Trim: 7in × 10in Top: 0.375in Gutter: 0.

875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

ADHESIVE PARTICLE FLOW

A Discrete-Element Approach

Adhesive Particle Flow: A Discrete-Element Approach offers a compre-

hensive treatment of adhesive particle flows at the particle level. This
book adopts a particle-level approach oriented toward directly sim-
ulating the various fluid, electric field, collision, and adhesion forces
and torques acting on the particles, within the framework of a discrete-
element model.
It is ideal for professionals and graduate students working in engi-
neering and atmospheric and condensed matter physics, materials sci-
ence, environmental science, and other disciplines where particulate
flows have a significant role. The presentation is applicable to a wide
range of flow fields, including aerosols, colloids, fluidized beds, and
granular flows.
It describes both physical models of the various forces and torques
on the particles as well as practical aspects necessary for efficient imple-
mentation of these models in a computational framework.

Jeffrey S. Marshall is a Professor in the School of Engineering at the

University of Vermont. He is a Fellow of the American Society of
Mechanical Engineers. He obtained a Ph.D. in Mechanical Engineering
from the University of California, Berkeley. Dr. Marshall taught at the
University of Iowa from 1993 to 2006, and was Chair of the Mechanical
and Industrial Engineering Department Chair from 2001 to 2005. He
is a recipient of the ASME Henry Hess Award and the U.S. Army
Research Office Young Investigator Award. He has authored more
than 95 journal articles and book chapters and one textbook, Inviscid
Incompressible Flow (2001).

Shuiqing Li is a Professor in the Department of Thermal Engineering

at Tsinghua University. He obtained a Ph.D. in Engineering Thermo-
physics from Zhejiang University. He was a visiting scholar at the Uni-
versity of Leeds in 2004–2005, at the University of Iowa in 2006, and at
Princeton University in 2010–2011. Dr. Li is a recipient of the National
Award for New Century Excellent Talents (2009) and the Tsinghua
University Award for Young Talents on Fundamental Studies (2011).
He shared a Chinese National Teaching Award on Combustion The-
ory. He has been awarded five fundamental grants from the Natural
Science Foundation of China and has authored more than 40 journal
articles.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Adhesive Particle Flow

A DISCRETE-ELEMENT APPROACH

Jeffrey S. Marshall
University of Vermont

Shuiqing Li
Tsinghua University
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

32 Avenue of the Americas, New York, NY 10013-2473, USA

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107032071

C Jeffrey S. Marshall, Shuiqing Li 2014

This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.

First published 2014

Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication Data
Marshall, Jeffrey S. (Jeffrey Scott), 1961–
Adhesive particle flow : a discrete-element approach / Jeffrey S. Marshall,
Shuiqing Li, University of Vermont, Tsinghua University.
pages cm
Includes bibliographical references.
ISBN 978-1-107-03207-1 (hardback)
1. Granular flow. 2. Adhesion. 3. Discrete element method. I. Li, Shuiqing,
1975– II. Title.
TA357.5.G47M37 2014
620.1 06–dc23 2013040678
ISBN 978-1-107-03207-1 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party Internet Web sites referred to in this publication
and does not guarantee that any content on such Web sites is, or will remain,
accurate or appropriate.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

To Marilyn and Yun,

and to Jodie, Eric, Emily, Paul, Jonathan, and Zelin
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Contents

Preface page xiii

Acknowledgments xvii

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1. Adhesive Particle Flow 1
1.2. Dimensionless Parameters and Related Simplifications 5
1.2.1. Stokes Number 5
1.2.2. Density Ratio 7
1.2.3. Length Scale Ratios 8
1.2.4. Particle Reynolds Number 10
1.2.5 Particle Concentration and Mass Loading 11
1.2.6. Bagnold Number 14
1.2.7. Adhesion Parameter 15
1.3. Applications 15
1.3.1. Fibrous Filtration Processes 15
1.3.2. Extraterrestrial Dust Fouling 18
1.3.3. Wet Granular Material 21
1.3.4. Blood Flow 23
1.3.5. Aerosol Reaction Engineering 25

2. Modeling Viewpoints and Approaches . . . . . . . . . . . . . . . . . . . . . . . . 29

2.1. A Question of Scale 29
2.2. Macroscale Particle Methods 30
2.2.1. Discrete Parcel Method 30
2.2.2. Population Balance Method 32
2.3. Mesoscale Particle Methods 34
2.3.1. Molecular Dynamics 36
2.3.2. Brownian Dynamics 37
2.3.3. Dissipative Particle Dynamics 38
2.3.4. Discrete Element Method 40
2.4. Microscale Dynamics of Elastohydrodynamic Particle
Collisions 41

vii
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

viii Contents

2.4.1. Microscale Simulations of Elastohydrodynamic

Interactions 42
2.4.2. Experimental Results for Two-Particle Collisions 44
2.4.3. Simplified Models for Restitution Coefficient in a
Viscous Fluid 46

3. Contact Mechanics without Adhesion . . . . . . . . . . . . . . . . . . . . . . . . 51

3.1. Basic Concepts 51
3.2. Hertz Theory: Normal Elastic Force 54
3.2.1. Derivation 55
3.2.2. Two-Particle Collision 56
3.3. Normal Dissipation Force 58
3.3.1. Physical Mechanisms 58
3.3.2. Models for Solid-Phase Dissipation Force 61
3.4. Hysteretic Models for Normal Contact with Plastic
Deformation 66
3.5. Sliding and Twisting Resistance 69
3.5.1. Physical Mechanisms of Sliding and Twisting Resistance 69
3.5.2. Sliding Resistance Model 72
3.5.3. Twisting Resistance Model 73
3.6. Rolling Resistance 74
3.6.1. Rolling Velocity 74
3.6.2. Physical Mechanism of Rolling Resistance 77
3.6.3. Model for Rolling Resistance 78

4. Contact Mechanics with Adhesion Forces . . . . . . . . . . . . . . . . . . . . . 81

4.1. Basic Concepts and the Surface Energy Density 82
4.2. Contact Mechanics with van der Waals Force 86
4.2.1. Models for Normal Contact Force 86
4.2.2 Normal Dissipation Force and Its Validation 96
4.2.3. Effect of Adhesion on Sliding and Twisting Resistance 98
4.2.4. Effect of Adhesion on Rolling Resistance 99
4.3. Electrical Double-Layer Force 100
4.3.1. Stern and Diffuse Layers 101
4.3.2. Ionic Shielding of Charged Particles 102
4.3.3. DLVO Theory 103
4.4. Protein Binding 107
4.5. Liquid Bridging Adhesion 111
4.5.1. Capillary Force 111
4.5.2. Effect of Roughness on Capillary Cohesion 116
4.5.3. Viscous Force 117
4.5.4. Rupture Distance 118
4.5.5. Capillary Torque on a Rolling Particle 118
4.6. Sintering Force 120
4.6.1. Sintering Regime Map 121
4.6.2. Approximate Sintering Models 123
4.6.3. Hysteretic Sintering Contact Model 124
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Contents ix

5. Fluid Forces on Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

5.1. Drag Force and Viscous Torque 131
5.1.1. Effect of Flow Nonuniformity 131
5.1.2. Effect of Fluid Inertia 132
5.1.3. Effect of Surface Slip 135
5.2. Lift Force 138
5.2.1. Saffman Lift Force 138
5.2.2. Magnus Lift Force 140
5.3. Forces in Unsteady Flows 141
5.3.1. Pressure-Gradient (Buoyancy) Force 141
5.3.2. Added Mass Force 142
5.3.3. History Force 143
5.4. Brownian Motion 145
5.5. Scaling Analysis 147
5.6. Near-Wall Effects 151
5.6.1. Drag Force 151
5.6.2. Lift Force 154
5.7. Effect of Surrounding Particles 156
5.7.1. Flow through Packed Beds 159
5.7.2. Flow through Fluidized Beds 159
5.7.3. Simulations 161
5.7.4. Effect of Particle Polydispersity 164
5.8. Stokesian Dynamics 165
5.8.1. Example for Falling Cluster of Particles 165
5.8.2. General Theory 169
5.9. Particle Interactions with Acoustic Fields 170
5.9.1. Orthokinetic Motion 172
5.9.2. Acoustic Wake Effect 173

6. Particle Dispersion in Turbulent Flows . . . . . . . . . . . . . . . . . . . . . . . 182

6.1. Particle Motion in Turbulent Flows 182
6.2. Particle Drift Measure 185
6.3. Particle Collision Models 188
6.3.1. Collision Mechanisms 188
6.3.2. Orthokinetic Collisions (Small Stokes Numbers) 190
6.3.3. Accelerative-Independent Collisions (Large Stokes
Numbers) 192
6.3.4. Accelerative-Correlative Collisions (Intermediate
Stokes Numbers) 192
6.4. Dynamic Models for Particle Dispersion 195
6.5. Dynamic Models for Particle Clustering 199

7. Ellipsoidal Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

7.1. Particle Dynamics 207
7.2. Fluid Forces 209
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

x Contents

7.3. Collision Detection and Contact Point Identification 211

7.3.1. Two-Dimensional Algorithms 212
7.3.2. Algorithms Based on a Common Normal Vector 213
7.3.3. Algorithms Based on Geometric Level Surfaces 214
7.4. Contact Forces 217
7.4.1. Geometry of Colliding Particles 217
7.4.2 Hertz Theory for Ellipsoidal Particles 218

8. Particle Interactions with Electric and Magnetic Fields . . . . . . . . . . . 223

8.1. Electric Field Forces and Torques 224
8.1.1. Coulomb Force and Dielectrophoresis 224
8.1.2. Dielectrophoresis in an AC Electric Field 227
8.1.3. Application to Particle Separation and Focusing 229
8.2. Mechanisms of Particle Charging 231
8.2.1. Field Charging 232
8.2.2. Diffusion Charging 233
8.2.3. Contact Electrification 235
8.3.4. Contact De-electrification 237
8.3. Magnetic Field Forces 237
8.4. Boundary Element Method 239
8.4.1. General Boundary Element Method 239
8.4.2. Pseudoimage Method for Particles near an Electrode
Surface 242
8.4.3. Problems with DEP Force near Panel Edges 243
8.5. Fast Multipole Method for Long-Range Forces 245
8.6. Electrostatic Agglomeration Processes 249
8.6.1. Relative Importance of Electrostatic and van der Waals
Adhesion Forces 249
8.6.2. Particle Chain Formation 250

9. Nanoscale Particle Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256

9.1. Continuum and Free-Molecular Regimes 257
9.1.1. Drag Force 258
9.1.2. Brownian Force 260
9.1.3. Mean-Free-Path of Nanoparticles 261
9.1.4. Thermophoretic Force 262
9.1.5. Competition between Diffusion and Thermophoresis
during Deposition 265
9.2. Nanoparticle Interactions 266
9.2.1. Collision of Large Nanoparticles 266
9.2.2. Collision of Small Nanoparticles 269
9.2.3. Long-Range Interparticle Electrostatic Forces 271
9.3. Time Scales of Nanoparticle Collision-Coalescence Mechanism 274
9.3.1. Time Scale of Particle Collisions 275
9.3.2. Time Scale of Nanoparticle Sintering 278
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Contents xi

10. Computer Implementation and Data Analysis . . . . . . . . . . . . . . . . . 286

10.1. Particle Time Stepping 286
10.1.1. Numerical Stability 287
10.1.2. Multiscale Time-Stepping Approaches 288
10.2. Flow in Complex Domains 289
10.2.1. Particle Search Algorithm 290
10.2.2. Level Set Distance Function 293
10.3. Measures of Local Concentration 294
10.4. Measures of Particle Agglomerates 297
10.4.1. Particle Count and Orientation Measures 297
10.4.2. Agglomerate Orientation Measures 298
10.4.3. Equivalent Agglomerate Ellipse 298
10.4.4. Agglomerate Fractal Dimension 300
10.4.5. Particle Packing Measures 302

11. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305

11.1. Particle Migration in Tube and Channel Flows 305
11.1.1. Inertial Particle Migration in Straight Tubes 306
11.1.2. Collision-Induced Particle Migration 307
11.1.3. Particle Migration in the Presence of Wavy Tube Walls 309
11.2. Particle Filtration 311
11.2.1. Fiber Filtration 312
11.2.2. Enhancement of Filtration Rate by Particle Mixtures 316
11.2.3. Enhancement of Filtration Rate by Electric Fields 318
11.3. Rotating Drum Mixing Processes 320
11.3.1. Flow Regimes 320
11.3.2. Mixing and Segregation 322
11.3.3. Cohesive Mixing and Segregation 326
11.4. Dust Removal Processes 328
11.4.1. Hydrodynamic Dust Mitigation 328
11.4.2. Electric Curtain Mitigation for Charged Particles 331
11.5. Final Comments 332

Index 339
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Preface

There has been a rapid increase in the number of research papers over the past
decade concerning flow of adhesive particles. Interest is driven in part by a focus on
particulate flow problems with small particle sizes, for which adhesive force becomes
increasingly important compared with particle inertia or gravity. Literature on flow
of adhesive particles is found both in the standard particulate flow and fluid mechan-
ics journals and in more specialized journals dealing with applications in areas such
as ash filtration, aerosol and cloud modeling, dust mitigation, nanoparticle depo-
sition, ceramics manufacturing, fouling of MEMS devices, food science, bioengi-
neering, microfluidics, sediment transport, and production of biofuels. Unlike pre-
vious research involving adhesive particles, which employed a population-balance
method, this recent work has adopted a mesoscale particle-level approach that simu-
lates the various fluid, electric field, collision, and adhesion forces and torques acting
on individual particles, enabling study of the collaborative dynamics governing the
interaction of groups of many agglomerates consisting of large numbers of particles.
Particle-level modeling of such problems is made possible both by improved physical
models of the various forces and torques acting on the particles, and by improved
computational algorithms for handling systems with a wide range of time scales.
Adhesive particulate flows arise in many applications in industry, nature, and life
sciences. In the field of manufacturing, applications include dust fouling of electronic
equipment, 3D printing, manufacturing and surface treatment of ceramic materials,
and electrospray processes. A variety of new microscale and nanoscale devices have
been designed whose manufacturing requires the precise placement of nanoparticles
and nanotubes onto a substrate using some type of dispersion process. Microfluidic
processes used for biological assay (“lab-on-a-chip”) rely on the ability to manipu-
late and sort particles and biological cells, which can be treated as particles. Algae
biofuel production requires the ability to process and optimize flows with suspended
algae cells, which respond to near-surface turbulent flow fields and light. Blood flow
involves not only interaction of red and white blood cells, but also interaction of
blood cells with platelets and other particles (e.g., liposomes or cancer cells) that
might be transported in the blood. Particulate pollution problems are of great con-
cern in many parts of the world due to ash from combustion processes that needs to
be captured before it escapes into the atmosphere. In many of these examples and

xiii
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

xiv Preface

in a wide variety of others, adhesion of particles (via a variety of mechanisms) plays

a critical role. The physics of these processes are often controlled by agglomerate
formation and breakup processes at the particle scale, and in cases such as nanopar-
ticle dispersion it is desirable to precisely control the motion of individual particles
in the presence of complex flow geometry.
The objective of this book is to provide a comprehensive account of mod-
ern particle-level approaches for analyzing and simulating particulate flows at the
mesoscale, with particular focus on flows involving adhesive particles. Although sev-
eral different modeling approaches are described, the book focuses specifically on
the soft-sphere discrete-element method (DEM), which is useful for a wide range of
particulate flow problems. DEM shares a similar Lagrangian computational method-
ology with molecular dynamics methods, but at the same time it makes use of exten-
sive modeling for fluid-induced forces on particles and for interparticle interaction
via collision, adhesion, and electric field effects. The book is structured accord-
ing to different types of approximations and computational models used for flow
simulation. The first chapter discusses various applications of flows with adhesive
particles and the associated dimensionless parameters governing them. This chapter
also introduces various approximations that are commonly made when analyzing
particulate fluids. The second chapter compares different modeling approaches for
adhesive particle flows as a function of length and time scales of the problem, and
examines different types of multiscale modeling approaches. This chapter introduces
the discrete-element model and compares it to other mesoscale and macroscale
models for particulate flows, such as molecular dynamics, Brownian dynamics, dis-
sipative particle dynamics, discrete parcel method, and population balance method.
The third chapter summarizes forces and torques that occur during particle col-
lision for cases with no adhesion forces, including elastic and dissipative normal
forces as well as resistance to sliding, rolling, and twisting motions. The effect of
adhesion on collision forces is discussed in the fourth chapter, including van der
Waals forces, electrical double-layer repulsion, protein binding forces often found
for cell interaction problems, liquid bridging, and sintering forces. Different fluid-
induced forces on particles are discussed in the fifth chapter, including a scaling
analysis to assess when different fluid forces can be neglected. This chapter also
discusses particle interaction with acoustic radiation. Turbulent dispersion models
are discussed in the sixth chapter, with an emphasis on accurate modeling of particle
collision rate in turbulent flows and its relationship to small-scale concentration field
heterogeneity. Chapter seven extends the discrete-element method to nonspherical
particles, which are common in applications such as blood flow, biofuel combus-
tion, and food processing. Particle interactions with electric and magnetic fields are
discussed in Chapter eight. These forces are important for many particle adhesion
problems as well as for control of particulate flows in many applications. Chap-
ter nine examines the differences between flows with micron-sized particles and
those with nanoscale particles. These differences arise from noncontinuum effects
due to the fact that nanoscale particles are often of a similar size as the mean-
free-path of the surrounding fluid. Chapter ten discusses issues that arise during
computer implementation of discrete-element methods, including numerical stiff-
ness, numerical instability, and challenges of computing particulate flows in complex
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Preface xv

domains. This chapter also discusses various measures used to characterize parti-
cle agglomerates. Chapter eleven describes select applications of discrete element
modeling of particulate flow problems, which are selected to illustrate interest-
ing physical phenomena exhibited by particles interacting with fluids and electric
fluids.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33

Acknowledgments

We would like to thank the many students and former students with whom we
have collaborated in studies of particulate flows, several of whose work played a
significant role in shaping this book. Particular thanks are extended to Jennifer
Chesnutt, Guanqing Liu, Kyle Sala, John Mousel, Greg Hewitt, Auston Maynard,
and Yiyang Zhang, each of whose work is featured in different sections of the book.
Comments on this research were provided by many colleagues, and we particu-
larly acknowledge valuable discussions with Charley Wu, Colin Thornton, Norman
Chigier, Chung K. Law, Aibing Yu, Jun-ru Wu, H.S. Udaykumar, Louis Rossi,
Stephen Tse, Albert Ratner, Yulong Ding, Jonathan Seville, Eric Loth, Stefan Lud-
ing, and Cetin Cetinkaya. Professors V.C. Patel and Qiang Yao and Dr. John R.
Grant are particularly acknowledged for their invaluable mentoring and friendship
throughout our careers.
Funding to support the work of JSM on particulate flow from NASA
(NNX12AI15A, NNX13AD40A, NNX08AZ07A), the U.S. National Science Foun-
dation (DGE-1144388, CBET-1332472), the U.S. Department of Energy (DE-
FG36 – 08G088182), the Caterpillar Corporation, and the University of Iowa Facil-
ities Management Group is greatly appreciated. SQL particularly acknowledges
support from the National Science Foundation of China (No. 50306012, 50776054,
50976058, and 51176094) in his early career, and from the National Key Basic
Research and Development Program (No. 2013CB228506) to work across disci-
plines.
Assistance in production of the book was provided by Runru Zhu, Wenwei Liu,
Melissa Faletra, and Yihua Ren in producing some of the figures; by Bing Chen
for code assistance; and by the students enrolled in SQL’s Introduction to Particle
Transport class at Tsinghua. Special thanks are extended to Emily Marshall for han-
dling all of the permissions, and to our editor Peter Gordon at Cambridge University
Press and our project manager Adrian Pereira at Aptara, Inc., for their enthusiasm,
encouragement, and professionalism throughout the writing and production of the
book.

xvii
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-FM CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:33
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1 Introduction

1.1. Adhesive Particle Flow

A particulate flow is one in which a moving fluid interacts with a large number of dis-
crete solid particles. The category is extraordinarily broad, encompassing everything
from suspended dust carried by atmospheric winds to avalanches of debris or snow
rolling down a hillside. Widely varying industrial, biological, and environmental pro-
cesses can be interpreted as particulate flows, encompassing areas of study such as
sediment transport by stream and coastal flows, aerosol dynamics, colloidal suspen-
sions, fluidized bed reactors, granular flows, slurries, and nanoparticle dispersions.
There are also many situations where a suspension of biological cells can be inter-
preted as a particulate fluid, which extends the notion of particulate flow to problems
such as blood flow and algal suspensions. Finally, there are many aspects of the meth-
ods used to analyze and model particulate flows that can be either directly applied
or applied with small modifications to other types of multiphase flows, including
droplet dispersions and bubbly flows, assuming that the deformation of the droplets
and bubbles is minimal.
Despite the many different forms in which we encounter them, there are a
number of characteristics that are shared by most particulate flows. Some of these
characteristics arise from the interaction of the individual particles with the sur-
rounding fluid. For instance, a particulate flow past a blunt body tends to exert a
higher drag force than the body would experience in a fluid with no particles. One
characteristic of particulate fluids associated with this increased drag is the fact that
the effective viscosity of the fluid increases due to the presence of particles, even if
the particles move with the same velocity as the surrounding fluid (Einstein, 1906). A
second characteristic of particulate fluids is the high degree of dissipation that occurs
due to the interfacial force caused by particle motion relative to the fluid velocity
field, for instance, as the fluid accelerates around a body. This relative particle-fluid
motion also causes oscillatory flows associated with sound waves and turbulent fluc-
tuations to attenuate rapidly in particulate fluids (Epstein and Carhart, 1953). A third
characteristic of particulate fluids is that they can be highly abrasive due to collision
of particles with a surface, for instance, in a sand blasting application. This abrasive
quality of particulate fluids arises from the fact that heavy particles with more inertia
than the surrounding fluid will drift relative to fluid streamlines as they curve around

1
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

2 Introduction

a body, leading particles to instead collide with the body surface, imparting their
momentum to the body. Of course, fluid flows in general, and multiphase flows in
particular, are highly nonlinear, and so none of these characteristics hold invariably.
For instance, Lashkov (1992) showed that at small concentrations the presence of
particles can serve to accelerate occurrence of drag crisis in flow past a bluff body,
thus moving the point of boundary layer separation backward on the body surface.
Although the local skin friction is increased by the presence of particles, the pressure
drag on the body is sufficiently decreased by the backward motion of the separation
point that the net drag on the body decreases due to the presence of particles in the
flow in a select range of Reynolds numbers.
Other characteristics of particulate flows are associated with the interaction of
the particles with each other. A common simplification in the fluid flow literature
is that of dilute particulate flow, in which case the particles are assumed to be so
few in number and so small as to never interact with each other. While from a
modeling point of view this is often a convenient assumption, in many applications
particle collisions play a vital role. Particle collisions occur due to two different
processes. Inertia-induced collisions occur due to the drift of particles relative to the
surrounding fluid that is caused by the particle inertia. The direction and degree of
particle drift varies across the flow field, and these differences give rise to particles
originating from different parts of the flow colliding with each other. The second
cause of particle collisions is fluid shear stress. Shear-induced collisions occur due to
the finite particle size as particles in a faster-moving layer of the shear flow move past
particles in a slower-moving layer. Collisions may also occur when particles exert
forces on each other, as might be caused by the long-range Coulomb force when
particles are electrically charged (Zhang et al., 2011).
Particle collisions have a number of important consequences. For instance, col-
lisions affect the particle transport by modifying the direction of particle velocity,
thereby increasing the rate of particle diffusion in the fluid. Collisions can cause trans-
fer of electrons between the particles, leading to accumulation of electrical charge
on the particles, a process known as tribocharging. This process is responsible, for
instance, for charging of the dust particles on Mars and associated radio and electri-
cal discharges emitted by the dust storms (Renno et al., 2003). Violent collisions can
cause fracture and breaking of the particles, or more moderate collisions can over
time cause a gradual smoothening of the particle surfaces. Most importantly, in the
presence of an adhesive force, collisions can cause particles to stick together.
Particle adhesion typical occurs for suspensions of small particles, with diameter
of about 10 microns or smaller, for which the adhesive force can overcome the
gravitational force and the elastic rebound force acting on the particles. Colliding
particles held together by relatively weak adhesive forces (such as van der Waals
attraction or capillary force) are generally called agglomerates, whereas aggregates
occur when particles adhere together by strong forces, for example by covalent,
ionic, or sintering bonds. This terminology, first established by Nichols et al. (2002),
is not universally accepted in the particulate flow literature, and we often find the
terms switched or referred to instead as soft agglomerate and hard agglomerate,
respectively.
There are a wide variety of different adhesive forces that act between particles
during and after particle collisions. Kinloch (1987) proposed six distinct categories of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.1 Adhesive Particle Flow 3

particle adhesion: dispersive adhesion due to van der Waals attraction of molecules
in the contact region between the particles; electrostatic adhesion due to electrical
attraction of charged or polarized particles; diffusive adhesion due to diffusion of
atoms or molecules from one particle to the next, such as occurs with sintering
of ceramic or metal powders; chemical adhesion due to swapping or sharing of
atoms of the two particles, as occurs with ionic and covalent bonding processes,
respectively; hydrogen-bond adhesion due to formation of a weak bond when atoms
such as oxygen, nitrogen, or fluorine share a hydrogen nucleus; and finally mechanical
adhesion, which occurs when the void or pore space between the particles is filled
with a third medium that mechanically holds the particles together.
A common example of mechanical adhesion occurs for slightly wet granular
particles in air, in which each particle is surrounded by a thin liquid film. The liquid
films on two colliding particles connect to form a liquid bridge, which fills the space
between the particles and holds the particles together by capillary force, modified
by a viscous effect within the liquid as the particles are moved toward or away from
each other (Ennis et al., 1990). Dry dust particles usually adhere to each other or to
the surfaces of a surrounding solid body by a combination of electrostatic and van
der Waals forces. Electrostatic forces are relatively weak, but since they act over a
relatively long range, electrostatic forces are effective at transporting dust particles
together or moving them toward the surface of a body immersed in the flow. Once
collision has occurred, the stronger (but relatively shorter-range) van der Waals
force typically takes over and is primarily responsible for holding the dust particles
together (Feng and Hays, 2003; Gady et al., 1996). Biological cells adhere to one
another by a type of chemical adhesion called ligand-receptor binding (Bell, 1978), in
which ligand proteins on the cell surface connect to specific receptor proteins on the
surface of the opposing cell to form a molecular link between the cells. This cellular
bonding is responsible for adhesion of white blood cells to endothelial cells on a
blood vessel wall prior to transport of the white blood cells into surrounding tissue,
among other things. These and other examples of flows with adhesive particles are
discussed in the final section of this chapter.
There exists extensive literature on particulate flows, including many excellent
books and review articles. Much of this literature, however, deals with either dilute
particulate flow or with granular flow of nonadhesive particles. Most existing lit-
erature on adhesive particles takes what is called a population-balance approach,
in which analytical models are used to cluster agglomerates of particles into larger
“effective particles” without taking into account the dynamics of the individual par-
ticles that make up the agglomerate (Friedlander, 2000). The population-balance
approach can be very useful for large-scale systems provided that the conditions
under which it is derived are satisfied. This requires, for instance, that the particle con-
centration is sufficiently small, that particles are spherical and have negligible inertia,
that the particle distribution is homogeneous, that collisions are binary, and that the
flow length scales are much larger than the particle agglomerates, so that we can
treat agglomerates as existing in a simple shear flow (Reinhold and Briesen, 2012).
The rise in prominence of microscale and nanoscale engineering in recent years,
including microscale description of biological systems, has resulted in a great deal
of increased focus on processes that are outside the scope of traditional population-
balance models. Such problems include processes where particles interact with
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

4 Introduction

microscale structures that have a length scale comparable to that of the particle
agglomerates, processes with high particle concentrations or with nonspherical par-
ticles, processes involving strong gradients in particle concentration, or processes
where particle dispersion relative to the flow streamlines plays an important role.
Such problems require a fundamentally different approach for modeling and sim-
ulation, based on tracking the dynamics of individual particles both while they are
traveling alone through the fluid and while they are colliding with and adhering to
other particles. At the same time, the modeling approach should allow a sufficiently
large number of particles in the simulation to enable examination of the collective
dynamics of adhesive particles as large agglomerates interact with each other and
with fluid flows in complex domains. Such collective dynamics often require simula-
tions ranging from tens of thousands to millions of particles. This requirement puts
such problems out of the scope of microscale approaches, which simulate the flow
around each individual particle.
Within the past few decades, a variety of computational methods have been
developed that model particle collective dynamics at an intermediate (or mesoscale)
level. These mesoscale methods, which we broadly classify as Lagrangian particle
methods, attempt to strike a middle ground in which the motions of individual parti-
cles are resolved, but sufficiently coarse modeling is used to describe the particle
interactions that large numbers of particles can be accommodated. Lagrangian
particle methods have developed rapidly in application areas such as atomic-level
simulation of molecular processes, colloidal fluids and aerosols, granular flows, and
polymeric fluids. A number of different variations of these methods have arisen,
examples being molecular dynamics (MD), Brownian dynamics (BD), dissipative
particle dynamics (DPD), and the discrete-element method (DEM). Lagrangian
particle methods share the feature that the computational elements (e.g., atoms,
molecules, particles) are allowed to interact for a period of time under prescribed
interaction laws.
These different methods have important differences in both approach and appli-
cability. MD deals with individual atoms with dimensions of roughly 0.1 nm by using
a computational time step on the order of 1 femtosecond (10–15 s), while the duration
of the overall computation typically ranges from picoseconds to nanoseconds. At
present, MD is too time consuming to apply for particles with sizes larger than about
10 nm. At a larger scale, DPD simulates the dynamic and rheological properties of
simple and complex fluids by using a computational element that represents a group
of fluid molecules, rather than individual molecules or atoms. Brownian dynamics
(BD), or the more general Langevin dynamics (LD), introduces the notion of a fluid
continuum surrounding a set of molecules or small particles, which is used to rep-
resent effects of a solvent or gas phase without simulating the individual molecular
dynamics of this surrounding fluid. The surrounding fluid continuum exerts both a
drag force and a random (Brownian) force on the computational elements. Use of
the fluid continuum approximation, together with neglect of particle inertia in BD,
enables the BD method to employ much higher time steps than the MD or DPD
methods.
This book focuses on the discrete element method, in which the motion, collision,
and adhesion of individual particles are resolved in time and space, but analytical
models are used to approximate the interaction of the particles with the surrounding
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.2 Dimensionless Parameters and Related Simplifications 5

fluid. The discrete element method was originally developed for nonadhesive granu-
lar flows, for which it is used extensively to simulate collisions of individual granules
with each other (Bertrand et al., 2005; Cundall and Strack, 1979). Although for
macroscopic granular particles the dynamics are governed mostly by gravity and col-
lisional and frictional forces, for adhesive microparticles the dominant interactions
include electrostatic (Coulomb) and van der Waals forces (Aranson and Tsimring,
2006). There has recently been rapid progress on understanding the physics related
to the intermolecular and surface forces at the microscale (Israelachvili, 2011), which
enable us to develop more rational adhesive contact models. In this book, we exam-
ine the discrete element method in a way that is applicable for a wide range of
particulate flows, from granular flows to nanoparticle flows. The book places par-
ticular emphasis on adhesive particle flows, which consist of suspensions of adhesive
particles immersed in a fluid flow. The modeling approach for such flows must deal
with collision and adhesion of the particles to each other and to surrounding sub-
strates, with interaction of the particles with the surrounding fluid, with complex
flow domains, and with the effect of external fields, such as electrostatic or acoustic
fields, on the particles.

1.2. Dimensionless Parameters and Related Simplifications

Dimensionless parameters are used in fluid mechanics to delineate regimes of motion
and to establish limitations of various assumptions and approximations used in
obtaining a problem solution. A large number of dimensionless parameters are used
in describing particulate flows; however, among these, a few play a particularly
important role in characterizing the flow type and the various approximations used
in flow analysis. While various other dimensionless parameters will be introduced
throughout the book, it is useful to discuss a few particular parameters early on so as
to introduce and characterize the different assumptions typically used in analyzing
particle flows.

1.2.1. Stokes Number

A particulate flow in which particles are strongly influenced by the surrounding
fluid is called a suspension, examples being colloids and aerosols. At the opposite
extreme, particulate flows where there is little influence between the particles and
the fluid are called granular flows. Where a specific particulate flow falls on this
spectrum from suspension to granular flow is governed largely by the Stokes num-
ber, denoted by St. The Stokes number is defined as the ratio of the characteristic
response time of a particle to the characteristic fluid time scale. By way of motiva-
tion in developing these time scales, we consider a simplified form of the momentum
equation for a single particle in which the particle momentum changes only due to
the drag force between the particle and the surrounding fluid. It is further assumed
that the particle is very small and slow-moving relative to the fluid and that it is
spherical in shape, so that the classic Stokes drag solution gives the drag force on the
particle as

Fd = −3π μd(v − u). (1.2.1)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

6 Introduction

In this equation, d is the particle diameter, μ is the fluid viscosity, and v and u are
the particle velocity and the fluid velocity evaluated at the particle centroid (in
the absence of the particle), respectively. With these approximations, the simplified
particle momentum equation becomes

dv
m = −3π μd(v − u), (1.2.2)
dt
where m is the mass of a single particle and d/dt denotes the rate of change in time
at a point following the particle. If the fluid flow in which the particle is immersed
has a characteristic velocity scale U and length scale L, dimensionless velocity and
time variables can be defined as

v = v/U, u = u/U , t  = tU/L. (1.2.3)

Substituting these dimensionless variables into (1.2.2) yields

mU dv
= −(v − u ). (1.2.4)
3π μdL dt 

The fluid advection time scale is τ f = L/U and the particle response time is τ p =
m/3πdμ. From these time scales, the Stokes number can be defined by
τp mU
St ≡ = , (1.2.5)
τf 3π μdL

which is equal to the coefficient multiplying the dimensionless particle acceleration

in (1.2.4). Writing particle mass m = π ρ p d3 /6 in terms of particle density ρ p gives
an alternative expression for Stokes number as

ρ p d 2U
St = . (1.2.6)
18μL

When the Stokes number is much smaller than unity, the particles nearly follow
the fluid streamlines and the magnitude of the particle drift velocity relative to the
fluid is small compared to the fluid velocity scale U. When the Stokes number is
much larger than unity, the particles respond only very slowly to the fluid force.
Some of the most interesting flows occur when the Stokes number is near unity, for
which the fluid flow has a strong effect on the particle transport but the particles
also disperse significantly relative to the fluid flow. An example illustrating particle
transport in the vortex street wake behind a square cylinder is given in Figure 1.1
from Jafari et al. (2010), using a two-dimensional lattice-Boltzmann method. In this
figure, the particles are initiated upstream in a random configuration and transported
past the cylinder by the fluid flow. Computations are shown with two values of the
particle Stokes number. In Figure 1.1a, with St = 0.001, the particles are advected
with the fluid flow and exhibit random positions independent of the fluid vortices. In
Figure 1.1b, with St = 1, the particles are expelled from the vortex cores by the action
of centrifugal force, forming thin particle sheets surrounding the vortex structures
in the wake.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.2 Dimensionless Parameters and Related Simplifications 7

(a)

(b)

Figure 1.1. Instantaneous particle dispersion patterns from a numerical simulation of a plane
wake behind a square cylinder at (a) St = 0.001 and (b) St = 1.0. [Reprinted with permission
from Jafari et al. (2010).]

1.2.2. Density Ratio

The ratio of the fluid density to the particle density, χ ≡ ρ f /ρ p , plays an important
role in determining how particles respond to curvature in the fluid streamlines. To
see this, we consider the problem of a particle placed in the flow field of a line vortex
of strength . In addition to the drag force given by (1.2.1), the particle is subject to
two other forces that act in the radial direction relative to the vortex axis. The first of
these forces is the centrifugal force, which acts to expel the particle from the vortex.
In an inertial reference frame, the centrifugal force is included as part of the dv/dt
term in (1.2.2), so to better highlight this effect we examine the particle in a frame
of motion that rotates at a rate F = uθ /r = /2π r2 , where r is the radial position
of the particle centroid and uθ = /2π r is the fluid azimuthal velocity at radius r. In
this frame, the centrifugal force is

Fc = m2F r, (1.2.7)

where r = rer is the particle radial position times the unit vector er in the radial
direction. The second force that acts on the particle in the radial direction arises
from the fact that the pressure has a minimum value at the vortex axis and increases
away from the axis, such that there exists a radial pressure gradient surrounding the
line vortex given by

ρ f 2
∇p = er . (1.2.8)
4π 2 r3
This pressure gradient induces a force on the particle, similar to the buoyancy
force that acts on a beach ball held under water in the presence of a hydrostatic
pressure gradient. The value of the resulting pressure-gradient force acting on the
particle is (Marshall, 2001)

ρ f V 2
F p = −V ∇ p = − er , (1.2.9)
4π 2 r3
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

8 Introduction

Figure 1.2. Bubble ring formed by a scuba diver by

bubble entrainment into a vortex ring core.

where V is the particle volume. The centrifugal force acts to throw the particle out
of the vortex and the pressure-gradient force acts to draw the particle inward toward
the vortex axis.
The ultimate fate of the particle is determined by the sum of these two radial
forces, given by

(ρ p − ρ f )V  2
Fc + F p = er . (1.2.10)
4π 2 r3
If the particle density is greater than the fluid density, the net force is oriented in the
positive radial direction and the particle drifts outward, away from the vortex axis.
In this case, the particles are referred to as “heavy” and exhibit patterns similar to
that shown in Figure 1.1b in which particles collect in high-concentration sheets that
surround the vortex structures. This effect leads to development of a high degree of
particle concentration intermittency in certain turbulent flows, and it also accounts
for the formation of streaky patterns in the near-wall region of turbulent boundary
layers where the particles are centrifuged out of the cores of the quasi-streamwise
vortices that form in the near-wall turbulent flow (Kaftori et al., 1995). If the particle
density is less than the fluid density, the net force in (1.2.10) is negative and the
particle is pulled toward the vortex axis. For example, cavitation bubbles formed
along a propeller surface are pulled into the trailing vortex behind the propeller
blade, leading to the formation of the distinctive gas-filled propeller trailing vortex.
Similarly, air bubbles blown by a scuba diver can be entrained into the core of
a vortex ring, forming the so-called “bubble-ring” seen in Figure 1.2. Whales and
dolphins have been observed spending a great deal of time forming and playing with
bubble rings.

