − κ(ṗ − Ω · p), (2.254)
where κ is a model parameter whose impact will be analyzed later. The fact of
considering the term involving the vorticity tensor Ω ensures the objectivity of the
resulting rotary velocity.
We can notice that if F applies at bead located at pL, then the force at the opposite
bead at −pL reads
F(−pL) = −F(pL), (2.255)
which ensures the equilibrium of forces. For the sake of notational simplicity, the
force acting on bead pL will be noted simply by F.
As the resulting torque must also vanish, the only possibility is that force F acts
along p, that is F = λp, with λ ∈ R. Thus, we can write
λp = τ · p − κ(ṗ − Ω · p). (2.256)
Premultiplying Eq. (2.256) by p and taking into account that p · p = 1 and

consequently p · ṗ = 0, and that Ω : (p ⊗ p) = 0 since Ω is skew-symmetric, we
have
λ = τ : (p ⊗ p), (2.257)
which leads to the extended Jeffery equation
1
ṗ = Ω · p + (τ · p − (τ : (p ⊗ p)) p) . (2.258)
κ
When considering τ = 2ηD and κ = 2η, Eq. (2.258) reduces to the standard
Jeffery equation.
The forces applied at the rod ends pL and −pL are respectively λp and −λp, i.e.
directed along the rod and in equilibrium by construction.
With λ given by Eq. (2.257), we have
F(pL) = (τ : (p ⊗ p)) p. (2.259)
By applying Kramers’ formula, the corresponding contribution to the stress is

given by
p = (τ : (p ⊗ p))p ⊗ p, (2.260)
which can be rewritten as
p = τ : (p ⊗ p ⊗ p ⊗ p). (2.261)
For considering the effect of all rods in the suspensions, it suffices to consider
the integral of each contribution averaged with the orientation distribution function,
adding the Newtonian matrix contribution 2ηD. The self-consistency implies:
τ = 2ηD + A : τ . (2.262)
We now consider the equation governing the evolution of the second-order

moment a. For this purpose, we take its time derivative

ȧ = (ṗ ⊗ p + p ⊗ ṗ) ψ dp, (2.263)
S
and consider the extended Jeffery expression (2.258) to express the rod rotary velocity
ṗ. Thus, we obtain
1 2
ȧ = Ω · a − a · Ω + (τ · a + a · τ ) − A : τ . (2.264)
κ κ
Remark 10 When considering again τ = 2ηD and κ = 2η, Eq. (2.264) reduces
to the standard orientation evolution equation. However, by considering the same
constitutive equation τ = 2ηD but κ = 2η, we obtain a sort of sliding while ensuring
the objectivity of the evolution equation that involves the so-called Johnson-Segalman
objective derivative. Some authors introduced such sliding affecting the gradient of
velocities, a choice that violates the objectivity of the resulting model.
2.5.5.1 Jeffery’s Model with Feedback
We consider in the case of dilute rod suspensions the Cauchy stress
τ = 2ηD + 2ηN p A : D, (2.265)
that was derived from the standard Jeffery model.

By considering the stress (2.265) in the extended Jeffery expression (2.258), we

obtain
2η
ṗ = Ω · p + (D · p − (D : (p ⊗ p)) p)
κ
2ηN p
− ((A : D) · p − ((A : D) : (p ⊗ p)) p) . (2.266)
κ
The associated orientation equation governing the evolution of ȧ (2.264) reads
2η
ȧ = Ω · a − a · Ω + (D · a + a · D − 2A : D)
κ
2ηN p
− ((A : D) · a + a · (A : D)T − 2A : (A : D)). (2.267)
κ
Remark 11 By considering κ = 2η and N p = 0 in Eqs. (2.266) and (2.267), we
obtain the standard orientation equations. The extra-term, the one affected by the
coefficient N p , is related to the feedback.
2.5.6 Orientation Induced by an Electric Field
In this section, we propose first the equation governing the orientation of a rod
immersed in a Newtonian fluid of viscosity η considering an electrical field ε(x, t)
and a velocity field v(x, t). Then, this model will be introduced in coarser descriptions
of the rod population within the kinetic theory framework.
We consider a suspending medium consisting of a Newtonian fluid in which there
are suspended N rigid slender rods (e.g. CNTs) of length 2L. As first approximation,
it is assumed that the rod presence and orientation do not affect the flow kinematics
that are defined by the velocity field v(x, t), with x ∈ Ω ∈ Rd .
The microstructure can be described at the microscopic scale by the unit vector
defining the orientation of each rod, i.e. pi , i = 1, . . . , N . In absence of electrical
field, one fiber can be defined by p or −p and a symmetry condition is retained in
the orientation distribution. However, when considering the electrical field induced
charges, that symmetry is broken and the orientation is defined univocally. We assume
that p points from the negative bead to the positive one.
If the suspension is dilute enough, rod interaction can be neglected and then a
micro-mechanical model can be derived by considering a single generic rod whose
orientation is defined by the unit vector p.
We consider the system illustrated in Fig. 2.15 consisting of a rod immersed in a
fluid flow and experiencing the effects of an external electric field.
The resulting forces applied at the rod beads, again taking into account that v0 =
vG , read:
F(pL) = E + ξ L (∇v · p − ṗ) . (2.268)
Fig. 2.15 Hydrodynamic

and electrostatic forces
applied on a rod immersed in
a Newtonian fluid
Obviously, if F applies on the bead pL, then at the opposite bead −pL, the
resulting force reads
F(−pL) = −E − ξ L (∇v · p − ṗ) , (2.269)
where both beads are assumed having an electrical charge of opposite sign. Thus,
the linear momentum balance is ensured. As the resulting torque must also vanish,
the only possibility is that force F acts along p, that is F = λp, with λ ∈ R. Thus we
can write
λp = E + ξ L (∇v · p − ṗ) . (2.270)
Premultiplying Eq. (2.270) by p and taking into account that p · p = 1 and

consequently p · ṗ = 0, we obtain:

λ = E · p + ξ L pT · ∇v · p , (2.271)
implying
F = (E · p) p + ξ L pT · ∇v · p p, (2.272)
and
1
ṗ = (E − (E · p)p) + ∇v · p − pT · ∇v · p p . (2.273)
ξL
This can be rewritten as:

1
ṗ = (I − p ⊗ p) · E + ∇v · p − pT · ∇v · p p
ξL
= ṗ E + ṗ J , (2.274)
where ṗ E stands for the rotary velocity due to the electrostatic forces and ṗ J repre-
sents the hydrodynamic contribution (described by the Jeffery equation).
As the rod population is very large, the description that we just proposed, despite
its conceptual simplicity, fails to address the situations usually encountered in prac-
tice. For this reason, coarser descriptions are preferred. The first plausible coarser
description applies a zoom-out, in which the rod individuality is lost in favour of
probability distribution function based descriptions.
In the case of rods, one could describe the microstructure at a certain point x and
time t from the orientation distribution function Ψ (x, t, p) that gives the fraction of
rods that at position x and time t are oriented in direction p. Obviously, the function
Ψ verifies the normality condition:

Ψ (x, t, p) dp = 1, ∀x, ∀t, (2.275)
S
where S is the surface of the unit ball that defines all possible rod orientations. The
balance ensuring the probability conservation reads:
∂Ψ
+ ∇x · (ẋ Ψ ) + ∇ p · (ṗ Ψ ) = 0, (2.276)
∂t
wherein for inertialess rods ẋ = v(x, t) and the rod rotary velocity is given by Eq.
(2.274),
1
ṗ = (I − p ⊗ p) · E + ∇v · p − pT · ∇v · p p . (2.277)
ξL
In this section, we illustrate the transition from the mesoscopic to the macroscopic
scale. At the macroscopic scale, the orientation distribution function is substituted
by its moments for describing the microstructure. We consider the first to fourth
orientation moments a(1) , a(2) , a(3) and a(4) , respectively defined from:

(1)
a = p Ψ dp, (2.278)
S

(2)
a = p ⊗ p Ψ dp, (2.279)
S

a(3) = p ⊗ p ⊗ p Ψ dp (2.280)
S
and
(4)
a = p ⊗ p ⊗ p ⊗ p Ψ dp. (2.281)
S
The time derivative of a(1) reads

ȧ(1) = ṗ Ψ dp = ṗ E + ṗ J Ψ dp
S S

1
= (I − p ⊗ p) · E Ψ dp + ∇v · p − pT · ∇v · p p Ψ dp
ξL S S
1
= E − a · E + ∇v · a − ∇v : a(3) .
(2) (1)
(2.282)
ξL
In the previous expression, a closure issue appears related to a(2) and a(3) . We
could evaluate the time derivative of the second moment, but again higher-order
moments are involved and the closure issue persists.
2.5.6.3 Introducing Randomizing Mechanisms
The main issue of the modeling approach just presented is that it tends to fully align
rods along a given direction, that is:
p(t → ∞) = p̂, (2.283)
and
a(1) (t → ∞) = p̂, (2.284)
where p̂ is a given orientation that results as a compromise between the ones enforced
by the flow and the electrical field.
In flowing systems, full alignment is prevented by rod-rod interactions that have
a randomizing effect. This effect can be accurately modeled with a diffusion term in
the Fokker-Planck equation, that moreover accounts for Brownian effects:
∂Ψ
+ ∇x · (vΨ ) + ∇ p · (ṗ Ψ ) = ∇ p · Dr ∇ p Ψ , (2.285)
∂t
where Dr quantifies the isotropic Brownian diffusion. The rotary velocity is given
again by:
1
ṗ = (I − p ⊗ p) · E + ∇v · p − pT · ∇v · p p . (2.286)
ξL
Equation (2.285) can be rewritten as:
∂Ψ
+ ∇x · (v Ψ ) + ∇ p · p̃˙ Ψ = 0, (2.287)
∂t
with the equivalent rotary velocity given by
∇pΨ
p̃˙ = ṗ Ψ − Dr . (2.288)
Ψ
Now, coming back to the macroscopic scale, the evolution of the first moment
of the orientation distribution function, following the rationale considered in [43]
(Sect. 3.3), is obtained:
1
ȧ(1) ≈ E − a(2) · E + ∇v · a(1) − ∇v : a(3) − Dr a(1) , (2.289)
ξL
which ensures, in absence of electrical field E = 0 and absence of flow v(x, t) = 0,