1.2.3. Length Scale Ratios

The dimensionless particle diameter is defined as the ratio of particle diameter d and
fluid length scale L. We use the term “bulk flow” to refer to the flow field minus the
perturbations caused by the particles. In particulate fluids, it is usual to assume that
the bulk flow seen by an individual particle is relatively simple, often consisting of just
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.2 Dimensionless Parameters and Related Simplifications 9

a uniform flow and a shear flow, so that simple analytical models can be constructed
for the fluid forces on the particles. This condition is fulfilled provided that the value
of d/L is sufficiently small compared to unity. If the bulk flow seen by an individual
particle is complicated (i.e., if d/L is not small compared to unity), then it is usually
necessary to numerically model the complete flow field around each particle, which
limits the number of particles that can be considered in the computation.
To give a specific example, consider the flow of red blood cells (the “particles”) of
“effective” diameter d in a blood vessel of diameter L. The value of the dimensionless
particle diameter, d/L, depends on which section of the cardiovascular system is
under consideration. In the large arteries d/L ≈ 0.001, so the changes in the bulk
flow occur on a much larger length scale than the particle diameter. In small arteries
and arterioles, d/L has values between 0.02 and about 0.08, which are still sufficiently
small that the particles can be regarded as seeing relatively simple bulk flows. On
the other hand, in the capillaries, d/L has values ranging from 0.5 to 1, for which the
simple drag and lift expressions used in particulate flow models break down and it is
necessary to numerically simulate the entire flow field around each red blood cell.
The dimensionless particle diameter is also an important parameter for studies
of turbulent flow of particulate fluids. For such problems, the fluid length scale L can
often be replaced by the turbulence integral length scale. For small values of d/L,
the particles are found to attenuate the turbulence. This attenuation arises from a
combination of increased mixture inertia due to addition of particles, particle drag
that occurs from relative motion of the particle through the fluid (including radial
drift of the particles in response to the turbulent eddies), and enhanced effective
viscosity of the fluid (Balachandar and Eaton, 2010). By contrast, for high values
of d/L the particles are found to accentuate the turbulence due to injection of new
turbulent eddy structures in the particle wakes. In this case, each particle acts to
transition kinetic energy from the bulk flow into turbulent kinetic energy within the
particle wake. Buoyancy-induced instabilities that arise from density variations in
the fluid as the particles cluster in certain preferential locations can also enhance the
turbulence intensity (Elghobashi and Truesdell, 1993). The transition from atten-
uation of turbulence to enhancement of turbulence was examined by a number of
investigators (Gore and Crowe, 1989; Hetsroni, 1989). Based on the plot reprinted
in Figure 1.3, Gore and Crowe (1989) proposed that turbulence is attenuated by the
presence of particles when d/L < 0.1 and that particles enhance turbulence when
d/L ≥ 0.1.
Another important length scale ratio for particle transport is the Knudsen num-
ber, defined as the ratio of mean free path λ of the molecules in the fluid to the
particle diameter, or
λ
Kn ≡ . (1.2.11)
d
The flow past a particle can be treated using the standard continuum assumption
if Kn < 10−3 (Crowe et al., 2012). For particles in gases, the mean free path of the
fluid can often be large enough that this criterion is exceeded. For instance, air is
primarily nitrogen, which has a mean free path of λ = 10−7 m. Dust particles have
a typical diameter of about 10 microns, or d = 10−5 m. The Knudsen number for a
dust particle in air is therefore approximately Kn ∼
= 0.01, which is about an order of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

10 Introduction

400

% CHANGE IN TURBULENT INTENSITY

300

+
+
200
× ×

100 +
×

×
+
+
0 × ×××× × × × ×
× ×× +
× +
+
×+
+
+
–50 +
0.0001 0.001 0.01 0.1 1
dp/le

Figure 1.3. Attenuation and enhancement of turbulence by particles as a function of the

ratio of dimensionless particle diameter to the turbulence integral length scale. The symbols
represent experimental data collected from many different researchers, as listed by Gore and
Crowe (1989). [Reprinted with permission from Gore and Crowe (1989).]

magnitude larger than the criterion given earlier for application of the continuum
approximation. The most common effect of violation of this criterion for particulate
flows is the presence of nonzero slip of the fluid past the particle, which modifies the
Stokes drag expression (1.2.1), as discussed in Section 5.1.

1.2.4. Particle Reynolds Number

The Reynolds number is a ratio of inertial to viscous force. The instantaneous particle
Reynolds number Re p is defined in terms of the particle instantaneous velocity
relative to the fluid as
ρ f d |v − u|
Re p ≡ , (1.2.12)
μ
where ρ f is the fluid density. The difference vd ≡ v − u is called the particle slip
velocity or drift velocity, depending on the context. Because |v − u| varies in time
as a particle moves about, Re p is also a function of time. However, it follows from
(1.2.4) and the scaling relationship (1.2.3) that v − u = O(St U ) for small values of
the Stokes number, so a characteristic particle Reynolds number Re p can be defined
which is a constant for the flow field as follows:
 
ρ f d(St U ) d
Re p ≡ = St Re f , (1.2.13)
μ L
where Re f ≡ ρ f LU/μ is the flow Reynolds number.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.2 Dimensionless Parameters and Related Simplifications 11

For small values of the particle Reynolds number, the simple Stokes drag force
(1.2.1) can be used for the particle drag, assuming that other particles are sufficiently
far away. For sufficiently small particles immersed in a gas, fluid slip at the particle
surface becomes important and other forces, such as Brownian motion, also play an
important role. Many particulate flows are characterized by small particle Reynolds
numbers, including most colloidal and aerosol suspensions. More complicated drag
expressions exist for cases where the particle Reynolds number is not small, which is
often the case in particulate flows that arise in applications such as biofuel combus-
tion, agricultural grain processing, bed-load sediment transport, and pneumatic coal
feeders, as well as in many granular flows in general. The range of particle sizes for
which adhesion forces play a significant role tend to have small particle Reynolds
number, although problems involving wet granular media might pose an exception.
The flow Reynolds number, Stokes number, density ratio, and dimensionless
particle diameter are not independent of each other, but can be related to each other
in a variety of ways. For instance, the Stokes number can be written in terms of the
flow Reynolds number as
 2
1 d
St = Re f . (1.2.14)
18χ L
Substituting (1.2.14) into (1.2.13) gives the averaged particle Reynolds number as
 3
1 d
Re p = Re2f . (1.2.15)
18χ L
Combining (1.2.15) and (1.2.14) gives the Stokes number in terms of the averaged
particle Reynolds number as
 
1 d
St =
2
Re p . (1.2.16)
18χ L

1.2.5. Particle Concentration and Mass Loading

The particle volume concentration, or simply particle concentration, φ is defined as
the ratio of the volume VS,p occupied by all of the particles in a system divided by
the total system volume VS, f + VS,p , or
VS,p
φ≡ , (1.2.17)
VS, f + VS,p
where VS, f is the part of the system volume occupied by the fluid. This definition
yields a single concentration value for the entire flow field. It is also possible to define
a local concentration field that varies as a function of position and time within a flow,
which is discussed in detail in Chapter 10.
A useful length scale ratio can be derived from the concentration value. Let us
assume that a system consists of spherical particles of diameter d that are arranged
in a uniform array with particle spacing distance between each pair of adjacent
particles. Each cubic cell in this array of volume V = 3 includes a particle volume
equal to the volume of a single particle, or Vp = π d3 /6. The particle concentration
is therefore given by
 
Vp π d 3
φ= = . (1.2.18)
V 6
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

12 Introduction

(a) (b)
Figure 1.4. Two popular close-packed arrangements of spheres: (a) the face-centered cubic
(fcc) and (b) hexagonal-centered cubic (hcc).

Solving (1.2.18) for the particle spacing ratio /d gives

 1/3
π
= . (1.2.19)
d 6φ
This result is useful for determining the significance of particle-particle collisions in
a particulate flow. If /d 1, the particles will rarely collide with each other and
the particulate flow can be considered to be dilute. As /d approaches unity, the
rate of particle collisions increases rapidly, and the particulate flow is said to be
dense. Of course, if the particles are adhesive then even infrequent collisions will
over sufficient time lead to formation of agglomerates, so even in the limit of large
/d particle collisions cannot be neglected in adhesive particle flows.
The volume concentration has an upper bound, φmax , when the particles are
in the maximum packing arrangement. The value of φmax depends on the particle
shape and arrangement in the maximum packing configuration. For spherical parti-
cles maximum packing occurs in a close-packed arrangement such as those shown
in Figure 1.4, in which each sphere is in contact with exactly 12 other spheres.
Among the popular close-packed arrangements are the face-centered cubic (fcc)
and hexagonal-centered cubic (hcc) arrangements. All close-packed arrangements
yield the same theoretical value of maximum concentration, given by
π
φmax = √ ∼ = 0.7405. (1.2.20)
18
In practice, particulate flows cannot reach this high value of volume concentration
because the particles are not arranged in a regular packing, but are instead randomly
carried along by the flow. The volume concentration for a randomly packed particle
bed varies from about 0.52 to 0.74, depending on the method used to deposit the
particles (Dullien, 1992).
The mass concentration, φM , is defined as the ratio of the mass of the particles
in a system to the total mass of the system, or
ρ pVS,p
φM ≡ . (1.2.21)
ρ f VS, f + ρ pVS,p
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.2 Dimensionless Parameters and Related Simplifications 13

In a dilute flow, the volume concentration is small, such that VS,p VS, f and φ ∼
=
VS,p /VS, f 1. In this case, the volume and mass concentrations can be related as
φ
φM ∼
= . (1.2.22)
χ +φ
We do not neglect the φ term in the denominator of (1.2.22), because the density
ratio χ can also be very small. Some authors define mass concentration differently
as the ratio of particle mass to fluid mass (e.g., Crowe et al., 2012), so that the
volume and mass concentration in a dilute flow will be related simply as φM ∼ = φ/χ .
In the present book we use the definition in (1.2.21), which is more in line with the
traditional interpretation of concentration.
The mass loading (or sometimes simply the loading), Z, is defined as the ratio
of mass flow rate ṁ p of the particles to the mass flow rate ṁ f of the fluid, or
ṁ p
Z≡ . (1.2.23)
ṁ f

If the fluid and particle velocities are approximately the same (e.g., for low Stokes
numbers), then the mass loading reduces to a ratio of the particle mass to the fluid
mass in the system, or
ρ pVS,p
Z∼
= . (1.2.24)
ρ f VS, f
For dilute flows, the volume concentration can be approximated as φ ∼
= VS,p /VS, f , so
(1.2.24) reduces to
φ
Z∼
= . (1.2.25)
χ
The mass loading plays a key role in determining the effect of the particles on the
fluid flow. In a solution of particulate flows with small concentration values, it is often
useful to make the assumption of one-way coupling, which assumes that the particles
do not affect the fluid flow. This assumption greatly simplifies the flow computation
because we can solve for the fluid flow independently of the particles, and then solve
for the particle motion in a known fluid flow field. In flows where this assumption is
not valid we must utilize full two-way coupling, in which the particle and fluid motion
are fully coupled to each other, often requiring an iterative solution approach.
Validity of the one-way coupling assumption requires that the drag exerted on
the fluid by the particles is much less than the momentum flux of the particulate flow.
For a dilute flow system (φ 1) with N particles, in which the particles are assumed
to have a sufficiently small Reynolds number so that the Stokes drag law (1.2.1) is
valid, a momentum coupling parameter can be defined as the ratio of net particle
drag force exerted on the fluid to the fluid momentum flux, giving
3π μdNvd
= , (1.2.26)
ρ f U 2 L2
where vd denotes the magnitude of the characteristic particle slip velocity vd ≡ v − u.
The one-way coupling approximation can be considered to be valid when the value
of is sufficiently small, typically less than about 10%.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

14 Introduction

The number of particles N in (1.2.26) can be written in terms of the volume

concentration as
6φL3
N= . (1.2.27)
π d3
Because the slip velocity has order of magnitude vd = O(St U ) for a flow at small
Stokes number, substituting these estimates for N and vd into (1.2.26) and using the
definition (1.2.6) for Stokes number gives the result that the momentum coupling
parameter at small Stokes numbers is approximately equal to the mass loading, or
φ ∼
= = Z. (1.2.28)
χ
In this limit, the fluid and particle velocities are equal to each other to first order
in the Stokes number, so the approximation (1.2.25) applies. An extension of this
estimate valid for finite Stokes numbers is proposed by Crowe et al. (2012) as
1 φ
= . (1.2.29)
1 + St χ
These results indicate that the assumption of one-way coupling is valid for small
mass loadings, and that this assumption becomes even more accurate as the Stokes
number increases with a fixed mass loading.

1.2.6. Bagnold Number

The Bagnold number, Ba, is used to measure the importance of particle collisions on
stress transmission in a particulate flow, and can loosely be interpreted as a ratio of
the order of magnitude of particle collision stress to that of viscous fluid stresses in
the flow field. The Bagnold number is defined mathematically by (Bagnold, 1954)
ρ p d2 λ1/2
B γ̇
Ba = , (1.2.30)
μ
where γ̇ is the fluid shear rate and λB is called the linear concentration, defined in
terms of the particle volume concentration φ and the maximum packing concentra-
tion φmax by
1
λB = . (1.2.31)
(φmax /φ)1/3 − 1
Writing the shear rate in terms of a fluid velocity scale U and length scale L as
γ̇ = U/L, the Bagnold number can be written as a function of the Stokes number
and the particle concentration as
Ba = 18 St λ1/2
B . (1.2.32)
Based on a series of experiments for the dynamics of large particles in Newton
shear flows, Bagnold separated the flow dynamics into a “viscous” regime in which
the flow is dominated by viscous fluid stresses when Ba is sufficiently small (typically,
Ba < 40) and a “grain-inertia” regime in which the stress is primarily transmitted
by particle collisions when Ba is sufficiently large (typically Ba > 400 to 450). In the
grain-inertia regime, the particle system behaves as a granular flow and the effect of
the interstitial fluid is relatively small.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.3 Applications 15

1.2.7. Adhesion Parameter

In flows of adhesive particles, the set of dimensionless parameters that govern the
particle transport must also include some parameter associated with the particle
adhesive force. This can be done in a variety of ways, but several recent studies
have defined an adhesion parameter as a ratio of the adhesive force and the particle
inertia (e.g., Li and Marshall, 2007). Transport of particles in the micrometer size
range is dominated by a balance between particle inertia and drag force, so use of
either viscous drag or particle inertia in the denominator of this parameter yields
an essentially equivalent parameter. In this book, the adhesion parameter, Ad, is
defined by

2γ
Ad = , (1.2.33)
ρ pU 2 d

where γ is the adhesive surface energy, which is equal to half the work required to
separate two surfaces that are adhesively bound per unit surface area. This parameter
has an obvious relationship to the Weber number, We = ρ f U 2 d/2σ , which is a
measure of the relative importance of the fluid inertia compared with its surface
tension σ . For large values of the adhesion parameter, particles tend to stick together
upon collision, forming particle agglomerates whose size is determined primarily by
the bending and shear forces exerted by the fluid drag force on the agglomerated
particles. For small values of the adhesion parameter, colliding particles tend to
separate from each other, and the primary role of the adhesive force is to reduce the
particle rebound velocity.

1.3. Applications
There are a large number of applications in nearly all areas of engineering, biology,
agriculture, and environmental science involving flow of adhesive particles. In this
section, five application areas are briefly described that illustrate different aspects
and interesting physics occurring in adhesive particle flows.

1.3.1. Fibrous Filtration Processes

Filtration is one of the primary processes used in many chemical and materials purifi-
cation applications. Filtration also plays a critical role in environmental processes
used for production of clean air and drinking water. Filtration is vital for clean room
operations used in electronics manufacturing, as well as in microfabrication and
nanofabrication industries. Filtration processes involve a mechanical separation of
particles from a fluid stream by collision and adhesion of particles onto the filtration
media. There are many types of filtration media, varying from a granular bed (such
as soil) to paper or cloth. All filtration media share the characteristic that the fluid
flow is forced to meander back and forth in a tortuous path as it passes through the
medium. Filtration differs from sieving in that a sieving process involves capture of
particles that are typically larger than the sieve hole size, whereas in filtration the
particles are much smaller than the channel or gap size in the filtration medium. The
example discussed in the current section deals with fibrous filters, which consist of a
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

16 Introduction

P1
10μm Figure 1.5. Photographs showing (a) fly ash particles
(a) recovered from the exhaust stream of a combustion pro-
cess and (b) a time series showing adhesion of the fly ash
particles onto a fiber, with flow orthogonal to the fiber.
[Reprinted with permission from Huang et al. (2006).]

P1

1 2 3 4 5 6 7 8
P1 Flow Direction 20μm
(b)

mesh of thin cylindrical structures (fibers) joined together to form a structure similar
to a cloth or porous paper.
A study of the micromechanics of fibrous filtration processes is reported by
Huang et al. (2006), with a subsequent computational study by Li and Marshall
(2007). Huang et al. (2006) used fly ash particles recovered from a combustion
process exhaust stream (Figure 1.5a), which were forced to flow orthogonally to a
single fiber. A time series of images showing capture of the particles by the fiber is
given in Figure 1.5b, starting with a relatively clean fiber (on the left) and ending
with a fiber that is nearly saturated with captured particles (on the right).
Particles entrained in the flow collide with the fiber by a combination of two
mechanisms, illustrated in Figure 1.6. The first mechanism involves the inertia of
the particles. Particles in a gas flow are much denser than the surrounding gas, and
tend to have Stokes number values near or above unity, thus allowing significant
inertial drift of the particles. As the fluid streamlines bend near the stagnation point
at the front of the fiber, the greater inertia of the particles carries them forward and
causes the particles to collide with the front surface of the fiber (Figure 1.6a). The
second mechanism involves the finite size of the particles in a shear flow, and can
allow for particle collision with the fiber even for cases of neutrally buoyant particles
in a liquid. Shear-induced collision becomes important as the distance between
streamlines gradually narrows as the fluid accelerates to move around the fiber
(Figure 1.6b). Particles with centroids positioned on streamlines lying very close to
the fiber surface can collide with the surface due to the finite particle size as the fluid
carries the particles close to the fiber surface. This type of collision tends to deposit
particles along the sides of the fiber. Particles can also collide with other particles
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.3 Applications 17

fluid streamline
4 collision
1 2 3
fiber
particle path

(a)
Figure 1.6. Schematic illustrating (a) inertia-induced col-
lision and (b) shear-induced collision of a particle onto a 6
cylindrical fiber. 5
fluid streamline
4
1 2 3 collision

fiber

(b)

that are already attached to the fiber, forming dendritic structures that reach out
into the flow field, as seen in the time series in Figure 1.5b.
The primary adhesion force involved in fibrous filters is the van der Waals
attraction force. The van der Waals force can be very strong, but it decays quickly with
distance away from the particle surface. For instance, the adhesive force between
two planar surfaces separated by a distance h can be written as
A B
F =− + 9, (1.3.1)
6π h3 h
where A and B are positive constants. The first term in this expression is the attractive
van der Waals force, and the second term is the short-range repulsive force. Typically,
the van der Waals force between two surfaces is small if the separation distance
between the surfaces is more than about δvdw ∼ = 10 nm. Particles used for filtration
processes, such as smoke or dust particles, have typical diameters d of 1–100 µm, or
between 2 and 4 orders of magnitude larger than the length scale of the adhesive
force. This observation has a number of consequences in modeling of the particle
adhesion process. First, it should be obvious that van der Waals adhesion is significant
only for colliding particles, as particles that do not collide are typically too far apart
for the van der Waals force to be significant. Second, the fact that δvdw /d is so small
implies that the particle adhesive force is highly sensitive to small amounts of particle
deformation. To illustrate this second point, consider a case where two colliding
particles are idealized as perfect spheres of equal diameter d that are touching at a
single point (the contact point). Elementary geometry indicates that the separation
distance between the particle surfaceswill exceed δvdw within a distance from the
contact point equal to approximately δvdw d, so only within this small region near
 adhesive force significant. For instance, for a 10 µm diameter
the contact point is the
particle, the value of δvdw d is about 0.3 µm, assuming δvdw ∼10 nm. Within this small
region around the contact point, the particle surfaces are pulled strongly toward each
other by the van der Waals force, whereas outside of this region there is little force
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

18 Introduction

particle contact
region Figure 1.7. Flattened contact region that forms near the
contact point of two otherwise spherical particles, show-
fluid δ ing a separation distance δ between the particle surfaces
within the contact region.
particle

between the particle surfaces. The attractive pull of the surfaces toward each other
within this region is balanced at very small separation distances by the short-range
repulsive force, so that the particle surfaces tend to adopt an equilibrium separation
distance δ at which these two forces are in balance. As a result of this combination
of a strong short-range force acting on the surfaces of much larger particles, the
particles tend to elastically deform within a small contact region surrounding the
contact point, forming a flattened region in which the particle surfaces are separated
by an approximately constant distance δ (Figure 1.7). In some models for contact
of adhesive particles, the van der Waals force is considered to act only within this
contact region.
Particles captured by a fiber attach onto other particles captured by the fiber
to form chain-like dendritic structures that project outward into the flow field (Tien
et al. 1977). As these fibers grow, they experience increased bending force due to
the increased fluid force acting on the particles within the dendrite. Eventually,
the dendrites become so long that the adhesion force between the particles can
no longer sustain the required bending force of the structure, and the dendritic
structures either break or to bend backwards onto the fiber (Figure 1.8). The latter
process often disturbs other dendritic structures, leading to breakage of structures
and release of particles downstream in a domino effect. These breaking and bending
processes lead over time to an equilibrium state, in which the rate of new particles
captured by the fiber is equal to the rate at which particles are lost from the fiber by
breakage of the dendritic structures.

1.3.2. Extraterrestrial Dust Fouling

One of the most severe challenges of performing manned or robotic operations on the
Moon or on planets such as Mars is the presence of fine dust. Neither the Moon nor

Breakage

Figure 1.8. Schematic showing attachment of particle

chains onto a fiber, and two mechanisms for chain failure
leading to an equilibrium state of captured particles.

Rotating
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.3 Applications 19

(a) (b)
Figure 1.9. Photographs showing (a) an astronaut in a dusty space suit during a moon walk
and (b) dust on astronaut Eugene Cernan’s suit inside the lunar capsule following a moon
walk during the Apollo 17 mission.

Mars has liquid water, so instead of forming clays, fine dust particles remain loose in
the soil. On the Moon, these dust particles become electrically charged by the action
of the solar wind, whereas on Mars the dust particles become electrically charged
primarily through triboelectric effects that result from the frequent planetary dust
storms. When viewed under a microscope, the dust particles are found to be very
sharp and irregular in shape.
The presence of very fine, charged dust particles leads to a wide range of prob-
lems in extraterrestrial exploration missions. For instance, during the Apollo mis-
sions to the moon in the 1960s and early 1970s, dust was found to quickly adhere onto
the astronauts’ space suits during moon walks, and was then carried back into the
astronauts’ living space when they returned to the lunar module (Figure 1.9). Once
tracked into the habitat, the fine dust particles formed an aerosol that produced
respiratory problems when breathed by the astronauts, even after only a few days’
residence. Being charged, the dust particles cling to everything within the habitat,
from electrical instrumentation to air recycling units. For longer space missions, the
dust fouling would, over time, be expected to cause instrument failure and severe
health problems for the astronauts unless adequate precautions are taken. An indi-
cation of the significance of dust fouling was given during astronaut Eugene Carnan’s
technical briefing following the Apollo 17 mission, in which he states: “I think dust is
probably one of our greatest inhibitors to a nominal operation on the Moon. I think
we can overcome other physiological or physical or mechanical problems except
dust.”
Dust fouling has also been a critical factor limiting operation of the rover mis-
sions to Mars. In this case the primary issue has been decreased efficiency of the
solar panels that provide energy to the rover due to covering of the panels by a dust
coating (Tanabe, 2008). Dust storms have been a particular threat to the rovers,
leading to significant deposition of dust layers on the solar panels. On the other
hand, dust devils and high wind events associated with dust storms have also helped
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

20 Introduction

electrodes particles
insulation
dielectric
phase 0 π/2 π 3π/2

(a) (b)
Figure 1.10. (a) Cross-section showing a 4-phase electric curtain design with plate-like elec-
trodes and (b) a sample of the transparent “dust shield” under development at NASA
Kennedy Space Center for dust mitigation on solar panels.

to remove dust from the rovers’ panels, resulting in large variations in system power
over time.
The characteristic feature of lunar and Martian dust that makes it so difficult to
manage is the fact that the dust particles are electrically charged. Applications with
charged dust particles in flowing fluids are also common on earth, for instance, in
electrostatic precipitators used to remove ash from the exhaust streams of coal power
plants and in control of particle mixing, separation, and transport in microfluidic sys-
tems (Markarian et al., 2003). Electrostatic interactions influence particle transport
both by the dipole field induced on the particle by an external electric field, leading to
dielectrophoretic (DEP) force on the particle, and by the monopole field associated
with charged particles, leading to a Coulomb force on the particle in the presence of
an external electric field. The Coulomb force between two charged particles decays
in proportion to distance squared, which is slow compared to the decay rate of the
van der Waals force. For this reason, charged particles have significant interaction
via their mutual electric fields over distances that are large compared to the particle
radius.
A wide variety of approaches have been proposed for dust mitigation in plane-
tary exploration on the Moon and Mars. One of the more interesting and potentially
effective approaches is the use of an electric curtain device (Calle et al., 2009). An
electric curtain consists of a series of many parallel electrodes embedded on a dielec-
tric surface and covered by a thin layer of insulation (Figure 1.10). Each electrode has
a high-voltage oscillating potential, such that the phase of the potential oscillation
differs by an amount 2π /n between neighboring electrodes, where n is the number
of electrodes in each group before the phase repeats. A standing wave is obtained
with n = 2, and all larger values of n correspond to traveling electrostatic waves
on the curtain. The traveling electrostatic waves can carry particles with them in a
number of different modes, including a mode where particles are levitated above the
surface, a mode where particles intermittently hop along the surface, and a mode
where particles roll back and forth on the surface. Particle interaction and adhe-
sion to the dielectric surface modifies these modes, leading to emergent behavior of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.3 Applications 21

Figure 1.11. Photographs showing streams of 130 µm copper

spheres both in a dry state (left) and mixed with a small amount
of oil (right) emitted from a small opening in the bottom of a Increasing
hopper. [Reprinted with permission from Royer et al. (2009).] cohesion

the particles in the system that can be significantly different than the behavior of
individual particles (Liu and Marshall, 2010a,b).

1.3.3. Wet Granular Material

Granular materials such as soil, rocks, snow, and volcanic debris are of obvious
importance in a large number of environmental and geophysical processes, as well
as in biological processes involving motion of organisms that live in the soil. Granular
mechanics also plays an important role in engineering processes in the pharmaceu-
tical, chemical, and materials processing industries. Agricultural grain sorting and
materials handling are dominated by granular flow mechanics. Although most of
the literature on mechanics of granular materials involves cohesionless particles,
in many practical applications the presence of moisture within the granular matrix
leads to particle adhesive force that plays an important role in the overall system
mechanics.
It was found that the addition of small amounts of a liquid, even with a vol-
ume fraction of 10–4  10–2 , into a dry granular system can lead to strong cohesion
force between particles with significant changes of the macroscopic system behavior
(Fiscina et al., 2010; Samadani and Kudrolli, 2000). For instance, Figure 1.11 shows a
stream of dry copper grains of diameter d = 130 ± 30 µm falling out of a nozzle with
a diameter of 4 mm (left). As a small amount (3 × 10–4 by volume) of mineral oil
is mixed with the copper grains, the uniform stream of dry granular material breaks
up into particle clusters. The break-up process appears to be qualitatively similar
to the capillary break-up of a liquid jet into droplets (Royer et al., 2009). Interpar-
ticle cohesive forces were measured in this study by recording force-displacement
curves of individual grains brought into contact and pulled apart using an atomic
force microscope. The histogram shows significantly larger pull-off forces for the
oil-wetted copper particles than for the dry particles.
The larger pull-off force of wet particles compared with dry particles is caused
by the interaction of thin liquid films that form around each particle in the presence
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

22 Introduction

Liquid Bridge

(a) (b)
Figure 1.12. Liquid bridge formed between two spherical particles: (a) photograph of liquid
bridge on particles pulled apart in normal direction and (b) schematic of liquid film capillary
deformation on a rolling sphere. [Figure (a) reprinted with permission from Soulié et al.
(2006). Figure (b) reprinted with permission from Schade and Marshall (2011).]

of a small amount of moisture. When particles move sufficiently close to each other
for a collision to occur, the liquid films of the colliding particles touch and merge
into one another. As the two particles are pulled away from each other following
collision, the capillary force between the liquid films forms a connection between
the particles that is known as a liquid bridge. An example of a liquid bridge is
shown in Figure 1.12a for two particles of different sizes pulled apart along a line
connecting the particle centroids. A liquid bridge exerts a force between the particles
due to both the interfacial surface tension at the liquid-gas-solid contact line and
the pressure decrease within the liquid that develops from interfacial curvature of
the liquid bridge. When the particles move relative to each other, viscous effects due
to liquid motion within the film introduce an additional viscous force that must be
added to the attractive capillary force between the particles (Ennis et al., 1990). The
result of this liquid bridging force is to cause particles to adhere to container walls
and to each other, forming particle agglomerates that resist gravitational and other
forces, such as fluid shear, that try to tear them apart.
A less-studied consequence of the presence of liquid films around particles in a
wet granular flow is the effect of film asymmetry on the rolling motion of particles
along container walls or during particle collisions with each other. Rolling is a critical
form of particle interaction, which is typically far more likely to occur for systems
of small particles than are sliding or twisting motions due to the lower threshold for
onset of rolling than for these other motions. It is well known that van der Waals
adhesion causes a resistance to particle rolling (Dominik and Tielens, 1995), which
can have important consequences on the motion of particulate systems. However, it
was not realized until a recent paper by Bico et al. (2009) that liquid films can also
cause a resistance to rolling motion of particles in a wet granular flow. As shown
in a study by Schade and Marshall (2011), rolling of a particle causes the liquid
bridge connecting the particle to a flat surface to become anisotropic, such that the
liquid-gas interface contact line moves inward on one side and outward on the other
side of the particle (Figure 1.12b). The net result of the contact line movement is to
induce a capillary torque on the particle in such a direction as to resist the rolling
motion.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.3 Applications 23

(a) (b)
Figure 1.13. (a) Scanning electron microscope image showing a red blood cell (left), an acti-
vated platelet (middle), and a white blood cell (T-lymphocyte) (right). (b) Rouleaux formed
of red blood cells at low shear stress levels. [Figure (b) reprinted with permission of 2009
Rector and Visitors of the University of Virginia, Charles E. Hess, M.D., and Lindsey Krstic,
B.A.]

1.3.4. Blood Flow

Fluid flows consisting of suspensions of biological cells can often be considered
to behave as if they were particulate flows. This idealization is useful for model-
ing environmental flows with microorganisms, such as to determine the response of
microalgae cell suspensions to turbulent mixing (Marshall and Sala, 2011) or to study
instabilities that can occur in a fluid with a suspension of self-propelled microorgan-
isms (Pahlavan and Saintillan, 2011). In such applications, it is necessary to endow
the particles with additional properties to account for processes such as cell growth,
cell division, microorganism competition, and bacterial locomotion.
An area in which particulate flow models are particularly useful to represent sus-
pensions of biological cells occurs in flows within the human cardiovascular system
(Chesnutt and Marshall, 2009), in which the particulate matter consists of a com-
bination of red blood cells (RBCs, or erythrocytes), white blood cells (leukocytes),
and platelets immersed in a fluid called the plasma (Figure 1.13a). Together, blood
cells constitute about 45% of total blood volume. The most numerous type of blood
cells are red blood cells, which have the form of biconcave disks with thickness of
about 2 µm and a maximum diameter between 6 and 8 µm. RBCs number between
4 and 6 million cells per cubic millimeter of blood. A mature RBC has no nucleus,
but instead has the form of a capsule that is filled with hemoglobin, which holds the
oxygen that is carried through the body by the cardiovascular flow. In the very small
capillary vessels, measuring only 8–10 µm in diameter, the RBCs must deform by
large amounts to pass through narrow constrictions. On the other hand, in the larger
arterioles (approximately 80 µm diameter), and throughout the different small and
large arteries (300 µm to 2 cm diameter), very little RBC deformation is observed. At
low values of shear stress, RBCs adhere to each other to form chain-like agglomer-
ates called rouleaux (Figure 1.13b). Formation of rouleaux is not common in normal
physiological blood flow conditions, but they can form under various disease states
such as diabetes, infections, and cancer when blood protein levels are high, as well
as in low-shear wake regions behind obstructions in the circulatory system.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

24 Introduction

Red blood cells exhibit a tendency to collect near the central part of a blood
vessel, leaving a region near the walls with low RBC concentration. This process,
known as marginalization, leads to a state where the RBCs flow primarily within
the central portion of a blood vessel and other blood cells (such as white blood cells
and platelets) collect within the near-wall (“cell-free”) regions with low numbers of
RBCs (Aarts et al., 1988). Marginalization is important for regulating transport of
white blood cells to the endothelial cells along the vessel walls, and it has a significant
effect on the radial distribution of effective blood viscosity, which varies as a function
of local RBC concentration. Marginalization also plays an important role on the flow
of blood through branching junctions, which are a common occurrence in the car-
diovascular system. When the flow rate through one of the outlet branches of the
junction is substantially less than that through the other outlet branch, it is observed
that the branch with lower flow rate has significantly lower RBC concentration than
the outlet with the higher flow rate. This phenomenon, called plasma skimming,
arises because the lower flow rate outlet primarily draws fluid from the near-wall
cell-free layer whereas the high flow rate outlet draws fluid primarily from the RBC-
rich central region of the inlet vessel. Significant differences in RBC concentration
(“hematocrit”) develop within the body as a result of plasma skimming, and the
resulting variation in blood viscosity is known to provide a number of important
physiological benefits (Jonsson et al., 1992).
White blood cells (leukocytes) are produced in the bone marrow and are a
key element of the immune system, responsible for protecting the body from infec-
tious diseases. There are several different types of white blood cells with different
functions and different sizes, although typical white blood cells are approximately
spherical in shape with diameter ranging from 7 to 15 µm. The number of white
blood cells can vary greatly with disease state, but under normal conditions there
are approximately 7,000 white blood cells per cubic millimeter of blood, or about
1 white blood cell for every 600–700 red blood cells. White blood cells can adhere
to endothelial cells along the blood vessel wall, which is sometimes followed by
passage of the white blood cell into the surrounding tissue. This adhesion process is
influenced by rolling of the white blood cells along the wall and by collision of white
blood cells with passing red blood cells during the adhesion process (Melder et al.,
2000).
Platelets make up the final category of major types of blood cells. Platelets are
irregularly shaped cell fragments which are approximately 2–3 µm in diameter, and
number about 1 platelet for every 20 red blood cells. Unactivated platelets travel in
the blood waiting for a signal to become activated. Once activated, platelets change
form and become highly adhesive, gathering at the site of an injury to the blood
vessel and acting to initiate and regularize the formation of a blood clot to seal the
vessel. Excessively high platelet levels are associated with development of strokes
and heart attacks. The platelet activation process occurs through a complex series
of chemical reactions, which can be triggered by a number of factors. Among these
triggering factors is exposure to high shear stress, which can occur both naturally
and during closing operations of mechanical heart valves. The high fluid shear levels
observed on mechanical heart valves can lead to thrombus formation, which limits
useful life span of the valves (Bluestein et al., 2004).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

1.3 Applications 25

1.3.5. Aerosol Reaction Engineering

In recent years, there has been a surge in development of tailored nanoparti-
cles (including nanowires, nanotubes, and nanofilms) because of their wide range
of industrial applications, such as selective catalysts, chemical sensors, optical
films/fibers, additives for nanofluids, and fuel additives for energetic chemical propul-
sion. As tailored nanoparticles come into increased industrial usage, it is necessary
to place greater emphasis on the problem of how to synthesize these particles on an
industrial scale with high throughput under carefully controlled conditions. Aerosol
reactors are technological systems used for the large-scale synthesis of nanoparti-
cles of designed properties (Friedlander, 2000). Within aerosol reactors, precursors
are converted to monomers by processes such as oxidation and pyrolysis, produc-
ing a supersaturated vapor that undergoes nucleation to form primary particles.
These particles then grow via the collision-sintering mechanism to form aggregate
nanoparticles. Aerosol reactors employ a vapor-phase approach that is different
from conventional wet-chemistry methods for nanoparticle production, and offers a
number of distinct advantages. According to the mode of heat supply to the process,
aerosol reactors can be classified as various types, such as flame, laser, plasma, hot-
wall, and electric-heating. Key governing factors are aerosol precursor properties
and reactor process conditions (e.g., temperature).
More recently, another widely used vapor-phase method, vapor deposition
(either chemical or physical), has received increased attention for production of
thin films, in which materials in a vapor state are directly condensed to form solid-
phase materials on a temperature-controlled substrate (Chen and Mao, 2007). By
combining advantages of both vapor-phase flame aerosol reactors and deposition
techniques, a one-step stagnation flame aerosol synthesis was developed, in which
high-quality nanostructured TiO2 films can be produced at large growth rates in
a single step (Zhang et al., 2012). The synthesis process can be classified into two
zones, consisting of a predeposition gaseous zone and a postdeposition zone. In the
former, the sintering occurs between two colliding particles induced by either the
inertial or the shear collision processes. In the latter, the sintering happens between
two quasistatic contacting particles on the substrate.
The roles of these two distinct sintering mechanisms in determining the mor-
phologies of TiO2 films, which can be inferred from the specific surface area (SSA)
of the films, are illustrated in Figure 1.14. When the substrate temperature is low
enough, the on-substrate sintering is quite weak due to the low local temperature,
with particles “frozen” upon deposition. The SSA dramatically decreases as the
precursor concentration increases. The SSA of produced films is mainly depen-
dent on the in-flame collision-sintering process prior to deposition, which we term
the “in-flame-collision-sintering controlled” regime. In contrast, when the substrate
temperature is high, SSA dependence on precursor concentration is much weaker,
with SSA varying in a narrow range from 91.4 m2 /g to 129.7 m2 /g. Subplots A and
B of Figure 1.14 display TEM images of TiO2 films for precursor concentrations
of 116.4 ppm and 291.0 ppm, respectively, with the substrate at 763 K. The size of
primary particles in subplot B (291.0 ppm) is only slightly larger than that in sub-
plot A (116.4 ppm), although precursor concentration is about 2.5 times larger. It
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

26 Introduction

Substrate temperature 763K

Substrate temperature 383K A

Specific surface area (m2/g)

300
In-flame-collision-sintering
controlled
240
20 nm

180
On-substrate-sintering B
A controlled
120
B
0 100 200 300
Precursor concentration (ppm) 20 nm

Figure 1.14. Specific surface area (SSA) of TiO2 nanofilms for different precursor concentra-
tions using a stagnation flame aerosol reactor. Open circles represent the data for a substrate
temperature of 383 K, while the filled squares correspond to a substrate temperature of 763 K.
The TEM images on the right-hand side show the morphologies of the nanofilms at (A) 116.4
ppm and (B) 291.0 ppm, for a substrate temperature of 763 K.

is concluded that on-substrate sintering of primary particles after deposition greatly

reconstructs the TiO2 film morphology.

REFERENCES

Aarts PAMM, van den Broek SAT, Prins GW, Kuiken GDC, Sixma JJ, Heethaar RM. Blood
platelets are concentrated near the wall and red blood cells, in the center in flowing blood.
Arterioscler 8, 819–824 (1988).
Aranson IS, Tsimring LS. Patterns and collective behavior in granular media: theoretical
concepts. Review of Modern Physics 78, 614–692 (2006).
Bagnold RA. Experiments on a gravity-free dispersion of large solid spheres in a Newtonian
fluid under shear. Proceedings of the Royal Society of London A 225(1160), 49–63 (1954).
Balachandar S, Eaton JK. Turbulent dispersed multiphase flow. Annual Review of Fluid
Mechanics 42, 111–133 (2010).
Bell GI. Models for the specific adhesion of cells to cells. Science 200, 618–627 (1978).
Bertrand F, Leclaire LA, Levecque G. DEM-based models for the mixing of granular mate-
rials. Chemical Engineering Science 60, 2517–2531 (2005).
Bico J, Ashmore-Chakrabarty J, McKinley GH, Stone HA. Rolling stones: The motion of a
sphere down an inclined plane coated with a thin liquid film. Physics of Fluids 21, 082103
(2009).
Bluestein D, Yin W, Affeld K, Jesty J. Flow-induced platelet activation in mechanical heart
valves. Journal of Heart Valve Disease 13(3), 5001–5008 (2004).
Calle CI, Buhler CR, McFall JL, Snyder SJ. Particle removal by electrostatic and dielec-
trophoretic forces for dust control during lunar exploration missions. Journal of Electro-
statics 67, 89–92 (2009).
Chen XB, Mao SS. Titanium dioxide nanomaterials: Synthesis, properties, modifications, and
applications. Chemical Reviews 107, 2891–2959 (2007).
Chesnutt JKW, Marshall JS. Blood cell transport and aggregation using discrete ellipsoidal
particles. Computers & Fluids 38, 1782–1794 (2009).
Crowe CT, Schwarzkopf JD, Sommerfeld M, Tsuji Y. Multiphase Flows with Droplets and
Particles, 2nd ed., CRC Press, Boca Raton, Florida (2012).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

References 27

Cundall PA, Strack ODL. A discrete numerical model for granular assemblies. Geotechnique
29, 47–65 (1979).
Dominik C, Tielens AGGM. Resistance to rolling in the adhesive contact of two elastic
spheres. Philosophical Magazine A 72(3), 783–803 (1995).
Dullien FAL. Porous Media: Fluid Transport and Pore Structure, 2nd ed., Academic Press,
San Diego (1992).
Einstein A. A new determination of the molecular dimensions. Annalen der Physik 19, 289–
306 (1906).
Elghobashi S, Truesdell GC. On the two-way interaction between homogeneous turbulence
and dispersed solid particles. I. Turbulence modification. Physics of Fluids A 5, 1790–1801
(1993).
Ennis BJ, Li J, Tardos GI, Pfeffer R. The influence of viscosity on the strength of an axially
strained pendular liquid bridge. Chemical Engineering Science 45(10), 3071–3088 (1990).
Epstein PS, Carhart RR. The absorption of sound in suspensions and emulsions. Journal of
the Acoustical Society of America 25(3), 553–565 (1953).
Feng JQ, Hays DA. Relative importance of electrostatic forces on powder particles. Powder
Technology 135, 65–75 (2003).
Fiscina JE, Lumay G, Ludewig F, Vandewalle N. Compaction dynamics of wet granular
assemblies. Physical Review Letters 105, 048001 (2010).
Friedlander SK. Smoke, Dust and Haze: Fundamentals of Aerosol Dynamics, Oxford Univer-
sity Press, New York (2000).
Gady B, Schleef D, Reifenberger R. Identification of electrostatic and van der Waals inter-
action forces between a micrometer-size sphere and a flat substrate. Physical Review B 53,
8065–8070 (1996).
Gore RA, Crowe CT. Effect of particle size on modulating turbulence intensity. International
Journal of Multiphase Flow 15, 279–285 (1989).
Hetsroni G. Particles-turbulence interaction. International Journal of Multiphase Flow 15,
735–746 (1989).
Huang B, Yao Q, Li SQ, Zhao HL, Song Q, You CF. Experimental investigation on the
particle capture by a single fiber using microscopic image technique. Powder Technology
163, 125–133 (2006).
Israelachvili JN. Intermolecular and Surface Forces. 3rd ed., Academic Press, Boston (2011).
Jafari S, Salmanzadeh M, Rahnama M, Ahmadi G. Investigation of particle dispersion and
deposition in a channel with a square cylinder using the lattice Boltzmann method. Journal
of Aerosol Science 41, 198–206 (2010).
Jonsson V, Bock JE, Nielsen JB. Significance of plasma skimming and plasma volume expan-
sion. Journal of Applied Physiology 72(6), 2047–2051 (1992).
Kaftori D, Hetsroni G, Banerjee S. Particle behavior in the turbulent boundary layer. 1.
Motion, deposition, and entrainment. Physics of Fluids 7(5), 1095–1106 (1995).
Kinloch AJ. Adhesion and Adhesives, Chapman and Hall, London (1987).
Lashkov VA. Drag of a cylinder in a two-phase flow. Fluid Dynamics 27(1), 93–97 (1992).
Li SQ, Marshall JS. Discrete-element simulation of micro-particle deposition on a cylindrical
fiber in an array. Journal of Aerosol Science 38, 1031–1046 (2007).
Li SQ, Marshall JS, Liu G, Yao Q. Adhesive particulate flow: The discrete element method
and its application in energy and environmental engineering. Progress in Energy and
Combustion Science 37, 633–668 (2011).
Liu G, Marshall JS. Effect of particle adhesion and interactions on motion by traveling waves
on an electric curtain. Journal of Electrostatics 68, 179–189 (2010a).
Liu G, Marshall JS. Particle transport by standing waves on an electric curtain. Journal of
Electrostatics 68, 289–298 (2010b).
Markarian JAP, Markarian N, Yeksel M, Khusid B, Farmer KR, Acrivos A. Particle motions
and segregation in dielectrophoretic microfluidics. Journal of Applied Physics 15, 4160–4169
(2003).
Marshall JS. Inviscid Incompressible Flow. John Wiley and Sons, New York (2001).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-01 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:43

28 Introduction

Marshall JS, Chesnutt JKW, Udaykumar HS. Mesoscale analysis of blood flow. In Image-based
Computational Modeling of the Human Circulatory and Pulmonary Systems, KB Chandran,
HS Udaykumar, J Reinhardt, editors, Springer Publications, New York, Chapter 6, pp. 235–
266 (2011).
Marshall JS, Sala K. A stochastic Lagrangian approach for simulating the effect of turbulent
mixing on algae growth rate in photobioreactors. Chemical Engineering Science 66, 384–392
(2011).
Melder RJ, Yuan J, Munn LL, Jain RK. Erythrocytes enhance lymphocyte rolling and arrest
in vivo. Microvascular Research 59, 316–322 (2000).
Nichols G, Byard S, Bloxham MJ, Botterill J, Dawson NJ, Dennis A, Diart V, North NC,
Sherwood JD. A review of the terms agglomerate and aggregate with a recommendation
for nomenclature used in powder and particle characterization. Journal of Pharmaceutical
Sciences 91(10), 2103–2109 (2002).
Pahlavan AA, Saintillan D. Instability regimes in flowing suspensions of swimming micro-
organisms. Physics of Fluids 23, 011901 (2011).
Reinhold A, Briesen H. Numerical behavior of a multiscale aggregation model – coupling
population balances and discrete element models. Chemical Engineering Science 70, 165–
175 (2012).
Renno NO, Wong A-S, Altreya SK. Electrical discharges and broadband radio emission by
Martian dust devils and dust storms. Geophysical Research Letters 30(22), 2140 (2003).
Royer JR, Evans DJ, Oyarte L, Guo Q, Kapit E, Mobius ME, Waitukaitis SR, Jaeger HM.
High-speed tracking of rupture and clustering in freely falling granular streams. Nature 459,
1110–1113 (2009).
Samadani A, Kudrolli A. Segregation transitions in wet granular matter. Physical Review
Letters 85, 5102–5105 (2000).
Schade P, Marshall JS. Capillary effects on a particle rolling on a plane surface in the presence
of a thin liquid film. Experiments in Fluids 51(6), 1645–1655 (2011).
Soulié F, Cherblanc F, El Youssoufi MS, Saix C. Influence of liquid bridges on the mechan-
ical behavior of polydisperse granular materials. International Journal for Numerical and
Analytical Methods in Geomechanics 30, 213–228 (2006).
Tanabe K. Modeling of airborne dust accumulation on solar cells at the Martian surface. Acta
Astronautica 62, 683–685 (2008).
Tien C, Wang, CS, Barot DT. Chainlike formation of particle deposits in fluid-particle sepa-
ration. Science 196, 983–985 (1977).
Zhang YY, Li SQ, Yan W, Yao Q, Tse SD. Role of dipole-dipole interaction on enhancing
Brownian coagulation of charge-neutral nanoparticles in the free molecular regime. Journal
of Chemical Physics 134, 084501(2011).
Zhang YY, Li SQ, Yan W, Yao Q, Tse SD. Direct synthesis of nanostructured TiO2 films with
controlled morphologies by stagnation swirl flames. Journal of Aerosol Science 44, 71–82
(2012).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2 Modeling Viewpoints and Approaches

2.1. A Question of Scale

Particulate flow problems are inherently multiscale. The need for a multiscale
approach is evident simply by estimating particle numbers in most applications,
such as in the adhesive particle applications described in Chapter 1. For instance,
a dust layer with a thickness of only one 10 µm diameter particle requires on the
order of 1010 particles to cover a solar panel with surface area of 1 m2 . As seen in
Figure 1.5a, fly ash particles measure about 2 µm in diameter. A coating one fly ash
particle thick covering a single 20 µm diameter fiber measuring 10 cm long would
contain about 1.5 million particles. A single cross-knit filter sheet measuring 10 cm
on each side, with a 50% void ratio, would contain approximately 7.5 × 109 particles.
There are on average about 5 million red blood cells per cubic millimeter of whole
blood. The human body contains about 5.6 liters of blood, or about 2.8 × 1013 red
blood cells. There are fewer platelets and white blood cells in the body, number-
ing about 1.4 × 1012 and 4 × 1010 , respectively. The number of particles involved in
wet granular flows varies widely depending on the specific application. However,
the common comparison of a value that is so large as to be essentially uncountable
as comparable to the “number of grains of sand in all the beaches of the world”
gives an indication that even wet granular flows often involve very large numbers of
particles. Somewhat surprisingly, a number of people have tried to determine even
this uncountable value, perhaps the most widely cited being a recent estimate by a
researcher at the University of Hawaii that there are 7.5 × 1018 grains of sand on all
the world’s beaches.1 On a more human scale, it was recently estimated that there
are approximately 5 × 1011 grains of sand on a regulation beach volleyball court.2
A full simulation of a multiphase flow would entail not only tracking such huge
numbers of particles, but also solving for the fluid flow field in the interparticle
region surrounding each particle. For particles that collide with each other, or that
reside in clusters or agglomerates, the fluid flow about even a single particle can
itself require a multiscale approach, particularly when one accounts for the large
difference in scale between the particle diameter and the small gap size within the
contact region separating colliding particles. Because of this large disparity in scales
1 http://www.hawaii.edu/suremath/jsand.html.
2 http://www.6footsix.com/my weblog/how-many-grains-of-sand-i.html.

29
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

30 Modeling Viewpoints and Approaches

and the large number of particles involved in most multiphase flow problems, a
wide range of different modeling approaches have been developed for simulation
of particulate fluids. Each of these modeling approaches seeks to describe a specific
scale of particulate flow, although there is a great deal of overlap between the
capabilities of the different approaches.
In this chapter, we simplify the various approaches for simulation of adhesive
particle flows as dealing with one of three basic scales of modeling. At the macroscale,
the simulation model does not follow individual particles, but instead invokes some
type of approximation to represent groups of particles by a single computational
element. There are various different approximations that are used for this purpose,
which are described in the second section of this chapter. At the mesoscale, the
individual particles are followed by the simulation method, but various analytical
and numerical approximations are invoked to accelerate the computation. The most
common of these approximations is to introduce analytical models for the different
fluid forces acting on the particles. Again, a wide range of computational approaches
fit into this general class of mesoscale models, a summary of which is given in the
third section of this chapter.
At the microscale, we attempt to fully model both the motion of individual
particles and the fluid flow about each particle. This involves not merely placing
grid points in the region between the particles, but properly demonstrating that the
solutions are independent of grid resolution and time step size used. Interparticle
flow fields in general are very difficult to accurately simulate, due to the moving
particle boundaries and the interaction of particles with the wakes of other particles.
Interparticle flows for adhesive particles are especially difficult to simulate due to
the wide variation in length scale between that associated with the average particle
spacing and the very small length scale associated with the gap separating the contact
surfaces of colliding particles. Adhesive particles are continuously in contact with
each other, with contacts forming and breaking in an ongoing process. Microscale
simulations are often used for describing flows in microscale geometries, such as a
red blood cell in a capillary vessel or a bubble in a microchannel, for which other
modeling approaches are not appropriate. Microscale simulations are also used
to understand particle interactions within small regions of a larger fluid flow, for
instance, in order to develop and validate models for particle interaction to be used
in mesoscale or macroscale simulation approaches.

2.2. Macroscale Particle Methods

2.2.1. Discrete Parcel Method

A common approach for modeling particulate flows with large numbers of particles
is to approximate groups, or parcels, of particles by a single representative particle
with some average velocity v (Figure 2.1). This method has been known by a variety
of names in the literature, including Particle-Source-in-Cell (Crowe et al., 1977) and
Multiphase-Particle-in-Cell (Andrews and O’Rouke, 1996). In models of plasmas
these particle parcels are sometimes referred to as super-particles (Dawson, 1983).
Methods in this general category were termed discrete parcel methods (DPMs) by
Crowe et al. (2012), and we have retained this terminology. In a discrete parcel
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.2. Macroscale Particle Methods 31

Figure 2.1. Schematic diagram illustrating the parti-

cle cloud method, which replaces a group of parti- R
cles by a single representative particle with average
velocity v, together with a cloud of virtual particles
v
having cloud radius R.

method, each representative particle carries along with it a group of NC virtual parti-
cles, where the characteristic length scale of the group is set equal to the group radius
R. The representative particles are transported in the same manner as any individ-
ual particle, by solution of the particle momentum equation balancing the rate of
particle momentum with fluid forces acting on the particle. Because the number
of particles per parcel can have any value, the discrete parcel method is capable
of simulating transport of practically unlimited numbers of particles, subject only
to resolution limitations. This is a common simulation approach for flows of dilute
particles, in which it is assumed that the particles are far enough apart so as not to
interact with each other.
A variety of extensions of the discrete parcel method have been proposed.
For instance, to represent the effect of turbulence on particle dispersion, Zhou
and Yao (1992) proposed increasing the parcel radius R as a function of time.
Sommerfeld (2001) proposed a stochastic method to account for effects of particle
collisions on the motion of the representative particles in the discrete parcel method,
in which a random fluctuation velocity that mimics the effects of collisions with other
(nonresolved, or fictitious) particles is added to the velocity of each representative
particle at each time step. In dense particle flows, multiple particle collisions can
result in transmission of stress through the two-phase medium. To account for this
effect within a discrete parcel method, Andrews and O’Rourke (1996) inserted
an additional term in the particle momentum equation that is proportional to the
gradient of the “solids stress,” given by −∇τ /ρ p φ, where τ = Ps φ β /(φcp − φ) is a
function of the particle concentration φ, φcp is the concentration in a close-packed
state, and Ps and β are empirical constants (see also Harris and Crighton, 1994).
Typical values proposed by Snider (2007) for these coefficients are Ps = 100 pa,
β = 2, and φcp = 0.7.
A second type of discrete parcel method for dense particle flows was proposed by
Patankar and Joseph (2001). This paper treated parcels as if there are large particles,
so that parcel-parcel collisions are modeled similarly to particle-particle collisions in
a discrete element model, with the difference that the parcels are larger and may have
different elastic and dissipation parameters than the constituent particles. Validation
studies for discrete parcel methods were reported by Snider (2007) and Benyahia and
Galvin (2010). Snider reported good agreement between the Andrews-O’Rourke
DPM model and experimental results for several different granular flow problems.
Benyahia and Galvin similarly reported good agreement for the concentration field
for both the Andrews-O’Rourke and the Patankar-Joseph DPM formulations for
several different validation problems in comparison to results from full discrete
element simulations. However, this paper noted significant differences in other flow
quantities, such as granular temperature and particle mean velocity fields. Some of
these differences can be mitigated by assigning parcels different properties than the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

32 Modeling Viewpoints and Approaches

Figure 2.2. Schematic diagram illustrating the popu-

lation balance method, which replaces each particle
agglomerate by a single representative particle with
diameter larger than that of the actual particles.

constituent particles, for instance, for restitution coefficient in the Patankar-Joseph

approach. A significant amount of research needs to be performed on validation and
extension of DPM under different flow conditions. DPM has not been frequently
used for flows with adhesive particles, due in part to the difficulty of accounting for
adhesion of particles belonging to different (and possibly overlapping) parcels.

2.2.2. Population Balance Method

The population balance method (PBM) is based on the idea of replacing an agglom-
erate of particles by a single effective particle with volume set equal to the sum
of the volumes of the constituent particles making up the agglomerate, as shown
schematically in Figure 2.2. Collision and adhesion of two agglomerates results in
the elimination of the effective particles representing the original two agglomerates
and the formation of a new effective particle with volume equal to the sum of that of
the original two agglomerates. The elimination and formation of effective particles
in a population balance approach may also depend on processes such as particle
growth, nucleation, and agglomerate break-up. If we consider the particle collision
and agglomeration process alone, the governing equation for the rate of change of
concentration ni of a particle with volume Vi due to coagulation with other particles
is given by a discrete form of the Smoluchowski equation (Smoluchowski, 1917)
as

  ∞
dnk 1  
= β(Vi , V j )ni n j − nk β(Vi , Vk )ni , k = 1, 2, . . . , (2.2.1)
dt coag 2
i+ j=k i=1

where β(Vi , V j ) represents a measure of collision/agglomeration frequency called

the coalescence kernel. The first term on the right-hand side in (2.2.1) represents the
increase of particles of size k due to collision and agglomeration of smaller particles
whose volumes sum up to Vk . The factor of 1/2 is used to avoid double-counting
when the same particle pair is considered with i and j interchanged. The second term
on the right-hand side of (2.2.1) represents the decrease in concentration of particles
of size k due to collision of these particles with other particles.
Summaries of various coalescence kernels in free-molecular, continuum, and
turbulent regimes, as well as under the effects of various external forces, are given by
Friedlander (2000) and Cameron et al. (2005). Extensions to the population balance
equation to account agglomerate breakage were reviewed by Ramkrishna (2000).
Particle growth in aerosol suspensions by gas-to-particle conversion was considered
in the context of population balance theory by Friedlander (2000) and McCoy (2002),
which is important for crystallization, granulation, and aerosol-reaction processes.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.2. Macroscale Particle Methods 33

The effect of convective and diffusive processes on the rate of change of nk is

governed by the advection-diffusion equation, given by
 
∂nk dnk
+ ∇ · (nk v) = ∇ · (D∇nk ) + − ∇ · (nk vt ). (2.2.2)
∂t dt coag

In this equation, v is the particle velocity, D is a diffusion coefficient, and vt is a

particle terminal velocity resulting from an external force field (e.g., gravity, electric
fields, thermophoresis, etc.). All of these terms depend on the PBM particle size. The
term in brackets on the right-hand side of (2.2.2) represents the rate of change of nk
due to particle agglomeration, which is given by (2.2.1). If mass transfer processes
are present that allow the particles to grow or evaporate with time, an additional
term must be added to the right-hand side of (2.2.1) to incorporate these effects.
The equation (2.2.2) is often called the general dynamic equation (GDE) for aerosol
flows.
A variety of different approaches have been used in the literature for solving
the GDE. Perhaps the most efficient approach is to a priori assume a particle size
distribution (e.g., log-normal) which is characterized by certain coefficients, and then
to use (2.2.2) to estimate the values of these coefficients as functions of time and
position. Alternatively, the size distribution might be characterized by moments of
the particle number density of the form
 ∞
μk ≡ r k g(r)dr, (2.2.3)
0

where r is the particle radius and g(r) is the size distribution function. The GDE
(2.2.2) is used to solve for the time evolution of these moments (Frenklach and
Harris, 1987). The most accurate method to solve the GDE is the sectional approach,
in which the particle size distribution is divided up into sections, or bins, and the
GDE is used to solve for the number of particles in each bin.
The population balance method is very popular for applications with adhesive
particles because it provides for rapid simulation of systems with large numbers
of agglomerated particles. However, it should be emphasized that accuracy of the
method requires satisfaction of a number of restrictions. In particular, the tradi-
tional analytic models used for the coalescence kernel require low particle mass
fraction, spherical particles, binary collisions, single-size component particles, and
homogeneous agglomerate size distribution. Significant errors in the collision model
assumed for the standard coalescence kernel used in population balance models can
arise at small particle volumetric concentrations. For instance, as shown in Figure 2.3,
computations reported by Heine and Pratsinis (2007) and Trzeciak et al. (2006) using
the Langevin dynamics method indicate that the dilute-flow approximation to the
particle collision rate is in error by between 10% to 400% as the particle con-
centration increases from 0.001 to 0.1, respectively. A number of researchers have
recently utilized variations of the discrete element method in efforts to motivate
new coalescence kernel expressions that can be used in situations with larger par-
ticle concentrations and with heterogeneous flows (Freireich et al., 2011; Reinhold
and Briesen, 2012; Cameron et al., 2005; Gantt et al., 2006).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

34 Modeling Viewpoints and Approaches

Collision frequency ratio, βLD / βdilute 40

Figure 2.3. Plot showing the ratio of

the computed collision frequency to the
10 analytical estimate for dilute suspen-
sions typically used in population bal-
ance models as a function of the particle
concentration. Circles and the solid line
represent data from a Langevin dynam-
ics calculation and a best fit curve by
Heine and Pratsinis (2007), and triangles
represent numerical data by Trzeciak
et al. (2006). [Reprinted with permission
from Heine and Pratsinis (2007).]
1
0.1 1.0 10
Particle concentration, %

2.3. Mesoscale Particle Methods

Mesoscale methods for particulate flow range from methods designed for simulation
of molecular interactions to methods developed specifically for particulate systems.
The various methods invoke different approximations to speed up the computation,
which are valid for different ranges of particle sizes. In the current section, we review
four mesoscale models, each of which invokes approximations that are significantly
different from the others. The molecular dynamics (MD) method was developed for
simulating interactions of individual atoms having prescribed potential fields with
each other. In particulate flows, MD is often used for problems such as simulating
growth of nanoparticles or interaction of nanoparticles with the surrounding solvent
(Tian, 2008). Although MD is more properly considered to be a microscale method
from the point of view of particulate systems, we have included it in this section on
mesoscale modeling because the basic MD formulation is the starting point used
in many mesoscale modeling approaches. The dissipative particle dynamics (DPD)
method was developed to simulate groups of molecules, in a manner similar in spirit
to how the discrete parcel method simulates groups of particles (Section 2.2.1). Both
due to this coarse-graining approach and due to the fact that DPD employs soft
potentials, instead of the sharp potentials typical of MD, DPD can be employed
for significantly larger time and spatial scales than MD. The Brownian dynamics
(BD) method replaces individual interactions between small colloidal particles and
the molecules of the surrounding solvent with a stochastic forcing term, achieving
considerable speed-up over methods such as MD and DPD that attempt to explicitly
resolve particle-molecular interactions. Similar to MD, Brownian dynamics approx-
imates adhesive and collision forces between particles by an interaction potential.
The MD, DPD, and BD methods were all designed either for molecules or
for particles with diameters in the nanoscale range, such as for a colloidal particle
whose size is on the same order of magnitude (3–10 nm) as the length scales of the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.3. Mesoscale Particle Methods 35

m
Biomass
Sand
mm
Coal dust
Length scale

DEM Blood cells

μm Clay

DPD BD Fumed silica

MD
nm
1A Atoms

fs ps ns μs ms s

Time scale
Figure 2.4. Diagram showing the approximate time and length scales for particle flow sim-
ulation methods, including molecular dynamics (MD), dissipative particle dynamics (DPD),
Brownian dynamics (BD), and discrete-element method (DEM).

electrical double-layer and van der Waals forces that govern its adhesion processes.
By contrast, the discrete element method (DEM) was developed for particles with
diameters significantly larger than the length scales of adhesive forces, typically
applying to cases with particle diameters above about 100 nm. The key change in
DEM compared to these other methods is that particles in this size range can no
longer be approximated as spheres, because the small elastic deformation of the
particle surface that occurs when two particles collide can be of a similar order
of magnitude to the length scale of the adhesive force. In such cases, the particle
deformation greatly modifies the area on the particle surface over which significant
adhesive interaction occurs.
A schematic diagram illustrating the particle length and time scales for which
each of these models is generally used is given in Figure 2.4. We note, however, that
there is a great deal of variation in the literature on the names used for different
methods and the types of approximations imposed. Regardless of the specific type of
computational element examined, MD, DPD, BD, and DEM are all computer simu-
lation approaches in which the elements are allowed to interact for a period of time
in a Lagrangian framework under various imposed laws of motion. A listing of some
of the models that will be discussed in this chapter along with their computational
elements is given in Table 2.1.
MD was developed for atoms or molecular groupings with sizes from 0.1 nm to
more than 10 nm, with the computational time interval below about 1 ns. The time
and length scales of the coarse-graining DPD approach are about one to two orders
of magnitude larger than the respective scales for traditional MD. Larger time scales
(milliseconds) can be used for BD simulations with particles such as rigid protein
cells as compared with MD simulations of macromolecules (typically nanoseconds),
even for identical length scales. The DEM approach can treat particles ranging
from about 100 nm to several millimeters. The common computational time step of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

36 Modeling Viewpoints and Approaches

Table 2.1. Listing of different mesoscale and macroscale computational

methods, along with their acronym and the typical computational element used

Method Acronym Computational element

Molecular Dynamics MD Single atom

Brownian Dynamics BD Atom or particle
Dissipative Particle Dynamics DPD Cluster of atoms
Discrete Element Method DEM Single particle
Population Balance Method PBM Particle agglomerate
Discrete Parcel Method DPM Cloud of particles

DEM with micrometer-size particles is on the order of 1 µs or below. Because of

the large differences in time scales associated with processes such as collision and
fluid transport in particulate systems with micron-size particles, DEM calculations
tend to be numerically stiff. Consequently, many DEM simulations utilize a multiple
time scale computational structure to separately resolve processes that occur at very
different time scales (Marshall, 2009).

2.3.1. Molecular Dynamics

In traditional molecular dynamics simulations, each computational element rep-
resents an individual atom. Interaction of the atoms is governed by an assumed
potential field representing nonbonded forces between the atoms, including elec-
trostatic (Coulomb) force, attractive (van der Waals) force, and repulsive (steric)
forces, together with expressions for the bonded forces that occur between atoms in
a molecule. Electrostatic force is important for ions, where the number of electrons
and protons are unequal, giving the atom a net charge. Van der Waals force arises
from either permanent or induced atomic polarization, giving rise to electric dipole-
dipole interaction between nearby molecules. The steric repulsion acts on a scale
on the order of the atomic radius, and is caused by interaction between the electron
clouds of two interacting atoms.
A popular potential for nonbonded interactions is the Lennard-Jones potential,
in which the interaction energy U (r) is given by
 12  6
rp rp
U (r) = 4εLJ − , (2.3.1)
r r

where r is the distance between two atoms, εLJ is the interaction energy parameter,
and r p ≡ d/2 is the atomic “radius.” The first term in (2.3.1) represents the short-
range repulsive force and the second term is the attractive van der Waals force.
The electron dynamics can be ignored when developing these force potentials based
on the Born-Oppenheimer approximation, which states that electron dynamics is
sufficiently rapid that electrons can be assumed to respond instantaneously relative
to the time scale for atomic motion. The interaction force acting between the atoms
is F = −∇U , which is oriented along the line passing through the atom centers.
The Lennard-Jones potential is called a pair-wise potential because it deals with
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.3. Mesoscale Particle Methods 37

only two atoms at a time. There are a wide range of other potentials designed for
modeling atomic interactions in various situations, which account for effects such as
electrostatic interaction and multiple-atom interactions. With the development of
quantum mechanical simulation tools, such as the ab initio method, it has become
increasingly common to utilize smaller-scale quantum mechanics simulations to
motivate the potential field used in subsequent MD simulations.
The time and length scale restrictions of classical MD severely limit the method.
For instance, in order to capture the thermal motion of the atoms as well as the
stiffness imposed by the forces representing molecular bonds, the time step for
MD must be very small, on the order of 1 femtosecond (10–15 s). With current
generation computer systems, MD computations are limited to computations with
total time duration of about 1 ns and length scales of about 10 nm. Some investigators
utilize MD for larger-scale systems using a coarse-graining approach, in which the
computational particles represent an entire molecule, a group of molecules, or even
a nanoparticle (Tian, 2008). Of course, accuracy of the potential function in such
approaches becomes an increasingly important issue as the MD particle structure
becomes more complex.

2.3.2. Brownian Dynamics

Brownian motion refers to the seemingly random motion of a small particle observed
when it is suspended in a fluid. Among the earliest written accounts of this motion
is that of Roman philosopher Lucretius (c. 60 B.C.), who described the random,
tumbling motion of dust particles in air as proof of the existence of atoms. Botanist
Robert Brown provided a detailed account of the random motion of pollen grains
when viewed under the microscope in 1827 while studying the plant life of the South
Seas (Brown, 1828). Einstein (1905) and Smoluchowski (1906) provided among the
first mathematical theories for Brownian motion, the former using the measured
mean value of the second moment of particle displacement < x2 > as a means
of deducing Avogadro’s number. A stochastic differential equation governing the
velocity v of a small particle in the presence of Brownian motion, with no mean fluid
velocity field, was proposed by Langevin (1908) in the form

dv
m = −ζ v + FR (t ), (2.3.2)
dt

where ζ = 3π μd for the Stokes drag law and FR (t ) is a rapidly fluctuating force
that represents the force imposed on the particle from impact of molecules of the
surrounding fluid. Instead of resolving each individual molecular impact, the forc-
ing term FR (t ) in the Langevin equation is chosen to be a random variable with
Gaussian probability distribution having zero mean and variance 2ζ kb T , where kB is
Boltzmann’s constant (the ratio of the ideal gas constant to Avogadro’s number) and
T is the absolute temperature. For this choice of random forcing function, it follows
that the autocorrelation of the random forcing is given by

F R (t )F R (t  ) = 2ζ kB T δ(t − t  ), (2.3.3)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

38 Modeling Viewpoints and Approaches

where δ(·) is the Dirac delta. Over long time periods, it can be shown from the system
of equations presented here that the mean value of the second moment of particle
displacement increases linearly in time, or

< x2 > = 2kB Tt/ζ . (2.3.4)

Brownian dynamics (BD) was introduced as a mesoscale computational method

by Ermak and McCammon (1978) as a simplification of the more general Langevin
dynamics (LD) method. In LD, it is typical to add a particle interaction term to the
Langevin equation (2.3.2), giving
dv 
m = −ζ v + FCj (t ) + FR (t ). (2.3.5)
dt
j

The particle interaction force FCj is determined using a conservative potential, such
as the Lennard-Jones potential (2.3.1), and the sum in (2.3.5) is over particles that
of nearby to the given particle. Langevin dynamics is often applied for applications
where the computational elements represent small colloidal particles, sections of a
long-chain molecule, or large-size atoms. The small solvent atoms are not explicitly
resolved in this method, but instead the effect of these solvent atoms on the compu-
tational particle is modeled by the sum of the Stokes drag and random forcing terms
in (2.3.5). This simplification yields a considerable speed-up for computations with
LD compared with those with MD, for which every atom of both the solvent and the
particles are resolved.
Brownian Dynamics (BD) invokes the further simplification that the particle
inertia in (2.3.5) is negligible. In the absence of particle collisions, BD thus assumes
that the Brownian motion of each particle is balanced by hydrodynamic drag. Use of
the fluid continuum approximation, together with neglect of particle inertia, enables
BD to use significantly larger time steps than MD, as indicated in Figure 2.4. More
information on Brownian dynamics can be obtained from the review article by Chen
and Kim (2004).

2.3.3. Dissipative Particle Dynamics

Dissipative particle dynamics (DPD) was introduced by Hoogerbrugge and Koelman
(1992) as a type of coarse-grained MD, in which each DPD computational particle
represents a collection of many molecules. The DPD approach was later extended
to colloidal fluids (Dzwinel and Yuen, 2002), complex fluids such as blood flow
(Dzwinel et al., 2003; Filipovic et al., 2008; Pivkin et al., 2009), and polymeric fluids
(Symeonidis et al., 2005; Wijmans et al., 2001). Español and Warren (1995) derived
the Fokker-Planck equation for the DPD method and the corresponding fluctuation-
dissipation theorem, demonstrating that the DPD model yields collective behavior
consistent with classical hydrodynamics. A review of the DPD approach is given by
Groot and Warren (1997).
Similar to the Langevin dynamics approach illustrated by (2.3.5), the DPD model
imposes three distinct types of interaction forces between any pair of computational
particles i and j, representing a dissipative force FD C
i j , a conservative force Fi j , and
R
a random force Fi j . However, there is considerable difference with LD regarding
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.3. Mesoscale Particle Methods 39

the expressions used for these forces. In particular, in DPD all forces act between
the computational particles, and there is no notion of a surrounding continuum fluid
representing a solvent or gaseous interstitial region. Galilean invariance requires that
these forces depend only on the vector separating their centroids r = ri − r j and the
relative centroid velocity vR = vi − v j , rather than the particle centroid positions and
velocities independently. It is also useful to define the unit vector n pointing from the
centroid of particle i to the centroid of particle j by n ≡ (ri − r j )/r, where r = |r j − ri |
is the distance between the particle centroids. Expressions for the dissipative and
random forces were proposed by Español and Warren (1995) for r < rC , where rC is
a prescribed cut-off distance, as

i j = −γdpd ωD (r)(n · vR )n,

FD i j = σdpd ωR (r)ζi j n.
FR (2.3.6)

In these equations, ζi j is a Gaussian random variable with unit variance, and γdpd and
σdpd are the friction coefficient and the amplitude of the random forcing, respectively.
The functions ωD (r) and ωR (r) govern how these forces decay as the distance between
the particles increases. The balance between the dissipative and random terms serves
as a thermostat for the DPD simulation, such that the dissipative force tends to cool
the system and the random force tends to heat the system. The fluctuation-dissipation
theorem for DPD requires that the coefficients of these two terms are related to the
absolute temperature T of the medium by

ωR2 (r) = ωD (r), σdpd

2
= 2kB T γdpd , (2.3.7)

where kB is Boltzmann’s constant.