a fully random (isotropic) distribution, i.e. a(1) (t → ∞; E = 0, v = 0) = 0.
References
1. G.B. Jeffery, The motion of ellipsoidal particles immersed in a viscous fluid. Proc. R. Soc.
Lond. A102, 161–179 (1922)
2. R. Keunings, in Micro-Macro Methods for the Multiscale Simulation of Viscoelastic Flow Using
Molecular Models of Kinetic Theory. eds. by D.M. Binding, K. Walters. Rheology Reviews
(British Society of Rheology, Durham, 2004), pp. 67–98
3. R.B. Bird, C.F. Curtiss, R.C. Armstrong, O. Hassager, in Dynamic of Polymeric Liquid. Kinetic
Theory, vol 2 (Wiley, New York, 1987)
4. M. Doi, S.F. Edwards, The Theory of Polymer Dynamics (Clarendon Press, Oxford, 1987)
5. A. Ammar, F. Chinesta, A particle strategy for solving the Fokker-Planck equation governing
the fibre orientation distribution in steady recirculating flows involving short fibre suspensions.
Lect. Notes Comput. Sci. Eng. 43, 1–16 (2005)
6. C.V. Chaubal, A. Srinivasan, O. Egecioglu, L.G. Leal, Smoothed particle hydrodynamics tech-
niques for the solution of kinetic theory problems. J. Non-Newtonian Fluid Mech. 70, 125–154
(1997)
7. C. Chauviere, A. Lozinski, Simulation of dilute polymer solutions using a Fokker-Planck
equation. Comput. Fluids. 33, 687–696 (2004)
8. F. Chinesta, G. Chaidron, A. Poitou, On the solution of the Fokker-Planck equation in steady
recirculating flows involving short fibre suspensions. J. Non-Newtonian Fluid Mech. 113, 97–
125 (2003)
9. C. Cruz, L. Illoul, F. Chinesta, G. Regnier, Effects of a bent structure on the linear viscoelastic
response of carbon nanotube diluted suspensions. Rheol. Acta 49, 1141–1155 (2010)
10. C. Cruz, F. Chinesta, G. Regnier, Review on the Brownian dynamics simulation of bead-rod-
spring models encountered in computational rheology. Arch. Comput. Methods Eng. 19(2),
227–259 (2012)
11. H.C. Öttinger, M. Laso, in Smart Polymers in Finite-Element calculations, eds. by P. Mold-
enaers, R. Keunings. Theoretical and Applied Rheology, Proceedings on XIth International
Congress on Rheology, vol. 1 (Elsevier, Amsterdam, 1992), pp. 286–288
12. P. Wapperom, R. Keunings, V. Legat, The backward-tracking Lagrangian particle method for
transient viscoelastic flows. J. Non-Newtonian Fluid Mech. 91, 273–295 (2000)
13. P. Wapperom, R. Keunings, Numerical simulation of branched polymer melts in transient
complex flows using pom-pom models. J. Non-Newtonian Fluid Mech. 97, 267–281 (2001)
14. A. Lozinski, C. Chauviere, A fast solver for Fokker-Planck equation applied to viscoelastic
flows calculations: 2D FENE model. J. Comput. Phys. 189, 607–625 (2003)
complex fluids. J. Non-Newtonian Fluid Mech. 139, 153–176 (2006)
complex fluids. Part II: transient simulation using space-time separated representations. J.
Non-Newtonian Fluid Mech. 144, 98–121 (2007)
17. F. Chinesta, A. Ammar, A. Leygue, R. Keunings, An overview of the proper generalized
decomposition with applications in computational rheology. J. Non-Newtonian Fluid Mech.
166, 578–592 (2011)
18. F. Chinesta, R. Keunings, A. Leygue, The Proper Generalized Decomposition for Advanced
Numerical Simulations (Springer, New York, 2014)
19. B. Mokdad, E. Pruliere, A. Ammar, F. Chinesta, On the simulation of kinetic theory models of
complex fluids using the Fokker-Planck approach. Appl. Rheol. 17(2), 1–14, 26494 (2007)
20. K. Chiba, A. Ammar, F. Chinesta, On the fiber orientation in steady recirculating flows involving
short fibers suspensions. Rheol. Acta 44, 406–417 (2005)
21. R. Keunings, On the Peterlin approximation for finitely extensible dumbbells. J. Non-
Newtonian Fluid Mech. 68, 85–100 (1997)
22. F. Folgar, C. Tucker, Orientation behavior of fibers in concentrated suspensions. J. Reinf. Plast.
Comp. 3, 98–119 (1984)
23. G.K. Batchelor, The stress system in a suspension of force-free particles. J. Fluid Mech. 41,
545–570 (1970)
24. G.L. Hand, A theory of anisotropic fluids. J. Fluid Mech. 13, 33–62 (1962)
25. J. Hinch, G. Leal, The effect of Brownian motion on the rheological properties of a suspension
of non-spherical particles. J. Fluid Mech. 52, 683–712 (1972)
26. J. Hinch, G. Leal, Constitutive equations in suspension mechanics. Part I. J. Fluid Mech. 71,
481–495 (1975)
27. J. Hinch, G. Leal, Constitutive equations in suspension mechanics. Part II. J. Fluid Mech. 76,
187–208 (1976)
28. S. Advani (ed.), Flow and Rheology in Polymer Composites Manufacturing (Elsevier, Amster-
dam, 1994)
29. J. Azaiez, K. Chiba, F. Chinesta, A. Poitou, State-of-the-art on numerical simulation of fiber-
reinforced thermoplastic forming processes. Arch. Comput. Methods Eng. 9(2), 141–198
(2002)
30. E. Cueto, R. Monge, F. Chinesta, A. Poitou, I. Alfaro, M. Mackley, Rheological modeling and
forming process simulation of CNT nanocomposites. Int. J. Mater. Form. 3(2), 1327–1338
(2010)
31. M.A. Martinez, E. Cueto, M. Doblar, F. Chinesta, Natural element meshless simulation of
injection processes involving short fiber suspensions. J. Non-Newtonian Fluid Mech. 115,
51–78 (2003)
32. C. Tucker, Flow regimes for fiber suspensions in narrow gaps. J. Non-Newtonian Fluid Mech.
39, 239–268 (1991)
33. A. Ma, F. Chinesta, M. Mackley, The rheology and modelling of chemically treated carbon
nanotube suspensions. J. Rheol. 53(3), 547–573 (2009)
34. C. Petrie, The rheology of fibre suspensions. J. Non-Newtonian Fluid Mech. 87, 369–402
(1999)
35. J. Ferec, G. Ausias, M.C. Heuzey, P. Carreau, Modeling fiber interactions in semiconcentrated
fiber suspensions. J. Rheol. 53(1), 49–72 (2009)
36. J. Wang, C.A. Silva, J.C. Viana, F.W.J. van Hattum, A.M. Cunha, C. Tucker, Prediction of fiber
orientation in a rotating compressing and expanding mold. Polym. Eng. Sci. 48(7), 1405–1413
(2008)
References 107
37. J. Wang, J. O’Gara, C. Tucker, An objective model for slow orientation kinetics in concentrated
fiber suspensions: theory and rheological evidence. J. Rheol. 52(5), 1179–1200 (2008)
38. J. Phelps, C. Tucker, An anisotropic rotary diffusion model for fiber orientation in short and
long fiber thermoplastics. J. Non-Newtonian Fluid Mech. 156(3), 165–176 (2009)
39. A. Ma, F. Chinesta, A. Ammar, M. Mackley, Rheological modelling of carbon nanotube aggre-
gate suspensions. J. Rheol. 52(6), 1311–1330 (2008)
40. G. Ausias, X.J. Fan, R. Tanner, Direct simulation for concentrated fibre suspensions in transient
and steady state shear flows. J. Non-Newtonian Fluid Mech. 135, 46–57 (2006)
41. S. Le Corre, D. Caillerie, L. Orgéas, D. Favier, Behavior of a net of fibers linked by viscous
interactions: theory and mechanical properties. J. Mech. Phys. Solids 52(2), 395–421 (2004)
42. S. Le Corre, P. Dumont, L. Orgéas, D. Favier, Rheology of highly concentrated planar fiber
suspensions. J. Rheol. 49(5), 1029 (2005)
43. F. Chinesta, From single-scale to two-scales kinetic theory descriptions of rods suspensions.
Arch. Comput. Methods Eng. 20(1), 1–29 (2013)
44. E. Abisset-Chavanne, R. Mezher, S. Le Corre, A. Ammar, F. Chinesta, Kinetic theory
microstructure modeling in concentrated suspensions. Entropy 15, 2805–2832 (2013)
45. E. Abisset-Chavanne, F. Chinesta, J. Ferec, G. Ausias, R. Keunings, On the multiscale descrip-
tion of dilute suspensions of non-Brownian rigid clusters composed of rods. J. Non-Newtonian
Fluid Mech. (in press)
46. S. Advani, C. Tucker, The use of tensors to describe and predict fiber orientation in short fiber
composites. J. Rheol. 31, 751–784 (1987)
47. S. Advani, C. Tucker, Closure approximations for three-dimensional structure tensors. J. Rheol.
34, 367–386 (1990)
48. F. Dupret, V. Verleye, in Modelling the Flow of Fibre Suspensions in Narrow Gaps, eds. by
D.A. Siginer, D. De Kee, R.P. Chabra. Advances in the Flow and Rheology of Non-Newtonian
Fluids (Elsevier, Rheology Series, Amsterdam, 1999), pp. 1347–1398
49. M. Kroger, A. Ammar, F. Chinesta, Consistent closure schemes for statistical models of
anisotropic fluids. J. Non-Newtonian Fluid Mech. 149, 40–55 (2008)
50. E. Pruliere, A. Ammar, N. El Kissi, F. Chinesta, Recirculating flows involving short fiber
suspensions: numerical difficulties and efficient advanced micro-macro solvers. Arch. Comput.
Methods Eng. State Art Rev. 16, 1–30 (2009)
51. P. Grassia, J. Hinch, L.C. Nitsche, Computer simulations of brownian motion of complex
systems. J. Fluid Mech. 282, 373–403 (1995)
52. P. Grassia, J. Hinch, Computer simulations of polymer chain relaxation via brownian motion.
J. Fluid Mech. 308, 255–288 (1996)
53. H.C. Öttinger, Stochastic Processes in Polymeric Fluids (Springer, Berlin, 1996)
54. P. Dumont, S. Le Corre, L. Orgeas, D. Favier, A numerical analysis of the evolution of bundle
orientation in concentrated fibre-bundle suspensions. J. Non-Newtonian Fluid Mech. 160, 76–
92 (2009)
55. J. Ferec, E. Abisset-Chavanne, G. Ausias, F. Chinesta, The use of interaction tensors to describe
and predict rod interactions in rod suspensions. Rheol. Acta 53(5–6), 445–456 (2014)
56. E. Pichelin, T. Coupez, Finite element solution of the 3D mold filling for viscous incompressible
fluid. Comput. Methods Appl. Mech. Eng. 163, 359 (1998)
57. D. Gonzalez, E. Cueto, F. Chinesta, M. Doblare, A natural element updated Lagrangian strategy
for free-surface fluid dynamics. J. Comput. Phys. 223(1), 127–150 (2007)
58. T. Belytschko, Y. Krongauz, D. Organ, M. Fleming, P. Krysl, Meshless methods: an overview
and recent developments. Comput. Methods Appl. Mech. Eng. 139, 3–47 (1998)
59. F. Chinesta, S. Cescotto, E. Cueto, P. Lorong, Natural Element Method for the Simulation of
Structures and Processes (ISTE Wiley, London, 2011)
60. R. Sibson, A vector identity for the Dirichlet tesselation. Math. Proc. Camb. Philos. Soc. 87,
151–155 (1980)
61. V. Belikov, V. Ivanov, V. Kontorovich, S. Korytnik, A. Semenov, The non-sibsonian interpo-
lation: a new method of interpolation of the values of a function on an arbitrary set of points.
Comput. Math. Math. Phys. 37(1), 9–15 (1997)
62. E. Cueto, M. Doblare, L. Gracia, Imposing essential boundary conditions in the natural element
method by means of density-scaled alpha-shapes. Int. J. Numer. Methods Eng. 49(4), 519–546
(2000)
63. J. Yvonnet, D. Ryckelynck, P. Lorong, F. Chinesta, A new extension of the natural element
method for non-convex and discontinuous problems: the constrained natural element method.
Int. J. Numer. Methods Eng. 60(8), 1452–1474 (2004)
64. J.A. Garcia, Ll. Gascon, E. Cueto, I. Ordeig, F. Chinesta, Meshless methods with application to
resin transfer molding simulation. Comput. Methods Appl. Mech. Eng. 198, 27002709 (2009)
65. E. Cueto, M. Laso, F. Chinesta, Meshless stochastic simulation of maicro-macro kinetic theory
models. Int. J. Multiscale Comput. Eng. 9(1), 1–16 (2011)
66. E. Cueto, A. Ma, F. Chinesta, M. Mackley, Numerical simulation of spin coating processes
involving functionalised carbon nanotube suspensions. Int. J. Mater. Form. 1(2), 89–99 (2008)
67. J. Yvonnet, P. Villon, F. Chinesta, Natural element approximations involving bubbles for treat-
ing mechanical models in incompressible media. Int. J. Numer. Methods Eng. 66(7), 1125–1152
(2006)
68. A. Ma, F. Chinesta, M. Mackley, A. Ammar, The rheological modelling of carbon nanotube
(CNT) suspensions in steady shear flows. Int. J. Mat. Form. 2, 83–88 (2008)
69. F. Chinesta, A. Ammar, A. Falco, M. Laso, On the reduction of stochastic kinetic theory models
of complex fluids. Model. Simul. Mater. Sci. Eng. 15, 639–652 (2007)
70. R. Shanker, J.W. Gillespie, S.I. Güçeri, On the effect of nonhomogeneous flow fields on the
orientation distribution and rheology of fiber suspensions. Polym. Eng. Sci. 31, 161–171 (1991)
71. R. Shanker, The effect of non homogeneous flow fields and hydrodynamic interactions on the
rheology of fiber suspensions, Ph.D. dissertation, University of Delaware, 1991
72. E. Fried, M.E. Gurtin, Tractions, balances, and boundary conditions for non-simple materials
with application to liquid flow at small length scales. Arch. Ration. Mech. Anal. 182, 513–554
(2006)
73. T. Kim, J. Dolbow, E. Fried, A numerical method for a second-gradient theory of incompressible
fluid flow. J. Comput. Phys. 223, 551–570 (2007)
74. E. Abisset-Chavanne, J. Ferec, G. Ausias, E. Cueto, F. Chinesta, R. Keunings, A second-
gradient theory of dilute suspensions of flexible rods in a Newtonian fluid. Arch. Comput.
Methods Eng. (in press)
Chapter 3
Flows of Simple Fluids in Complex
Microstructures: Composite Processing
of Structural Polymer Composites
Abstract Continuous fiber reinforced polymer (CFRP) always involves a fluid flow
through the fibrous medium, whether to create a semi-product like a prepreg that will
be further processed to make the final part or to directly manufacture a part from
dry fiber reinforcement. This chapter discusses the physics and modeling of flow of
simple fluids within complex microstructure. Microstructure refers here to features
of internal structure of engineered fibrous materials used to reinforce polymer com-
posites. As already mentioned earlier in this book, the flow is greatly influenced by
the type of fibers being used and is viewed as key for ensuring successful fabrication.
This chapter has a complementary focus as compared to Chap. 2. Here only media
made of continuous fibers are considered. An important characteristic of continuous
fibers is that they cannot flow with the polymer, even if the viscosity of the polymer
is low. This chapter reviews the modeling of viscous resin into a porous network of
stationary fibers, at a variety of length scales. As far engineered fibrous materials
are concerned, the appropriate length scales are those that reveal the partition of the
fibrous media into regions of nearly continuous phase such as continuous fibers and
cluster of fibers called fiber bundles.
3.1 Introduction
Polymer composites are usually divided into two groups, advanced (or high perfor-
mance) composites and engineering composites. They differ principally in the type
and length of the fiber reinforcement and in the characteristics of the resins used.
Polymer reinforced by a high content of continuous fiber (carbon or glass) belongs to
the second one. However there is not sharp division boundary between both groups,
one type of composite changes gradually into the other. An important characteristic
of continuous fibers is that they cannot flow with the polymer, even if the viscosity of
the polymer is low. Continuous fiber reinforced polymer (CFRP) always involves a
fluid flow through the fibrous medium, whether to create a semi-product like a prepreg
that will be further processed to make the final part or to directly manufacture a part
from dry fiber reinforcement. To design more efficiently and cost effectively and
to develop improved composite materials, it is necessary to understand and model
DOI 10.1007/978-3-319-16757-2_3
110 3 Flows of Simple Fluids in Complex Microstructures: Composite Processing …
several important aspects of composite manufacturing processes. It is the purpose

of composite manufacturing science to develop a scientific basis for understanding
how process variables affect final properties of composite parts. The primary goal
of composite manufacturing science is to develop the knowledge and engineering
tools to control the manufacturing process to ensure complete and uniform curing or
solidification, minimize chemical and thermal residual stresses, control fiber/resin
content and ensure accurate fiber placement. This requires models that can predict
the influences of key process variables and techniques for monitoring these variables
so that pressure and temperature can be adjusted accordingly. Such models would
also provide useful guidelines for tooling design. Modeling and simulation tools for
processes can speed up the development cycle for new manufacturing processes and
innovative designs.
In this chapter we will focus on flow phenomena in fibrous microstructures
encountered in structural composites, which are viewed as key for ensuring success-
ful fabrication. Interested readers can refer to [1, 2, 9] for more detailed information
about composite manufacturing processes that involve flow of liquid resin within
fibrous microstructures and their modeling.
The fluid flow through the fibrous medium can be viewed as flow through tortuous
channels (also called capillaries). Figure 3.1 (right) shows an example of channels
in an assembly of woven fabrics. When the porosity increases, it can be viewed as
flow over objects (Fig. 3.1–left) where drag models can be applied.
Most processes for structural composites can be classified as either contact mold-
ing using single-sided molds and low pressures, or closed molding using a matched
pair of molds and higher pressure. Contact molding processes include hand lay-up,
liquid resin infusion in LCM technologies, in or out of autoclave molding and filament
winding. Closed molding processes include Resin Transfer Molding, compression
molding, and pultrusion.
Accordingly there are many material forms used in composite structures (Fig. 3.2).
The fibers can be continuous or discontinuous, oriented or disoriented (random).
They can be found as dry fibers or pre-impregnated with resin (prepreg or premix).
Fig. 3.1 Two cases of flow through fibrous medium: left Flow over isolated fibers (cross-sectional
view) and right Tortuous channels formed by the textile structure (image obtained by X-ray
tomography)
3.1 Introduction 111
Fig. 3.2 Material/process for structural composites manufacturing
If structural performance is an important design parameter, then continuous rein-

forcements are usually selected because discontinuous fibers yield lower mechanical
properties (Fig. 3.3). Rovings and tows are an unidirectional assembly of continuous
fiber. The roving or yarn is the basic material form that can be chopped, transformed
by textile operations to create woven, stitched or braided forms or prepregged into
other product forms.
This section provides a brief overview of primary manufacturing techniques that
are currently used to make structural composites, in which flow of resin occurs.
Although almost any part can be fabricated by hand layup, this time-consuming
and hand labor-intensive process is not considered here because flow phenomena
are controlled by fabrication workers and not by a mechanical device. Indeed, after
several dry fabrics have been laid on the tool and resin spread on them, the plies are
debulked by applying pressure to the laminate pile with a roller to remove voids and
to ensure that the stacked plies are sufficiently compacted.
Fig. 3.3 Influence of fiber

orientation and length on
processing and composite
performance
In autoclave molding, after several plies have been laid (manually or with a
machine) on the tool, pressure is applied to the laminate pile to remove voids and
to ensure that the stacked plies are sufficiently compacted. Then, the part is bagged
and sealed before being cured in an autoclave. The bagging process involves placing
materials such as peel ply, release fabric, bleeder ply, breather ply, a caul plate, and
a flexible heat-resistant membrane over the tool and part. The main challenge is to
bleed out some excess resin contained in the laminate before the resin gels. If the
bleeding is insufficient, the excess resin degrades final part properties.
Automated Tape Placement (ATP) is the process by which plies of composite mate-
rial (often carbon fiber reinforced thermoplastic tapes) are laid down with pressure
on a tool surface in their proper position by an automated machine. The dispens-
ing head is equipped with a heating source (Laser, IR) and a compaction roller to
debulk the composite and promote bonding between successive layers. The main
challenge regarding flow phenomena is to ensure a good polymer healing at the
interface between the former and the new deposited tapes. 3D squeeze flow is the
relevant mechanism to study.
In filament winding dry fiber bundles or dry narrow tapes are pulled through a
resin bath (wetting stage) and then immediately wound onto a rotating mandrel (tool).
These impregnated layers are then left to cure. A prepreg roving can also be used in
this process, which eliminates the need for the resin bath. The applications of this
process are limited to cylindrical parts such as tubes, pressure vessels and cylinders.
In the pultrusion process, a continuous bundle of dry fiber is pulled through a resin
bath to be wetted and then fed into a die where they are subjected to compaction
and heating. The cross-sectional shape of the pulled fiber bundles is formed by these
dies. When the assembly of fibers and resin exit the die, they are compacted and
cured. This process is limited to straight parts with a constant cross section (beams,
frame sections). The main challenge regarding flow phenomena is to understand the
relation between axial tension, pulling speed and impregnation quality.
Liquid Composite Molding (LCM) refers those processes that involve infiltration
of a dry fibrous preform with a liquid (low viscosity) polymeric resin. The most
3.1 Introduction 113
popular technique is the Resin Transfer Molding (RTM). In RTM, a low viscosity
catalyzed resin matrix is injected into a mold containing a fibrous preform, and heat
is then applied to cure the part. The preforming stage, prior to injection, is also a
process. The preform is usually created by laying up layers of fabric in a preforming
tool, either by hand (draping) or by vacuum bagging techniques or matched-die molds
(preforming). The main challenge in RTM is to predict the flow pattern, impregnation
quality linked to the local saturation of fiber bundles and pressure field. The saturation
of a fluid is the volume fraction of the void space filled by that fluid.
There are several variations of this process. A very popular technology for the
manufacturing of large parts is called Liquid Resin Infusion (LRI). LRI is very similar
to the RTM process except that the resin is drawn into the preform and mold only
with vacuum pressure. Generally, fiber preform is placed on a single-sided mold
and is covered with a flexible membrane and vacuum-sealed. The flow issues to be
addressed in this process include impregnation of the resin inside the fiber preform
along with the progressive decompaction of the fiber preform. Actually during the
impregnation stage, the liquid pressure of the flowing resin balances off some of the
compacting atmospheric pressure thus leading to a dynamic increase of the wetted
fabrics. A second variation is Resin Film Infusion (RFI). In RFI, a dry preform is
placed in a mold on top of a solid resin film. Heat and pressure are then applied
so that the resin viscosity decreases and infuses throughout the preform. A third
variation is Compression-RTM (CRTM). In CRTM, the impregnation stage comprises
a combination of injection and compression driven flows. Unlike RTM, the mold is
not completely closed during the early stages of the injection, which results in lower
overall fiber volume fraction, and hence a lower global resin flow resistance of the
preform. In the above three variant processes the flow concern to be addressed is
the modeling of the flow in a deformable preform. Difference between all the above
mentioned process are essentially in the boundary conditions.
3.2 Description of Fibrous Microstructures in Composites
Microstructure refers to features of internal structure of heterogeneous materials