The conservative force FiCj in DPD has the form of a soft potential, and thus
avoids the numerical stiffness associated with rapid variation of FiCj with r in potentials
such as the Lennard-Jones potential (2.3.1). A simple linear form is often assumed
for both the conservative force and for the spatial decay function ωR (r) of the
form

a[1 − r/rC ]n for r < rC 1 − r/rC for r < rC

FiCj = , ωD (r) = ωR2 (r) = (2.3.8)
0 for r > rC 0 for r > rC

where a is an adjustable parameter related to the material compressibility. The

resulting momentum equation for a DPD particle i thus has the form

dvi N

m = i j + Fi j .
FiCj + FD R
(2.3.9)
dt
j=1
j=i

Because DPD particles represent a group of molecules, instead of individual

atoms, and because of the use of soft potentials and lack of covalent bonds between
DPD particles, the DPD method can be used for problems involving length and
time scales that are one to two orders of magnitude greater than can be used for
MD. A time step limitation for DPD is presented by Español and Warren (1995)
as t/tc 1, where tC = rC /vrms and vrms = (3kB T/m)1/2 is the root-mean square
particle velocity.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

40 Modeling Viewpoints and Approaches

2.3.4. Discrete Element Method

The discrete element method (DEM) is similar to BD in that it solves for motion
of individual particles using a continuum approximation of the surrounding fluid.
The most significant difference between DEM and BD concerns the model used
for interparticle contact forces. Whereas it is typical in BD simulations to employ
simple conservative potentials, such as the Lennard-Jones potential, to describe
interactions of very small particles (on the nanometer scale), such potentials do
not adequately represent the interparticle forces present for collision of the larger
(micron-size) particles used for DEM simulations. Instead, DEM is founded on
the assumption that the particle size is substantially larger than the characteristic
length scale of van der Waals force or other adhesion force, which for van der Waals
interaction is on the order of a nanometer. One consequence of this assumption is
that in DEM, particle adhesion force has no effect until two particles collide. This
situation is unlike the MD, DPD, or BD models, which are intended for smaller
particles for which adhesion forces act over length scales on the order of the particle
diameter. These larger particles also have more inertia than the nanometer scale
particles treated with the MD, DPD, and BD schemes, so it is typically necessary
to account for resistance to sliding, rolling, and twisting motions of the particles in
addition to the interaction force directed along the particle normal vector n. Since
the various modes of frictional resistance depend on the relative tangential motion
between the particle surfaces at the contact point, which in turn depends on both
the particle rotation rate and centroid velocity, in DEM simulations it is necessary
to solve governing equations for both particle linear and angular momentum.
Two very different variations of DEM models exist, called the hard-sphere and
soft-sphere models. In the soft-sphere model, the particle momentum and angular
momentum equations, given by
dv d
m = FF + FA + FE , I = MF + MA + ME , (2.3.10)
dt dt
are solved during the entire period in which two colliding particles are touching each
other. In these equations, m and I are the particle mass and moment of inertia, v and 
are the particle velocity and rotation rate, FF and MF are the force and torque exerted
by the surrounding fluid, FA and MA are the force and torque due to the collision
and adhesion forces, and FE and ME are the force and torque induced by interaction
with the surrounding electric field. The time derivative in (2.3.10) is evaluated on a
moving particle. The soft-sphere model requires that the computational time step
be sufficiently small to resolve the period during which the collision occurs. Because
particle elastic collisions tend to occur over very small time scales, soft-sphere DEM
computations tend to be numerically stiff. This difficulty is handled in practice by use
of multiple time-step algorithms, and also in some cases by use of elastic moduli for
the particles that are smaller than the actual elastic moduli. By contrast, the hard-
sphere model assumes that collisions between particles are pair-wise additive and
occur instantaneously. The particle collisions are thus accounted for using particle
impulse equations. The hard-sphere model removes the stiffness associated with the
need to resolve the time interval during which particles collide, allowing this method
to use much larger time steps than the soft-sphere model. However, the hard-sphere
model cannot handle problems with collisions of three or more particles or problems
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.4. Microscale Dynamics of Elastohydrodynamic Particle Collisions 41

in which particles remain in contact for a prolonged time period. Consequently, most
studies of adhesive particle flows utilize the soft-sphere model. Those studies that
do use the hard-sphere model for adhesive particles typically assume that particles
that adhere together are “frozen” upon collision with each other. However, this
assumption ignores deformation of particle agglomerates by fluid-induced forces
and by collisions with other agglomerates, which have been found to be a common
and important process governing the physics of the agglomerated particles.
As discussed in Section 1.3.1, adhesive particles in the micron size range cannot
be treated as spheres when they collide, because the length scale associated with par-
ticle adhesive force is typically on the same order of magnitude, or smaller, than that
associated with particle elastic deformation. A common type of soft-sphere DEM
model makes a very different assumption – that the particle adhesive forces only act
within the flattened part of the colliding particle surfaces called the contact region
(Figure 1.7). For this model, the adhesive interaction between two micrometer-sized
particles is approximated by the adhesion force between two parallel flat plates,
rather than between two spheres. This observation is the basis of the JKR the-
ory (Johnson, Kendell, and Roberts, 1971) of elastic adhesion. Since the particle
deformation is caused by the force acting between the particles, which is subse-
quently influenced by the extent of particle deformation, the interaction of adhesive
micron-size particles is a highly nonlinear process. As a consequence, adhesion and
collision forces between the particles cannot simply be added together, as they can
for two spherical nanoparticles, but instead a new model must be formulated for
the particle contact forces that includes the combination of elastic and adhesive
effects.
There has recently been quite a lot of activity with DEM for adhesive particles,
both in terms of applying the method to a wide variety of problems and in extending
the method to deal with particle interactions with different field effects, such as
acoustic and electric fields, and to handle different types of adhesive forces. These
extensions of DEM are the focus of most of the remainder of this book, so more
detailed discussion of these developments is deferred until later chapters.

2.4. Microscale Dynamics of Elastohydrodynamic Particle Collisions

Microscale methods simulate the detailed interaction of particles with the surround-
ing fluid flow, or with other fields such as electric, acoustic, or thermal fields. There
might be some confusion as to what exactly constitutes a microscale simulation
given the broad range of scales covered by the mesoscale methods discussed in
the previous section. In general, a mesoscale method employs a relatively simple
approximate model for the interaction of the computational particles, whereas a
microscale method directly simulates the surrounding field distortions that give
rise to the interaction. For example, molecular dynamics is a mesoscale method
for simulating interaction of atoms that uses a simple potential model (e.g., the
Lennard-Jones potential) to approximate the interaction of atoms. By contrast, a
direct simulation of the quantum field around the atoms using the ab initio method
would represent a microscale simulation of atomic interaction. For particulate fluids,
Brownian dynamics or the discrete element method represent mesoscale methods
that employ simple models for particle interactions, whereas a microscale simulation
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

42 Modeling Viewpoints and Approaches

would directly simulate the flow around each particle and the particle deformation
during collisions.
There is a broad literature devoted to simulation of detailed flow fields about
groups of spherical particles, droplets, and bubbles, as well as the various compu-
tational methods used for such problems. A thorough review of this literature is
outside the scope of what we wish to cover in this book. Instead, in the current
section we consider a specific example of a microscale simulation by examining the
problem of particle elastohydrodynamic interaction, which deals with the combined
elastic deformation and the squeeze-film fluid dynamics that occurs in the contact
region between two colliding particles. The fluid dynamics in this region is of great
importance in constructing physical scaling parameters for DEM models, since this
thin region of the flow field governs the extent to which the two particle surfaces
come together in a dynamic collision process.

2.4.1. Microscale Simulations of Elastohydrodynamic Interactions

A classic series of papers on the elastohydrodynamics of the head-on collision of
two particles immersed in a viscous fluid was published in the 1980s by Robert
Davis and his students at the University of Colorado, along with E.J. Hinch from
Cambridge University. In this work, the particle elastic deformation was computed
using a boundary-element formulation, and the fluid flow within the thin gap region
separating the particles was computed using a lubrication equation of the form
 
∂h 1 ∂ 3∂p
= rh . (2.4.1)
∂t 12μr ∂r ∂r
The gap thickness h(r, t ) and the averaged pressure p(r, t ) over the gap are both
functions of the radial coordinate r and time t. The lubrication theory assumes that
the thickness of the gap between the particles is much smaller than the contact region
radius and that the flow within this squeeze-film region has a low Reynolds number,
evaluated based on the gap thickness.
The boundary-element theory assumes small elastic displacement of the particle.
An effective radius R and a parameter θ are defined by
 −1 1 − σ12 1 − σ22
R ≡ r1−1 + r2−1 , θ≡ + , (2.4.2)
π E1 π E2
where r1 and r2 are the radii, σ1 and σ2 are the Poisson ratios, and E1 and E2 are
the elastic moduli of the two colliding particles. The extent of elastic deformation
can be characterized based on the dimensionless elasticity parameter εE , which is
defined as
4θ μv0 R3/2
εE ≡ , (2.4.3)
x5/2
0

where μ is the fluid viscosity and the particle impact velocity v0 is evaluated when
the minimum gap thickness has a value x0 . The parameter x0 is a measure of the
gap thickness between the particles, but its value is somewhat arbitrary subject to
the restrictions x0 /R 1, εE 1, and Re x0 /R 1, where Re = ρv0 R/μ is the
particle Reynolds number. The computations presented by Davis et al. (1986) select
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.4. Microscale Dynamics of Elastohydrodynamic Particle Collisions 43

v0 t / x0 = 0

–2
εE=10

hm / x1
h / x0

–4
10
–6
10
–8
10

1/2
r / (ax0 )
(a) (b)
Figure 2.5. Plots showing details of particle collision with no adhesion: (a) sequence of profiles
of the dimensionless gap thickness as a function of dimensionless radius for different times for
St = 5 and εE = 0.01 and (b) plot of minimum gap half-thickness divided by the axial length
scale x1 as a function of the Stokes number for different values of the elasticity parameter.
[Reprinted with permission from Davis et al. (1986).]

x0 = 0.01 R. Using this value for x0 , typical values of εE for collisions of 1–100 µm
diameter particles in air and water lie in the range 10–7 –10–5 . An alternative length
scale characterizing the gap thickness is motivated by noting that the numerator of
(2.4.3) has dimensions of length to the 5/2 power, so that a length scale x1 can be
defined by

x1 ≡ (4θ μR3/2 v0 )2/5 . (2.4.4)

Numerical results from Davis et al. (1986) are plotted in Figure 2.5. Figure 2.5a
plots the ratio of the particle gap thickness h(r, t) to the axial length scale x0 as a
function of the ratio of radius r to the radial length scale (Rx0 )1/2 . Curves are drawn at
different times during the particle collision. It is of note that the separation distance
between the particles decreases to some minimum value, but for times after this
minimum value is reached the particle deformation forms a flattened contact region
between the two particles. Motion of the particles toward each other at times after
this minimum separation is achieved results in growth of the contact region, but not
in decrease of the particle minimum separation distance. Figure 2.5b plots the ratio of
the minimum gap thickness achieved during the collision, hm , to the alternative axial
length scale x1 as a function of Stokes number for different values of the elasticity
parameter εE . This plot demonstrates that at a sufficiently high Stokes number, the
ratio of the minimum gap thickness to the length scale x1 approaches a value that
is independent of the elasticity parameter. This result provides strong evidence that
the parameter x1 gives a correct scaling for the minimum gap thickness for collision
of smooth particles in the absence of adhesive force.
In later work, Barnocky and Davis (1989) examined effects of density and vis-
cosity variation that occurs due to high pressures within the contact region. Serayssol
and Davis (1986) examined the effect of adhesive forces on the elastohydrodynamics
of particle collisions. An interesting result of the latter paper is shown in Figure 2.6,
which plots h/x0 as a function of r/(Rx0 )1/2 for a case with adhesion force acting
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

44 Modeling Viewpoints and Approaches

v0 t / x0 = 0
Figure 2.6. Sequence of profiles showing dimension-
less gap thickness as a function of dimensionless
h / x0

radius for different times during normal collision of

two particles, for a case with adhesive force acting
between the particles for St = 5 and εE = 0.01. The
curves are plotted for different values of the dimen-
sionless time, as indicated in the figure. [Reprinted
with permission from Serayssol and Davis (1986).]

1/2
r / (ax0)

between the particle surfaces. Similar to the case without adhesion shown in Fig-
ure 2.5a, the particle surfaces form a flattened contact region as the particles approach
each other. However, for the adhesion case, it is observed that as the particles begin
moving apart during the rebound phase of the collision, an interesting necking phe-
nomenon occurs (at times 5.0–5.6 for the example shown in Figure 2.6) during which
the surface of each particle is pulled outward toward the opposing particle by the
adhesion force. This adhesion-induced necking behavior plays an important role in
the contact dynamics of adhesive particles.

2.4.2. Experimental Results for Two-Particle Collisions

Extensive experimental studies have been performed to examine the effect of inter-
stitial fluid on particle-particle and particle-wall interactions. In the simplest experi-
ments, in which two particles collide with each other with velocities oriented along
the normal direction n relative to the other particle, the experiments measure the
relative velocity of the particles before and after collision, given by

vI ≡ |v2 (tC − t0 ) − v1 (tC − t0 )|, vR ≡ |v2 (tC + t0 ) − v1 (tC + t0 )|, (2.4.5)

where tC is the time at which the collision occurs and t0 is a sufficiently large time
interval that for t < tC − t0 the particles have negligible effect on each other. The
ratio of the relative normal particle speed after collision to that before collision is
called the restitution coefficient, e, and is given by

e ≡ vR /vI . (2.4.6)

In general, restitution coefficient is a function of all parameters associated with

processes that lead to energy dissipation during the collision process. This includes
both energy loss due to viscous dissipation within the fluid as well as energy loss
due to inelastic deformation of the solid particles. In granular flows, which typically
involve relatively large particles in air, the majority of the energy dissipation occurs
due to solid phase energy loss, and the fluid loss is often neglected. On the other
hand, for the small particles that are typically of concern for adhesive particle flows,
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.4. Microscale Dynamics of Elastohydrodynamic Particle Collisions 45

Restitution Coefficient, e
0.8
Figure 2.7. Comparison of predicted total restitu-
tion coefficient etot (solid line) and spherical par-
ticle restitution coefficient esph (dashed line) with 0.6
the experimental data of Yang and Hunt (2006)
for collision of identical spherical particles, made
from steel (squares), glass (circles), and Delrin 0.4
(triangles). The predictions are performed with
parameter values εE = 10−7 , δ/r p = 5 × 10−5 , and 0.2
h0,init /r p ∼
= 0.02. [Reprinted with permission from
Marshall (2011).]
0 0
10 101 102 103
StC

the collision velocity is relatively small and the fluid losses often dominate over the
solid losses.
Estimates for the fluid losses are usually expressed as functions of the Stokes
number, defined in (1.2.5). It was demonstrated by Yang and Hunt (2006) that
restitution coefficients for collisions between particles of two different radii, r1 and
r2 , with two different masses, m1 and m2 , can be approximately collapsed onto a
single curve using the effective radius R defined in (2.4.2) and an effective mass M
defined by

 
−1 −1
M ≡ m−1
1 + m2 . (2.4.7)

The Stokes number for collision processes is defined with particle mass replaced by
M, the length scale L set equal to R, and the velocity scale U is set equal to the
impact velocity vI , such that (1.2.5) becomes

MvI
StC = . (2.4.8)
6π μR2

The data collapse presented by Yang and Hunt (2006) includes data for two
equal-size colliding spheres made of different materials and with different masses,
unequal-size spheres, and sphere collision with a wall, where the wall is treated as a
sphere of infinite radius. Yang and Hunt (2006) even found that this data collapse
can be applied to oblique collisions if the restitution coefficient is defined only in
terms of the normal component of the particle velocities before and after collision.
The full plot is rather cluttered, so we present a reduced plot with data only for two
equal-size spheres in Figure 2.7. A particularly interesting feature of this plot is that
the restitution coefficient is essential equal to zero if the collision Stokes number
is less than 10. This case applies to many applications involving adhesive particles,
which tend to involve particles of a sufficiently small size that the adhesive forces are
of a comparable magnitude to other forces that might exist, such as the gravitational
and fluid drag forces and the particle inertia.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

46 Modeling Viewpoints and Approaches

2.4.3. Simplified Models for Restitution Coefficient in a Viscous Fluid

Integrating the lubrication equation (2.4.1) in the radial direction over the interval
(0, r) gives the radial pressure gradient within the squeeze-film as
∂p 6μr ∂h
= 3 . (2.4.9)
∂r h ∂t
Integrating (2.4.9) a second time in the radial direction, this time over the interval
(r, r p ), where r p is the particle radius, gives the squeeze-film pressure distribution as

 rp
ξ ∂h
p(r, t ) − p0 = −6μ dξ , (2.4.10)
r h3 ∂t
where ξ is a dummy variable of integration and the pressure is assumed to approach
an ambient value p0 as r → r p . The normal damping force Fnd exerted by the squeeze-
film fluid along the axial direction n is obtained by integrating the pressure field from
(2.4.10) over the circular area covered by the film, which after using integration by
parts yields
 r  r   r 3
p p ξ ∂h p r ∂h
Fnd = −12π μ dξ rdr = −6π μ dr. (2.4.11)
0 r h ∂t
3
0 h3 ∂t
The damping force between two spherical particles that are approaching each
other can be obtained from (2.4.11) using
∂h dh0 r2
= = f (t ), h − h0 ∼
= , (2.4.12)
∂t dt rp

where h0 (t ) ≡ h(0, t ) is the gap thickness alongthe line separating the particle cen-
troids and we assume h/r p 1. Defining η ≡ r/ r p h0 and using (2.4.12), the integral
in (2.4.11) becomes
 (h0 /r p )−1/2
6π μr2p dh0 η3
Fnd = − dη. (2.4.13)
h0 dt 0 (1 + η2 )3
Since h0 /r p 1, the upper bound for the integration in (2.4.13) can be approximated
by infinity, for which case the integral is equal to 14. The resulting expression for
the normal damping force on the spherical particle becomes
3π μr2p dh0
Fnd (t ) = − . (2.4.14)
2h0 dt
The normal force expression (2.4.14) applies only prior to and following collision
of two particles. During the collision itself, the viscous damping force is dominated
by the fluid flow near the outer edge of the contact region as this region grows and
contracts with motion of the particle centroids toward or away from each other.
An expression for normal damping force during collision was derived by Marshall
(2011) using lubrication theory as
 
3π μr2p dδN δN
Fnd = 1+ , (2.4.15)
2δ dt δ
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

2.4. Microscale Dynamics of Elastohydrodynamic Particle Collisions 47

where δN (t ) is the “normal overlap” of the particles and δ is the gap thickness between
the particle surfaces within the contact region. The normal overlap for collision of
two equal spheres is given by

δN = 2r p − |xi − x j |, (2.4.16)

where |xi − x j | denotes the distance between the sphere centroids. At the instant of
particle collision, the variables in (2.4.15) have the values h0 (t ) = δ, δN (t ) = 0, and
δ̇N (t ) = −ḣ0 (t ). Substituting these values into (2.4.15) yields the equation (2.4.14) for
two spherical particles, thus guaranteeing that the normal force expressions (2.4.14)
and (2.4.15) are continuous the throughout the collision process.
Head-on collision of two identical, nonadhesive particles of radius r p and equal
mass m, with centroid positions ±[r p + x(t )] and velocities ∓v(t ), is governed by the
equations
dx dv
= v, m = Fnd + Fne , (2.4.17)
dt dt
where Fnd is the normal damping force, given by (2.4.14) before and after the particles
collide and by (2.4.15) during the collision, and Fne is the elastic rebound force that
acts to push the particles apart during the collision. The two particles are assumed
to have initial centroid positions ±x0 and to travel with initial velocity ∓v0 toward
each other.
It is helpful to define dimensionless variables v  = v/v0 , x = x/x0 , and t  =
tv0 /x0 . Prior to the collision, substituting (2.4.14) in (2.4.17) with h0 = 2x and
dh0 /dt = −2v, the governing equation can be written in dimensionless form as

dv  1 v
v = − , (2.4.18)
dx StC x
where StC is the collision Stokes number, with M = m/2, R = r p /2, and vI = 2v0 for
two equal colliding spheres. The left-hand side of (2.4.18) is obtained using chain
rule by writing v  = v  (x ) and using (2.4.17)1 . Dividing (2.4.18) by v  and integrating
over x yields
v(x) 1
=1− ln(x0 /x). (2.4.19)
v0 StC
This equation describes the decrease in velocity of the spheres as they approach
each other. If we assume that the spheres collide with no loss in energy during the
collision, then the resulting restitution coefficient esph is obtained from (2.4.19) as
2
esph = 1 − ln(x0 /δ), (2.4.20)
StC
where δ is the equilibrium gap thickness between the particle surfaces within the
contact region. The factor of 2 in the last term in (2.4.20) accounts for the motion of
the two spheres both toward and away from each other. On the other hand, energy
might also be lost during the collision process itself, either within the solid phase
due to deformation of the particles or within the fluid phase due to viscous flow
associated with change in the size of the contact region, with an expression for the
latter damping force given by (2.4.15). If the restitution coefficient associated with
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

48 Modeling Viewpoints and Approaches

the collision itself, not including the motion of the two spheres toward each other
prior to the collision or away from each other following the collision, is denoted by
ecol , then the equation for the total restitution coefficient becomes
1 + ecol
etot = ecol − ln(x0 /δ). (2.4.21)
StC
The predicted values of esph from (2.4.20) and of etot from (2.4.21) as functions of
the Stokes number are plotted in Figure 2.7 as dashed and solid lines, respectively, in
comparison with the experimental data of Yang and Hunt (2006) for collision of two
equal-size spheres. The value of ecol in (2.4.21) was obtained by a numerical integra-
tion of the equations of motion using the expression (2.4.15) for the normal damping
force and the classical Hertz law for the elastic force. Although both expressions
agree reasonably well with the data for small Stokes numbers, the total restitution
coefficient given by (2.4.21) fits the data considerably better as the Stokes number
increases.

REFERENCES

Andrews MJ, O’Rourke PJ. The multiphase particle-in-cell method for dense particulate
flows. International Journal of Multiphase Flow 22, 379–402 (1996).
Barnocky G, Davis RH. The influence of pressure-dependent density and viscosity on the
elastohydrodynamic collision and rebound of two spheres. Journal of Fluid Mechanics 209,
501–519 (1989).
Benyahia S, Galvin JE. Estimation of numerical errors related to some basic assumptions in
discrete particle methods. Industrial & Engineering Chemistry Research 49, 10588–10605
(2010).
Brown R. A brief account of microscopical observations made in the months of June, July
and August, 1827, on the particles contained in the pollen of plants; and on the general
existence of active molecules in organic and inorganic bodies. Philosophical Magazine 4,
161–173 (1828).
Cameron IT, Wang FY, Immanuel CD, Stepanek F. Process systems modelling and applica-
tions in granulation, A review. Chemical Engineering Science 60(14), 3723–3750 (2005).
Chen JC, Kim AS. Brownian dynamics, molecular dynamics, and Monte Carlo modeling of
colloidal system. Advances in Colloid and Interface Science 112, 159–173 (2004).
Crowe CT, Sharma MP, Stock DE. The particle-source-in-cell method for gas droplet flow.
Journal of Fluids Engineering 99, 325–332 (1977).
Crowe CT, Schwarzkopf JD, Sommerfeld M, Tsuji Y. Multiphase Flows with Droplets and
Particles, CRC Press, Boca Raton, Florida (2012).
Davis RH, Serayssol J-M, Hinch EJ. The elastohydrodynamic collision of two spheres. Journal
of Fluid Mechanics 163, 479–497 (1986).
Dawson JM. Particle simulation of plasmas. Reviews of Modern Physics 55(2), 403–447 (1983).
Dzwinel W, Yuen DA. Mesoscopic dispersion of colloidal agglomerate in a complex fluid
modelled by a hybrid fluid-particle model. Journal of Colloid and Interface Science 247,
463–480 (2002).
Dzwinel W, Boryczko K, Yuen DA. A discrete-particle model of blood dynamics in capillary
vessels. Journal of Colloid and Interface Science 258, 163–173 (2003).
Einstein A. On the movement of small particles suspended in stationary liquids required by
the molecular-kinetic theory of heat. Annalen der Physik 17, 549–560 (1905).
Ermak DL, McCammon JA. Brownian dynamics with hydrodynamic interactions. Journal of
Chemical Physics 69(4), 1352–1360 (1978).
Español P, Warren P. Statistical mechanics of dissipative particle dynamics. Europhysics
Letters 30(4), 191–196 (1995).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

References 49

Filipovic N, Kojic M, Tsuda A. Modelling thrombosis using dissipative particle dynamics

method. Philosophical Transactions of the Royal Society of London A 366, 3265–3279
(2008).
Freireich B, Li J, Litster J, Wassgren C. Incorporating particle-flow information from discrete-
element simulations in population balance models of mixer-coaters. Chemical Engineering
Science 66, 3592–3604 (2011).
Frenklach M, Harris SJ. Aerosol dynamics modeling using the method of moments. Journal
of Colloidal and Interface Science 118(1), 252–261 (1987).
Friedlander AK. Smoke, Dust and Haze: Fundamentals of Aerosol Dynamics, 2nd ed., Oxford
University Press, U.S.A. (2000).
Gantt JA, Cameron IT, Litster JD, Gatzke EP. Determination of coalescence kernels
for high-shear granulation using DEM simulations. Powder Technology 170(2), 53–63
(2006).
Groot RD, Warren PB. Dissipative particle dynamics: bridging the gap between atomistic and
mesoscopic simulation. Journal of Chemical Physics 107(11), 4423–4435 (1997).
Harris SE, Crighton DG. Solitons, solitary waves, and voidage disturbances in gas-fluidized
beds. Journal of Fluid Mechanics 266, 243–276 (1994).
Heine MC, Pratsinis SE. Brownian coagulation at high concentration. Langmuir 23, 9882–9890
(2007).
Hoogerbrugge PJ, Koelman JMVA. Simulating microscopic hydrodynamic phenomena with
dissipative particle dynamics. Europhysics Letters 19(3), 155–160 (1992).
Johnson KL, Kendall K, Roberts AD. Surface energy and the contact of elastic solids. Pro-
ceedings of the Royal Society of London A 324, 301–313 (1971).
Langevin P. On the theory of brownian motion 1908. C. R. Acad. Sci. (Paris) 146, 530–533
(1908).
Li S-Q, Marshall JS, Liu G, Yao Q. Adhesive particulate flow: The discrete element method
and its application in energy and environmental engineering. Progress in Energy and Com-
bustion Science 37(6), 633–668 (2011).
Marshall JS. Discrete-element modeling of particulate aerosol flows. Journal of Computational
Physics 228, 1541–1561 (2009).
Marshall JS. Viscous damping force during head-on collision of two spherical particles. Physics
of Fluids 23(1), 013305 (2011).
McCoy BJ. A population balance framework for nucleation, growth, and aggregation. Chem-
ical Engineering Science 57, 2279–2285 (2002).
Patankar NA, Joseph DD. Lagrangian numerical simulation of particulate flows. International
Journal of Multiphase Flow 27, 1685–1706 (2001).
Pivkin IV, Richardson PD, Karniadakis GE. Effect of red blood cells on platelet aggregation.
IEEE Engineering in Medicine and Biology Magizine 28(2), 32–37 (2009).
Ramkrishna D. Population Balances – Theory and Applications to Particulate Systems in
Engineering. Academic Press, San Diego (2000).
Reinhold A, Briesen H. Numerical behavior of a multiscale aggregation model: Coupling
population balances and discrete element models. Chemical Engineering Science 70, 165–
175 (2012).
Serayssol J-M, Davis RH. The influence of surface interactions on the elastohydrody-
namic collision of two spheres. Journal of Colloid and Interface Science 114(1), 54–66
(1986).
Smoluchowski M von. Zur kinetischen Theorie der Brownschen Molekularbewe-gung und
der Suspensionen. Annalen der Physik 21, 756–780 (1906).
Smoluchowski M von. Versuch einer mathematischen Theorie der Koagulationkinetik kol-
lider lösungen. Z. Phys. Chem. 92, 129–168 (1917).
Snider DM. Three fundamental granular flow experiments and CPFD predictions. Powder
Technology 176, 36–46 (2007).
Sommerfeld M. Validation of stochastic Lagrangian modeling approach for inter-particle
collisions in homogeneous turbulence. International Journal of Multiphase Flow 27, 1829–
1858 (2001).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-02 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:47

50 Modeling Viewpoints and Approaches

Symeonidis V, Karniadakis GE, Caswell B. Dissipative particle dynamics simulations of poly-

mer chains: scaling laws and shearing response compared to DNA experiments. Physical
Review Letters 95, 076001 (2005).
Tian P. Molecular dynamics simulations of nanoparticles. Annual Reports on the Progress of
Chemistry, Section C Physical Chemistry 104, 142–164 (2008).
Trzeciak TM, Podgόrski A, Marijnissen JCM. Stochastic calculation of collision kernels: Brow-
nian coagulation in concentrated systems. In: 5th World Congress on Particle Technology
(WCPT5), Orlando, Florida, p. 202d (2006).
Wijmans CM, Smit B, Groot RD. Phase behavior of monomeric mixtures and polymer solu-
tions with soft interactions potentials. Journal of Chemical Physics 114(17), 7644–7654
(2001).
Yang F-L, Hunt ML. Dynamics of particle-particle collisions in a viscous liquid. Physics of
Fluids 18, 121506 (2006).
Zhou Q, Yao SC. Group modeling of impacting spray dynamics. International Journal of Heat
and Mass Transfer 35, 121–129 (1992).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3 Contact Mechanics without Adhesion

Contact mechanics refers to the solid phase stresses and deformation that occur
when two bodies collide, resulting in the formation of a contact force and torque
exerted on the bodies. As discussed in Chapter 1, for micrometer-sized particles
the contact forces are changed in a nonlinear manner by the presence of adhesive
forces. For this reason, the contact force and the adhesive force are not additive, but
instead collision with adhesion must be examined as a combined mechanical theory.
Although the theory of particle contact in the presence of adhesion reduces to that
with no adhesion as a limiting case, it is nevertheless of value to consider the problem
of contact with no adhesion by itself both as an introduction to the broader theory
and because adhesionless contact has a large number of applications for problems
involving granular materials.

3.1. Basic Concepts

Two spherical particles with radii ri and r j , elastic moduli Ei and E j , Poisson ratios
νi and ν j , and shear moduli Gi = Ei /2(1 + νi ) and G j = E j /2(1 + ν j ) are in contact
with each other. Typical values of elastic modulus and Poisson ratio for different
materials are shown in Table 3.1. An effective particle radius R and effective elastic
and shear moduli, E and G, are defined by
1 1 1 1 1 − νi2 1 − ν 2j 1 2 − νi 2 − νj
≡ + , ≡ + , ≡ + . (3.1.1)
R ri rj E Ei Ej G Gi Gj
If no force is applied between the two particles, the contact occurs only at a
point. In the presence of a force Fn along the direction of the vector −n normal
to the particle surface, each particle is deformed in the region near the contact
point such that the particles contact each other over a finite region of circular shape
called the contact region. The point at the center of the contact region, which is
usually the first point to make contact as two spherical particles collide, is called
the contact point. The radius a of the contact region is called the contact radius.
An important parameter characterizing the particle deformation within the contact
region is the normal overlap δN , defined as the distance between the contact points on
the particles if the particles had remained undeformed (Figure 3.1). If xi and x j denote

51
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

52 Contact Mechanics without Adhesion

Table 3.1. Typical values for Young’s modulus and Poisson’s ratio of
representative materials. [Extracted from data in Engineering Tool Box
(http://www.engineeringtoolbox.com).]

Materials Young’s modulus, Ei (GPa) Poisson’s ratio, σi

Metals
Aluminum 69–120 0.33
Cast Iron 200 0.21–0.26
Copper 120 0.33
Gold 75 0.42
Magnesium 45 0.35
Stainless Steel 180 0.30–0.31
Steel 200 0.27–0.30
Titanium 110 0.34
Soil
Clay 5–50 0.30–0.45
Concrete 30 0.2
Sand 10–80 0.20–0.45
Other Materials
Glass 50–90 0.18–0.3
Foam 0.01–0.15 0.10–0.40
Rubber 0.01–0.1 0.50

the centroid positions of the two spherical particles, such that n = (x j − xi )/|x j − xi |
is a unit vector pointing from the centroid of particle i to the centroid of particle j,
then the normal overlap is given by
δN = ri + r j − |xi − x j |. (3.1.2)
If the contact region radius a is small compared to the particle radius, the contact
region radius and the normal overlap can be related geometrically as

a = RδN . (3.1.3)

Particle i

z(j)

δΝ uzj r-θ
uzi plane

z(i)
a a

Particle j
Figure 3.1. Geometry of two spherical particles in contact, illustrating the normal overlap and
the contact radius. The dashed lines denote what the particle surfaces would have been if they
had remained spherical, and the solid lines denote the deformed particle surfaces.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.1 Basic Concepts 53

(a) (b) (c) (d) (e) (f)

Figure 3.2. Schematics of classic rheological models in continuum mechanics. (a) Elastic;
(b) viscous; (c) slider; (d) Maxwell; (e) Kelvin; (f) Burger’s.

For two spherical particles, the contact region forms a circle centered at the contact
point. At any point on the contact surface with radial distance r from the circle center,
the corresponding elastic displacements of the two bodies in the direction normal to
the contact surface (±z), uzi and uz j , must satisfy the kinematic relationship

r2
uzi + uz j = δN − . (3.1.4)
2R
Even for two particles that are considered to be in contact, there still exists an
equilibrium gap between the particle surfaces with thickness δ, which has a value
ranging between 0.16 and 0.40 nm for smooth surfaces. For rough surfaces, the gap
size is on the order of the roughness length scale. Due to the small size of this
gap, it is generally negligible for the purpose of modeling elastic deformation for
nonadhesive surfaces, but it plays a major role in the contact mechanics of adhesive
surfaces discussed in Chapter 4.
There are many different rheological models relating force and displacement
in modeling deformation of solids, which in the current context relate the normal
contact force Fn and the normal overlap δN . Several of these models are shown
schematically in Figure 3.2. The simplest such relationship is that of a purely elastic
response, as is typical of Hooke’s law for elastic deformation, which is represented
in Figure 3.2a by a spring. The elastic model does not dissipate energy or lead
to plastic material deformation, enabling the corresponding Fne − δN curve to be
the same for loading and unloading processes. The purely viscous response, as is
typical of a Newtonian viscous fluid, is represented in Figure 3.2b by a damping
dashpot, such that the force is a function of the rate of deformation, for example,
Fn ∝ dδN /dt. The rigid–perfectly plastic response, represented as a slider in Fig-
ure 3.2c, indicates that there is no deformation (δN = 0) until a yield point is reached,
and after that the deformation continues at a constant loading force. The Maxwell
response model in Figure 3.2d consists of a spring and a dashpot arranged in series,
which implies that the elastic and viscous elements experience the same force but
exhibit different displacements. In the Maxwell model, the total displacement is the
sum of elastic and viscous displacements. The Kelvin response model shown in Fig-
ure 3.2e comprises a spring and dashpot acting in parallel. In the Kelvin model, the
elastic and viscous forces are different but both the spring and dashpot experience the
same displacement. Finally, Burger’s model, shown in Figure 3.2f, is a combination
of the Kelvin and Maxwell models.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

54 Contact Mechanics without Adhesion

(a) (b) (c) (d)

Figure 3.3. Modes of particle interaction: (a) normal impact; (b) sliding; (c) twisting; and
(d) rolling.