at a variety of length scales. As far engineered fibrous materials are concerned, the
appropriate length scales are those that reveal the partition of the material into regions
of nearly continuous phase such as fibers and cluster of fibers called fiber bundles.
Fibrous media with a large number of fibers and inter-fiber volumes cannot be
described by considering each of the fibers (view as heterogeneities here) or inter-
fiber volumes, which would lead to intractable boundary value problems to solve for
the flow. In addition the geometry of the bounding surfaces at microscale, where the
boundary conditions are defined are usually too complex to describe. Then, another
continuum approach at a coarser level is required. The classical method is to replace
this heterogeneous medium by a continuous equivalent (also called homogeneous)
one, whose description is valid at the macroscopic scale.
This macroscopic scale is relevant for manufacturing engineers because it is the

scale where materials and processing parameters are defined and measured. The
change of scale from small scale to larger scale is called upscaling. Its aim is to
describe the evolution of a given physical mechanism in a given heterogeneous
medium by an equivalent macroscopic continuous behavior. The derived macroscopic
behavior should be intrinsic to the heterogeneous medium and has to be independent
of the macroscopic boundary conditions.
This raises the question of the definition of the Representative Volume Element
(RVE). Practical limitations in both numerical and real representations of microstruc-
ture require studying samples of minimal size that are supposed to represent the
overall material [19].
The RVE is defined as an arbitrary volume chosen large enough to be statistically
representative of all the physical properties in the vicinity of a point and much smaller
than the exterior domain of the porous medium. The average of a physical property
taken with respect to the REV represents the property itself at the centroid of the RVE.
In other words, it is the smallest region that one can characterize to ensure that the
randomly chosen volume fully captures the important features of the microstructure.
The order of magnitude of the RVE is related to the local average porosities and
local gradients of the average porosities in the considered fibrous medium. Note that
for non-homogeneous fibrous structures where the spatial variance in the porosity
(inter-fiber space) may be quite significant, an upper limit for the RVE is of extreme
importance. Intuitively, since the description of flow phenomena in a fibrous medium
has to be, in most cases, at the macroscopic level, the average of each term (taken with
respect to the RVE) in the microscopic governing equations has to be determined.
This intuitive technique is known as volume averaging.
In a recent publication Zhang et al. analyze real engineered fibrous preforms
with random clustering of fibers [22]. For such a two-phase heterogeneous fibrous
medium, the size, orientation and distribution of phase clusters are significant to the
macroscopic properties. For example, the average size of clusters has influence on
the magnitude of overall property, while the orientation and distribution of clusters
control the anisotropy. The weighted influence of each phase at any location can
be represented by the volume fraction and clusters, while the directional cluster-
ing effect may lead to anisotropy. For this type of fibrous material with large scale
randomness with spatial correlation, the statistical continuum approach can be used
with 2-point statistics [19]. It consists in deriving the up-scaled property from the
governing constitutive relationship with the fluctuation of local property incorpo-
rated. Actually 2-point statistics may be inserted naturally into averaging schemes
by virtue of the ergodic assumption introduced hereafter [11]. Relations are derived
through perturbation expansions where the Green’s function solution are used to
provide a link between terms of the series expansion and correlation functions that
contain different orders of geometrical information. Since the effective coefficients
are defined by the local correlation functions, it was concluded that probability func-
tions play a critical role in describing the statistical distribution of different phases
in this heterogeneous fibrous medium [22].
3.2 Description of Fibrous Microstructures in Composites 115
Due to some long-range heterogeneity in engineered fabrics, the defined RVE

size may be too large for numerical analysis or experimental characterization. Thus
a smaller region, often named a statistical volume element (SVE), is used instead.
The overall fibrous structure statistics or response is then determined by ensemble
averaging over a set of SVEs. The size of this set must be large enough to fully
sample the fibrous material characteristics. When using SVEs rather than RVE, one
generally employs the ergodic hypothesis. This mathematical principle assumes that
the ensemble average of integrals taken over a large number of SVEs is equivalent
to an integral over a single large RVE.
The volume fraction occupied by voids or polymer, i.e., the total void/polymer
volume divided by the total volume occupied by the fibers and void/polymer volumes,
is called the porosity. In fibrous media each void is connected to more than one other
pore, pores are always interconnected and fluid flow is always possible through the
interconnected pores. Finally, the voids are usually non-uniform in their size and in
their distribution.
3.2.1 Monodisperse Porosity Medium
The monodisperse porosity distribution is rigorously encountered in fibrous compos-

ites in two cases: in pure unidirectional (UD) fabrics and in fiber bundle (or rovings).
The Chopped Strand Mat (CSM), which is manufactured by random distribution of
chopped strands of glass fiber (usually 10–100 mm in length) into a uniform mat,
has a porous structure close to be monodisperse. Compared to pure UD fabrics or
bidirectional (BD) fabrics composed of orthogonal yarns, the complexity of the CSM
lies in the random microstructure containing irregular pores with a large distribu-
tion of density, size and form. Although the porous structure appears highly random
and disordered, the statistical properties, such as the distribution of pore sizes and
shapes, the degree of correlation and the interconnectivity of pores, govern the flow
process and the macroscopic material properties. Therefore, accurate prediction of
flow properties depends on appropriate description and representation of the random
microstructure.
The microstructure depicted in Fig. 3.4 represents the cross-section of UD array
of fibers in polymer composites. It is an unidirectional array of discrete cylindri-
cal solids. It is random in the ensemble sense, and then the mesoscale response of
this microstructure subjected to a given physics is statistical, but tends to become
deterministic as the scale separation LM /lm grows. If we take the fibrous medium
as a set F = (F(ω); ω ∈ Ω), defined over the sample space Ω then for an event
ω ∈ Ω, we get a realization F(ω) of infinite extension. However, for practical reason
already mentioned, a mesoscale sub-domain of finite size Lm is cut out to compute
the response of a given flow. The question is to know whether this sub-domain is
large enough to get an intrinsic response, i.e. independent on the location of the
mesoscale sub-domain. If yes, this sub-domain is a RVE of the microstructure. If
this property is not verified, the flow response is said apparent. As a consequence, for
Fig. 3.4 Spherical RVE,

characteristic lengths and
different interfaces in a
monodisperse fibrous lm
medium
Sff
Lm
Sss
Sfs
fibrous material with large scale randomness with spatial correlation, the RVE can
be several order bigger than the size of the microscale heterogeneity. If the ergodic
hypothesis holds, it is acceptable to replace analysis (averaging) over a large enough
RVE by the analysis (ensemble average) over a large number of sub-domains.
In practice, since this microstructure is discrete and almost always measured
digitally on a discrete grid of spatial locations, local material state descriptors are
discrete. The random microstructure can be described by a phase function (also
indicator function) Xi (x). It is used to identify a random point x, located inside or
outside of phase i:
⎧
⎨1 if x lies in phase i
Xi (x) = . (3.1)
⎩0 otherwise
By definition volume fraction for phase i (here phases are fiber and inter-fiber
volumes), Vi , is the one-point probability function (also called one-point correla-
tion function ) Vi = P(Xi (x) = 1) if the material is statistically homogenous. If the
material is not statistically homogenous then the one-point probability function can
be interpreted as a position dependent volume fraction.
Volume fraction alone cannot capture the whole complexity of morphology in ran-
dom heterogeneous fibrous media when studying effective properties. More details
of the shape, orientation distribution and morphology of the microstructure including
the spatial correlation of the components in the media should be considered in order to
give an accurate prediction of effective properties. This can only be realized by using
higher order distribution functions. The statistical details of a heterogeneous fibrous
medium can be represented by an n-point probability distribution function (PDF).
Volume fraction, commonly used to capture the complexity of a microstructure, is
actually one-point probability distribution function.
The n-point probability function is defined as the probability of n points existing in

a certain combination of phases [19]. For a two-phase heterogeneous fibrous medium,
the size, orientation and distribution of phase clusters are significant to the effective
or macroscopic properties. Obviously, the average size of clusters influences the
overall property magnitude, while the orientation and distribution of clusters control
the anisotropy. As such information is not contained in the one-point probability
function, the two-point probability function has to be considered.
For any two points x and x in a domain composed of two phases Xi (i = 0, 1), the
two-point probability function Pi j (x, x ) is the probability of the vector x − x with
its head x ∈ Xi and tail x ∈ Xj (i, j = 0, 1). For the sake of convenience, the one-
dimensional Pi j (x, x ) for different orientation θ (0 ≤ θ < π ) can be used instead.
If the medium is isotropic, Pi j (x, x ) is identical for any value of θ . For a given θ ,
if the material is statistically homogenous then the function should only depend on
the relative positions and not on the absolute position, Pi j (x, x ) is a function of the
distance r = |x − x |. Using a given two-phase medium, Pi j (x, x ) is evaluated by
accumulating the number of events of each group at a certain distance r . Finally, the
condition
Pi j (r ) = 1, (3.2)
i, j=0,1
is imposed to normalize Pi j (r )(i, j = 0, 1).