The Kelvin model is widely used for describing the normal contact force between
particles in contact. In this model, the total normal force Fn is the sum of the elastic
normal force Fne and the dissipative normal force Fnd . In addition to the normal
forces, solids in contact can also exhibit resistance to other motions associated with
relative tangential displacement of their surfaces. Among these motions we include
sliding, twisting, and rolling, shown schematically in Figure 3.3. The resistance to
normal displacement consists of a force acting in the −n direction passing through
the particle centroids, so that for spherical particles this force exerts no torque on
the particles. The particle resistance to other types of motions might have the form
of a force, a torque, or even a combination of a force and a torque. The sliding
resistance exerts a force Fs acting in a direction tS , corresponding to the direction of
relative motion of the particle surfaces at the contact point projected onto a plane
orthogonal to n. The sliding resistance also imposes a torque on the particle in the
n × tS direction with magnitude ri Fs , where ri is the radius of particle i. The twisting
resistance exerts a torque Mt on the particle in the n direction, but no force. Rolling
resistance exerts a torque Mr on the particle in the tR × n direction, where tR is
the direction of the “rolling velocity” vL , which is defined in Section 3.5. The total
collision force FA and torque MA acting on particle i can be written as the sum of
these various components as

FA = −Fn n + Fs tS , MA = ri Fs (n × tS ) + Mr (tR × n) + Mt n. (3.1.5)

The negative sign before Fn in (3.1.5) is used in order to conform with the convention
that a positive value of Fn corresponds to a repulsive force and a negative value
corresponds to an attractive force. The contact mechanics of nonspherical particles
is somewhat different from that for spheres. This matter is addressed in Chapter 7.

3.2. Hertz Theory: Normal Elastic Force

The foundational theory of nonadhesive elastic contact of two elastic bodies was
developed over a century ago by Hertz (1882). Interestingly, his initial research was
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.2. Hertz Theory: Normal Elastic Force 55

driven not by an effort to develop a theory for colliding particles, or even colliding
elastic bodies. He was instead trying to determine how the optical properties of
stacked glass lenses are modified due to slight deformation of their surfaces by elastic
deformation. Despite various extensions, the Hertz theory has stood the test of time
relatively unchanged and has found application in a wide variety of engineering
problems.

3.2.1. Derivation
The basic approximation introduced by Hertz to make an analytical solution of
the contact problem tractable is to treat each of the colliding bodies as an elastic
half-space loaded within the small contact region in such a manner as to produce
displacements satisfying the kinematic relationship (3.1.4). This simplification allows
us to use the well-developed methods available for solving elasticity problems in an
elastic half-space for contact problems (Love, 1952), without having to worry about
satisfaction of boundary conditions on the particle surface outside of the contact
region. In order for this simplification to be valid, it is necessary that the contact area
radius is small compared to the particle radius. This restriction is usually well satisfied
by particles made of metals, hard plastic, glass, or ceramics operating within their
elastic limits, but caution must be taken when applying contact mechanics models to
particles formed of soft materials, such as biological cells. A second approximation
made in the Hertz theory is that the contact surfaces are frictionless, so that only
normal stress is transmitted between them. This approximation is discussed in more
detail in Sections 3.3 and 3.4.
We consider a cylindrical polar coordinate system with coordinates (r, θ, z),
such that z = 0 corresponds to the top boundary of an elastic half-space that fills the
region z < 0. It is known from the kinematic result (3.1.4) that the displacement uz
in the normal (z) direction within a circle of radius a on the surface z = 0 must vary
in proportion to r2 . Hertz observed that a displacement of this form can be obtained
by imposing a surface pressure distribution of the form

p = p0 [1 − (r/a)2 ]1/2 , (3.2.1)

where the coefficient p0 is related to the total elastic normal force Fne on the particle
by
 a
2
Fne = 2π p(r)rdr = p0 π a2 . (3.2.2)
0 3
A solution for the problem of a point force P oriented in the normal direction
on the surface of an elastic half-space is given by Timoshenko and Goodier (1970,
402), based on the original solution by Boussinesq (1885). This solution gives the
surface displacement as
 
 P 1 − νi2
uz z=0 = , (3.2.3)
π Ei r

where νi is the Poisson’s ratio and Ei is the elastic modulus of the material. If we now
consider a distributed load with pressure p(r) over a circle of radius a on the surface
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

56 Contact Mechanics without Adhesion

of the elastic half-space, the load P in (3.2.3) can be set equal to an infinitesimal load
pda applied at a position x on the half-space surface. By superposition, the resulting
displacement at any point x on the surface is given by

1 − νi2 p(x )
uz (x) = da , (3.2.4)
π Ei S |x − x|
where S denotes the interior of the circle of radius a on the surface of the half-space.
Substituting (3.2.1) into (3.2.4) and performing the integration yields a solution for
normal displacement within the contact region of particle i of the form (Johnson,
1985, 92)

1 − νi2 π p0
uzi = (2a2 − r2 ). (3.2.5)
Ei 4a
Writing a similar expression for the displacement of the surface of particle j and
substituting into the kinematic requirement (3.1.4) yields

π p0 r2
uzi + uz j = (2a2 − r2 ) = δN − . (3.2.6)
4aE 2R
From (3.2.6), we can write

a = πRp0 /2E, δN = π ap0 /2E. (3.2.7)

Solving for p0 in (3.2.7) and substituting into (3.2.2) gives

4Ea3 3/2
Fne = = KδN , (3.2.8)
3R
where the stiffness coefficient K is given by
√
K = 43 E R. (3.2.9)

The equation (3.2.8) is the classical Hertz force expression for elastic contact of
colliding bodies. A useful alternative form for this expression is given by

Fne = kN δN , (3.2.10)

where the stiffness kN is written in terms of contact region radius a as

kN = 43 Ea. (3.2.11)

It should be noted that (3.2.10) does not yield a linear dependence between elastic
force and the normal overlap, because the contact region radius a in (3.2.11) varies

in proportion to δN from (3.1.3).

3.2.2. Two-Particle Collision

To illustrate the results of the normal elastic force model derived earlier, the problem
of head-on elastic collision of two particles, labeled Particle 1 and Particle 2, traveling
in a vacuum in the horizontal (x) direction is considered. The particle centroid posi-
tions and velocities at the initial time satisfy x1 < x2 and v1 > v2 . The faster-moving
particle on the left catches up with the slower-moving particle on the right, leading
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.2. Hertz Theory: Normal Elastic Force 57

to a collision between the two particles. During the collision, the displacement of
each particle is described by the particle momentum equation
d2 x1 d2 x2
m1 = −Fne , m2 = Fne . (3.2.12)
dt 2 dt 2
Substituting (3.2.12) into the time derivative of the definition (3.1.2) of the
normal overlap, δN = r1 + r2 − (x2 − x1 ), gives
d 2 δN d2 (x1 − x2 ) F
= = − ne , (3.2.13)
dt 2 dt 2 M
where M = (m−1 1 + m2 )
−1 −1
is the effective mass of the particles.
Substituting the Hertz expression for normal elastic force (3.2.8) into (3.2.13)
yields a differential equation for the normal overlap as
d 2 δN K 3/2
+ δN = 0. (3.2.14)
dt 2 M
Multiplying (3.2.14) by dδN /dt, we can write the result as
 
1 d dδN 2 2K d 5/2
+ (δ ) = 0. (3.2.15)
2 dt dt 5M dt N
Integrating (3.2.15) over time gives
 
1 dδN 2 2K 5/2
+ δ = C, (3.2.16)
2 dt 5M N
where C is a constant of integration. If v10 and v20 are the velocities of Particle 1 and
2, respectively, at the instant of time t = 0 just prior to collision and we define an
approach velocity v0 ≡ v10 − v20 , then the constant of integration is given by
C = 12 v02 . (3.2.17)
The coefficient C is equal to the kinetic energy of the two-particle system divided
by the effective mass M when measured in a coordinate system in which the origin
coincides with the center of inertia.
Both the maximum overlap and the maximum repulsive elastic force are attained
when the particle approach velocity vanishes, such that dδN /dt = 0. Solving for the
maximum value of δN from (3.2.16) and (3.2.17) yields
 2/5  2/5
5Mv02 15Mv02
δmax = = , (3.2.18)
4K 16ER1/2

where the last expression is obtained using (3.2.9). Substituting (3.2.18) into (3.1.3)
gives the maximum contact region radius as
 1/5  1/5
 5Mv02 15Mv02 R2
amax = Rδmax = R =
1/2
. (3.2.19)
4K 16E

Equation (3.2.16) is of a separable form, and can be rearranged as

dδN
dt =   . (3.2.20)
v02 − 16 ER1/2 δ 5/2 /M
15 N
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

58 Contact Mechanics without Adhesion

Defining y ≡ δN /δmax and integrating (3.2.20) between the time t = 0 of initial

contact and the time τC /2 of maximum overlap, the total contact time τC is obtained
as
 √
2δmax 1 dy 4 π δmax (2/5) ∼ δ
τC =  = = 2.943 max , (3.2.21)
v0 0 1−y 5/2 5 v0 (2/5 + 1/2) v0

where (·) is the gamma function. Substituting the result (3.2.18) for the maximum
overlap gives the contact time as
 1/5
M2
τC ∼
= 2.868 . (3.2.22)
E 2 Rv0

The prediction that contact time decreases with increase in impact velocity v0 ,
as predicted by (3.2.22), is verified by experimental measurements (e.g., Goldsmith,
1960; Stevens and Hrenya, 2005).
As a specific example, we consider the impact of 1 mm diameter glass spheres.
It is assumed that the approach velocity v0 = 1 m/s, and the elastic modulus, Poisson
ratio, and density for glass are E1 = E2 = 55 GPa, ν1 = ν2 = 0.3, and ρ p = 2500
kg/m3 . For identical spheres, R = r1 /2, M = m1 /2 = (2/3)πr13 ρ p = (16/3)πR3 ρ p ,
and E = E1 /2(1 − ν12 ). Substituting these values into (3.2.18), (3.2.19) and (3.2.21)
gives δmax = 2.22 µm, amax = 33.3 µm, and τC = 6.53 µs. The extremely short con-
tact time obtained in this example illustrates some of the challenges involved in
computing flows with colliding particles, for which the time scale associated with
particle collisions is often six or more orders of magnitude smaller than the time
scale associated with the fluid flow in which the particles are entrained.

3.3. Normal Dissipation Force

In an elastic contact model, the force-displacement response is the same during load-
ing and unloading, implying that the net kinetic energy of the particles is conserved
during the collision. In the practice, there is a loss of kinetic energy to heat during
collision processes that can be characterized by the restitution coefficient e defined in
(2.4.6). A completely elastic collision, such as one governed by the Hertz expression
in the momentum equation (3.2.13), results in a restitution coefficient of unity. A
completely inelastic collision, for which there is no particle rebound, corresponds to
a restitution coefficient of zero. All real collisions occur with restitution coefficients
between these two extremes.

3.3.1. Physical Mechanisms

The kinetic energy loss is in general due to a combination of dissipation within the
particle (solid-phase dissipation) and dissipation within the fluid surrounding the
particle (fluid-phase dissipation). The solid-phase dissipation can be attributed to
a wide range of factors. For instance, in high-velocity impacts, plastic deformation
of the material or localized cracking near the contact region may occur. Because
the current study is concerned with motion of adhesive particles immersed in fluid
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.3. Normal Dissipation Force 59

Fn Fn
A A
Approach Approach

Figure 3.4. Hysteretic behavior of the

load-displacement curve for (a) visco- Rebound Rebound
O O
elastic and (b) plastic material behavior. Β
δN Β δres δN

(a) (b)

flows, high-velocity impacts are generally outside the scope of this text. Interfacial
slip within the contact region during normal impact can lead to kinetic energy loss;
however, as noted by Johnson (1985, 119), such slip only occurs in the normal
collision problem if the elastic moduli of the two colliding particles are different.
Solid-phase dissipation occurring in low-velocity collision of two identical parti-
cles is generally associated with what is called viscoelastic response. Classical elasticity
theory is developed utilizing an assumption that the material deformations change
so slowly that the material is approximately static. Although this quasi-static approx-
imation yields an expression for the normal elastic force of two colliding particles
that is in reasonable agreement with experimental findings, in reality the collisions
occur over relatively short time intervals. The rapid change in particle deformation
during collisions gives rise to a viscous response caused by internal friction within
the solid, which is proportional to the rate of change of the material deformation.
The sum of this elastic response and the viscous response characterizes a viscoelastic
material response, which can be visually identified by the presence of hysteresis in
the force–displacement curve (i.e., the plot of Fn versus δN ). Hysteresis causes the
Fn − δN relationship to be different as the two particles approach each other com-
pared with when they move apart from each other, as illustrated in Figure 3.4.
Differences in viscoelastic and plastic material behavior are apparent in this curve
by the fact that the deformation returns to zero as the load vanishes in a viscoelastic
material, whereas it does not in a plastically deforming material.
The Hertz theory of normal elastic collision also assumes that the particle sur-
faces are perfectly smooth. Of course, all materials are formed of molecules, which
cluster into crystals, grains, and so on, so no surface can be perfectly smooth. Instead,
the notion of smoothness implies that the length scale associated with the roughness
elements on the surface is much smaller than some comparable length scale in the
problem. For the problem of collision of two particles, a natural comparison length
scale is the maximum particle overlap δmax , given for Hertz theory by (3.2.18). It
has previously been discussed that δmax is typically much smaller than the particle
diameter. For instance, in the example given in Section 3.2 for collision of two 1 mm
diameter glass particles, we found that δmax ∼ = 2.2 µm. Because of the small size of
δmax , the contact mechanics of surfaces can be substantially influenced by roughness
even for surfaces with very small roughness elements. For instance, energy dissipa-
tion during a collision process can be caused by sliding friction and localized plastic
deformation as the highest roughness elements on each surface contact each other at
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

60 Contact Mechanics without Adhesion

v2 Sliding
Figure 3.5. Schematic diagram showing the rough-
ness elements on the surfaces of two colliding parti-
cles interacting with each other.
Crushing v1

various angles and are either crushed by high local stresses or slide relative to each
other as the surfaces move closer together (Figure 3.5).
Another important mechanism for reduction of the restitution coefficient is the
presence of adhesion forces between the particle surfaces. This mechanism becomes
important for low-velocity collisions of relatively small particles. Even for cases
where the adhesion forces are far smaller than is necessary to stop the particle
rebound completely, adhesion still exerts a force on particles which decreases the
restitution coefficient. There are many types of adhesive forces that can act between
particles, which are discussed in detail in Chapter 4. Since in models of adhesive
particle flows the adhesion force is treated explicitly in a combined elastic-adhesion
force model (see Chapter 4), we do not in this section include effects of adhesion
forces in modeling Fnd . Nevertheless, adhesion forces are present in nature and
may influence experimental data used for validation of normal dissipative force
expressions.
In addition to the solid-phase dissipation mechanisms discussed earlier, fluid-
phase friction is an important source of dissipation for the low-velocity impacts of
small particles that are typically of concern in adhesive particle flows. The fluid-
phase dissipation is typically dominated by the high shear stress motions within the
lubrication film separating the two particles as they approach or move away from
each other, or else from the fluid motion near the outer edges of the contact region
as the contact region area changes during the collision. Fluid dissipative forces are
discussed in detail in Section 2.4, including forces prior to, during, and following
collision of the particles. The fluid-phase dissipation often dominates effects of the
solid-phase dissipation for low-velocity impacts of small particles, particularly in
cases where the collision Stokes number StC (defined in (2.4.8)) is small. As a
consequence, it is often possible to predict the restitution coefficient quite well in
such cases using the fluid dissipation force alone, as shown in Figure 2.7.
Figure 3.6 presents a compilation of experimental data for restitution coeffi-
cient for particle collision studies with different materials. These experiments were
conducted in air with relatively high impact velocities and sufficiently large parti-
cles so that the fluid-phase dissipation force is small compared with the solid-phase
dissipation force. The relevant mechanical and geometrical properties used for the
data in Figure 3.6 are listed in Table 3.2, along with sources of the data. The data
exhibit a general trend for the restitution coefficient to decrease with increase in
the particle approach velocity ν0 . It is noted from (2.4.8) that the collision Stokes
number StC is proportional to particle approach velocity, and so that the expression
(2.4.21) for restitution coefficient under purely fluid dissipation force is consistent
with this trend. However, as shown in Section 3.3.2, this trend is also consistent with
certain models used for describing the solid-phase dissipation force.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.3. Normal Dissipation Force 61

Table 3.2. Material properties used for measuring the coefficient of restitution in the previous literature
(corresponding to data in Figure 3.6)

Material Ei (GPa) σi ρ p (kg/m3 ) ri (mm)a Investigator

1 Stainless Steel (Grade 316) 193 0.35 7830 12.7 Stevens & Hrenya (2005)
2 High Carbon Steel 211 0.295 8184 16.5 Kuwabara Kono & (1987)
3 Brass(Cu70/Zn30) 110 0.331 8418 15.0
5 Al2 O3 /Ala 360/70 0.23/0.35 3900/6500 5.00 Gorham & Kharaz (2000)
4 Brass 96 0.36 8522 20.0
6 Aluminum 68.5 0.36 2707 20.0 Goldsmith (1960)
7 Lead 16.2 0.45 11373 20.0
8 Marble 71.0 0.30 2900 20.0
9 Cork 0.05 0 1654 16.6 Kuwabara & Kono (1987)
25.4
10 Nylon 0.35 0.3 1140 12.7 Labous et al. (1997)
6.35
a Case of Al2 O3 sphere impacting with Al alloy plate, instead of two-sphere impact.

3.3.2. Models for Solid-Phase Dissipation Force

We consider two colliding particles in the viscoelastic regime. The normal force on
the particles is modeled by the Kelvin rheological model shown schematically in
Figure 3.2, consisting of a spring and a dashpot acting in parallel, such that the spring
and dashpot experience different forces but an equal deformation. The total normal
force on each particle is the sum of the Hertzian elastic force Fne and the normal
dissipative force Fnd . The dissipation force Fnd for a dashpot is proportional to the
approach velocity of the particles, and so is given by
Fnd = ηN vR · n. (3.3.1)
In this equation, vR is the relative particle surface velocity at the contact point and
ηN is the normal dissipation coefficient. The surface velocity of particle i at the
contact point is given for a spherical particle by vC,i = vi + i × ri , where ri = ri n is
Restitution coefficient

Restitution coefficient

Impact velocity v0 (m/s) Impact velocity v0 (m/s)

(a) (b)
Figure 3.6. Experimental data for restitution coefficient as a function of the initial normal approach
velocity, from sources listed in Table 3.2. (a) Viscoelastic regime; (b) viscoelastic and plastic deformation
regimes.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

62 Contact Mechanics without Adhesion

the vector from the particle center to the contact point, vi = dxi /dt is the particle
centroid velocity, and i is the particle angular velocity. For the opposing particle j,
we can write r j = −r j n. The relative particle surface velocity at the contact point is
then given by

vR = vC,i − vC, j (3.3.2)

The normal component of vR can also be written in terms of the normal overlap δN
as
dδN
vR · n = , (3.3.3)
dt
such that vR · n is positive for particles that are approaching each other and negative
for particles that are moving away from each other. Using the result (3.3.3), it is
noted that the expressions for fluid dissipation force both before collision (2.4.14)
and during collision (2.4.15) can be written in the form (3.3.1).
Using both the Hertz theory for the elastic normal force and the expression
(3.3.1) for the dissipative normal force, the governing equation for the normal over-
lap becomes
d 2 δN η dδ K 3/2
+ N N + δN = 0. (3.3.4)
dt 2 M dt M
The restitution coefficient can be written in terms of the time derivative δ̇N of the
normal overlap at time t = 0, just after onset of collision, and at time t = TC , just
before the end of the collision, as
δ̇N (τC )
e=− . (3.3.5)
δ̇N (0)
A wide variety of approaches have been taken in the literature to determine
expressions for the normal dissipation coefficient ηN . The existence of many different
approaches for modeling normal dissipation is not surprising given the different
physical mechanisms that give rise to the solid-phase dissipation. Lee and Herrmann
(1993) assumed that ηN is constant and proportional to the reduced mass M, such
that

ηN = αLH M, (3.3.6)

where αLH is a constant. This model predicts that the coefficient of restitution will
increase with increase in impact velocity δN , which is opposite to the experimental
trend shown in Figure 3.6.
Kuwabara and Kono (1987) and Brilliantov et al. (1996) examined the normal
dissipation force from first principles by making the assumption that the deformation
in the constitutive equation for the dissipative part of the stress tensor can be replaced
by the deformation predicted by the purely elastic Hertz theory. This quasi-static
approximation is valid for values of restitution coefficient close to unity, and leads to
the conclusion that the normal dissipation coefficient is proportional to the contact
1/2
area radius a(t ), or ηN ∝ δN . Brilliantov et al. (1996) derived an expression for the
normal dissipation coefficient as
1/2
ηN = αKKB KδN , (3.3.7)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.3. Normal Dissipation Force 63

K/M=98360 K/M=98360

Scaled overlap, velocity,

Scaled overlap, velocity,

and acceleration
and acceleration

t /τ0 t /τ0
(a) (b)
Figure 3.7. Time variation of the scaled overlap, velocity, and acceleration for two colliding particles,
computed using the Kuwabara-Kono-Brilliantov model for the normal dissipation coefficient. (a) Weak
damping; (b) strong damping.

where K is the Hertz stiffness coefficient given by (3.2.9) and αKKB is a material
constant with units of time. Substituting (3.3.7) into (3.3.4) gives
d 2 δN K 1/2 dδN K 3/2
2
+ αKKB δN + δN = 0. (3.3.8)
dt M dt M
For initial conditions given by δN (0) = 0 and δ̇N (0) = v0 , the Hertz theory for two
purely elastic spheres yields a length scale δ0 and a time scale τ0 with order of
magnitude
 −2/5  −2/5
K K
δ0 = v04/5 , τ0 = v0−1/5 . (3.3.9)
M M
A plot showing the time variation of δN , δ̇N , and δ̈N during the collision process,
made dimensionless using the length and time scales in (3.3.9), is given in Figure 3.7.
The computations are performed using material constants for Al2 O3 /Al, which are
given in Table 3.2, and particles with diameter 10 mm, impact velocity v0 = 2 m/s,
and K/M = 98,360 m−1/2 s−2 . Results are given in Figure 3.7 for two cases, one with
fairly weak dissipation (αKKB = 2.5 × 10−6 s) and restitution coefficient e = 0.73, and
another with strong dissipation (αKKB = 1.0 × 10−5 s) and a much smaller restitution
coefficient of 0.29. The time required for the collision is somewhat delayed by the
dissipative term compared with what it would have been in a purely elastic case,
which is indicated by a vertical dashed line in Figure 3.7.
These restitution coefficients are typically obtained using the assumption that
the beginning and end of the collision correspond to times at which δN = 0. However,
it is observed in Figure 3.7 that the second derivative term δ̈N becomes positive near
the end of the collision. The total normal force Fn is proportional to −δ̈N , so a change
from negative to positive values of δ̈N implies a change in sign in Fn from positive
to negative. It is recalled that a positive value of Fn acts to push the particles apart,
whereas a negative value of Fn acts to pull the particles together. This change in sign
of Fn occurs when the magnitude of the normal dissipation force term Fnd exceeds
that of the normal elastic force term Fne . As the particles move toward each other
both Fne and Fnd have a positive sign, resisting the particle motion toward each
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

64 Contact Mechanics without Adhesion

other. However, as the particles move away from each other the sign of Fne remains
positive whereas that of Fnd becomes negative, such that the two normal forces
oppose each other. As δN approaches zero near the end of the collision process, the
magnitude of the elastic force Fne approaches zero more quickly than does that of
the dissipative force Fnd because of the larger 3/2 power of δN in (3.2.8) compared to
the 1/2 power in (3.3.7). Schwager and Pöschel (2008) argue that this state in which the
total normal force is negative is nonphysical, and that in reality the particle surfaces
would separate with the particle still in a deformed state, so that the collision should
end at the time at which Fn = 0 rather than the time at which δN = 0.
Determination of the time corresponding to the end of the collision process
is significant for determining the value of the restitution coefficient. Schwager and
Pöschel (2008) derived an approximate analytical solution of (3.3.8) by expanding

the equation in powers of t/τ0 , where τ0 is the Hertzian contact time scale given
in (3.3.9). This expansion yields a series approximation for restitution coefficient
e, based on the usual assumption that the collision terminates at a time for which
δN = 0, as
∞
  n
e=1+ cn αKKB (K/M)2/5 v0n/5 . (3.3.10)
n=0

An alternative expression for e was derived using the observation that the particles
will actually lose physical contact while still in a deformed state (such that δN > 0)
at a time for which Fn = 0, which yields the restitution coefficient as
∞
  n/2 n/10
e=1+ hn αKKB (K/M)2/5 v0 . (3.3.11)
n=0

The values of the coefficients cn and hn are listed in Table 3.3. A plot com-
paring restitution coefficient results obtained from these analytical solutions with
results from a numerical solution of (3.3.8) and from experimental data is given in
Figure 3.8. The KKB model qualitatively predicts the trend of decreasing e with
increasing impact velocity. However, the model significantly underpredicts the slope
of the experimental data. Stevens and Hrenya (2005) observed a similar trend for
predictions of the KKB model for impact of stainless steel spheres.
Numerous other expressions for ηN have been proposed, but of the available
alternative expressions that proposed by Tsuji et al. (1992) is of particular interest
because it leads to a value for the restitution coefficient that is independent of impact
velocity. In this expression, an assumption is made that ηN ∝ (MkN )1/2 , where kN is
the normal stiffness coefficient, such that
√ 1/4
ηN = αT T I MKδN , (3.3.12)
where αT T I is a dimensionless constant. Substituting this expression into (3.3.4) gives

d 2 δN K 1/4 dδN K 3/2
+ αT T I δ + δN = 0. (3.3.13)
dt 2 M N dt M
Setting tˆ = t/τ0 and δ̂ = δN /δ0 gives
d2 δ̂ dδ̂
+ αT T I δ̂ 1/4 + δ̂ 3/2 = 0. (3.3.14)
dtˆ2 dtˆ
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.3. Normal Dissipation Force 65

Table 3.3. Table of the first ten values of cn and the first twenty
values of hn for the Schwager and Pöschel (2008) analytic
solution of the restitution coefficient from the KKM model

n cn hn

1 −1.153 448 854 0

2 0.798 266 555 3 −1.153 448 856
3 −0.522 882 560 9 0
4 0.348 742 667 8 0.798 266 558 1
5 −0.233 098 126 0 0.266 666 666 7
6 0.156 682 147 7 −0.522 882 565 7
7 −0.105 818 782 8 −0.461 379 542 4
8 0.071 765 282 42 0.348 742 673 7
9 −0.048 857 172 37 0.452 351 049 6
10 0.033 373 471 94 −0.146 431 464 4
11 −0.367 728 299 2
12 −0.043 248 9833
13 0.281 804 232 5
14 0.147 852 5872
15 −0.179 442 059 0
16 −0.178 466 032 6
17 0.065 933 588 82
18 0.171 358 617 8
19 0.025 249 822 3
20 −0.137 923 498 6

The restitution coefficient e is therefore affected only by the coefficient αT T I in

(3.3.14), as all other coefficients in both the governing equation and the initial
condition have been removed by scaling. The TTI model is particularly suitable for
materials in which the restitution coefficient exhibits weak dependence on impact
velocity over the range of interest in the problem, for which it is simplest to prescribe
a fixed value for the restitution coefficient.
Restitution coefficient

Impact velocity v0 (m/s)

Figure 3.8. Comparison of predictions of the KKB model for normal dissipation from Equa-
tion (3.3.8) (solid line) with experimental data of Gorham and Kharaz (2000) (symbols).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

66 Contact Mechanics without Adhesion

α TTI

Restitution coefficient, e
Figure 3.9. The restitution coefficient as a function of the damping coefficient αT T I for two-
particle collision predicted by the TTI model.

Equation (3.3.14) was solved numerically using a fourth-order Runge-Kutta

method. In each case, the overlap increases to some maximum value δ̂ = δ̂max at a
time tˆ = tˆmax , and then decreases again to zero as the two particles rebound from
each other. In Figure 3.9, the values of δ̂max , tˆmax , and e are plotted as functions of
αT T I both for cases where the collision is assumed to end at δ̂ = 0 and for cases where
it is assumed to end at Fn = 0.
Despite these various developments, many issues remain in selecting an accu-
rate, simple model for the normal dissipation force that includes all of the various
mechanisms that may be significant for a given problem. Part of the challenge is that
models for solid-phase dissipation force are typically developed for large particles
colliding at high velocities, as might be typical of granular flows, whereas those for
fluid-phase dissipation force are often suited more for low-velocity collisions. Simi-
larly, models developed for the normal dissipation force often account for only one
of the various mechanisms discussed here and neglect the others, making validation
difficult when multiple mechanisms are significant. Models for fluid-phase dissipation
force can give rise to severe stiffness issues in the numerical analysis, so even if the
models are accurate they can be challenging to implement in practical models with
large particle numbers. For all of these reasons, research into improved expressions
for normal dissipation force of colliding particles continues to be an active research
area.

3.4. Hysteretic Models for Normal Contact with Plastic Deformation

An alternative approach for modeling the normal elastic and dissipative forces is
the so-called hysteretic model, in which the unloading spring stiffness kU is much
larger than the loading stiffness kL , with the difference presumably due to plastic
deformation. This difference in spring stiffness results in hysteresis in the force-
deformation diagram, leading to energy loss during the collision. The important
contributions to this class of model were pioneered by Walton and Braun (1986),
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.4. Hysteretic Models for Normal Contact with Plastic Deformation 67

Scaled normal overlap and velocity

1/2
δN(k/M) / v0

(dδN /dt) / v0

1/2
t / (M/kL )
Figure 3.10. Time variation of the scaled overlap and velocity for two colliding particles,
computed using the Walton-Braun (1986) model with contact e.

followed by the work by Sadd and Tai (1993), Thornton (1997), and Vu-Quoc and
Zhang (1999). Both linear and nonlinear (i.e., Hertzian) stiffness coefficients have
been used in the literature.
The simple linear hysteretic model, first proposed by Walton and Braun (1986),
is illustrated in Figure 3.4b. The normal contact force is assumed in this model to
have the form

kL δN Loading
Fn = (3.4.1)
kU (δN − δres ) Unloading

where δres is the residual overlap at the end of the unloading period, known as
the final resulting plastic deformation between two colliding particles. Substituting
this linear hysteretic model into the momentum equation (3.2.13) for two-particle
collision, an analytic solution can be obtained. Results of this solution for the scaled
normal overlap and impact velocity are shown in Figure 3.10. As two particles initially
collide and move from the origin O to point A in Figure 3.4b under the stiffness kL ,
this solution gives the maximum overlap and normal force as δmax = v0 (M/kL )1/2 and
Fn,max = v0 (MkL )1/2 , respectively. As the particles move apart, going from point A to
point B in Figure 3.4b under a stiffness kU , the residual normal overlap is obtained
as δres = (1 − kL /kU )δmax = v0 (1 − kL /kU )(M/kL )1/2 . The linear hysteretic model
leads to a coefficient of restitution given by

e = kL /kU , (3.4.2)

such that the duration of contact is

π 1/2  −1/2 −1/2 
τ0 = M kL + kU . (3.4.3)
2
Because it is not practicable to keep track of the effect of plastic deformation on
the shape of each particle over time, the model assumes that particles start off as
spherical at each new contact. Put another way, as Fn approaches zero during the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

68 Contact Mechanics without Adhesion

unloading process, the model “forgets” the plastic deformation, so that point B
approaches the origin O.
This model leads to a constant value of e regardless of the value of the approach
velocity v0 , which is inconsistent with the experimental trend of the e − v0 curve
shown in Figure 3.6b in the plastic deformation regime. To address this issue, Walton
and Braun (1986) modified their model by relating the unloading stiffness kU to both
kL and Fn,max in a linear manner,

kU = kL + sFn,max , (3.4.4)

where s is an empirical parameter with units of m−1 . Using Fn,max = v0 (MkL )1/2 , this
assumption yields an expression for restitution coefficient as
 1
e= kL /kU =  . (3.4.5)
1 + sv0 (M/kL )1/2

Stevens and Hrenya (2005) showed that the Walton-Braun model with variable e
exhibits reasonably good predictions for restitution coefficient when compared with
experiment data for impact of stainless steel spheres, but that this model underpre-
dicts the influence of the impact velocity on the contact time scale. The experimental
data reported in this paper for restitution coefficient generally lie between the pre-
dictions of this model and those of the KKB model discussed in Section 3.3. Of
course, it should be kept in mind that these comparisons were performed with two
empirical fitting coefficients, kL and s.
The existing experimental data for particle collisions (Stevens and Hrenya, 2005;
Kruggel-Emden et al., 2007) indicate that although linear models can be tuned to
provide reasonable predictions for restitution coefficients, a full nonlinear (Hertzian)
model is necessary to also obtain accurate predictions of contact duration time.
Thornton (1997) proposed a nonlinear hysteretic model for the contact of two elastic,
perfectly plastic spheres. In this model, the collision of two approaching particles is
decomposed into three parts. In the first part (elastic loading), the particle loading
behaves like an elastic material as governed by the Hertz theory until the material
yield point is reached. In the second part, the yield point is reached and the material
behaves in a perfectly plastic manner as the loading is continued. The third part
of the process involves unloading as the particles move away from each other. The
normal force on the particle is assumed to have the following form in each of these
three parts of the collision process:
⎧
⎪
⎨ KL δN3/2
δN < δY for v0 ≥ 0
Fn = KL δY3/2 + π py R(δN − δY ). δN ≥ δY for v0 ≥ 0 (3.4.6)
⎪
⎩ KU (δN − δY )3/2 for v0 < 0

where KL and KU are the coefficients of stiffness during loading and unloading (given
by (3.2.9)), py is the material yield point (a material property), and δY is the normal
overlap at the yield point, which can be calculated by δY = R(π py /2E)2 . Starting
from Hertzian contact theory, the Thornton model provides very good estimates for
contact time but it is reported to overpredict the dependency of e on v0 compared
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.5. Sliding and Twisting Resistance 69

with the experimental data. An expression for the restitution coefficient can be
obtained from the model as
⎡ ⎤1/4
1/2
(vy /v0 )2 vy /v0
e = 1.442 1 − ⎣  ⎦ . (3.4.7)
6 vy /v0 + 2 [6 − (vy /v0 )2 ]/5

Because it does not include viscoelastic effects, the Thornton model predicts that
e = 1 for v0 < vy , where vy refers to the yield velocity for plastic deformation esti-
mated by vy = 3.194p5/2 y R
3/2
/E 2 M1/2 (e.g., for stainless steel vy is about 0.3 m/s).
This prediction is at odds with experimental data, such as those shown in Figure 3.6b,
which exhibit substantial reduction in restitution coefficient by viscoelastic effects
prior to onset of the plastic deformation regime.

3.5. Sliding and Twisting Resistance

Sliding occurs when the surface of one particle slips relative to that of another particle
during contact, due to relative tangential motion of the particle surfaces within the
contact region. Dissipation of kinetic energy is associated with this slip, where the
rate of energy dissipation is equal to the slip velocity times the slip resistance force
Fs . The slip force acts in the direction of the slip velocity vS (t ), which is the tangent
projection of the relative surface velocity vR to the particle surface at the contact
point, or
vS = vR − (vR · n)n, (3.5.1)
where vR is defined in (3.3.2). The slip direction is given by the unit vector tS =
vS /|vS |. Because the sliding force acts at the contact point in a direction that does
not pass through the particle centroid, there is an associated sliding torque, given for
particle i by ri Fs (n × tS ).
Twisting is closely related to sliding, with the difference that in a pure twisting
motion the slip velocity at the contact point vanishes, but there is slip at surrounding
points within the contact region due to difference in the rotation rate of the particles
in a direction along the normal vector n. For rigid-body motions, in sliding the slip
velocity within the contact region is equal to that at the contact point, whereas in
a twisting motion the slip velocity increases linearly within the contact region with
distance from the contact point. The resulting particle response to the twist motion
is a torque Mt oriented parallel to the unit normal vector n. As discussed in the next
section, elastic deformation of the material modifies this observation, allowing slip
between the two surfaces in only part of the contact region.

3.5.1. Physical Mechanisms of Sliding and Twisting Resistance

Sliding resistance arises from energy dissipation in both the fluid and the solid phases.
In the solid phase, sliding is resisted by impact of the roughness elements on the two
surfaces as they move relative to each other. The presence of a fluid within the gap
between the sliding surfaces introduces additional friction within the fluid phase. For
sliding with perfectly smooth surfaces, the sliding resistance would be entirely due
to the fluid friction, the magnitude of which increases with decrease in gap thickness
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

70 Contact Mechanics without Adhesion

Table 3.4. Typical values of static friction coefficient for

collision of two bodies made of the same material. [Extracted
from data in Engineering Tool Box (http://www.
engineeringtoolbox.com).]