In practice it is often assumed that the ergodic hypothesis holds and it is acceptable
to replace averaging over the ensemble with averaging over a large enough volume.
It can be shown that
⎧
⎪
⎪ Pi j (r → ∞) = pi p j
⎪
⎪
⎨
Pii (r → ∞) = pi2 = Vi 2 , (3.3)
⎪
⎪
⎪
⎪
⎩
Pii (r → 0) = pi = Vi
and ⎧ i
⎪
⎪ S
− 2v , d = 1
⎪
⎪
⎪
⎪
d Pii ⎨
i
(r → 0) = − Sv , d = 2 , (3.4)
dr ⎪
⎪ π
⎪
⎪
⎪
⎪
⎩ Svi
−4,d=3
where d is the dimension of the space, pi and p j are respectively the one-point
probability function of phases Xi and Xj , and Svi is the specific surface or interface
area per unit volume associated to phase i. The specific surface area is defined as
the ratio of the internal surface area of the voids over the bulk volume of the porous
medium.
Fig. 3.5 a Two-phase sampling region; b Two-point probability function in the x-direction and
c two-point probability function in the y-direction (solid curves real data; dotted line fitted analytical
model)
Figure 3.5a shows a sampling of a two-phase heterogeneous material. The phase

distribution is analyzed using the two-point probability function Pi j . The corre-
sponding values of the Pi j are reported in graphs 3.5b, c for the x and y directions
respectively (solid lines). The dotted lines are a fitted analytical model.
There is a large body of literature devoted to possible definitions of the RVE for
heterogeneous media. Generally speaking, the RVE size is considered as function of
many parameters, e.g. sample size, allowable error, the property of interest, etc. A
generic definition of RVE size is proposed in [10, 11]. Consider a random function
Q(x) as a mapping from the phase function Xi (x) to a phase property. Q(x) can be
the fiber volume fraction field
Q(x) = V(x) = Vi x ∈ Xi , (3.5)
and the permeability field:
D(x) = k(x) = ki x ∈ Xi , (3.6)

where Vi the volume fraction associated to phase Xi and ki is the phase permeability.
The point variance of the random process Q(x) is defined as [10]
Dk2 = p0 p1 (Q0 − Q1 )2 , (3.7)
where pi and Qi denote the one-point PDF and the phase property (Vi or ki ). In
addition, an integral range, An , is defined for Q(x) as [16]

1
Ad = (C(Q, r ) − C(Q, 0)2 ) dr, (3.8)
C(Q, 0) − C(Q, 0)2 Rd
where d denotes the dimension of space, C(Q, r ) is the covariance function with the
same meaning as the two-point PDF P00 (r ) or P11 (r ) (depending on the phase in
consideration) with r as the lag distance. Based on the framework of [13], the length
of RVE (L RV E ) for a given relative error ε (i.e. ratio of the absolute error to the mean
value K̄ ) and number of realizations (N ) can be derived as [10]
4Dk2 Ad
L RV E = 2
. (3.9)
ε2 K N
Assuming the statistics of the fibrous medium properties are statistically homo-
geneous and mean is ergodic, then the ensemble average equals the spatial average.

1
Q(ω) ≡ lim V →Rd Q(x, ω)dx = Q(x, ω)d P(ω) ≡ Q(x) . (3.10)
V Ω
3.2.2 Bidisperse Porosity Medium
Fabrics used as reinforcement in structural composites are usually made of assembled

fiber tows. As a consequence, fabrics have a dual-scale porous medium (two-phase
fibrous medium), i.e. porosity structure is composed of mesoscopic gaps between
fiber tows and microscopic pores between fibers inside each tow. The meso-volumes
offer low flow resistance compared to the fiber tow with high flow resistance.
When the resin flows at a high velocity, i.e. when the pressure gradient in the
vicinity of the flow front is the dominant driving forces compared to capillary pressure
at the front, the microscopic pores are difficult to saturate due to the high resistance
inside the bundle, and air can be entrapped inside the bundles [6–8, 12, 16, 17]. On
the contrary, at a low flow speed, due to the relative stronger capillary force inside
the tow, flow front direction is distorted and may form a void between fiber tows.
Both the macroscopic dry spots and micro-voids are detrimental to the strength of the
composite parts, and the reduction of these defects in processing is a main objective
of manufacturing engineers. Thus, it is evident that the dual-scale flow is a local
phenomenon highly dependent of the fibrous microstructure.
Fig. 3.6 Dual scale porosity distribution in a woven composite: left cross section of a woven
composite; right extraction of the meso-porosity
An example of such a porous structure is shown in Fig. 3.6 (left). It is a cross section
of a composite laminate made of unbalanced woven fabrics. The darkest regions are
the meso-volumes (volumes between fiber tows), while the grey scale zones are
fiber bundles containing the micro-volumes. Figure 3.6 (right) shows an extraction
of the meso-volumes from images obtained by X-ray tomography. It demonstrates
the complex morphology of the meso-volumes. One can notice some isolated meso-
volumes, however then cannot be considered as dead-end pores since meso-volumes
are always connected to micro-volumes.
Fiber bundles can be viewed as cluster of fibers and meso-volumes as channels
between fiber bundles, as depicted in Fig. 3.7. The characteristic length of the pores
within the tows (∼10 µm) is typically one to two order of magnitude smaller than the
size of pores (meso-volumes) outside tows. For such a two-phase fibrous medium, the
size, orientation and distribution of phase clusters are significant to the macroscopic
Averaging volume
St
lm
Lm
LM Close-up
nt
Inter-tow volume
Fig. 3.7 Characteristic lengths in a fibrous medium with a bidisperse porosity and averaging volume
used to upscale the flow variables. Close-up view of the inter-tow/tow boundary
microscale mesoscale macroscale
Fig. 3.8 Three scales associated to a bidisperse porosity medium. left microscopic view, center
mesoscopic view where tows are considered as continuum media, right macroscopic view where
the fabric is considered as a homogeneous porous material
properties. For example, the average size of clusters has influence on the magnitude
of overall property, while the orientation and distribution of clusters control the
anisotropy. All this information can be described by the N -point probability function,
defined as the probability of N points existing in a certain combination of the two-
phase porous medium. As the clustering effect and anisotropy cannot be accounted
for by the one-point PDF, the two-point PDF should be considered.
As it will be seen later in the next section, the mass and momentum balance
equations are usually derived at the mesoscopic scale, i.e. the microscopic flow
within the tows is not considered in detail. The averaging volume encompasses both
the fibrous tows and inter-tow volumes (Fig. 3.8).
3.3 Governing Equations for Flows in Porous Media
The macroscale perspective was introduced in the previous sections with the motiva-
tion to get macroscale variables helpful for manufacturing engineers. The definition
of the macroscale physical quantities must not be just a mathematical construct for
convenience but must also be a quantity that is generally measurable at the macro-
scopic level.
This section starts with the microscale description of the mass and momentum
balance equations. Then we will present how to transform the mass balance equations
in terms of microscale variables to mass and momentum balance equations in terms
of macroscale variables. This transformation will be accomplished by averaging the
point microscale conservation equation over a REV. This procedure also involves
averaging of the time and space derivatives of those physical quantities. Therefore,
additional tools needed to facilitate the transformation to the macroscale will be
introduced; the divergence and transport theorems that convert the average of a
derivative to a derivative of the average.
3.3.1 Pore-Scale Isothermal Flow Description
• Continuity equation (microscopic mass balance equation)

The equation of conservation for the fluid valid at all points in the fluid is
∂ρf
+ ∇ · (ρf vf ) = 0. (3.11)
∂t
This equation describes the conservation of total mass at a microscale point [20].
The first term accounts for the rate of change of mass at the point while the second
term is the net outward flow of mass per unit time from the point. In particular for
incompressible fluid with constant density this equations simplifies in
∇ · vf = 0. (3.12)
• The momentum equation given for an incompressible and Newtonian fluid is the
Navier-Stokes that writes
∂(ρf vf )
+ ∇ · (ρvf vf ) = ∇ · σ + ρg, (3.13)
∂t
where σ is the second-order total stress tensor and g is the field of the external
forces per unit mass. σ = −pI + τ for an incompressible and Newtonian fluid and
τ = 2ηD, where D is the deformation rate tensor.
3.3.2 Macroscopic Scale: Upscaling
Upscaling techniques are continuous approaches that allow the derivation of an equiv-
alent macroscopic continuous description from the description at the REV scale.
There are essentially two methods to up-scale flow equations valid at the microscale
to equations valid at the larger scale: multiscale homogenization and volume averag-
ing. Homogenization methods can derive for coefficients in the up-scaled equations,
whereas volume averaging methods give the form of the up-scaled equations but
generally must be accompanied with physical arguments to determine some coef-
ficients (see for instance [18] for the introduction of the notion of permeability in
macroscale momentum equation). This is due to the complexity of the flow paths
and the interpore and pore-to-pore fluid dynamic interactions.
Homogenization is often performed on periodic models, while volume averaging
does not require any assumption of periodicity and therefore is close to the reality
of fibrous materials. As previously seen, other methods can take the details of the
spatial statistics of some complex fibrous media directly into account, or provide
additional information about important corrections to the average equations.
3.3 Governing Equations for Flows in Porous Media 123
The averaging theorem is based on Green’s theorem and the divergence theorem
and supposes that in relatively small regions volume averages of spatial gradients in
statistically homogeneous media are close to gradients of volume averages. However,
looking at the typical double-scale porosity medium in Fig. 3.7, great care has to be
taken to calculate averaged quantities at the tow/inter-tow interface where abrupt
changes occur. We could be tempted to use small averaging volumes to better cap-
ture transitions in rapidly changing microstructure. However a too small averaging
volume implies rapid fluctuations in the quantities of interest (like the fluid velocity).
On the contrary, a too large averaging volume will smooth out the effects of local
inhomogeneities and the coefficients in the equations will tend to be constant. Then
the average quantities are functions of the particular choice of averaging volume.
Now let Ψ represent any physical quantity that has a value at each point. In
general, Ψ can be a scalar, a spatial vector, or a second-order tensor. The spatial
average, denoted by angle brackets is defined as

1
Ψ = Ψ dV. (3.14)
V
V
Using this definition, the volume fraction of both solid (s) and fluid (f) phases are
respectively

Vf 1
φf = = X f dV, (3.15)
V V
V
and
Vs 1
φs = = X s dV, (3.16)
V V
V
where volume fraction of the fluid φf is the porosity of the fibrous medium when not
impregnated yet.
The phase average of a physical quantity related to the fluid phase with respect to
a REV is given by

1 1
Ψf = Ψ X f dV = Ψf dV. (3.17)
V V
V Vf
The phase average of the interstitial velocity (within the pores) with respect to
a REV is called superficial velocity and corresponds to the Darcy velocity. It is the
average velocity determined by measuring the volumetric flow rate and dividing it
by the cross-sectional area of the total domain. This velocity is also referred to as
the filter velocity or the specific discharge.
The intrinsic phase average a physical quantity is defined as

Ψ X f dV

1 V
Ψf f = Ψf dV = . (3.18)
Vf
Vf X f dV
V
The phase and intrinsic averages are related by:
φf Ψf f
= Ψf . (3.19)
In order to locally average the Navier-Stokes and mass balance equations, the aver-
age of a gradient (or divergence) must be replaced with the gradient (or divergence)
of an average. Then, the macroscopic balance equations for mass and momentum
are derived by taking the term-wise average of the corresponding balance equations.
To do so, two useful relationships needed are given here.
• The general transport theorem, which is the general form of the Reynolds’ transport
theorem, states that

d ∂Ψ
Ψ dV = dV + Ψ v · n dS, (3.20)
dt ∂t
V (t) V (t) S(t)
where Ψ is a function with a continuous first derivative in time within the volume,
v is the velocity of the surface S(t) which encompasses the volume of interest V(t)
and n is the unit vector perpendicular to S(t) oriented to be positive outward from
the volume. The velocity v can be different than the fluid velocity. When these
two velocities are the same, then the general and Reynolds’ transport theorems
are identical. Physically this equation states that a rate of change of the integral
of a quantity over a volume is related to the change in the value of the quantity in
the volume and any change in the size of the volume due to the movement of its
boundaries (due to expansion or contraction of the volume).
In particular, applying this theorem to the phase function shows that the partial
derivative of the fluid volume fraction respect to time is

∂φf 1
= vf · n dS, (3.21)
∂t V
Sfs
where Sfs is the interfacial area between fluid and solid phases within the averaging
volume V .
• The second relationship is the divergence theorem (the Green’s transformation),
which states that

∇ · Ψ dV = Ψ v · n dS. (3.22)
V (t) S(t)
The theorem for the volume average of a gradient reads

1
∇Ψ = ∇Ψ + Ψ nfs dS, (3.23)
V
Sfs
where nfs is the unit normal to that surface, directed from the fluid towards the solid.
We can express the gradient of the volume fraction in terms of an integral over
the fluid/solid interface.

1
∇φf = − nfs dS. (3.24)
V
Sfs
A similar relation can easily be developed for vector and tensor quantities, referred
to as the theorem for the volume average of a divergence:

1
∇ · Ψ = ∇ · Ψ + Ψ · nfs dS. (3.25)
V
Sfs
The temporal averaging theorem relates the average of the partial time derivative
of a microscale quantity to the partial time derivative of an averaged macroscale
quantity
∂Ψ ∂
= Ψ , (3.26)
∂t ∂t
where the scalar Ψ can be replaced by a vector or a second-order tensor.
3.3.2.1 Macroscopic Mass Balance Equation
For the fluid flowing between fibers, the mass balance equation at each point is given
by the equation
∇ · v = 0. (3.27)
Taking the volume (phase) average of this equation gives for a single-scale poros-
ity medium with a stationary solid phase the macroscopic mass balance equation
∇ · v = 0. (3.28)
Now, considering fibrous medium featuring a double-scale porosity structure as

depicted in Fig. 3.7, the mass balance equation is derived at the mesoscopic scale,
i.e. the microscopic flow within the tows is not considered in detail. The averaging
volume encompasses both the fibrous tows and inter-tow volumes (Figs. 3.7 and 3.8).
The volume (phase) average of Eq. (3.27) for an incompressible fluid flowing in the
inter-tow volumes is

1
∇ · vit = − vit · nt dS, (3.29)
V
St
where vit is the fluid velocity in the inter-tow volumes and St the tows/fluid interphase.
The right-hand side of Eq. (3.29) represents the rate per unit volume of resin
absorption by the tows bounding the inter-tow volumes. This term is generally
referred to as the sink term Stow in the macroscopic mass balance equation. Then,
using this definition, Eq. (3.29) can be rewritten as
∇ · vit = Stow , (3.30)
where the sink term is by definition

1
Stow =− vit · nt dS. (3.31)
V
St
Equation (3.29) shows that the flow in the inter-tow region slows down because
of the absorption of resin by the tows, as long as tows are not completely filled up.
This transition zone where the fluid is impregnating the tows is generally referred to
as the unsaturated region of the fibrous medium.
3.3.2.2 Macroscopic Momentum Balance Equation
The microscopic momentum balance equation for the fluid flowing between fibers is
∂(ρv)
+ ∇ · (ρvv) = ∇ · σ + ρg, (3.32)
∂t
where σ is the total stress tensor and g represents the body force per unit mass.
Assuming the solid does not move and that there is no exchange of matter between
the fluid and fibers, the volume (phase) average of Eq. (3.32) is obtained by applying
the averaging theorem

∂ρf vf 1
+ ∇ · ρf vf vf = ∇ · σf + ρf g + σf · nfs dS. (3.33)
∂t V Sfs
Now assuming that the following conditions apply: (i) Constant density, (ii) New-
tonian fluid, (iii) Fully saturated medium, (iv) No surface tension between the fluid
and the solid, (v) Porosity is uniform in space (no gradient of porosity) and (vi)
Isothermal conditions; then the volume-averaged macroscopic momentum equation
writes
∂ρf vf
+ ∇ · ρf vf vf = φf ∇Pf f
+ ∇ · τf − fd , (3.34)
∂t
where τf is the extra stress tensor (also called deviatoric fluid stress tensor), defined
for a Newtonian fluid of constant viscosity η by
τf = η ∇vf + (∇vf )T = 2ηDf , (3.35)
where the superscript T indicates a tensor transpose and fd is the fluid-solid drag
force defined as

1 1 f 1
fd = − σf · nfs dS + ρf ghnfs ds + Pf nfs dS. (3.36)
V V V
Sfs Sfs Sfs
The first integral in the right-hand side of Eq. (3.36) represents the total force
per unit volume (other than those associated with hydrostatic and ambient pressures)
exerted by the fluid on fibers or pore walls. This force is associated to the moving
fluid, whereas a hydrostatic pressure produces no relative motion between fiber and
fluid.
The second integral is the force per unit volume exerted on the fiber-tow by a
pressure gradient due to the gravity effect in the fluid. The third integral represents
the force exerted on the fiber tow due to the average intrinsic liquid pressure. With
the help of Eq. (3.24), this third integral can be expressed

1
− Pf f nfs dS = ∇φf Pf f . (3.37)
V
Sfs
When a gradient of porosity (or fiber volume fraction) exists at the scale of the
RVE, then a uniform pressure in the fluid will produce of force on the fiber tows pro-
portional to the intrinsic average fluid pressure. This hydrostatic force is antiparallel
to the vector ∇φf . If the fibrous preform is homogeneous at the scale of the RVE,
then the integral in Eq. (3.37) equals to zero.
In most of composite manufacturing processes gravity effect can be neglected,
the macroscopic fluid-solid drag force reduces to the force associated to the moving
fluid

1
fd = − σf · nfs dS. (3.38)
V
Sfs
3.4 Flow of Simple Fluids in Non-deformable Fibrous

Microstructures
For highly compacted array of fibers, the porosity does not change in the presence of
a pressure gradient and the fibrous medium is then stationary. In deformable fibrous
media the porosity depends on the pressure from the mechanical properties of the
fiber bed.
3.4.1 Negligible Inertia and Single-Scale Porosity Medium
For most of manufacturing processes of structural composites, the Reynolds’ number

Re < 1, the flow of polymer is a creeping-flow regime, also called Darcy regime.
Then viscous forces dominate over the inertia forces and only the local (pore-level)
geometry influences the flow. Then the left-hand side of Eq. (3.34) can be neglected,
giving the macroscopic momentum equation
− φf ∇Pf f
+ ∇ · τf − fd = 0. (3.39)
For a Newtonian fluid of constant viscosity η, the extra stress tensor τf is defined
in Eq. (3.35). With the help of the averaging theorem, the volume average of ∇vf
gives

1
∇vf = ∇vf + vf nfs dS. (3.40)
V
Sfs
If fibers do not move and the fluid does not slip at the fiber wall, the integral in
Eq. (3.40) vanishes. Then the volume average of the extra stress is simply expressed
by the gradient of the average velocity
τf = η ∇vf + (∇vf )T . (3.41)
Taking the divergence of Eq. (3.41) it results
∇ · τf = η ∇ · ∇vf + ∇ · (∇vf )T . (3.42)
Knowing that
∇ · (∇vf )T = ∇(∇ · vf ), (3.43)

3.4 Flow of Simple Fluids in Non-deformable Fibrous Microstructures 129
and using the incompressibility condition in Eq. (3.28), it comes
∇ · τf = η∇ · ∇vf = η∇ 2 vf . (3.44)
Finally, the difficulty is in relating the fluid-solid drag force to the fluid velocity
(relative to the fiber) and the geometrical properties of the fibrous medium. Using
dimensional analysis, it has been seen the drag force can be written as [18]
φf η
fd = vf , (3.45)
K
where K is the rank two intrinsic permeability tensor of the fibrous medium. The
permeability is an average medium property that measures the ability of the porous
medium to transmit fluid. The components depend on the physical system as well as
the orientation of the porous structure relative to the coordinate system. The tensor
component Kij is the intrinsic permeability that allows for flow to take place in
coordinate direction i as a result of a pressure gradient in coordinate direction j.
When the coordinate directions are aligned with the principal directions of flow, the
off-diagonal components of the intrinsic permeability tensor are zero.
Within the frame of the above assumptions, it can be proved that permeability is
determined only by the geometry of the porous phase and not by the fluid properties.
These assumptions are recalled here: (i) Constant density, (ii) Newtonian fluid, (iii)
Fully saturated medium, (iv) No surface tension between the fluid and the solid, (v)
Porosity is uniform in space (no gradient of porosity), (vi) Isothermal conditions,
(vii) Laminar flow and (viii) Stationary fibrous medium.
Dimensional analysis shows that permeability has the unit of m2 and permeability
reflects both the magnitude and anisotropy of the fluid/fiber resistance. The more
contact a fluid has with the fiber surface of the porous medium, the more resistance
to flow it will encounter.
Inserting Eqs. (3.44) and (3.45) into Eq. (3.39) gives the Brinkman’s equation
− φf ∇Pf f
+ η∇ 2 vf − φf ηK−1 · vf = 0. (3.46)
The second term of this equation is the Stokes term and is expected to be significant in
areas of large averaged velocity gradients. In a typical single-scale porosity medium
defined in Fig. 3.4, an order-of-magnitude analysis shows that this term becomes
negligible with respect to the fluid-solid drag force. Then the macroscopic momentum
balance equation reduces to
K
vf = − · ∇Pf f . (3.47)
η
This equation is known as the Darcy equation. If the fibrous medium is isotropic,
the permeability tensor K reduces to a scalar value K. Note that Darcy’s velocity is a
superficial velocity and not the interstitial velocity at which the fluid moves between
∂ω
microscale ω macroscale Ω
Fig. 3.9 Microscopic view of a unidirectional fiber array and the macroscopic homogenized struc-
ture
fibers. Then Darcy’s velocity is not a true velocity of the fluid but represents an
effective flow rate through the porous medium. Chemical species carried with the
flow move only in the pores and thus are transported relative to the fiber phase at an
average pore velocity vf f .
The interstitial velocity vf f is deduced by combining Eq. (3.19) and the Darcy
equation (3.47)
K
vf f
=− · ∇Pf f . (3.48)
ηφf
Combining Darcy’s equation with the incompressibility condition allows to elim-

inate the velocity to get the pressure equation
K
∇· · ∇Pf f
= 0. (3.49)
η
In practice Eq. (3.49) is first solved to get the pressure field, then pressure gradients
are computed and substituted into Eq. (3.47) to find the velocities.
Beyond dimensional analysis used historically to show that permeability is a
geometrical parameter attached to porous structure, some theoretical arguments can
be provided on the basis of the theory of homogenization to prove this result. In
addition, the following shows that the computation of pressure is not required to
predict the permeability of a fibrous medium.
Without loss of generality, let’s consider the fibrous microstructure depicted in
Fig. 3.9.
The total power dissipated by friction is

Pmicr o = σ : ∇vf dx, (3.50)
ω
that can be rewritten taking into account the linear momentum balance ∇ · σ = 0
and applying the divergence theorem

Pmicr o = vf · (σf · n) dx, (3.51)
∂ω
where ∂ω is the boundary of the microscale domain ω. The power dissipation cal-
culated at the microscale must be equal to the power dissipation obtained from the
volume integration over that element. Then the dissipated power Pmacr o reads
Pmacr o = |ω|∇Pf f
· vf , (3.52)
and by enforcing the equality of both dissipated powers it results
σf : Df |ω| = |ω|∇Pf f