Static friction coefficient

Materials Dry Lubricated

Metals
Aluminum 1.05–1.35 0.3
Iron 1.0 0.15–0.2
Copper 1.0 0.08
Graphite 0.1 0.1
Magnesium 0.6 0.08
Platinum 1.2 0.25
Silver 1.4 0.55
Steel 0.8 0.16
Zinc 0.6 0.04
Other Materials
Glass 0.9–1.0 0.1–0.6
Nylon 0.15–0.25
Polystyrene 0.5 0.5
Teflon 0.04 0.04

and approaches infinity as the gap thickness approaches zero. However, for a finite
sliding velocity the fluid pressure within the contact region also increases as the
gap thickness decreases. As a result, the presence of the fluid tends to increase the
gap thickness between the surfaces, thereby decreasing the sliding resistance and
providing a lubricating effect.
The problem of solid-phase sliding resistance between two bodies with uniform
pressure distribution between the body surfaces is a classical problem of mechanics,
which was largely solved in the eighteenth century by scientists such as Amontons,
Euler, Belidor, and Coulomb. As taught in elementary physics books, two bodies
in contact will not slide relative to each other until the external tangential force
Ftang between the two bodies exceeds a critical value Fcrit , equal to the product of
the normal force Fn and a static friction coefficient μS . The value of the friction
coefficient depends on the materials under consideration, the nature of the surface
(e.g., whether it is polished or rough), the gap thickness between the surfaces, and
the properties of any fluid contained within this gap. Since it depends on so many
different factors, determination of friction coefficient can often be quite uncertain.
A listing of typical friction coefficients is given in Table 3.4, but the values listed
in this table are based only on the type of material and do not include important
features such as surface preparation and fluid lubrication.
The case where the external tangential force between the bodies, Ftang, is less
than the critical force Fcrit = μS Fn is called static friction, because in this state the
two surfaces are at rest relative to each other. In this state, the response force Fs
acting on each body is equal to the imposed tangential force Ftang and the surfaces
are in equilibrium. If Ftang exceeds this critical force, then the surfaces begin to
move relative to each other and the surfaces are in a state called dynamic friction,
or sometimes kinetic friction. The response force during dynamic friction can be
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.5. Sliding and Twisting Resistance 71

μp qx

Shear stress
Figure 3.11. Plot showing the variation of shear
stress for the case with no slip (qx ) and the criti-
cal shear stress (μp) as functions of radius within
a contact region of radius a, for particle collision
subject to a tangential force Ftang. The shear stress
for the case with partial slip is indicated by a solid
line, where slip starts at radius r = c.

0 c a r
written as Fs = μ f Fn , where the dynamic friction coefficient μ f is generally smaller
than the static friction coefficient μS for a given problem. The dynamic friction
coefficient is often taken as a constant, particularly for cases in which the range of
the sliding velocity is fairly narrow. However, in general, dynamic friction coefficient
is a function of surface roughness, temperature, and slip velocity (Sang et al., 2008).
It was pointed out by Mindlin (1949) and Mindlin and Deresiewicz (1953) that
because the Hertz pressure field (3.2.1) is not uniform over the contact region,
different regions of the contact surface begin to slip at different times. A given point
on the contact surface will start slipping when the local shear stress in the direction
of tangential loading, qx , exceeds the product of the static friction coefficient and
the local pressure, or μS p. According to the Hertz pressure distribution (3.2.1), the
lowest pressure value occurs at the outer edge of the contact region and the pressure
increases as one progresses inward, achieving a maximum at the contact point. If a
tangential force difference with magnitude Ftang is imposed on two colliding particles
with no slipping, a shear stress qx forms within the contact region with the radial
distribution
qx = q0 (1 − r2 /a2 )−1/2 , (3.5.2)
where q0 = Ftang/2πa2 . Thus, the largest shear stress occurs on the outer part of the
contact region. As a consequence, slip starts within a ring near the outer part of
the contact region and progresses inward, with decreasing radius c as the tangential
force is increased. The ratio of the slip ring radius c to the contact region radius a is
given by (Johnson, 1985, 218)
c  1/3
= 1 − Ftang/μS Fn . (3.5.3)
a
As shown in Figure 3.11, in the partial slip case the shear stress increases with
radius from the contact point and the pressure decreases with radius until the slip
ring is reached, at which point qx = μS p. For radii greater than this value, the shear
stress remains proportional to the pressure and hence decreases with radius. As
the value of Ftang is increased, the radius of the slip ring grows smaller, until at
Ftang = μS Fn the slip ring radius decreases to zero and the two surfaces are free to
translate relative to each other in a state of dynamic friction. A plot of the tangential
displacement obtained from the solution with partial slip is given in Figure 3.12, in
comparison to the linear displacement variation that would be obtained with no slip.
For small values of tangential force the tangential displacement predicted by the two
approaches are similar, but as the tangential force increases to its maximum value
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

72 Contact Mechanics without Adhesion

1.0

Fs / μFn
Figure 3.12. Plot of the tangential displacement δT
of two spheres in static contact as a function of the
ratio of tangential force to critical tangential force
for cases with no slip (dashed line) and partial slip
(solid line).

0
0 1.0

16aδ T / 3μFn

the displacement for the partial slip case increases to nearly twice the value for the
no-slip case.

3.5.2. Sliding Resistance Model

In computations with the discrete element method, the tangential force on two
colliding particles is not generally known. It is therefore necessary to introduce
some type of model that relates the tangential force to known quantities, such as
the particle centroid positions or velocity. A variety of options have been proposed
in the literature, including the assumption
  that the tangential force is proportional
to the normal force, or Fs = −μ f Fn  (Haff and Werner, 1986; Schäfer and Wolf,
1995), and the assumption that the tangential force is proportional to the sliding
velocity, or Fs = −ηT vS · tS (Taguchi, 1992; Gallas et al., 1992; Luding et al., 1994).
However, as discussed in the review by Schäfer et al. (1996), both of these choices
lead to qualitative inconsistencies with experimental data for sliding. One of the most
successful models for sliding resistance is the spring-dashpot-slider model proposed
by Cundall and Strack (1979) and subsequently used in a large number of studies.
In this model, shown schematically in Figure 3.13, the sliding resistance in the static
friction state is absorbed by a combination of a spring and a dashpot. The force due
to the spring is proportional to the particle tangential displacement δT , given by
 t
δT = vS (ξ ) · tS dξ , (3.5.4)
t0

kS ηS
Figure 3.13. Schematic diagram of the spring-
kN dashpot-slider model proposed by Cundall and
Strack (1979) for the sliding resistance two col-
liding particles.
ηN
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.5. Sliding and Twisting Resistance 73

where t0 is the time of initial particle impact. The dashpot force is linearly propor-
tional to the sliding velocity vS . The complete Cundall-Strack model thus has the
form

Fs = −kT δT · tS − ηT vS · tS . (3.5.5)

An expression for the tangential stiffness coefficient kT is derived by Mindlin

(1949) and can be written in terms of the effective shear modulus G, defined in
(3.1.1), and the contact region radius a(t ) as

kT = 8Ga(t ). (3.5.6)

Many investigators omit the dashpot term in (3.5.5), or else assign ηT a constant
value. Other authors (e.g., Tsuji et al., 1992; Cleary et al., 1998) assume that ηT is
approximately equal to the normal viscous damping coefficient ηN .
When the magnitude of the tangential force reaches a critical value Fcrit = μS |Fn |,
the surfaces of the two particles start to slide relative to each other and the system
enters into a condition of dynamic friction. This value for the critical force is the
same as the value of Ftang for which the radius c of the slip ring, given in (3.5.3),
goes to zero. If |Ftan g| > Fcrit , the particle surfaces slip relative to each other, which
is represented by the slider in Figure 3.11. In this dynamic friction state, the sliding
response force is given by
 
Fs = −μ f Fn  . (3.5.7)

3.5.3. Twisting Resistance Model

Twisting occurs when the two colliding particles have different rotation rate along
the normal direction n, as illustrated in Figure 3.3c. A relative twisting rate T is
defined by

T = (i −  j ) · n. (3.5.8)

A rotational spring-dashpot-slider model for twisting resistance has a form analogous

to (3.5.5), giving the twisting resistance torque in the normal direction as
 t
Mt = −kQ T (τ )dτ − ηQ T . (3.5.9)
t0

The time integral in this expression is equal to the angular displacement of the particle
since the time of initial particle impact (t0 ). The torsional stiffness and dissipation
coefficients can be related to the linear stiffness and dissipation coefficients for sliding
as

kQ = kT a2 /2, ηQ = ηT a2 /2. (3.5.10)

Similar to the case for tangential force, slip within the contact region first occurs
within a radial ring starting at the outer edge of the contact region and progressing
inward as the imposed twisting torque is increased. The radius of the slip ring shrinks
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

74 Contact Mechanics without Adhesion

to zero when the magnitude of the imposed twisting moment equals a critical value
Mt,crit , whose value is given by (Johnson, 1985, 233)
3π
Mt,crit = aF , (3.5.11)
16 crit
 
where Fcrit = μS Fn  is the critical linear tangential force for sliding of the particle
surfaces. For imposed twisting moments with magnitude greater than Mt,crit , the
torsional resistance is given by
3π    
Mt = − aμ f Fn  T / T  . (3.5.12)
16

3.6. Rolling Resistance

Rolling occurs when particle rotation rates and centroid translation velocities are
set such that the particles can move tangentially relative to each other without
slipping. Rolling motion is known to be a primary mechanism by which assemblages
of particles deform under a shearing load, particularly in cases with densely packed
particles such as in granular materials or agglomerates of adhesive particles (Oda
et al., 1982; Bardet, 1994; Iwashita and Oda, 1998). Although the rolling motion of
a sphere generally has lower energy dissipation than an equivalent sliding motion,
there is nevertheless a resistance to rolling that takes the form of a torque imposed on
the particle in such a direction as to decrease the rolling velocity. As a consequence,
a spherical particle set into a rolling motion on a horizontal planar substrate will
not continue to roll forever, but will gradually slow down and eventually come to
rest. There are two separate challenges associated with incorporation of the rolling
resistance in discrete element models. The first is the challenge of clearly defining
the rolling velocity for arbitrary motions of a particle assemblage. Although it is a
simple thing to identify rolling of a particle on a plane, it is no simple matter to
clearly distinguish between rolling, sliding, and rigid body rotation of a group of
colliding particles moving in different directions. Once the rolling velocity is clearly
defined, the second challenge is to understand the physical mechanism that gives
rise to rolling resistance and to introduce a simple model that accurately represents
this mechanism. This also is not a simple matter, as multiple mechanisms play a role
in determining the rolling resistance.

3.6.1. Rolling Velocity

It is desired to define a velocity that characterizes rolling motion during collision
of two spherical particles, labeled Particle 1 and Particle 2, with radii r1 and r2 ,
respectively. To simplify the problem, we first consider the case of two-dimensional
motion in which all motion lies in the x-y plane. The positions of the two particles
are sketched in Figure 3.14 at times t and t + t. The centroid position vectors are
given by x1 and x2 at time t and by x1 and x2 at time t + t, so the displacement
vectors for each particle over this time interval are given by

du1 = x1 − x1 , du2 = x2 − x2 . (3.6.1)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.6. Rolling Resistance 75

x2 x'2
+ + A'2
dθ2
t n
dβ n'
A
Figure 3.14. Schematic diagram showing rolling B
dβ dθ1 A'1
rotation of two particles with radii r1 and r2 at times
(a) t and (b) t + t. +
x1 +x'
1

(a) (b)

The unit normal vector pointing from the centroid of Particle 1 to that of Particle
2 is denoted by n at time t and by n at time t + t. The angle between vector
n and vector n is denoted by dβ. A unit tangent vector t is defined in the x-y
plane in a direction tangent to the particle surfaces at the contact point, and a third
unit vector b = n × t is defined in a direction orthogonal to the plane of motion.
Vectors r1 and r2 extending from the particle centroids to the contact point are
defined by

r1 = r1 n, r2 = −r2 n, (3.6.2)

so dβ can be written as
(du2 − du1 ) · t (du2 − du1 ) · t
dβ = = . (3.6.3)
r 1 + r2 (r1 − r2 ) · n
The numerator in (3.6.3) denotes the relative tangential displacement of the particle
centroids during the time interval of length t.
In addition to translation of the centroids, each particle rotates about its centroid
by an angle dθ1 and dθ2 , respectively, during the time interval, where a positive angle
indicates rotation in the counterclockwise direction. In Figure 3.14a, the contact point
at time t is denoted by A. The material point on each particle moves during the time
interval to occupy positions denoted by A1 and A2 on the two particles at time t + t.
The new contact point at time t + t is denoted by B in Figure 3.14b. The arc length
along the surface of Particle 1 from point B to point A1 at time t + t is denoted by
da1 . Similarly, the arc length from B to point A2 along the surface of Particle 2 is
denoted by da2 . We take the arc length as being positive if the corresponding angle
is in the counterclockwise direction, and negative otherwise. Referring to Figure
3.14b, the arc lengths can be written as

da1 = r1 (dθ1 − dβ ), da2 = r2 (dθ2 − dβ ). (3.6.4)

Pure rolling motion occurs at the contact point if da1 = −da2 , whereas pure
sliding motion occurs when da1 = da2 . In a rigid-body rotation, da1 = da2 = 0.
Following Iwashita and Oda (1998), we define a rolling displacement duL by

duL = 12 (da2 − da1 )t. (3.6.5)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

76 Contact Mechanics without Adhesion

For either pure sliding motion or pure rigid body rotation duL = 0, whereas for pure
rolling motion duL = da2 t. Substituting (3.6.4) into (3.6.5) gives

duL = 12 [r2 dθ2 − r1 dθ1 − (r2 − r1 )dβ]t. (3.6.6)

Substituting the expression (3.6.3) for dβ into (3.6.6) and using the notation (3.6.2),
we can write the rolling displacement as

1 (r + r2 ) · n
duL = − dθ1 × r1 + dθ2 × r2 − 1 [(du2 − du1 ) · t]t . (3.6.7)
2 (r1 − r2 ) · n
The vectors dθ1 and dθ2 are oriented in the b direction, so that the cross products
dθ1 × r1 and dθ2 × r2 in (3.6.7) are oriented in the t direction.
The problem of defining rolling displacement for arbitrary motion of two collid-
ing particles of arbitrary shape was examined by Bagi and Kuhn (2004). They found
that for three-dimensional motion of two colliding spherical particles, the result
(3.6.7) applies provided that the product [(du2 − du1 ) · t]t in the last term is replaced
by the projection of the difference du2 − du1 in the plane normal to n, which is given
by [n × (du2 − du1 )] × n. The rolling displacement for general three-dimensional
collisions of spherical particles is then given by

1 (r + r2 ) · n
duL = − dθ1 × r1 + dθ2 × r2 − 1 [n × (du2 − du1 )] × n . (3.6.8)
2 (r1 − r2 ) · n
The rolling velocity vL can be obtained by dividing duL by the time increment
t and taking the limit t → 0, which yields

1 (r + r2 ) · n
vL = − 1 × r1 + 2 × r2 − 1 [n × (v2 − v1 )] × n . (3.6.9)
2 (r1 − r2 ) · n
An alternative expression for the rolling velocity in terms of the slip velocity vS ,
defined in (3.5.1), can be obtained by writing

vS = vR − (vR · n)n = (n × vR ) × n. (3.6.10)

Recalling that the relative particle surface velocity vR = vC,i − vC, j , where vC,i =
vi + i × ri is the surface velocity of particle i at the contact point, we can write for
the two-particle collision problem

vR = vC,1 − vC,2 = v1 − v2 + 1 × r1 − 2 × r2 . (3.6.11)

Because the cross-product terms in (3.6.11) are already tangent to the unit normal
vector n, substituting (3.6.11) into (3.6.10) gives

vS = [n × (v1 − v2 )] × n + 1 × r1 − 2 × r2 . (3.6.12)

Solving for [n × (v2 − v1 )] × n = −vS + 1 × r1 − 2 × r2 from (3.6.12) and substi-

tuting into (3.6.9) yields an expression for the rolling velocity as
1 r2 − r1
vL = R(2 − 1 ) × n + v , (3.6.13)
2 r 1 + r2 S

where R is the effective radius defined in (3.1.1).

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

3.6. Rolling Resistance 77

Ω2 = Ω
+ y=2r

n
Figure 3.15. Sketch of the collision of two equal-size particles, as used for the t
example in Section 3.6.1.
y

0+ x
Ω 1 = −Ω

For the special case of two equal-size spheres, (3.6.13) reduces to the simple
expression
vL = R(2 − 1 ) × n, (3.6.14)
indicating that in this case rolling velocity is proportional to the difference in rotation
rate of the two colliding particles. As an example, we consider a case with two equal-
size particles with radius r p = 2R in a state of pure rolling with the centroid of particle
1 at the origin of a Cartesian coordinate system and that of particle 2 at (x, y, z) =
(0, 2r p , 0), as sketched in Figure 3.15. The rotation rates of the particles are given by
1 = −e2 and 2 = −ez , thus satisfying the condition r1 1 + r2 2 = 0 for pure
rolling in two dimensions. Then, n = ey and (2 − 1 ) × n = −2ex , so (3.6.15)
gives vL = −r p ex . Letting b = ez gives t = b × n = −ex for this example, so we can
write vL = r p t. The rolling velocity in this pure-rolling example is therefore equal
to the difference between the particle centroid velocity and the particle material
velocity at the contact point.

3.6.2. Physical Mechanism of Rolling Resistance

The primary physical mechanisms underlying rolling resistance in low-velocity col-
lisions of nonadhesive particles are the solid-phase viscoelastic resistance and the
viscous fluid resistance. Adhesive force introduces an important third mechanism
for rolling resistance, which is discussed in Section 4.2. A detailed examination of the
solid-phase rolling resistance is given by Tabor (1955), who concluded that although
slip does occur during rolling motion in the outer part of the contact region, the mag-
nitude of the resistance torque that would be caused by slip is insufficient to account
for the observed rolling resistance. The relationship between rolling resistance and
the normal dissipative force that occurs during particle impact was discussed by Bril-
liantov and Pöschel (1998, 1999). A schematic diagram showing the different solid
stresses acting on a particle rolling in the counterclockwise direction along a flat
plane is given in Figure 3.16. Both elastic and dissipative stresses act in the direction
normal to the planar surface of the contact region, as discussed in Sections 3.2 and
3.3. The elastic stress depends only on the displacement of the particle surface and
it is therefore the same on both sides of the rolling particles. Consequently, the
elastic stress produces a normal force in the vertical direction in Figure 3.16, but
no net moment on the particle. As discussed in Section 3.3, the dissipative stress
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

78 Contact Mechanics without Adhesion

Particle
Figure 3.16. Schematic diagram showing a particle
Ω rolling to the left (with rotation in the counterclock-
+ wise direction) on a plane surface. The direction of the
elastic stress on each side of the particle is indicated by
black arrows, and the direction of the dissipative stress
Elastic stress is indicated by white arrows.

Dissipative stress

acts in a direction so as to resist the relative motion of the particle and the plate. For
the rolling particle shown in Figure 3.16, the left-hand side of the particle is moving
downward into the plate, so that the dissipative stress is positive upward on the
left-hand side. The right-hand side of the particle is moving upward away from the
planar surface, and so the dissipative stress on the right-hand side is downward.
The dissipative stresses on the two sides of the particle cancel out one another, and
therefore produce zero net vertical force. However, the dissipative stresses do pro-
duce a torque on the particle acting in the clockwise direction, resisting the rolling
motion of the particle.
For very small particles immersed in a fluid medium, energy dissipation due
to the fluid viscous shear also makes a significant contribution to the overall rolling
resistance. For the problem of a sphere in contact with a plane wall, application of the
incompressible Navier-Stokes equations together with the usual no-slip boundary
conditions to the fluid surrounding a perfectly smooth sphere results in the conclu-
sion that an infinite force on the particle is required in order for the particle to roll!
This singularity is resolved in practice by small deformation of the particle and micro-
scopic roughness of the particle surface, both of which lead to formation of a small gap
between the particle surface and the plane wall within the contact region. A detailed
examination of the effect of a wall on the forces and torque acting on a nearby
particle, including the case of a particle rolling along a wall, is given in Section 5.6.

3.6.3. Model for Rolling Resistance

Iwashita and Oda (1998) propose a rotational spring-dashpot-slider model for rolling
resistance similar to the Cundell-Strack model for sliding resistance. In this model,
the rolling torque Mr is set equal to the sum
Mr = −kR ξR − ηR vL · tR , (3.6.15)
!t
where ξR = t vL (τ ) · tR dτ is the rolling displacement of the particle and tR = vL /|vL |
0
is the direction of rolling. The first term in (3.6.15) represents a spring force and the
second term accounts for the effect of damping. The rolling torque is given by (3.6.15)
if |Mr | is less than a critical value Mr,crit , beyond which Mr = −Mr,crit .
This basic framework of the model given by (3.6.15) has been used by a large
number of investigators, for cases with no adhesion as well as for cases with adhesion.
However, not all investigators use all of the terms in this model, and there are
a number of differences in how the coefficients are set. Iwashita and Oda (1998)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

References 79

assume that the rolling stiffness and the sliding stiffness are of the same order of
magnitude, and thus set kR = ri kT . As discussed in Section 4.2, both the spring and
the slider mechanisms in (3.6.15) play a particularly important role when accounting
for the effect of adhesion on the rolling resistance. Focusing on the nonadhesive
case, Brilliantov and Pöschel (1998) omit the spring term in (3.6.15) and derive an
expression for the damping coefficient ηR using a quasi-static approximation in which
the purely elastic (Hertz) solution uel is used for the displacement u and the rate of
change of displacement u̇ is determined from

u̇(r) =  · (r × ∇ )uel (r). (3.6.16)

These expressions for u and u̇ are substituted into the constitutive equation for the
dissipative stress, which is then integrated to obtain the rolling resistance torque. The
resulting expression has the form of the damping term in (3.6.15) with the coefficient
ηR given by

ηR = μR |Fne |. (3.6.17)

Using the analogy with the normal dissipation force Fnd , Brilliantov and Pöschel
(1999) show analytically that the coefficient μR can be related to the same constant
coefficient αKKB that was introduced in the expression (3.3.7) for the Fnd , giving

μR = 23 αKKB . (3.6.18)

REFERENCES

Bagi K, Kuhn M. A definition of particle rolling in a granular assembly in terms of particle

translations and rotations. Journal of Applied Mechanics 71, 493–501 (2004).
Bardet JP. Observations on the effects of particle rotations on the failure of idealized granular
materials. Mechanics of Materials 18, 159–182 (1994).
Boussinesq J. Application des Potentiels a l’Etude de l’Equilibre et du Mouvement des Solides
Elastiques, Gauthier-Villars, Paris (1885).
Brilliantov NV, Pöschel T. Rolling friction of a viscous sphere on a hard plane. Europhysics
Letters 42(5), 511–516 (1998).
Brilliantov NV, Pöschel T. Rolling as a “continuing collision.” The European Physical Journal
B 12, 299–301 (1999).
Brilliantov NV, Spahn F, Hertzsch J-M, Pöschel T. Model for collisions in granular gases.
Physical Review E 53(5), 5382–5392 (1996).
Cleary PW, Metcalfe G, Liffman K. How well do discrete element granular flow models
capture the essentials of mixing processes? Applied Mathematical Modeling 22, 995–1008
(1998).
Cundall PA, Strack ODL. A discrete numerical model for granular assembles. Geotechnique
29(1), 47–65 (1979).
Gallas JAC, Hermann HJ, Sokolowski S. Molecular dynamics simulation of powder fluidiza-
tion in two dimensions. Physica A 189, 437–446 (1992).
Gorham DA, Kharaz AH. The measurement of particle rebound characteristics. Powder
Technology 112, 193–202 (2000).
Goldsmith, W. Impact: The Theory and Physical Behaviour of Colliding Solids, Arnold,
London (1960).
Haff PK, Werner BT. Computer simulation of the mechanical sorting of grains. Powder
Technology 48(3), 239–245 (1986).
Hertz H. Über die Berührung fester elastischer Körper. J. reine und angewandte Mathematik
92, 156–171 (1882).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-03 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 9:53

80 Contact Mechanics without Adhesion

Iwashita K, Oda M. Rolling resistance at contacts in simulation of shear band development

by DEM. Journal of Engineering Mechanics 124(3), 285–292 (1998).
Johnson, KL. Contact Mechanics, Cambridge University Press (1985).
Kruggel-Emden H, Simsek E, Rickelt S, Wirtz S, Scherer V. Review and extension of normal
force models for the discrete element method. Powder Technology 171, 157–173 (2007).
Kuwabara G, Kono K. Restitution coefficient in a collision between two spheres. Japanese
Journal of Applied Physics 26(8), 1230–1233 (1987).
Labous L, Rosato AD, Dave RN. Measurements of collisional properties of spheres using
high-speed video analysis. Physical Review E 56(5), 5717–5725 (1997).
Lee J, Herrmann HJ. Angle of repose and angle of marginal stability: Molecular dynamics of
granular particles. J. Phys. A: Math. Gen. 26, 373–383 (1993).
Love AEH. A Treatise on the Mathematical Theory of Elasticity, 4th ed., Cambridge University
Press, Cambridge (1952).
Luding S, Clement E, Blumen A, Rajchenbach J, Duran J. Onset of convection in molecular
dynamics simulations of grains. Physical Review E 50(3), R1762–R1765 (1994).
Marshall JS. Discrete-element modeling of particulate aerosol flows. Journal of Computational
Physics 228, 1541–1561 (2009).
Mindlin RD. Compliance of elastic bodies in contact. Journal of Applied Mechanics 16, 259–
268 (1949).
Mindlin RD, Deresiewicz H. Elastic spheres in contact under varying oblique forces. Journal
of Applied Mechanics 20, 327–344 (1953).
Oda M, Konishi J, Nemat-Nasser S. Experimental micromechanical evaluation of strength of
granular materials: Effects of particle rolling. Mechanics of Materials 1, 269–283 (1982).
Sadd MH, Tai QM. A contact law effects on wave-propagation in particulate materials using
distinct element modeling. International Journal of Non-Linear Mechanics 28(2), 251–265
(1993).
Sang Y, Dubé M, Grant M. Dependence of friction on roughness, velocity, and temperature.
Physical Review E 77, 036123 (2008).
Schäfer J, Wolf DE. Bistability in simulated granular flow along corrugated walls. Physical
Review E 51(6), 6154–6157 (1995).
Schäfer J, Dippel S, Wolf DE. Force schemes in simulations of granular materials. J. Phys. I
France 6, 5–20 (1996).
Schwager T, Pöschel T. Coefficient of restitution and linear-dashpot model revisited. Granular
Matter 9, 465–469 (2007).
Schwager T, Pöschel T. Coefficient of restitution for viscoelastic spheres: The effect of delayed
recovery. Physical Review E 78, 051304 (2008).
Stevens AB, Hrenya CM. Comparison of soft-sphere models to measurements of collision
properties during normal impacts. Powder Technology 154, 99–109 (2005).
Tabor D. The mechanism of rolling friction. 2. The elastic range. Proceedings of the Royal
Society of London A 229(1177), 198–220 (1955).
Taguchi R-H. New origin on a convective motion: Elastically induced convection in granular
materials. Physical Review Letters 69(9), 1367–1370 (1992).
Thornton C. Coefficient of restitution for collinear collisions of elastic perfectly plastic spheres.
Journal of Applied Mechanics 64, 383–386 (1997).
Timoshenko, SP, Goodier, JN, Theory of Elasticity, 3rd Ed., McGraw-Hill Publishers, New
York (1970).
Tsuji Y, Tanaka T, Ishida T. Lagrangian numerical simulation of plug flow of cohesionless
particles in a horizontal pipe. Powder Technology 71, 239–250 (1992).
Vu-Quoc L, Zhang X. An elastoplastic force-displacement model in the normal direc-
tion: displacement-driven version. Proceedings of the Royal Society of London. Series A,
Mathematical and Physical Sciences 455, 4013–4044 (1999).
Walton OR, Braun RL. Viscosity, granular temperature, and stress calculations for shearing
assemblies of inelastic, frictional disks. Journal of Rheology 30, 949–980 (1986).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4 Contact Mechanics with Adhesion Forces

As particles become sufficiently small, adhesion phenomena that are insignificant at

larger scales begin to become important. The specific particle length scale at which
adhesion becomes significant depends on the type of adhesive force. For instance,
van der Waals adhesion generally becomes important for particles with diameter on
the order of magnitude of a micrometer, whereas liquid bridging can be important
for particles as large as a millimeter or more (Li et al., 2011). The presence of field
forces that bring particles toward each other, such as electric or magnetic fields,
can enhance the effect of adhesive forces such as the van der Waals force, causing
them to be significant even for particle sizes for which they would not otherwise
have been important. The presence of significant adhesive forces can be established
experimentally by the observation that the area of the contact region is larger than
predicted by the Hertz theory and the observation that the contact region does not
vanish when the load is removed. The significance of adhesion also depends on the
magnitude of the applied load and on the roughness and cleanliness of the surface.
Within the context of DEM, we typically assume that adhesive forces between
particles act on length scales that are much smaller than the particle size, such that
in most cases the adhesive forces have negligible effect until two or more particles
collide. An exception occurs for the case of capillary forces between particles, which
typically act on a scale comparable to the particle diameter.
This chapter discusses various forces that give rise to adhesion between par-
ticles, including van der Waals force, electric double-layer repulsion force, ligand-
receptor binding of biological cells, liquid bridging forces of particles surrounded
by a thin liquid film in a gaseous environment, and sintering forces at relatively
higher temperature. Several of these adhesion forces arise from relatively weak
bonds, leading to formation of particle agglomerates, whereas the sintering forces
give rise to much stronger bonds leading to formation of particle aggregates (also
sometimes called “hard agglomerates”). These five different types of adhesive forces
are by no means exhaustive, but there are representative of the modeling involved
in adhesive DEM and are involved in numerous problems of interest in engineering
applications.

81
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

82 Contact Mechanics with Adhesion Forces

h h

(a) (b)
Figure 4.1. Diagrams illustrating (a) a single molecule or atom at a distance h from the
surface of a semi-infinite body and (b) two semi-infinite bodies with surfaces separated by a
distance h.

4.1. Basic Concepts and the Surface Energy Density

We consider two spherical particles approaching each other with minimum separa-
tion distance h0 (t ) between the particle surfaces. As the particles collide with each
other they deform within a small contact region of radius a(t ), where it is assumed
that the separation distance between the particle surfaces within the contact region
has a constant value δ. The gap thickness is assumed to remain uniform throughout
the contact region and to remain constant in time, so that movement of the particle
centroids toward or away from each other results only in a change of the contact
region radius a(t ).
The interaction potential between two microscopic elements, such as atoms or
molecules, separated by a distance r is assumed to have a power law form, given by

wm−m (r) = −C/rn . (4.1.1)

For instance, n = 6 corresponds to the potential associated with the long-range van
der Waals attraction force, in which the constant C includes the sum of disper-
sion, induction, and orientation forces between molecules. The corresponding force
between the microscopic elements is given by F (r) = −dwm−m (r)/dr, where a posi-
tive value of F corresponds to a repulsive force. Using the assumption of additivity,
the net adhesive force between two bodies is equal to the sum of the adhesive forces
caused by the interaction potential of all microscopic elements within the two bod-
ies. The net interaction energy between a “free” molecule located a distance h away
from the surface of a semi-infinite solid (Figure 4.1a) with molecule number density
ρ within the solid can be derived by summing all pair interactions (4.1.1) between the
free molecule and the molecules within the solid, giving the equation (Israelachvili,
1991, 156)
 ∞ ∞
r
wm−p (h) = −2πCρ dr dz
h 0 (r + z2 )n/2
2

= −2πCρ/(n − 2)(n − 3)hn−3 for n > 3. (4.1.2)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.1. Basic Concepts and the Surface Energy Density 83

The interaction energy dw p−p (z) between a thin sheet of material of width dz
having unit area in the x-y plane and molecule number density ρ1 located a distance
z from a semi-infinite solid with molecule number density ρ2 can be obtained from
(4.1.2) as dw p−p (z) = −2πCρ2 ρ1 dz/(n − 2)(n − 3)zn−3 . Integrating this expression
from a distance h to infinity gives the interaction energy per unit area between two
solids with planar surfaces separated by a distance h (Figure 4.1b) as
 ∞
2πCρ2 ρ1 dz 2πCρ1 ρ2
w p−p (h) = − =− . (4.1.3)
(n − 2)(n − 3) h z n−3 (n − 2)(n − 3)(n − 4)hn−4
The normal force per unit area (i.e., the pressure) between the two materials is
related to the derivative of the interaction energy by
p(h) = −dw p−p /dh. (4.1.4)
In the case of the long-range van der Waals force (with n = 6), (4.1.3) reduces
to
w p−p (h) = −πCρ1 ρ2 /12h2 . (4.1.5)
The Hamaker coefficient A12 for medium 1 interacting with medium 2 across a
vacuum is defined as A12 = π 2Cρ1 ρ2 , so that (4.1.5) becomes
w p−p (h) = −A12 /12πh2 . (4.1.6)
The Hamaker coefficient A12 can be approximately written in terms of the Hamaker
coefficient A11 for medium 1 interacting with itself and the respective coefficient A22
for medium 2 interacting with itself as

A12 ∼= A11 A22 . (4.1.7)
If media 1 and 2 interact across a space filled by a third medium 3, the resulting
Hamaker coefficient can be approximated by

A123 ∼= A131 A232 . (4.1.8)
where A131 can be written in terms of A11 and A33 as
  2
A131 ∼
= A11 + A33 − 2A13 ∼ = A11 − A33 . (4.1.9)

The final expression in (4.1.9) is obtained with use of (4.1.7) to write A11 A33 ∼
= A13 .
Using (4.1.9) for both A131 and A232 and substituting into (4.1.8) yields the useful
approximation
    
A123 ∼= A11 − A33 A22 − A33 . (4.1.10)
Expressions (4.1.7), (4.1.8), (4.1.9), and (4.1.10) are examples of what is called
combining relations. Such approximations are reasonably accurate, to within 15%
or so, provided that dispersion forces dominate the van der Waals force and that the
dielectric constants are not especially high. More accurate equations for calculation
of Hamaker coefficients are given by Israelachvili (1991). The value of the Hamaker
coefficient of most condensed phases, either solid or liquid, is typically in the range
of 0.4 × 10–19 J to 4 × 10–19 J in a vacuum. Table 4.1 summarizes typical values of
Hamaker constant, A131 , for two identical macroscopic bodies (identified as medium
1) interacting in the vacuum (air) or across the water (identified as medium 3).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

84 Contact Mechanics with Adhesion Forces

Table 4.1. Hamaker coefficients for two identical macroscopic bodies (media 1 and 2)
interacting across a third medium

Interacting media (symmetry A131 = A313 ) Hamaker constant A (unit: 10−20 J)

Medium 1 Medium 3
Water/Water Air 3.7–5.5
Pentane/Pentane Air 3.75
Pentane/Pentane Water 0.34
Dodecane/Dodecane Air 5.0
Dodecane/Dodecane Water 0.4–0.5

SiO2
Silica/Silica Air 6.5
Silica/Silica Dodecane 0.10–0.15
Fused quartz/quartz Water 0.5–1.0
Fused quartz/quartz Octane –
Mica/Mica Air 7–10
Mica/Mica Water 1.–2.9
α-Alumina/α-Alumina Air 15
α-Alumina/α-Alumina Water 2.7–5.2

Metals
Ag/Ag, Au/Au or Cu/Cu Air 20–50
Ag/Ag, Au/Au or Cu/Cu Water 10–40
Polymers
Polystyrene/Polystyrene Air 6.6–7.9
Polystyrene/Polystyrene Water 0.95–1.3
Teflon/Teflon Air 3.8
Teflon/Teflon Water 0.33

Forces between atoms and molecules often have a long-range attractive (van
der Waals) force and a short-range repulsive force. This behavior is apparent, for
instance, in the Lennard-Jones potential given in (2.3.1), in which the exponent
in (4.1.1) is equal to 6 for the attractive force and 12 for the repulsive force. The
resulting variation of interaction potential between two parallel surfaces can be
computed using (4.1.3) for each of these terms individually and then adding to
obtain the total value as a function of the distance h between the surfaces. Using
(4.1.4), this procedure results in a pressure distribution of the form
p(h) = −Ah−n + Bh−m , (4.1.11)
where A and B are positive constants and m > n, yielding a pressure variation curve
similar to that shown Figure 4.2. For instance, for the case where the molecules
of the two parallel surfaces interact via the Lennard-Jones potential, the resulting
exponents in (4.1.11) are obtained from (4.1.3) and (4.1.4) as n = 3 and m = 9.
The equilibrium gap thickness δ for two particles colliding in a vacuum is defined
as the value of separation distance h at which the attractive and repulsive (steric)
forces between the surfaces are equal to each other, corresponding to the point
p(δ) = 0 in Figure 4.2. Aside from setting the value of δ, the steric repulsive force
is not usually included in contact mechanics calculations. The reason for this is that
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.1. Basic Concepts and the Surface Energy Density 85

Repulsion
Figure 4.2. Plot showing the variation of pressure
p as a function of distance h between two surfaces
caused by the sum of the repulsive and attrac-
tive van der Waals forces. The equilibrium gap δ h
thickness δ is equal to the value of h at which p 0
vanishes, and the cross-hatched area above the

Attraction
curve to the right of h = δ is equal to twice the
2γ
surface energy density γ .

the steric force decays so quickly that if the separation distance h is even slightly
larger than δ the steric force will be very small, whereas the long-range attractive
force may have changed very little. The surface energy density γ is defined as half
the work required to separate two surfaces with initial separation distance h per unit
surface area. In Figure 4.2, the surface energy density corresponds to half the area
under the curve of p versus h to the right of the equilibrium point h = δ, as indicated
by the cross-hatched region. For a liquid, the surface energy density is the same
as the surface tension. The surface energy density of a solid reflects the nature of
the bonds, either weak or strong, between the atoms which constitute it. For strong
bonds between atoms or molecules, typical values of surface energy density in solids
range between 100 and 500 mJ/m2 for ionic crystals to 1,000–3,000 mJ/m2 for many
metals. For the weak van der Waals interactions between two surfaces separated
by a vacuum-filled gap with separation distance δ, surface energy density can be
obtained from (4.1.6) as

γ = 12 [w p−p (∞) − w p−p (δ)] = A/24π δ 2 , (4.1.12)

where A is the Hamaker coefficient for the given material. For problems involving
interaction of solids made of two different materials over a vacuum, the surface
energy density can be approximated using the relationship (4.1.7) for the Hamaker
√
coefficient as γ ∼= γ1 γ2 , where γ1 and γ2 are the surface energy densities of the
individual materials.
The minimum separation distance δ for two macroscopic surfaces in contact
should be substantially smaller than the intermolecular (or interatomic) center-to-
center distance. For most materials, experimental studies have recommended values
for δ ranging between 0.15 and 0.40 nm. However, use of the value δ = 0.165 nm in
(4.1.12) predicts the surface energy density in good agreement with measured values
for many different solids and liquids. For example, for polystyrene and Teflon, using
the Hamaker coefficients listed in Table 4.1 and δ = 0.165 nm, the surface energy
density is calculated as 32.1 and 18.5 mJ/m2 , respectively, and the corresponding
experimental values are 33.0 and 18.3 mJ/m2 . This simple approximation works well
for most materials, except strongly polar or hydrogen-bonding materials.
Interestingly, the surface energy of a material is dependent on the same inter-
molecular forces that determine the material’s latent heat and boiling point. As a
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

86 Contact Mechanics with Adhesion Forces

consequence, materials such as metals with high boiling points (TB > 2,000ºC) com-
monly have high surface energies, larger than about 1,000 mJ/m2 ; meanwhile, water
with a low boiling point (100°C) has a much lower surface energy density value of
73 mJ/m2 .

4.2. Contact Mechanics with van der Waals Force

The interaction forces between two noncharged contacting surfaces includes a long-
range attractive component with a length scale of approximately 10 nm and a short-
range repulsive component with a length scale of about 0.1 nm. The long-range (van
der Waals) force is associated with a variety of effects, typically referred to as induc-
tion, orientation, and dispersion force, all of which are associated with molecular or
atomic polarization effects. The dominant long-range force is the dispersion force,
which is associated with the fact that the instantaneous positions of the electrons
surrounding an atomic nucleus give rise to a polarized electric dipole field at any
instant of time. This electric dipole field induces surrounding atoms to also become
polarized, which leads these other atoms also to be surrounded by electric dipole
fields. The dispersion force is associated with the electrostatic interaction of these
induced dipoles with each other, leading to a dielectrophoretic attraction between
the dipoles. The short-range repulsive force, often called steric or exchange repul-
sion, is associated with overlap of the electron clouds of two different atoms, which
leads to strong electrostatic repulsion. This force is nearly zero outside of a distance
approximately equal to the atomic radius.

4.2.1. Models for Normal Contact Force

A variety of simplified models exist to account for the effect of van der Waals
adhesion on the combined adhesion-elastic force during collision of two particles.
The appropriateness of these models depends primarily on the size of the length
scale associated with elastic deformation of the particle relative to that associated
with the adhesive force. From the Hertz solution, Equations (3.2.8) and (3.2.9), the
normal overlap δN can be written in terms of the normal elastic force Fne in the
absence of adhesion as
 1/3
9 Fne2
δN = , (4.2.1)
16 E 2 R

where E and R are the effective elastic modulus and radius defined in (3.1.1). For
purposes of scaling analysis, the elastic normal force Fne in (4.2.1) is taken as being
of the same order of magnitude as the critical pull-off force FC , which is shown in
the following to be proportional to the product of R and the surface energy density
γ . Setting Fne ∝ Rγ in (4.2.1) and ignoring the constants (because this is a scaling
analysis) gives an estimate for the order of magnitude of the normal overlap caused
by adhesive force between two particles as
 1/3
Rγ 2
δN ∝ . (4.2.2)
E2
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 87

Figure 4.3. Schematic diagrams showing regions with

strong adhesion forces acting between surfaces of col- a
liding particles: (a) particles remain spherical and Sphere 1 Sphere 2 Contact
adhesion force acts along the close parts of the region
spherical surface (DMT approximation); (b) particles
develop flattened contact region and adhesion forces
act only within this region (JKR approximation).

(a) (b)

The result (4.2.2) yields a scaling for the elastic deformation of the interface
of a spherical particle in contact with a second particle due to the adhesive force
between the particles. A scaling for the characteristic length scale of the van der
Waals adhesion force is given by the equilibrium gap thickness δ defined by (4.1.12).
The ratio of these two length scales yields a dimensionless parameter called the
Tabor parameter (Tabor, 1977), defined as
 1/3
4Rγ 2
λT ≡ . (4.2.3)
E2δ3

The Tabor parameter is important for determining the extent to which elastic
deformation of the colliding particles influences the adhesive force. Specifically, if
λT 1 it can be assumed that there is no effect of elastic deformation on the adhesive
force, and so the Hertz elastic rebound force and the adhesive force for two spherical
particles in contact can simply be added together to obtain the total force on the
particle (Figure 4.3a). In the other extreme, if λT 1 then the length scale of the
adhesive force will be much smaller than that of the particle elastic deformation. In
this case, the particles cannot be regarded as spheres when determining the adhesive
force, but instead it can be assumed that the adhesive force only acts within the
flattened contact region, as shown in Figure 4.3b. Within this contact region, the
particle surfaces can be approximated as two infinite flat planes for purposes of
determining the adhesive force. These two extremes are the basis of two leading
theories of adhesive particle contact dynamics, the theory for which was developed
respectively by Derjaguin, Muller, and Toporov (1975) for small λT (the DMT
model) and by Johnson, Kendall, and Roberts (1971) for large λT (the JKR model).
A model appropriate for intermediate values of λT was developed by Maugis (1992)
based on an approximation to the adhesive potential similar to that suggested by
Dugdale. The resulting Maugis-Dugdale (M-D) model includes the DMT model as
a special case. In the extreme case of very small particles where the particle size
approaches the length scale of the adhesive force, given by the equilibrium gap
thickness δ, all of the models break down. In this case, particles can be assumed to be
rigid (so that there is no elastic rebound force) and the full Lennard-Jones potential
can be applied to each spherical particle as a function of distance away from the
opposing particle. This extreme, called the Bradley model, violates our previous
assumption that adhesion forces are not important until after particle collision. The
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

88 Contact Mechanics with Adhesion Forces

4
10
Dimensionless load, Fn / 2πγR
3
10
Figure 4.4. Adhesion map showing re-
2 gimes of different adhesive contact mod-
10 JKR els as a function of the dimensionless nor-
10
1 mal force and the Maugis parameter λM ,
DMT M-D which can be related to the Tabor param-
0
10 Bradley eter defined in (4.2.3) by λM = 1.16λT .
[Based on data of Johnson and Green-
-1
10 -3 -2 -1 0 1 2 wood (1997). Reprinted with permission.]
10 10 10 10 10 10

λ Μ = 1.16 λΤ

conditions required for the Bradley model to be valid are difficult to satisfy, even for
very small particles in the nanoscale size range.
A map showing the regimes in which each of these adhesive contact models
apply was developed by Johnson and Greenwood (1997), and a version of this
map is shown in Figure 4.4. The regimes are organized in terms of the normal
load Fn , normalized by the characteristic pull-off adhesive force 2πRγ , and a
parameter λM = σ0 (9R/4π γ E 2 )1/3 , where σ0 is an upper limit on the stress. Fol-
lowing Maugis (1992), Johnson and Greenwood (1997) chose the maximum stress
as σ0 = 1.03(2γ /δ), where δ is the equilibrium gap thickness obtained from the
Lennard-Jones potential. This choice results in a relationship between λM and the
Tabor parameter λT as λM = 1.16λT . The Hertz zone, in which adhesion is ignored,
is defined on this map as the region where the adhesive force is less than 5% of the
maximum elastic rebound force. The JKR region is taken to be that region where the
material displacement δa caused by adhesive forces is large compared to the equilib-
rium gap thickness δ, where for the map in Figure 4.4 it was assumed that δa /δ > 20
in the JKR region. Similarly, the DMT region was characterized by small values
of δa /δ, in this case δa /δ < 0.05. The Bradley model is assumed to apply when the
elastic displacement δh associated with Hertz theory is much less than δ, where the
criterion was selected as δh /δ < 0.05. Details of the DMT, M-D, and JKR adhesive
contact models are discussed in the following sections.

DMT Model
The Derjaguin approximation (Derjaguin, 1934) relates the adhesion force Fs−s (h)
between two rigid spherical particles with radii r1 and r2 separated by a distance h
to the interaction potential per unit area w p−p (h) between two planar surfaces with
the same separation distance h, as given by (4.1.6). The approximation is based on
discretizing the surface of each sphere into a series of concentric rings, each with
width dr in the plane P normal to the line connecting the sphere centers, as shown
in the close-up image in Figure 4.5. It is assumed that each of these rings, when
projected onto the cross-sectional plane P, interacts with the opposing sphere in the
same way as would a surface separated from an infinite plane by a distance Z, where
Z is equal to the distance between the spheres at the ring centerline when evaluated
along a line oriented normal to the plane P. If Fp−p (Z) denotes the force per unit
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 89

P
Figure 4.5. Close-up figure showing the region
Incremental Z
between two spheres separated by a distance h. Indi- ring
cated is a ring on the surface of Sphere 1 which when h
projected onto a plane P has width dr. The distance
between the sphere surfaces at the centerline of the n
ring is denoted by Z(r).
Sphere 1 r Sphere 2
dr

area exerted by a ring at separation distance Z(r), then the total attractive force
between the spheres can be written as
 Z=∞
Fs−s (h) = 2π rFp−p (Z)dr, (4.2.4)
Z=h

where r is the radial coordinate on the plane P. This approximation is valid provided
that the distance h between the spheres is much less than either sphere radii. This
limit will also allow us to approximate the separation distance Z(r) between the
spheres at radial distance r as
 
r2 1 1 r2
Z=h+ + =h+ , (4.2.5)
2 r1 r2 2R
which is similar to the expression (3.1.4) but with the sign on the second term
switched due to the fact that the particles do not yet overlap each other. Discretizing
this expression with respect to r gives

dZ = (r/R)dr. (4.2.6)

Substituting (4.2.6) into (4.2.4) gives

 Z=∞
Fs−s (h) = 2π R Fp−p (Z)dZ. (4.2.7)
Z=h

Because Fp−p (Z) = −dw p−p /dZ and w p−p (∞) = 0, we can integrate (4.2.7) to obtain
the final expression

Fs−s (h) = 2π Rw p−p (h). (4.2.8)

If the separation distance is set equal to the equilibrium gap thickness δ for two
spheres in contact, (4.1.12) and (4.2.8) can be used to write the sphere-sphere adhe-
sion force as

Fs−s (δ) = −4π Rγ . (4.2.9)

It is useful to define an effective radius c on the particle surface over which

significant adhesion force acts. The force per unit area between two plane surfaces
can be obtained from (4.1.6) as
w p−p A12
p=− =− . (4.2.10)
dh 6π h3
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

90 Contact Mechanics with Adhesion Forces

r1 r1 r1
Hertzian Hertzian Hertzian

a=ah ah ah ah ah
a a a a
r2 r2 r2
c c
C
B
A A Compressive

Adhesive

(a) (b) (c)

Figure 4.6. Schematic diagrams illustrating the stress distributions for various sphere-sphere
adhesive contact models, including (a) DMT model; (b) JKR model; and (c) M-D model. The
contact region radius is denoted by a and the radius that would be obtained with only the
Hertz elastic forces is denoted by ah . The radius of the region over which adhesive forces act
is denoted by c.

Substituting (4.1.6) into (4.2.8) gives

A12
Fs−s = −2π R . (4.2.11)
12ρh2
The effective area over which adhesion force acts is then obtained as
Fs−s
Aeff = = π Rh. (4.2.12)
p
Setting Aeff = π c2 gives the effective adhesion force radius as
√
c = Rh. (4.2.13)

Recalling
 the solution (3.1.3) for contact region radius from the Hertz elastic theory
as a = RδN and taking the sphere separation distance h for contacting spheres to
be on the order of magnitude of the equilibrium gap thickness δ gives
 1/2
c δ  
=O = O λ−1/2 , (4.2.14)
a δN T

where we recall that the Tabor parameter λT = O(δN /δ).

In Figure 4.6, different force profiles are sketched for the DMT, JKR, and M-D
models of adhesive contact. The upper plots in this figure show the particle surface
shape near the contact point (solid line) and the shape that the particle would have
with only elastic forces (dashed line). The lower plots illustrate variation of the
surface pressure as a function of radius in each model. The DMT model makes
the assumption that the standard Hertz pressure distribution (3.2.1) applies within
the contact region r < a and that the adhesive force applies only outside of the
contact region, within a region a ≥ r ≥ c. The net adhesive force can be obtained
by evaluating an integral similar to (4.2.7) but with the lower limit of integration
adjusted to equal the separation distance at the edge of the contact region, or Z(a).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 91

However, because (4.2.14) requires that c/a 1 for cases where λT 1, it follows
that the contact region is very small compared to the region over which the adhesion
force acts. Consequently, to first approximation the adhesive force can be evaluated
for the entire sphere in (4.2.7). If we then add the Hertz elastic force to the adhesive
force between two spheres, ignoring the elastic deformation in computation of the
adhesion force, the total force acting on each particle of a colliding pair is given by

Fne = 43 Ea3 /R − 4π Rγ , (4.2.15)


where the contact radius a = δN R in the Hertzian elastic rebound term. The
pull-off force Fne = −FC corresponds to that for which a = 0, which from (4.2.15)
gives

FC = 4π Rγ . (4.2.16)

Equation (4.2.15) gives the relationship for the combined elastic and adhesive normal
force in DMT theory. In comparison to numerical solutions based on integration
of the Lennard-Jones potential, Muller et al. (1980) found that the DMT force
expression compares well with the more exact results for λT < 0.1.

JKR Model
One of the earliest theoretical treatments of adhesive particle contact was proposed
by Johnson, Kendall, and Roberts (1971) at the Cavendish Lab in Cambridge. The
model is appropriate for relatively large, compliant particles for which the Tabor
parameter λT is large compared to unity, implying that the length scale of elastic
deformation is large compared to the length scale of the adhesive force. As a conse-
quence, the JKR model assumes that the adhesive force acts only inside the contact
region, such that c ≈ a, as shown schematically in Figure 4.6b. The size of the contact
region in the JKR model is larger than that in the nonadhesive Hertz model due
to the additional compression caused by the adhesive force. The original derivation
of JKR theory follows a type of thermodynamic approach. This approach involves
estimation of the total energy UT in the system as a function of contact radius a,
and then determining the equilibrium conditions by setting dUT /da = 0. The total
energy of the system is made up of three terms – the stored elastic energy UE , the
mechanical energy UM , and the surface energy US .
As shown in Figure 4.6b, under a normal loading Fne = F , without the adhesive
force there is a contact radius ah predicted by the Hertz equation. The resulting
expression for F from (3.2.8) and (3.2.9) is

4Ea3h κa3 √
F= = h = κ Rδh3/2 , (4.2.17)
3R R
where δh = a2h /R and
√ the coefficient κ (≡ 4E/3) is related to the stiffness K given in
(3.2.9) by κ = K/ R. This condition corresponds to point A in the plot of normal
load versus normal overlap in Figure 4.7.
If adhesive force is now introduced, the contact radius in the equilibrium condi-
tion will increase to a value a, which is larger than ah , as shown in Figure 4.6b. The
combined contact stresses at this new position arise from both adhesive and elastic
forces, corresponding to point B in Figure 4.7. Finally, we can associate an effective
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

92 Contact Mechanics with Adhesion Forces

C
F1

0
=
W

t
ns
Figure 4.7. The loading path of JKR

co
=
model. Hertz loading from O to C with

a
F A W = 0, and then from C to B by keeping
B
constant contact area a.
O δ0 δ1 δN δ1 = δH
3

–F1
2

Hertz loading F1 with the contact radius a (associated with point C in Figure 4.7),
given by
κa3 √
F1 = = κ Rδ13/2 (4.2.18)
R
where δ1 = a2 /R. The coordinates of points A, B, C in Figure 4.7 are (F, ah , δ h ), (F, a,
δ N ), and (F1 , a, δ 1 ), respectively. The stress profiles at points A, B, and C are plotted
in Figure 4.6b, indicated by the different letters.
According to Figure 4.7, the JKR loading path can be decomposed into two
steps: (1) a Hertz-type loading from point O to C with no adhesion (γ = 0), which
requires energy UO−C , and (2) unloading from point C to point B by keeping constant
contact area a, releasing energy UC−B . The energy required for the first step involving
loading from O−C is given by
 δ
1 2 2 a5
UO−C = F dδN = κR1/2 δ15/2 = κ 2 . (4.2.19)
0 5 5 R
The unloading process of B-C is similar to the problem of a flat punch that can
be solved by Boussinesq theory (Maugis, 1999), giving
2dF
dδN = . (4.2.20)
3aκ
Integrating (4.2.20) and using (4.2.17) and (4.2.18) gives the overlap δN at Point B
as
2F − 2F1 a2 2a3 a−1
δN = δ1 + = − h . (4.2.21)
3aκ 3R 3R
Substituting (4.2.20) into (4.2.19), UC−B can be calculated as
 δ  F
2F F 2 − F12 1 a6 a−1 − a5
UC−B = F dδN = dF = = κ h . (4.2.22)
δ1 F1 3aκ 3aκ 3 R2
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 93

Using UE = UO−C + UC−B , we obtain

1 a5 1 a6 a−1
UE = κ 2+ κ h 2 . (4.2.23)
15 R 3 R
The mechanical potential energy at point B under the normal force F is given
by UM = −F δN . Applying (4.2.17) and (4.2.21) gives

1 a3 a2 + 2a6h a−1
UM = − κ h . (4.2.24)
3 R2
The surface energy US is given by

US = −2π γ a2 . (4.2.25)

Adding the three different components of the energy, the total energy UT is obtained
as
1 a5 1 a6 a−1 1 a3 a2
UT = κ 2 − κ h 2 − κ h 2 − 2π γ a2 . (4.2.26)
15 R 3 R 3 R
Applying the equilibrium condition dUT /da = 0 gives
   
dUT κ R 3
= 2 2 a6 − 2 a3h + 6π Rγ a + a6h = 0 (4.2.27)
da 3a R κ
Using F = a3h κ/R and F1 = a3 κ/R, (4.2.27) can be written as

F12 − 2(F + 6π Rγ )F1 + F 2 = 0. (4.2.28)

Solving Equation (4.2.28) for F1 and taking the positive sign gives

F1 = F + 6π Rγ + 12πRγ F + (6π Rγ )2 . (4.2.29)

Setting Fne = F and F1 = a3 κ/R gives an equation for the contact radius in the
presence of adhesion as

3R
a3 = [Fne + 6π Rγ + 12πRγ Fne + (6π Rγ )2 ] (4.2.30)
4E
These results can be rearranged to obtain an expression for the normal overlap
by dividing (4.2.28) by 6πRγ to write
 2
F1 F F1
− = . (4.2.31)
6πRγ 6π Rγ 3π Rγ
Solving for F − F1 from this equation and substituting into (4.2.21) gives

2F − 2F1 a2 16πaγ
δN = δ1 + = − . (4.2.32)
3aκ R 3κ
The relation between the normal overlap and the contact radius can thus be written
as

a2 4π aγ
δN = − . (4.2.33)
R E
With zero applied load, the two colliding particles will approach an equilibrium
state in which the elastic repulsion is balanced by the adhesive attraction of the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

94 Contact Mechanics with Adhesion Forces

particles. In this zero-load equilibrium state, the radius a(t ) of the contact region can
be obtained by setting Fne = 0 from (4.2.30) to obtain
 1/3
9π γ R2
a0 = . (4.2.34)
E
For example, for two contacting particles with 2 µm radius (R = 1 µm), E = 30 GPa
and γ = 15 mJ/m2 , the equilibrium contact radius a0 is about 24 nm, accounting for
only 2.4% of R.
Modified forms of these results, which are more convenient for DEM computa-
tions, were obtained by Chokshi et al. (1993) in which expressions for Fne and δN are
written in terms of the ratio a/a0 . The expression for Fne is obtained by solving for F
from (4.2.28) with F1 = a3 κ/R and F = Fne . The expression for δN is obtained simply
by writing (4.2.33) in terms of a/a0 . The resulting system of equations is
 3  3/2
Fne a a
=4 −4 (4.2.35)
FC a0 a0
and
 2  
δN a 4 a 1/2
=61/3
2 − . (4.2.36)
δC a0 3 a0

In these equations, FC and δC are the critical force and overlap at the pull-off point
assuming constant pulling force (or constant load). As the two particles move away,
the contact will be maintained even for negative values of δN via necking of the
particle material, until the critical pull-off point is reached, at which Fne = −FC and
δN = −δC . As the particles are pulled further apart, the contact will suddenly break.
FC and δC are given as
a20
FC = 3π Rγ , δC = . (4.2.37)
2(6)1/3 R
The critical pull-off force FC predicted by the JKR model is 3π Rγ , which is slightly
lower than the value 4π Rγ predicted by the DMT model. In the example of the
2 µm radius particles mentioned earlier, δC is only 0.16 nm, which is close to
the minimum intermolecular center-to-center distance. Thus, although necking of
the particle material occurs, it is in practice quite small for relatively stiff particles.

M-D Model
We have previously noted that the JKR model is suitable for “soft” materials with
large surface energy and large particle size, corresponding to large values of the
Tabor parameter λT . The DMT model is suitable for “hard” materials with a low
surface energy and much smaller radius, corresponding to small values of the Tabor
parameter. The adhesion map proposed by Johnson and Greenwood (1997) indicates
roughly that the DMT model applies for λT < 0.1 while the JKR model applies for
λT > 3.
An approximate theory valid for intermediate values of λT was developed by
Maugis (1992), which includes the DMT and JKR theories as limiting cases. Due
to the complexity of the Lennard-Jones potential for theoretical simulation, Maugis
(1992) chose to use the much simpler Dugdale approximation of van der Waals
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 95

Figure 4.8. Relation between normal

Μ
Fn /2πRγ
force and normal overlap for various
adhesive contact models including JKR,
DMT, and M-D models.
Μ
Μ

1/3
δΝ / [9π2Rγ 2/4Ε2]

interactions. The resulting Maugis-Dugdale (M-D) model assumes a stress profile

(shown in Figure 4.6c) that admits both repulsive and adhesive stresses in the contact
region (like JKR), but assumes that the adhesive stress has a constant value σ 0
within the region outside of the contact region with radius a ≤ r ≤ c. Applying a step
distribution of the adhesive stress within this outer region, the dimensionless normal
force, normal overlap, and contact region radius in the M-D model are obtained as

 
F̄ = F̄h + F̄ad = ā3 − λM ā2 m2 − 1 + m2 cos−1 (1/m) (4.2.38a)

4 
δ̄N = δ̄h + δ̄ad = ā2 − āλM m2 − 1 (4.2.38b)
3
λM ā2  2 
m − 1 + (m2 − 2) cos−1 (1/m)
2
4λ2 ā  2 
+ M m − 1 cos−1 (1/m) − m + 1 = 1 (4.2.38c)
3

The normalization is done as F̄ = F/2πRγ , ā = a[2E/3πR2 γ ]1/3 , and δ̄N =

δN [4E 2 /9π 2 Rγ 2 ]1/3 , and the subscripts h and ad represent contributions from the
Hertzian elastic stress and the adhesive stress, respectively. The variable m ≡ c/a,
where c is the radius over which adhesive force acts, and λM is the Maugis parameter,
defined at the beginning of Section 4.2.1.
For given values of λM and ā, we can solve for m from (4.2.38c), and then
substitute it into (4.2.38a) and (4.2.38b) to obtain F̄ and δ̄. A plot of F̄ versus δ̄ is
given in Figure 4.8. When the pulling force is held constant, particle detachment
occurs at FC = 3π Rγ with δ̄N = 0.43, corresponding to Point III in Figure 4.8. When
the displacement is held constant, particle detachment occurs at δ̄N = 0.91 with
Fne = −1.66πRγ , corresponding to Point IV in Figure 4.8. The force-displacement
relationship for λM ≥ 0.5 predicted by the M-D model agrees well with the prediction
of the JKR model, while prediction of the M-D model for λM ≤ 0.01 is close to the
prediction of the DMT model. The M-D model applies for all values of λM , but
it is not commonly used in DEM computations because the solution of (4.2.38c)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

96 Contact Mechanics with Adhesion Forces

F
II

g
in
Figure 4.9. Loading curve illustrating sources

ad
Damping loss

Lo
First-contact of energy dissipation during the normal con-
loss O tact of adhesive spheres.
I
Adhesion
Unloading δN
Rebound III

for m requires a numerical iteration procedure. However, an approximate fit to the

numerical solutions of the M-D model has been proposed by Carpick et al. (1999),
which gives analytical expressions for the force and contact area accurate to within
about 1% of the M-D predictions for arbitrary values of λM .
For common materials, such as quartz in air, the work of adhesion due to van
der Waals interaction is about 0.05J/m2 , the modulus parameter E is about 1.0 Gpa
and the minimum separation δ is 0.165 nm, so the Tabor parameter λT is about 8.2
for R = 1.0 µm and 14.1 for R = 5.0 µm. This result suggests that the JKR model
is most applicable for particles in the micrometer size range. Since λM ∝ R1/3 , the
Tabor (or Maugis) parameter decreases very slowly with particle radius and can still
be fairly large even for particles in the nanometer size range. Application of the
continuum JKR model (or any of these continuum models) becomes questionable
as the particle size approaches 1 nm due to quantum effects for nanoparticles.

4.2.2 Normal Dissipation Force and Its Validation

The various normal contact models discussed in the previous section are devel-
oped for quasi-static deformations. Because particle collision processes are highly
dynamic, the energy dissipation that arises due to dynamic effects must be incorpo-
rated into the normal elastic-adhesive model. Figure 4.7 illustrates the main energy
dissipation processes that occur during the collision of adhesive particles. Both the
viscoelastic damping force and the adhesive force contribute to energy loss. The
energy loss from adhesion comes in part from the jump-on–pull-off behavior of the
adhesive contact, which occurs at points with δN = 0 and δN = −δC , respectively. At
the jump-on point (Point I in Figure 4.9), the contact region area suddenly goes from
zero to a finite value and the contact force suddenly goes from zero to a negative
(adhesive) value. This sudden change leads to a first-contact energy loss, as indicated
in Figure 4.9. The necking behavior of the material when the particles are pulled
away from each other allows the adhesive force to act even when the overlap δN < 0,
leading to an energy loss indicated by the shaded region in Figure 4.9. The second
part of the energy loss arises from the dynamic effect due to the viscoelasticity of
materials, as discussed in Chapter 3.
Liu et al. (2011) examined the energy losses due to both adhesion and viscoelastic
effects for an elastic material using the JKR approximation for normal pressure
together with a two-component linear dashpot approximation, such as is given by
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 97

Coefficient of restitution

Incident Velocity, m/s

Figure 4.10. Coefficient of restitution versus normal incident velocity: comparison between
the experiment (symbol) of Wall et al. (1990) for ammonium fluorescent microsphere impact
with a flat surface and the prediction by the JKR model with linear damping given by Equation
(4.2.40). [Reprinted with permission from Li et al. (2011).]

(3.3.1), for the attractive and repulsive components of the damping forces. This
simple JKR-based damping model has been validated using classic particle/surface
impact experiments such as those of Dahneke (1973) and Wall et al. (1990), in
which both the coefficient of restitution and the critical sticking velocity of bouncing
particles are well predicted. On the basis of the normal pressure distribution over the
contact region predicted by the JKR model, Thornton and Ning (1998) used a cut-
off corresponding to a limiting contact pressure for plastic deformation in order to
derive expressions for the effect of plastic deformation on coefficient of restitution,
critical sticking velocity, and so on for adhesive particle collision.
A particularly simple description of the normal collision of two particles with
equal radius r1 was given by Marshall (2009), which uses the JKR model result
(4.2.35) for Fne and the Tsuji et al. (1992) model given by (3.3.1) and (3.3.12) for the
damping force Fnd . The governing equation for the normal overlap becomes
  3  3/2
d 2 δN K 1/4 dδN 4 a a
2
+ αT T I δN + − =0 (4.2.39)
dt M dt M a0 a0

Dimensionless time, normal overlap, and contact region radius are defined by tˆ =
t(r11/2 K/M)1/2 , δ̂ = δ/r1 and â = a/r1 , so that (4.2.39) and (4.2.36) reduce to

d2 δ̂   dδ̂    
+ 2 21/4 αT T I â1/2 + 4 21/2 â3 1 − (â0 /â)3/2 = 0, (4.2.40a)
dtˆ2 dtˆ
 
δ̂ = 2â2 1 − 23 (â0 /â)3/2 . (4.2.40b)
The predictions of (4.2.40) are compared in Figure 4.10 with the experimental
data reported by Wall et al. (1990) for ammonium fluorescein microspheres with
diameters of 2–7 µm transported in an air jet at velocities ranging from 1 to 100 m/s
and impacting on a flat surface (Li et al., 2011). The laser-Doppler technique was
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

98 Contact Mechanics with Adhesion Forces

used for measuring both incident and rebound velocities of particles before and after
impact. The relationship between restitution coefficient and incident velocity (i.e.,
the e-vi curve) was obtained, and a critical velocity (vc ) for sticking was determined.
In the experiments, the modulus parameter K (= 4/3E) is 1.80 GPa and the adhesive
√
surface energy γ (≈ γ1 γ2 ) is 0.17 J/m2 . The Tabor parameter λT is estimated to be
greater than 20 for all particle sizes used, so we are well within the range of validity
of the JKR model. For each particle size examined, the damping parameter αT T I
is adjusted to best fit the experimental results. As particle diameter increases from
2.58 µm to 6.89 µm, the value of damping parameter αT T I decreases from 0.206 to
0.095, corresponding to the fact that small particles in dispersed aerosol flow have a
relatively lower restitution coefficient than larger particles. Even though this is a very
simple theory for normal damping, the predictions of (4.2.40) remain in reasonably
good agreement with the experimental results and follow the expected trend in the
e-vi curve. Particle impact with incident velocity vi lower than the critical velocity vc
leads to a final capture of the particle (i.e., e = 0). The larger the particle size, the
lower the value of vc . For vi values that are larger than, but still close to, vc , the e-vi
curve exhibits a steep increase, but thereafter its slope rapidly decreases as vi further
increases until finally the restitution coefficient e asymptotes to a constant value.

4.2.3. Effect of Adhesion on Sliding and Twisting Resistance

Both sliding and twisting motions are relatively rare for small adhesive particles –
rolling is generally the preferred deformation mode for agglomerates of adhesive
particles (Dominik and Tielens, 1995, 1996; Marshall, 2007). It is therefore desirable
to introduce relatively simple expressions for sliding and twisting resistance, similar
to those used in the DEM model for nonadhesive particles introduced in Chap-
ter 3. The standard sliding model for the case without adhesion is the spring-dashpot
model proposed by Cundall and Strack (1979), for which the sliding force Fs is given
by (3.5.5) when |Fs | < Fcrit and by the Amonton friction expression Fs = −Fcrit when
|Fs | ≥ Fcrit . A detailed study of the effect of van der Waals adhesion on tangential
sliding is given by Savkoor and Briggs (1977), and a relatively simple model was
proposed by Thornton (1991) and Thornton and Yin (1991) which appears to agree
reasonably well with experimental data. In this model, the only influence of van der
Waals adhesion on sliding force is to modify the critical force Fcrit at which sliding
occurs, which in the presence of adhesion is given by

Fcrit = μF |Fne + 2FC |, (4.2.41)

where FC is the critical normal force given by (4.2.37) and μF is a friction coefficient.
When particles are being pulled apart, the normal force approaches −FC at the point
of separation, at which point the critical sliding force in (4.2.41) approaches μF FC .
The same model with twisting resistance as described in Chapter 3 can be used in
the presence of adhesion, with the critical force Fcrit given by the expression (4.2.41)
used to obtain Mt,crit = 3π 16
aFcrit . For twisting moments with magnitude greater than
Mt,crit , the torsional resistance is given by

3π
Mt = − aμ f |Fn + 2FC |T /|T |. (4.2.42)
16
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.2. Contact Mechanics with van der Waals Force 99

Ω
Particle

R
Figure 4.11. Schematic diagram illustrating lagging of the con- O
tact region behind a rolling particle due to adhesion force, such v
that the particle centroid O moves ahead of the centroid C of the
ξ
contact region by an amount equal to the rolling displacement ξ .
R F
C
Contact
region Fa

4.2.4. Effect of Adhesion on Rolling Resistance

In Chapter 3, it was mentioned that the rolling resistance in low-velocity collisions of
nonadhesive particles (e.g., granular materials) results mainly from the solid-phase
viscoelastic deformation energy loss and the viscous fluid resistance. However, these
kinds of classical sources of rolling friction such as microslip at the interface, inelastic
or viscoelastic deformation of the involved materials, and large surface irregularities
turn out to be less important for the micron or submicron particle sizes that are of
primary concern in the fields of aerosol and colloidal science. The primary physical
mechanism underlying rolling resistance for micron-sized and smaller particles arises
from the asymmetry of the stress distribution resulting from the adhesive force.
A simple, but highly effective, model for adhesive rolling resistance was devel-
oped by Dominik and Tielens (1995). Although this model can be applied to the
problem of any two particles colliding, it is most easily to explain for the problem
of a particle in contact with a plane surface. We recall with reference to Figure 4.6b
that the stress distribution in the JKR model for normal collision of a particle with
a flat surface is symmetric about the contact point, which is defined as the point on
the surface directly below the particle centroid. As the particle starts to roll on the
surface (illustrated in Figure 4.11), the part of the particle surface on the front (F)
side of the particle moves downward and that on the rear (R) side of the particle
moves upward. Due to the presence of adhesive force, the material points on the
front side of the particle are continually “jumping on” to a state of contact with
the planar surface, whereas material points on the rear side are continually “pulling
off” of the surface. However, we know from Figure 4.8 that jump-on and pull-off
do not occur at the same value of the normal overlap. Specifically, jump-on occurs
approximately at δN = 0, whereas pull-off occurs at δN = −δC . The fact that pull-off
is delayed relative to jump-on due to the phenomenon of material necking gives
rise to an asymmetry in the contact region during particle rolling. Specifically, the
necking phenomenon allows the particle surface to remain in contact with the plate
longer in the rear side of the particle than on the front side, causing the horizontal
position of the center of the contact region to lag behind the horizontal position of
the particle centroid by a distance ξ , called the rolling displacement, as illustrated in
Figure 4.11.
The result (4.2.35) gives the total force in the JKR model as the sum of two
terms, the first of which is the Hertz stress and the second of which is associated with
the adhesive force, given by Fa = 4FC (a/a0 )3/2 , where FC is the critical pull-off force
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

100 Contact Mechanics with Adhesion Forces

given by (4.2.37). The Dominik-Tielens model of adhesive rolling resistance gives

the rolling resistance moment Mr as the product of the adhesive force Fa and the
rolling displacement ξ , which serves as the lever arm, giving

Mr = −kR ξ , (4.2.43)

where the rolling coefficient kR is equal to Fa , or

kR = 4FC (a/a0 )3/2 . (4.2.44)

When a particle starts rolling from a state of rest, the contact region will initially
remain fixed as the particle centroid moves by a distance ξ . When ξ increases to a
critical value ξcrit , the contact region will start moving at the same velocity as the
particle centroid, but with the contact region centroid remaining behind the particle
controid by a horizontal displacement equal to ξcrit . Consequently, once ξ ≥ ξcrit , the
rolling moment is given by

Mr = −kR ξcrit . (4.2.45)

The critical rolling displacement resistance ξcrit can also be expressed in terms of a
critical rolling angle θcrit = ξcrit /R, where, for the example of a particle on a plane
surface, R is equal to the particle radius.
The rolling displacement can be written in terms of a rolling velocity vL , defined
in (3.6.13), as
 
t
ξ= vL (τ )dτ · tR , (4.2.46)
t0

where tR = vL /|vL | is the direction of rolling. Experimental investigations of the

effect of adhesion on particle rolling are presented by Peri and Cetinkaya (2005)
and Ding et al. (2008). These demonstrate good agreement with the theoretical
expressions and can be used to estimate typical values for the maximum rolling
displacement ξcrit for different materials. For instance, for polymer microspheres
with diameter of nominally 8 µm, Ding et al. (2008) report values of ξcrit ranging
between 70 to 245 nm, with corresponding values for the critical rolling angle ranging
between 19 and 64 × 10−3 rad.