· vf , (3.53)
where Df is the symmetric component of the velocity gradient

1
Df = (∇vf + (∇vf )T ). (3.54)
2
The fluid being incompressible, the total fluid stress reduces to the extra stress
tensor, then Eq. (3.53) simplifies in
τf : Df = ∇Pf f
· vf . (3.55)
Replacing the extra stress tensor by its definition in Eq. (3.35), Eq. (3.55) can be
rewritten as
2ηDf : Df = ∇Pf f
· vf . (3.56)
Using Darcy’s law in Eq. (3.47) to eliminate the pressure gradient term in the
RHS of Eq. (3.56) it comes
2ηDf : Df = ηvf T
· K−1 · vf . (3.57)
Introducing Eq. (3.54) in the LHS of Eq. (3.57) and developing the contraction
product
1
Df : Df = Di j · Di j = (vi, j + v j,i ) · (vi, j + v j,i )
4
1
= vi, j · vi, j + vi, j · v j,i + v j,i · vi, j + v j,i · v j,i , (3.58)
4
where the subscript f in the components Di j is dropped for the sake of clarity. We
note vf = V .
It can be shown that vf and vf are continuous and that the transformation of the
average velocity into the local velocity is possible through the localization tensor L
defined by
vf = L · vf = L · V. (3.59)
This closure relation is often used to relate the deviations at the microscale to the
averaged quantities. Thus the local fluid velocity is a linear vector function of the
average velocity. Then, the velocity components are related by
vi = L ik · Vk . (3.60)
Developing the products of Eq. (3.58)

⎧
⎪
⎪ vi, j · vi, j = (L ik, j · Vk ) · (L im, j · Vm ) = Ukm · Vk · Vm
⎪
⎪
⎪
⎪
⎪
⎪
⎨ vi, j · v j,i = (L ik, j · Vk ) · (L jm,i · Vm ) = Wkm · Vk · Vm
(3.61)
⎪
⎪
⎪
⎪ v j,i · vi, j = (L jk,i · Vk ) · (L im, j · Vm ) = Ykm · Vk · Vm
⎪
⎪
⎪
⎪
⎩
v j,i · v j,i = (L jk,i · Vk ) · (L jm,i · Vm ) = Z km · Vk · Vm
Then Eq. (3.57) can be rewritten as

1 T
ηV · U + W + Y + Z · V = ηVT · K−1 · V, (3.62)
2
from which the inverse of the macroscopic permeability tensor can be identified when
enforcing that powers equality applies for any velocity V:
1
K−1 = U + W + Y + Z . (3.63)
2
Equation (3.63) shows the permeability only depends on the geometrical micro-
structural features through the localization tensor components U, W, Y and Z and
reflects both the magnitude and anisotropy of the fluid/fiber resistance.
3.4.2 Negligible Inertia and Double-Scale Porosity Medium
In the double scale porosity medium depicted in Fig. 3.7, the two interconnected
phases considered are the fiber tows and the inter-tow volumes. They exhibit very
distinct porous networks that can exchange fluid mass between them in a short-term
range [3, 4, 14, 15].
The fiber tows play the role of the solid fiber in the single scale porosity medium.
We still assume fibers do not move and that the fluid does not slip at the fiber wall.
However the no-slip boundary condition at the surface of the tows is not applicable
because the fluid can flow inside the tows. Then the integral in Eq. (3.40) does not
vanish. This latter represents the rate per unit volume of fluid absorption by the tows,
whose definition given in Eqs. (3.30) and (3.31) is recalled here

1
∇ · vit = − vit · nt dS = −Stow . (3.64)
V
St
To express the average viscous stress τit defined in Eq. (3.41), the transpose of
the average velocity gradient is first expressed

1
∇vit = (∇vit ) +
T T
nt vit dS. (3.65)
V
St
It comes

1
∇ · τit = η ∇ · ∇vit + ∇ · (∇vit ) + ∇ ·
T
(vit nt + nt vit )dS . (3.66)
V
St
Using the equality in Eq. (3.43) and the definition of the sink term in Eq. (3.64),
we get
∇ · (∇vit )T = ∇(∇ · vit ) = −∇ Stow . (3.67)
Then, Eq. (3.66) can be written as

1
∇ · τit = η ∇ 2 vit − ∇ Stow + ∇· (vit nt + nt vit )dS , (3.68)
V
St
or equivalently

1
∇ · τit = η ∇ 2 vit + ∇· [vit nt + nt vit − (vit · nt )I]dS , (3.69)
V
St
where I is the unit tensor.

The second member of the RHS of Eq. (3.69) represents the resulting force acting
on the fiber tows due to the non-zero velocity on the tow surface. It is expected to be
significant near the flow front where the normal component of the inter-tow velocity
is important as a result of the tow impregnation mechanism. When fiber tows are
fully impregnated, i.e. behind the flow front region, this second member tends to
zero. As a result Eq. (3.69) reduces to
∇ · τit = η∇ 2 vit , (3.70)
which is identical to the expression in Eq. (3.44) for the single-scale porosity medium.
Substituting expressions for the average viscous resistance, Eq. (3.69) and viscous
drag, Eq. (3.45), into the macroscopic momentum balance equation (3.39) gives

1
−φit ∇Pit it +η ∇ 2 vit + ∇· vit nt +nt vit −(vit ·nt )I dS −φit ηK−1 ·vit = 0,
V
St
(3.71)
where φit is the inter-tow porosity defined as the ratio of inter-tow volume over the
total averaging volume.
This expression is the equivalent of the Brinkman equation for a single-scale
porosity medium, extended here to the case where the tow saturation process occurs.
3.4.3 Fluid Inertia
It is generally believed that Darcy’s law is applicable as long as the Reynolds number
Re < 1. This is the flow regime in which viscous forces are predominant. Re is defined
in
√ terms of a characteristic length of the porous medium in which fluid flow occurs.
K is usually used as the characteristic length. Using permeability as the length
scale, Re is given by
√
ρf V K
Re = . (3.72)
η
Beyond Re = 1, there is a transition zone in which inertial forces begin to be

significant and eventually predominant. If the velocity of a fluid flowing in a porous
medium is high enough, Darcy’s law breaks down. Then one must add some correc-
tion terms to the basic Darcy law in order to take into account the effect of inertia.
The Forchheimer’s equation adds a term that is proportional to the square of the fluid
velocity to account for the extra pressure drop at higher Reynolds numbers
∂ η ρf b
− Pf f
= vf + √ vf ||vf ||, (3.73)
∂x Kx Kx
where b is a constant, ||vf || is the norm of the volume average fluid velocity and
and x is the direction of macroscopic flow. The dependence of the parameter b on
the porous media characteristics has been studied in granular media, for instance to
obtain the Ergun’s equation
∂ c1 (1 − φf )2 η c2 ρf (1 − φf )
− Pf f
= vx + vx 2 , (3.74)
∂x φf dp
3 2 φf3 dp
where dp is the average particle diameter, c1 and c2 are constants. The first term in the
RHS of Eq. (3.74) is the Blake-Kozeny-Carman model for predicting permeability
φf3 dp2
K = , (3.75)
c1 (1 − φf )2
while the second term (microscopic inertia term) contains a dependence on porosity
c2 −3/2
b = √ φf . (3.76)
c1
3.5 Flow of Complex Fluids in Non-deformable Fibrous

Microstructures
3.5.1 Slightly Compressible Fluids
The compressibility of viscous polymer is due to volatiles or air entrapped in the resin
that does not escape from the composite before resin cures or thermoplastic polymer
solidifies. For instance, impregnation of double scale porosity fabrics may lead to
the creation of micro voids according to mechanism depicted in Fig. 3.10. The fluid
flow in such dual scale porous media will impregnate the larger pores faster than the
smaller pores within fiber tows. The partially saturated region is created due to this
heterogeneity in the pore distribution. It refers to the region where the macroscopic
t1 t2 > t1
Fig. 3.10 Microvoid formation mechanism in a dual scale porosity fabric [4]
front has reached all the spaces between the fiber tows but the pores between the
fibers inside a tow are not yet completely filled.
The voids may migrate due to the pressure gradient in the fluid or other mechanical
action. Another source of void formation is a lack of consolidation in during processes
like robotic tape placement. If the resin pressure is too low, insufficient consolidation
could occur and the contact between the individual plies is not intimate. In addition,
if the pressure in the polymer is too low, it cannot prevent voids from growing.
Physically it implies that there are small voids trapped in the viscous fluid. Consid-
ering the fluid with entrapped voids as a continuum, the compound structure becomes
slightly compressible as long as voids are not squeezed out. The continuity equation
for the fluid has to be modified accordingly.
Substituting Eq. (3.11) into Eq. (3.47) yields

∂φf ρf ρf
−∇ · K · ∇ P = 0. (3.77)
∂t η
After carrying out the time differentiation of the LHS of Eq. (3.77) it comes

∂ρf ∂φf ∂ P ρf
φf + ρf −∇ · K · ∇ P = 0. (3.78)
∂P ∂ P ∂t η
If the fibrous medium is supposed incompressible, then

∂φf
= 0. (3.79)
∂P
The fluid compressibility C f is defined by the equation of state at a fixed temper-
ature T
1 ∂ V 1 ∂ρf
Cf = − = , (3.80)
V ∂P T ρf ∂ P T
where V stands for the volume occupied by the fluid.

Combining Eqs.(3.78)–(3.80) gives a relationship for the unknowns P and ρf . In
the following the case of a slightly compressible fluid is investigated so as to have
a linear relationship between P and ρf . We assume that the fluid compressibility Cf
is constant over a certain range of pressures. Then the fluid compressibility can be
written as
ρf = ρ0 eCf (P−P0 ) , (3.81)
where ρ0 is the density at the reference pressure P0 . Expressing Eq. (3.81) with a
Taylor expansion, it comes
1 2
ρf = ρ0 1 + Cf (P − P0 ) + C (P − P0 )2 + · · · , (3.82)
2! f
3.5 Flow of Complex Fluids in Non-deformable Fibrous Microstructures 137
which can approximated to
ρf ≈ ρ0 1 + Cf (P − P0 ) . (3.83)
Then it follows that

∂ρf
= ρ0 Cf . (3.84)
∂P
Substituting Eqs. (3.84) and (3.79) into Eq. (3.78) gives

∂P ρf
φf ρ0 Cf −∇ · K · ∇ P = 0, (3.85)
∂t η
which is a parabolic equation in P.
3.5.2 Reactive Fluids
We consider a fluid that flows at velocity vf between stationary fibers in the elemen-
tary volume depicted in Fig. 3.11. Suppose it reacts at the contact of fibers to create
solid matter surrounding each fiber wetted by the fluid. Then the fluid/fiber interface
S depicted in Fig. 3.11 evolves. We note W the absolute velocity of this interface.
We neglect the fluctuations of vf inside the elementary volume.
Fig. 3.11 Flow in an

evolving microstructure
U
U
S U
U
The conservation of mass in the elementary volume reads

∂ρf 1 1
+ ∇ · ρf vf = ρf W · ndS − ρf vf · ndS. (3.86)
∂t V V
S S
U is introduced as the relative velocity of the interface. If the representative vol-

ume element is non deformable, U is the relative velocity due to changes between
constituents (chemical reactions, phase changes).
U = W − vf . (3.87)
Then the continuity equation reads

∂ρf 1
+ ∇ · ρf vf = ρf U · n dS. (3.88)
∂t V
S
3.6 Numerical Simulation
The numerical simulation of the flow of a viscous resin through a complex array of
fibers can be developed by discretizing the governing partial differential equation pre-
sented in this chapter with numerical methods such as finite difference, finite element
or other techniques like boundary elements, meshless or element-free methods.
The goal of numerical simulation is to help to understand the flow behavior inside
fabrics especially when the medium to impregnate exhibits large spatial variation of
permeability or when the geometry of the part to manufacture is complex. As the
impregnation of the porous medium involves a moving boundary to track the flow
front, the Control Volume Finite Element Method (CVFEM) also called the Finite
Volume Method (FVM)—is often used to simulate the flow at the macroscale [1]
(alternative method using Level Set method, motivated by the multi-phase flow prob-
lems begins to be adopted [21, 22]). It is a hybrid method that combines the advan-
tages of the finite element and the finite volume methods. The CVFEM approach
does not require remeshing at each time step. For the discretization, the approach
follows that of the finite volume method, whereas the conservation laws are applied
to the control volumes of the FVM, such that the local conservation of the properties
is satisfied. Control volumes are created around the node points on a fixed finite
element mesh that conforms to an arbitrarily shaped domain. The fluxes across con-
trol volume faces are approximated using finite element interpolation. Each control
volume has a fill factor (f) associated with it. It represents the ratio of the volume of
fluid to the total volume of the control volume. The control volumes can be empty (f
= 0), partially filled (0 < f < 1), or completely filled (f = 1). The flow front is tracked
by determining which control volumes are partially filled so as to connect them to
obtain the flow front pattern. The numerical flow front is constituted of nodes with
3.6 Numerical Simulation 139
partially filled control volumes. As the exact location of the fluid inside the control
volume cannot be determined, the exact flow pattern front is not known. Thus, a too
coarse mesh can lead to a poor prediction of the flow pattern.
The pressure in the filled control volumes is usually evaluated by the Galerkin finite
element method, solving the pressure equation Eq. (3.49). Pressures are calculated
where nodes are full, whereas only partially filled nodes are assumed to be close to the
flow front and the flow front boundary condition is applied there. After pressures at
the nodes are computed, the velocities are calculated at the centroid of each element
using the Darcy law. It is assumed that the velocity of the fluid is constant throughout
each element. Then the flow rate is computed, and the front is advanced with the
control volume approach. The flow front is advanced at each time step by updating the
fill factors of control volumes, using the flow rates between the connecting nodes.
The time step is chosen so as to fill one additional control volume. A larger time
step induces the flow front to go beyond the control volume and yields to a loss
of mass. Once the flow front is advanced, the fluid domain is changed. Then the
pressure solution is computed for the new domain and this process is repeated until
the complete domain is filled up. More details about the VOF method are given in
Chap. 2 of this book.
References
1. S. Advani, E. Murat Sozer (eds.), Process Modeling in Composites Manufacturing, vol. 59

(CRC Press, Boca Raton, 2002)
2. S. Advani, K-T Hsiao (eds.), Manufacturing Techniques for Polymer Matrix composites (PMCs)
(Elsevier, Amsterdam, 2012)
3. C. Binetruy, B. Hilaire, J. Pabiot, The interactions between flows occurring inside and outside
fabric tows during RTM. Compos. Sci. Technol. 57(5), 587–596 (1997)
4. C. Binetruy, B. Hilaire, J. Pabiot, Tow impregnation model and void formation mechanisms
during RTM. J. Compos. Mater. 32(3), 223–245 (1998)
5. S. Comas-Cardona, B. Cosson, S. Bickerton, C. Binetruy, An optically-based inverse method
to measure in-plane permeability fields of fibrous reinforcements. Compos Part A: Appl. Sci.
Manuf. 57, 41–48 (2014)
6. B. Gourichon, C. Binetruy, P. Krawczak, A new numerical procedure to predict dynamic void
content in liquid composite molding. Compos. Part A: Appl. Sci. Manuf. 37(11), 1961–1969
(2006)
7. B. Gourichon, C. Binetruy, P. Krawczak, Experimental investigation of high fiber tow count
fabric unsaturation during RTM. Compos. Sci. Technol. 66(7), 976–982 (2006)
8. B. Gourichon, Dynamic void content prediction during radial injection in liquid composite
molding. Compos. Part A: Appl. Sci. Manuf. 39(1), 46–55 (2008)
9. S. Hoa, Principles of the Manufacturing of Composite Materials (DEStech Publications Inc,
Lancaster, 2009)
10. T. Kanit, S. Forest, I. Galliet, V. Mounoury, Determination of the size of the representative
volume element for random composites: statistical and numerical approach. Int. J. Solids.
Struct. 40, 3647–3679 (2003)
11. M. Ostoja-Starzewski, Microstructural Randomness and Scaling in Mechanics of Materials
(CRC Press, Boca Raton, 2010)
12. C.H. Park, A. Lebel, A. Saouab, J. Breard, W.I. Lee, Modeling and simulation of voids and
saturation in liquid composite molding processes. Compos Part A: Appl. Sci. Manuf. 42(6),
658–668 (2011)
13. C. Pelissou, J. Baccou, Y. Monerie, F. Perales, Determination of the size of the representative
volume element for random quasi-brittle composites. Int. J. Solids. Struct. 46, 2842–2855
(2009)
14. K.M. Pillai, S. Advani, Numerical and analytical study to estimate the effect of two length
scales upon the permeability of a fibrous porous medium. Transp. Porous Media 21(1), 1–17
(1995)
15. K.M. Pillai. Governing equations for unsaturated flow through woven fiber mats. Part 1. Isother-
mal flows. Compos. Part A, 33/7, 1007–1019 (2002)
16. J.S.U. Schell, M. Deleglise, C. Binetruy, P. Krawczak, P. Ermanni, Numerical prediction and
experimental characterisation of meso-scale-voids in liquid composite moulding. Compos. Part
A: Appl. Sci. Manuf. 38(12), 2460–2470 (2007)
17. J. Slade, K.M. Pillai, S. Advani, Investigation of unsaturated flow in woven, braided and stitched
fiber mats during mold-filling in resin transfer molding. Polym. Compos. 22(4), 491–505 (2001)
18. J.C. Slattery, Energy Momentum and Mass Transfer in Continua (Krieger, Huntington, 1981)
19. S. Torquato, Random Heterogeneous Materials: Microstructure and Macroscopic Properties
(Springer, New York, 2002)
20. C.L. Tucker III, R.B. Dessenberger, Governing Equations for Flow and Heat Transfer in
Stationary Fiber Beds. Chapter 8, ed. by S.G. Advani, Flow and rheology in polymer composites
manufacturing (Elsevier, Amsterdam, 1994), pp. 257–323
21. F. Zhang, B. Cosson, S. Comas-Cardona, C. Binetruy, Efficient stochastic simulation approach
for RTM process with random fibrous permeability. Compos. Sci. Technol. 71(12), 1478–1485
(2011)
22. F. Zhang, S. Comas-Cardona, C. Binetruy, Statistical modeling of in-plane permeability of
non-woven random fibrous reinforcement. Compos. Sci. Technol. 72(12), 1368–1379 (2012)

Flows in Polymers, Reinforced Polymers and Composites - A Multi-Scale Approach-Spr

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Flows in Polymers, Reinforced Polymers and Composites - A Multi-Scale Approach-Spr

Uploaded by

Copyright:

Available Formats

SPRINGER BRIEFS IN

APPLIED SCIENCES AND TECHNOLOGY

ISSN 2191-530X ISSN 2191-5318 (electronic)

Library of Congress Control Number: 2015934680

Springer Cham Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

January 2015 Christophe Binetruy

1 Multi-scale Modeling and Simulation of Polymer Flow . . . . . . . . . . 1

2 Complex Flows of Micro/Nano Structured Fluids:

3 Flows of Simple Fluids in Complex Microstructures:

3.4 Flow of Simple Fluids in Non-deformable

Abstract The flow of polymeric fluids in complex geometries relevant to

1.1.1 Rheology and Structure

© The Author(s) 2015 1

1.1.2 A Hierarchy of Modeling Approaches

Modern research in computational materials science exploits, sometimes in a coupled

1.1.3 The Role of Computational Rheology

suitable closure approximations that are invariably needed to derive a macroscopic

1.2 Macroscopic Modeling and Simulations

1.2.1 Governing Equations and Numerical Challenges

is neglected, a spectrum of relaxation times is readily taken into account by defining

where m is the memory function of linear viscoelasticity, and S is a model-dependent

1.2.2 A Spectrum of Numerical Methods and Applications

Standard Computational Fluid Dynamics (CFD) technology offers a wide spectrum

rheometry, or use of numerical simulation to aid the experimentalist in reducing his

1.3 Multi-scale Kinetic Theory of Viscoelastic Flow

Micro-macro methods of computational rheology [31] couple the coarse-grained,

1.3.1 Probability Distribution of Molecular Configurations

1.3.2 Fokker-Planck Equation and Kramers Relation

an NC -dimensional vector that defines the drift or deterministic component of the

Here, g is a model-dependent tensorial function of configuration. Thus, for a given

1.3.3 Exploiting the Multi-scale Model: Three Basic

where D = B · BT and W is a Wiener stochastic process of dimension NC . In a

In the sequel, we focus on the micro-macro methods needed to implement the

1.3.4 Illustrative Examples of Kinetic Theory Models

1.3.4.1 The FENE Dumbbell Model

where κ = ∇v T is the transpose of the velocity gradient, ζ is a friction coefficient,

where H is a spring constant. The stress-configuration relation (1.8) gives

τ p = n QFc (Q), (1.13)

where W is the three-dimensional Wiener process, namely a Gaussian stochastic

supplemented by an algebraic expression relating stress to average configuration,

Note that the FENE-P constitutive equation is of the form,

1.3.4.2 The Doi-Edwards Model with Independent Alignment

Fig. 1.3 a Test chain within

where W is the one-dimensional Wiener process. The evolution equation for u is

and Q is a strain measure defined as

(F(t, t ) · u(t ))(F(t, t ) · u(t ))

1.4 Micro-Macro Simulations Using the Stochastic

k molecules that are currently

1.4.2 Numerical Challenges

1.4.3 Brownian Configuration Fields

1.4.4 Lagrangian Particle Methods

ensembles of molecules can be viewed as a (Lagrangian) method of characteristics

1.4.5 Variance Reduction

The stochastic formulation of viscoelastic flows raises several mathematical issues of

1.5 Micro-Macro Simulations Using the Fokker-Planck

1.5.1 Proper Generalized Decomposition

Problems defined in high-dimensional spaces abound. For example, we have seen

The PGD yields an approximate solution in the separated form:

The PGD approximation is thus a sum of N functional products involving each

1.5.2 PGD Simulations of the Multi-bead-Spring FENE Model

governed by a Fokker-Planck equation of the general form (1.5). In order to demon-

τ p = n QFc (Q), (1.13)

This can be written as