4.3. Electrical Double-Layer Force

The electrostatic force between two charged particles in a vacuum or in air obeys
the inverse-square decay behavior expressed by Coulomb’s law. Since this force
decays slowly with distance and can act over distances of many times the particle
diameter, it is considered a long-range interaction rather than an adhesive force.
On the other hand, a charged particle immersed in an electrolyte (i.e., a solution
of positive and negative ions) attracts a cloud of ions of opposite sign (counter-
ions) around the particle, which act to screen the particle’s charge. The resulting
electrostatic potential decays rapidly to nearly zero within a few nanometers of
the particle surface. If the size of the particle is on the micron scale, the presence
of a particle charge together with electrolytic screening results in a short-range
force with length scale much smaller than the particle diameter, which can easily be
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.3. Electrical Double-Layer Force 101

Particle 1
Bulk of solution

Particle 2

Particle
Stern layer Diffuse Layer

σ σ
Plane of shear
Diffuse Layer
Stern layer
Surface ε
potential Surface Ψs
potential
Stern layer
potential Stern layer
Ψd
potential

ρS ρS
Concentration of positive ions
Counterions (ρZ+)
ρ∞=ρ0
Concentration of negative ions
Coions (ρZ–)

Stern layer

(a) (b)
Figure 4.12. Electrical double layer theory: (a) isolated particle in solution; (b) two approach-
ing particles in solution.

incorporated with the framework of adhesive contact models such as JKR or DMT.
This current section, therefore, deals only with electrostatic force on a charged
particle in an electrolytic solution, the so-called electrical double-layer force, and
defers discussion of more general electrostatic forces until Chapter 7. Additional
discussions of electrical double-layer forces can be found in books on colloid science
and intermolecular forces, including those by Russel et al. (1989), Probstein (2003),
Masliyah and Bhattacharjee (2006), and Israelachvili (1991).

4.3.1. Stern and Diffuse Layers

The free ions in the solution are either attracted to or repelled from the surface
of a charged particle depending on the sign of the surface charge. For instance, if
the particle surface is negatively charged, then positive ions within the solution will
drift toward the particle, as shown in Figure 4.12. The motion of these positive ions
toward the surface is limited by the finite ion size, so that only a fixed number of
ions can fit in the layer immediately next to the surface. This layer of essentially
motionless ions just outside of a charged surface is called the Stern layer. The width
of the Stern layer is approximately equal to one ionic radius. Because the ions within
the Stern layer are all of one sign, opposite to the sign of the particle surface charge,
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

102 Contact Mechanics with Adhesion Forces

the electrostatic potential ψ decreases rapidly within the Stern layer from a value ψs
at the particle surface to a value ψd at the outer edge of the Stern layer, called the
Stern plane. The value of the electrostatic potential just outside the Stern layer (ψd )
is known as the zeta potential.
The electrostatic potential is not reduced to zero by the fixed ions within the
Stern layer, and as a consequence other counter-ions within the solution are also
attracted to the particle surface. However, within this outer region the attraction of
these counter-ions is mitigated by the thermal motion of the ions. As a result, a layer
of mobile ions is formed in what is called the diffuse layer, located just outside of the
Stern layer. The diffuse layer consists of ions of both signs, but with a dominance of
counter-ions, that move under a combination of the bulk fluid motion, the thermal
motion of the ions, and the attractive force driving the ions to and from the surface.
The electrostatic potential decays approximately exponentially across the diffuse
layer, approaching zero at the outer edge of this layer. The thickness of the diffuse
layer is of the order of the Debye length λD , which is derived in the next section.
The diffuse layer and the Stern layer together constitute the electric double layer.

4.3.2. Ionic Shielding of Charged Particles

Under equilibrium conditions, the ion concentration of the ith ion species follows
the Boltzmann equation
 
z eψ
ni = ni∞ exp − i , (4.3.1)
kB T
where zi is the valence of the ith ion species, e is the elementary charge, ψ is the local
value of the electrostatic potential, kB is Boltzmann constant, and T is the absolute
temperature. As an example, for the case of CaCl2 salt, z1 for the calcium ion is +2
and z2 for the chloride ion is −1. The free charge density in the solution ρ e can be
expressed in terms of the ionic concentrations as

ρe = e ni zi . (4.3.2)
i

Equations (4.3.1) and (4.3.2) dictate how the electrostatic potential affects the
free charge density ρ e . However, the free charge density in turn changes the potential
via Possion’s equation
ε f ∇ 2 ψ = −ρe , (4.3.3)
where ε f is the fluid electrical permittivity. Combining the three equations thus
presented, we obtain the well-known Poisson–Boltzmann equation
  
z eψ
ε f ∇ 2ψ = − zi eni∞ exp − i . (4.3.4)
i
kB T

When the surface potential is small, that is, ψs is below about 25 mV, the zi eψ/kB T
term in (4.3.4) is much smaller than unity. The exponential term on the right-hand
side of (4.3.4) can be linearized, which is known as the Debye–Hückel approximation,
to obtain
λ2D ∇ 2 ψ − ψ = 0. (4.3.5)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.3. Electrical Double-Layer Force 103

The parameter λD has units of length and is known as the Debye length scale,
defined by
 −1/2

λD = ni∞ z2i e2 /ε f kB T . (4.3.6)
i

The value of the Debye length scale is solely dependent on the properties of the
solution, and it is independent of the surface charge or potential of the particle. For
instance, for a symmetric (z:z) electrolyte solution (e.g., NaCl) with molar concen-
tration M (mol/L), we obtain n+ = n− = n∞ = 1000Av M = 6.022 × 1026 M. Knowing
e to be equal to 1.602 × 10−19 C and for water at T = 298 K and ε f = 6.95 × 10−10
C2 N−1 m−2 (corresponding to a relative permittivity of εr = 78.5), the Debye length
scale is obtained as
0.304
λD = √ × 10−9 m. (4.3.7)
z M

Similarly, for (1:z) or (z:1) electrolytes with M (e.g. CaCl2 or Na2 CO3 ), we obtain

0.304
λD = √ √ × 10−9 m. (4.3.8)
1+z M

Consequently, the value of λD for a 0.01 M NaCl solution is 3.04 nm, and that of a
10–4 M NaCl solution increases to 30.4 nm. On the other hand, the value of λD for
pure water at a pH of 7 is approximately 960 nm, which is sufficiently large that it
might start to approach the size of the particles.

4.3.3. DLVO Theory

DLVO theory, which was developed independently during the Second World War
by Derjaguin and Landau (1941) and Verwey and Overbeek (1948), examines the
adhesive forces that arise when two charged particles collide with each other in
an electrolytic solution. These forces result from a combination of van der Waals
attractive force and an electrostatic repulsive force resulting from the overlap of the
ionic diffuse layers of each particle. Because colloidal solutions usually deal with
very small particles, with particle sizes ranging from approximately 0.1 to 1.0 µm, the
treatment of adhesive and elastic forces in the standard DLVO theory follows the
DMT model for adhesive contact mechanics. However, it is not difficult to extend
the theory to apply to the JKR model for larger particles as needed.
We recall that the DMT model for adhesive contact is based on the Derjaguin
approximation (4.2.8), which relates the adhesive force between two spheres with
effective radius R and separation distance h to the interaction potential per unit area
w p−p (h) between two planar surfaces separated by the same distance h, resulting
in the expression Fs−s (h) = 2π Rw p−p (h). This approximation is valid for arbitrary
adhesive forces provided that the separation distance h satisfies h R. In order
to obtain w p−p (h), we consider the problem of an electrolyte in a gap of width h
between two infinite planar surfaces, where the two surfaces are located at x = 0 and
x = h. If we further consider the special case of an electrolyte consisting of two types
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

104 Contact Mechanics with Adhesion Forces

of ions having equal and opposite valences (i.e., a z-z electrolyte), such as NaCl or
KCl, the Poisson–Boltzmann equation (4.3.4) reduces to
 
d2 ψ 2ez zeψ
= n sinh , (4.3.9)
dx2 εf ∞ kB T

where z and n∞ denote the valance and the ambient ion concentration of the
cation, respectively. When the dimensionless potential  ≡ ezψ/kB T 1, the
sinh(·) term in (4.3.9) can be linearized to obtain the one-dimensional Debye–Hückel
approximation
d2 ψ 2e2 z2 n∞
= ψ = κ 2 ψ, (4.3.10)
dx2 ε f kB T

where the inverse Debye length scale κ ≡ 1/λD is a constant.

In solving these equations, we first consider the case of the diffuse layer
along a single surface in an unbounded space (i.e., letting h → ∞). In this case,
the electrostatic potential within the diffuse layer is solved using the boundary
conditions

ψ (0) = ψs , ψ (∞) = 0. (4.3.11)

An exact solution to (4.3.9) is given by

 
1 + exp(−κx) tanh(s /4)
(x) = 2 ln , (4.3.12)
1 − exp(−κx) tanh(s /4)
where s is the dimensionless surface potential. For small surface potentials (s
1), (4.3.12) reduces to the solution of the Debye–Hückel approximation (4.3.10),
given by

(x) ∼ s exp(−κx). (4.3.13)

For large positive surface potentials (s 1), tanh(s /4) ∼

= 1 and (4.3.12) can
be approximated by
 
1 + exp(−κx)
(x) ∼ 2 ln , (4.3.14)
1 − exp(−κx)
from which it follows that the potential is independent of the surface potential at a
sufficient distance away from the surface. For this problem, a near-surface correction
must be introduced in order to satisfy the boundary condition at x = 0.
For large values of x, such that κx 1, the general solution (4.3.12) can be
approximated using the binomial theorem and the Taylor series approximation
ln(1 + x) ∼ x for small x as

 ∼ 4 exp(−κx) tanh(s /4). (4.3.15)

This result demonstrates that the electrostatic potential always decays exponentially
with x sufficiently far from the surface.
We now turn to the problem of determining the force per unit area that is
exerted on two infinite parallel planar surfaces separated by a distance h whose
diffuse layers overlap each other, as shown in Figure 4.13. The momentum equation
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.3. Electrical Double-Layer Force 105

ψ ψ
s s

Figure 4.13. Sketch of the gap between two parallel

charged surfaces in an electrolyte, showing the midplane
and the variation of the electric potential ψ.

x=0 x=h/2 x=h

for this one-dimensional problem in the absence of fluid motion or gravity reduces
to
dp
= ρe Ex , (4.3.16)
dx
where ρe is the free charge density defined by (4.3.2) and Ex = −dψ/dx is the x-
component of the electric field vector. The product on the right-hand side of (4.3.16)
is simply the electric field body force per unit mass acting on the fluid. Substituting
(4.3.1) and (4.3.2) for ρe into (4.3.16) and writing Ex = −dψ/dx gives
 
dp  zi eψ dψ
+ ezi ni∞ exp − = 0. (4.3.17)
dx i
kB T dx

Integrating this equation over x and using a point in the bulk solution at which ψ → 0
as a boundary condition yields a solution for pressure of the form
    
z eψ
p(x) = kB T ni∞ exp − i −1 . (4.3.18)
i
kB T

The result (4.3.18), when combined with the expression (4.3.1) for ion concen-
tration, leads to the contact-value theorem (Israelachvili, 1991, 225), which gives the
pressure within the gap as proportional to the ionic concentration evaluated at the
midplane x = h/2, or
 
 
p = kB T nmi (h) − nmi (∞) , (4.3.19)
i i

where the first term in parentheses in (4.3.19) is the sum of the midplane ionic
concentration when the plates are a distance h apart and the second term is the
sum of the ionic concentrations in the bulk electrolyte. The ionic concentration is
related to the electric potential ψ by the Boltzmann equation (4.3.1), so for a (z:z)
electrolyte we can write
       
ezψm ezψm z2 e2 ψm2 n∞
p = kB T n∞ exp − − 1 + exp −1 ∼ = , (4.3.20)
kB T kB T kB T
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

106 Contact Mechanics with Adhesion Forces

where the last expression assumes that the dimensionless electrostatic potential at
the midplane satisfies m 1. The first term in brackets in (4.3.20) represents the
cations and the second term is the anions, and the symbols z and n∞ in this equation
denote the valence and ionic concentration of the cations. Substituting the far-field
solution (4.3.15) for ψ (x) into (4.3.20) gives
 
ezψs
p = 64n∞ kB T e−κh tanh2 . (4.3.21)
4kB T
This expression is known as the weak overlap approximation because it uses the
far-field approximation for the electrostatic potential.
The interaction energy of the surfaces w p−p (h) can be obtained by integrating p
as indicated in (4.1.4), giving
 ∞  
ezψs
w p−p (h) = p(h)dh = (64n∞ kB T/κ )e−κh tanh2 . (4.3.22)
h 4kB T
This result can be used to obtain the electrical repulsion force during collision of two
charged spherical particles with effective radius R and separated by a distance h in
an electrolyte simply by applying the Derjaguin approximation (4.2.8) to write
 
−κh ezψs
Fs−s (h) = 2πRw p−p (h) = (128π Rn∞ kB T/κ )e tanh 2
. (4.3.23)
4kB T
For small values of the dimensionless surface potential s (corresponding to
approximately ψ s < 25 mV), a simpler form of (4.3.23) can be derived by linearizing
the tanh(·) term and using (4.3.6) to write κ 2 = 2n∞ z2 e2 /ε f kB T , giving

w p−p (h) ∼
= 2ε f ψs2 κ exp(−κh). (4.3.24)

The corresponding sphere-sphere interaction force for this case is given by

Fs−s (h) ∼
= 4π ε f ψs2 κR exp(−κh). (4.3.25)

The electric double-layer interaction is quite different from the van der Waals
interaction discussed in the previous section. Whereas the interaction potential
between two plane surfaces for van der Waals force decays with distance in pro-
portion to h−2 , the double-layer interaction potential decays exponentially on a
length scale equal to the Debye length λD . Figure 4.14 plots the interaction energy
per unit area for two parallel plane surfaces due to the electrical double layer alone,
the van der Waals force plus the short-range repulsive force alone, and the combi-
nation of all three forces in a (1:1) electrolytic solution for different values of the
surface electrical potential ψs . Because λD ∝ M−1/2 , the DLVO interaction energy
depends mainly on the electrolyte concentration M and the surface potential ψs .
The upper solid curve in Figure 4.14 represents a well-stabilized colloidal system,
typical of a high surface potential in a dilute electrolyte. The long Debye length causes
a strong long-range repulsive force that may extend to several nanometers, as far
as 10 nm. The large energy barrier caused by the electrical repulsive force prevents
the approach of two colloidal particles from short-range contact. If this barrier is
larger than the thermal energy kB T of the colloidal particles, the system should be
always stable in the absence of external forces (e.g, fluid shear forces), and there
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.4. Protein Binding 107

1.5

1.0 Energy barrier Electric double-layer repulsion

Interaction energy

0.5 High ψs

Medium ψs
0

Low ψs Secondary minimum

-0.5
van der Waals adhesion

-1.0
0 1 2 3 4 5 6 7 8 9 10

Dimensionless distance, κx
Figure 4.14. Schematic showing interaction energy per unit area between two parallel plates
as a function of separation distance for the electrical double-layer alone (upper dashed line),
for van der Waals attraction force alone (lower dashed line), and for the combined DLVO
theory with different values of the surface potential.

will be little particle agglomeration. In such cases, computational methods such as

Brownian or Langevin dynamics might be appropriate for modeling the system.
The lower solid curve in Figure 4.14 represents cases with low surface potential
values in which the energy barrier falls below zero. In these cases, the electrical
double-layer repulsive barrier is so small that the suspension will be unstable. In
this regime, the colloidal particles will coagulate rapidly, as they will be attracted
to the deep attractive “primary” energy minimum occurring at separation h = δ
(not shown) where the attractive van der Waals force and the short-range repulsive
force are in balance. DEM is suitable for prediction of the coagulation rate of
the particles and for examining the microstructures of particle agglomerates or
aggregates, provided that the condition λD R is met. In the presence of a fluid
flow, the agglomerates will be broken up by fluid shear at the same time as they
agglomerate following collisions.
There is also a transition region between the extremes of a stable and an unstable
suspension in which a second energy minimum occurs, as illustrated by the middle
solid curve in Figure 4.14. This secondary energy minimum is usually located several
nanometers away from the peak of the energy barrier, and it is typically observed in
cases where both the surface potential and the electrolyte concentration are relative
large. If the energy barrier is too high for a given colloidal particle to overcome, it may
either be attracted to the weaker secondary minimum or remain totally dispersed in
the solution. This condition is known as kinetically stable.

4.4. Protein Binding

A particularly interesting area of application of the DEM approach is in the area of
transport of biological flows containing cells, such as blood. In addition to plasma
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

108 Contact Mechanics with Adhesion Forces

and red blood cells (RBCs), which are the most numerous components, blood also
contains various types of white blood cells, platelets, and occasional foreign bodies
such as cancerous cells and liposomes used for drug delivery. Blood flow may be
examined at a mesoscale level using a particle flow framework in which each of these
discrete components is modeled as a type of particle (Longest and Kleinstreuer, 2003;
Chesnutt and Marshall, 2009). Adhesion of these different cells and particles to each
other and to the endothelial cells lining the blood vessel walls is of critical importance
for many biological processes. For instance, in blood flows cell adhesion governs
processes such as clotting of red blood cells, thrombosis formation due to platelet
activation, white cell adhesion to the endothelial wall prior to extravasation into
surrounding tissues, and formation of red blood cell rouleaux. Diseases such as sickle-
cell disease or high shear flow due to the presence of artificial elements such as stints
or artificial heart valves can adversely modify cell vessel wall adhesion properties,
leading to serious health problems. Adhesion of bacterial cells is necessary for the
growth of bacterial diseases in surroundings where surfaces are exposed to fluid
flow. This is of particular importance in biofilms, which contribute to infections
related to prosthetic heart valves, artificial joints, and catheters (Smith, 2005). The
redistribution of cancer cells from one location to another, known as metastasis, is
correlated with a reduction in the ability of tumor cells to adhere to each other
(Cavallaro and Christofori, 2001).
Since blood particles (as we will loosely refer to the discrete entities suspended
in blood) typically have radii in the range of 1 to 5 µm and are fairly compliant, the
basic framework for the adhesive contact described by the JKR model is applicable
to these problems. Although both van der Waals and electrical double-layer forces
are present for these biological problems, the dominant adhesion force typically
arises from phenomena such as entanglement in polymeric networks, osmotic forc-
ing and protein binding. One prevalent mechanism for cell adhesion is the so-called
ligand-receptor binding process, in which ligand proteins attached to the membrane
of one cell bond to receptor proteins attached to the membrane of a second cell. The
bonding occurs by intermolecular forces, such as ionic or hydrogen bonds and van
der Waals forces, rather than by covalent bonds. The characteristic length scale over
which protein binding occurs is on the order of 10 nm, which is similar to the length
scale for van der Waals adhesion. Unlike many other adhesion forces, the protein
connections involved in ligand-receptor binding are often specific to particular cell
types. Opposing the attractive force from the ligand-receptor bonds is a nonspecific
repulsive force due both to the fact that cell surfaces are negatively charged (and
therefore repel each other) and to the steric repulsion caused by the presence of a
layer of negatively charged glycoproteins coating the cell membranes (called the gly-
cocalyx). An illustration of these components of cell adhesion is shown in Figure 4.15.
An early model of cell adhesion by ligand-receptor binding was the kinetic
model proposed by Bell (Bell, 1978; Bell et al., 1984), which assumes that the force
exerted by each protein bond can be modeled as a spring with spring constant σ and
equilibrium length xo. The energy released by compressing or stretching a bond to a
length xb is
σ
Es = − (xb − x0 )2 . (4.4.1)
2
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.4. Protein Binding 109

Cell 1

Figure 4.15. Schematic diagram showing ligand and

receptor proteins (in black), and long-chain molecules
that make up the glycocalyx layer (in grey).
Cell 2

The protein bonds are assumed to break at some maximum bond length, such
that the adhesive force vanishes when the bond length exceeds the contact region
gap size δ. The total energy associated with receptor-ligand bonds per unit area of the
membrane is the product Nb Es of the single-bond energy Es and the number density
Nb (t ) of bonds joining the cell membranes. The formation and elimination of bonds
is modeled by a kinetic equation for the bond number density Nb (t ) given by

dNb
= k f (N 0 − Nb )(Nr0 − Nb ) − kr Nb , (4.4.2)
dt

where Nr0 and N 0 are the initial receptor and ligand densities on the membrane,
respectively, and k f and kr are the forward and reverse reaction rate coefficients.
The reaction rate coefficients vary with the length of the bond in accordance with
(Dembo et al., 1988)
   
σts (xb − x0 )2 (σ − σts )(xb − x0 )2
k f = k f 0 exp − , kr = kr0 exp , (4.4.3)
2kB T 2kB T

where k f 0 and kr0 are the initial forward and reverse equilibrium reaction rates, σts
is the “transition state spring constant,” kB is the Boltzmann constant, and T is the
absolute temperature.
Because the gap thickness in the JKR model is assumed to remain approximately
constant during the binding process, the rate coefficients in (4.4.3) are constants and
the rate equation (4.4.2) has the form as a Ricotti equation for the number density
Nb (t ). If an initial condition is assumed to be a state where the two cells have just
collided, such that there are no bonds at the initial time t0 , an analytical solution for
Nb (t ) is obtained as (Chesnutt and Marshall, 2009)
 
2A tanh 12 (B2 − 4AC)1/2 (t − t0 )
Nb (t ) =   , (4.4.4)
−B tanh 12 (B2 − 4AC)1/2 (t − t0 ) + (B2 − 4AC)1/2

where A ≡ k f N 0 Nr0 , B ≡ −[k f (N 0 + Nr0 ) + kr ], and C ≡ k f . The coefficients are

restricted by the condition B2 − 4AC > 0, which is satisfied for all cases provided
that N 0 = Nr0 . The value of Nb (t ) asymptotes at long time to an equilibrium value
Nb (∞) given by
 
Nb (∞) = 2A/ − B + (B2 − 4AC)1/2 (4.4.5)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

110 Contact Mechanics with Adhesion Forces

We make the usual JKR approximation that the adhesive force acts only over a
circular contact region of radius a(t ), over which the surface separation thickness is
taken as constant. Integration of the receptor-ligand binding energy per unit area,
Nb Es , over the contact region yields the net adhesion binding energy for a particle-
particle collision event as
 a(t )  a
US (t ) = 2π Es Nb (t, r)rdr = −π σ (xb − x0 )2 Nb (t, r)rdr. (4.4.6)
0 0

The bond number density depends on radius r through the impact time t0 (r),
which is defined as the time at which the particle surfaces at radius r within the contact
region first approach to within the equilibrium separation distance δ. The time scale
for bond formation TB can be approximated from (4.4.4) as TB = 2/(B2 − 4AC)1/2 .
When TB is much smaller than the contact time scale, the additional approximation
can be made that the surface impact time t0 (r) is nearly uniform over the contact
region, such that (4.4.6) reduces to
πσ
US = − (x − x0 )2 Nb (t )a2 (t ). (4.4.7)
2 b
Recalling that 2γ is the work per unit area required to separate two surfaces,
where γ is the surface energy density, it follows that US = π a2 (2γ ). Consequently,
(4.4.6) is equivalent to assuming a time-dependent effective surface energy density
γ (t ) as
 1
1
γ (t ) = − σ (xb − x0 )2 Nb (t − t0 (s))sds , (4.4.8)
2 0

which for TB TC reduces to the equilibrium value

γ = − 14 σ (xb − x0 )2 Nb (∞). (4.4.9)

Using the modified surface energy density expression (4.4.8) or (4.4.9), the expres-
sions for contact forces of two cells colliding under ligand-receptor binding are the
same as the expressions for normal force and sliding, twisting, and rolling resistance
discussed in Section 4.2. Ligand-receptor binding can therefore be treated mathe-
matically as a van der Waals adhesive contact process with time-dependent surface
energy density, under the context of the JKR adhesive contact model.
There have been a variety of extensions to the above ligand-receptor binding
model. A number of investigators have noted that the receptor proteins can move
along the interface during the bonding process. A model for this motion was devel-
oped by Tozeren et al. (1989), who add a diffusive term to the right-hand side of
(4.4.2). A model of this type was used by Agresar (1996) for two-dimensional simula-
tions of the collision of two cells, whereas other simulation approaches neglect pro-
tein mobility (N’Dri et al., 2003). Experimental work on measurement of receptor-
ligand binding kinetics is reported by Chesla et al. (1998) and Tachev et al. (2000).
Competition between multiple ligand-receptor species is discussed by Zhu and
Williams (2000) and Coombs et al. (2004), and a more comprehensive review of
receptor-ligand binding is given by Zhu (2000).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.5. Liquid Bridging Adhesion 111

ρ2
Figure 4.16. Liquid bridge forces for two h ρ1

elastic spheres: (a) capillary cohesion; (b)

viscous effect due to squeezing; (c) viscous
effect due to shearing.

(a) (b) (c)

4.5. Liquid Bridging Adhesion

In a humid environment, each particle of an aerosol or granular flow is surrounded by
a thin liquid film due to condensation of water from the atmosphere. The addition of
even a small amount of liquid in a dry particle system can lead to strong cohesive force
between the particles, significantly changing the particle system behavior (Halsey and
Levine, 1998; Fiscina et al., 2010). When two particles with thin liquid films collide,
the liquid films join to form a “liquid bridge” stretching between the particles. This
liquid bridge introduces a capillary force Fcap that pulls the two particles toward each
other, leading to adhesion of the particles, as shown Figure 4.16a. In addition, the
liquid film introduces an enhanced frictional force Fvisc between the particles in both
the normal and tangential directions due to the higher viscosity of the liquid filling the
contact region compared to the surrounding gas, as illustrated in Figure 4.16b and c.
The total normal liquid-bridge force is given by the sum of the capillary force and
the viscous force. Computation of particle collision with liquid-bridging adhesion
requires models for both the capillary and viscous forces on the particles, as well as a
criterion for rupture of the liquid bridge at a critical separation distance. Much of the
earlier literature examined the capillary force and bridge rupture condition for two
static spheres, good summaries of which are given by Mehrotra and Sastry (1980)
and Lian et al. (1993). The significance of the viscous force arising from relative
motion of the two particles on the normal force and rupture condition was pointed
out by Ennis et al. (1990).

4.5.1. Capillary Force

The problem of the capillary force Fcap acting between two equally sized spherical
particles with radius r p is examined. It is assumed that an axisymmetric liquid bridge
connects the two spheres, with separation distance h(t ) between the spheres. The
problem is described in terms of a cylindrical polar coordinate system (r, z). The
radius of curvature of the liquid-gas interface profile in the r-z plane is denoted by
ρ1 , which is assumed to be constant along the liquid-gas interface, and the radial
position of the liquid bridge at the “neck” (or mid-plane between the spheres) is ρ2 .
The solid-liquid contact angle is θ and the so-called half-filling angle is φ, as shown
in Figure 4.17a.
In a static condition, an equal and opposite force acts on each particle which
arises from two distinct effects: (1) the vertical component of the surface tension force
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

112 Contact Mechanics with Adhesion Forces

rp

φ z z
V Figure 4.17. Close-up sketches showing (a)
θ r r
geometrical parameters for liquid bridge
ρ1 between two spherical particles of the same
h ρ2 midplane
radius and (b) control volume V used for the
“gorge approximation.”

(a) (b)

on each particle and (2) the difference in hydrostatic pressure across the liquid-gas
interface due to the interface curvature. We refer to the sum of these two effects as
the capillary force, which is given by

Fcap = −2π r p σ sin φ sin(φ + θ ) − π r2p pI sin2 φ, (4.5.1)

where pI = pliq − pgas is the pressure jump over the gas-liquid interface and σ
is the liquid-gas surface tension. We follow the convention that a positive force is
repulsive as has been used in the rest of the book. The Young-Laplace formula
gives

pI = 2σ κ̄, (4.5.2)

where κ̄ ≡ 0.5(ρ1−1 − ρ2−1 ) is the mean curvature of the surface. The negative sign
before the ρ2−1 term in this equation results from the fact that the center of curvature
for the tangent circles corresponding to ρ1 and ρ2 are on opposite sides of the
interface. An approximate solution to (4.5.1) can be obtained using the so-called
“gorge approximation” (Hotta et al., 1974; Lian et al., 1993). This approximation
uses the control volume V shown in Figure 4.17b, which passes through the midplane
between the spheres and encloses one of the spheres. Using this control volume, the
force on the sphere can be written in terms of the pressure force and surface tension
force exerted on the midplane as

Fcap = −2π σ ρ2 − π ρ22 pI . (4.5.3)

Substituting (4.5.2) into (4.5.3) gives

 
ρ2
Fcap = −π σ ρ2 1+ . (4.5.4)
ρ1
The radii of curvature at the midplane can be written in terms of the angles φ
and θ as (Pitois et al., 2000)
h/2 + r p (1 − cos φ)
ρ1 = , ρ2 = r p sin φ − [1 − sin(φ + θ )]ρ1 . (4.5.5)
cos(φ + θ )
The static contact angle θ is a material property of the liquid-gas-solid system. The
half-fill angle φ depends on the liquid bridge volume. For the general case of liquid
bridges of arbitrary volume, it is necessary to use an iterative method to solve for φ
from the given liquid volume, which can be used in (4.5.5) to obtain ρ1 and ρ2 , and
then substituted into (4.5.4) to obtain the capillary force. Mathematical details are
provided by Lian et al. (1993) and Hotta et al. (1974).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.5. Liquid Bridging Adhesion 113

sphere center

sphere 1
Β
φ rp θ ρ1
θ
2s + h liquid
b bridge C Α
s
liquid bridge triple point
contact point sphere 2
(a) (b)
Figure 4.18. Sketches illustrating geometrical arguments for liquid bridge force derivation
with small fill angle. (a) Relationship between length scales s and b. (b) Close-up of region
between spheres, with sphere surfaces approximated as parallel lines.

A particularly useful limiting case is that where the volume of the liquid bridge is
very small, so that the radial location r = b of the liquid-solid-gas triple point on the
particle satisfies b r p . This case corresponds to the small fill-angle approximation
φ 1, and consequently ρ2 r p . With reference to Figure 4.18a, the triple-point
radius b and height s relative to the contact point can be written in general as
b = r p sin φ and s = r p (1 − cos φ). For φ 1, a Taylor series expansion can be used
to approximate these quantities to leading order in φ as

b = r p φ, s = r p φ 2 /2. (4.5.6)

Eliminating φ from these two equations gives the useful formula

π b2 ∼
= 2π r p s. (4.5.7)

A second useful formula can be obtained using the geometrical argument illus-
trated in Figure 4.18b. The top and bottom horizontal lines in this figure each repre-
sent a horizontal line passing through the triple point on each particle. The distance
between these parallel lines is equal to 2s + h. For small φ, the horizontal line is
nearly tangent with the sphere surface at the triple point, so the angle between the
horizontal line and the liquid-gas interface is approximately equal to the contact
angle θ. Under this approximation, the angle at vertex B of the right triangle ABC
shown in Figure 4.18b must also be equal to θ, with the consequence that the distance
between the horizontal lines is given by

2s + h ∼
= 2ρ1 cos θ . (4.5.8)

For sufficiently small separation distance h = O(s), it follows from (4.5.7) that

ρ1 /ρ2 = O(s/b) = O(φ). (4.5.9)

Consequently, for small liquid volumes with φ 1, the force arising from the pres-
sure reduction pI ∼= σ /ρ1 dominates the surface tension force, as well as the pres-
sure reduction force associated with the azimuthal curvature 1/ρ2 , and the capillary
force expression (4.5.4) reduces to

Fcap = −π σ ρ22 /ρ1 . (4.5.10)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

114 Contact Mechanics with Adhesion Forces

From (4.5.5) with ρ1 ρ2 , we can write ρ2 ∼= r p sin φ = b, so that using (4.5.7) and
(4.5.8) an expression for capillary force with small fill angle is obtained as
 
h −1
Fcap = −2π r p σ cos θ 1 + . (4.5.11)
2s
This is the same expression as that derived by Maugis (1999, 315); however, Maugis
used a derivation based on a thermodynamic approach related to crack propagation
theory (see also Maugis, 1987).
It is convenient to write the parameter s in (4.5.11) in terms of the liquid volume
VL per contact. If Z(r) denotes the height between the sphere surfaces at a given
radial position r, given by (4.2.5) for h r p , then VL can be written as
 b  b
π b4
VL = 2π Z(r)rdr = 2π (h + r2 /r p )rdr = π b2 h + , (4.5.12)
0 0 2r p

where we have used (4.5.7) to write s = b2 /2r p . Equation (4.5.12) is a quadratic

equation for b2 , which yields the solution
⎡  1/2 ⎤
2VL
b2 = hr p ⎣−1 + 1 + ⎦. (4.5.13)
π h2 r p

Substituting s = b2 /2r p into (4.5.11) with b2 given by (4.5.13), an expression for

capillary force between two equal spheres is obtained as
Fcap = −2π r p σ G f cos θ , (4.5.14)
where the coefficient G f is defined by
" #−1/2
G f ≡ 1 − 1 + 2VL /π r p h2 . (4.5.15)

The term in brackets in (4.5.15) approaches zero as h → 0, implying that G f → 1

for cases where the particles are touching. While we have derived the expression
(4.5.14) for interaction of two equal spheres, it can readily be generalized to the
case of interaction of two arbitrarily sized spheres simply by replacing 2s in (4.5.8)
with the sum s1 + s2 of the triple-point heights from the two spheres. The resulting
general expression has the same form as (4.5.14) and (4.5.15) with r p replaced by
2R, where R is the effective radius defined by (3.1.1). The liquid bridge volume can
be normalized as V ∗ = VL /R312 , where the harmonic mean diameter R12 is equal
to twice the effective radius (R12 = 2R). Figure 4.19 compares the prediction from
(4.5.14) with experimental data from Pitois et al. (2000) and Willett et al. (2000). Both
theory and experiment show a monotonic decay in the capillary force as the sepa-
ration distance h(t) is increased. The expression (4.5.14) is found to be reasonably
accurate both for collisions of particles with equal radii and for collisions of parti-
cles with unequal size over the full range of separation distances, up to the rupture
distance.
Adding the capillary force (4.5.14), the Hertz contact force (3.2.8), and the
dissipation force (3.3.1), the total normal contact force is given by
√
Fn = 4πRσ G f cos θ + 43 xE Rδ 3/2 − ηN vR · n. (4.5.16)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.5. Liquid Bridging Adhesion 115

1.0
Data from Pitois et al. (2000)
V*=0.0230;
− Fcap/(4πRσcosθ) 0.8 V*=0.0010
Data from Willett et al. (2000)
V*=0.0055;
0.6 V*=0.0023;
V*=0.0010

0.4 Model prediction

0.2

0.0

0.00 0.05 0.10 0.15 0.20 0.25

h/2R
Figure 4.19. Comparison of predictions from Equation (4.5.14) with experimental data from
Pitois et al. (2000) and Willett et al. (2000) for normalized liquid bridging force as a function
of separation distance between two spheres. [Reprinted with permission from Li et al., 2011.]

The prediction (4.5.16) for normal contact force in the presence of a liquid bridge is
plotted in Figure 4.19. The parameters used in the simulation are consistent with a
case examined experimentally by Willett et al. (2000). The particle has a diameter
of 2.4 mm, which is sufficiently large that van der Waals force is negligible. The
liquid surface tension σ = 0.206 N/m, and nondimensional liquid volume V ∗ = 0.001.
Considering that the modulus of most solids is of magnitude 109 Pa, the difference
of length scale between deformation and adhesion is several orders of magnitude.
In order to compare liquid bridge adhesion with van der Waals adhesion, we also
predict the JKR model results using the force expression (4.2.35) with the surface
energy γ = 0.206 J/m2 and no liquid bridge adhesion (Li et al., 2011). Results of two
computations are shown in Figure 4.20, one with E = 105 Pa and one with E = 106
Pa. It is found that for δN > 0, the force predictions with the combined theory given
by (4.5.16) are very close to the JKR results from (4.2.35) with no liquid bridge. This
agreement arises because the Hertz elastic rebound force dominates either the van
der Waals or liquid-bridge adhesion forces in this regime. On the other hand, for
δN < 0, indicating that the spherical shapes of the particles do not overlap, the JKR

5
+ Liquid bridge, E=10 Pa
2 5
JKR theory, E=10 Pa
6
Liquid bridge, E=10 Pa
6
JKR theory, E=10 Pa
Fn / 2πRσ

1
Figure 4.20. Prediction for the normal
contact force in the presence of a liquid
bridge using (4.5.16), with d = 2.4 mm and 0
σ = 0.206 N/m. [Reprinted with permission
from Li et al., 2011.]
-1

-2
-0.25 -0.20 -0.15 -0.10 -0.05 0.00 0.05

δN /R
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

116 Contact Mechanics with Adhesion Forces

(I) (II) (III)

–Fcap Figure 4.21. The behavior of the liquid bridge

force between two particles with rough surfaces,
Fcap = –2πσR cosθ
shown on a plot of capillary force versus liquid
fraction φLB . The three regimes include: I, asper-
ity regime; II, roughness regime; and III, spher-
ical regime. The capillary force asymptotes to
a constant value in regime III. [Reprinted with
permission from Halsey and Levine (1998).]
φLB
φ1 φ2

predictions and the predictions from the combined theory given by (4.5.16) deviate
strongly. Specifically, JKR theory predicts that the particles will reach a critical state
and snap apart from each other at a relatively small separation distance on the order
of h/R = −δN /R ∼ = 0.01 − 0.02. The case with liquid bridging, on the other hand,
exhibits significant force out to h/R = −δN /R ∼= 0.2.

4.5.2. Effect of Roughness on Capillary Cohesion

The capillary force acting between two particles with rough boundaries is more
complex than for smooth particles for cases where the liquid film thickness is less than
the size of the surface asperities. The particle surface roughness can be characterized
by a mean radial height r and lateral separation distance d of the surface asperities.
We denote by φLB the ratio of the volume of liquid to the particle volume. Three
kinds of liquid-bridge regimes are identified as a function of φLB : (1) the asperity
regime (φLB < φ1 ), where the cohesive force depends primarily on surface roughness
in the neighborhood of the contact point; (2) the roughness regime (φ1 ≤ φLB < φ2 ),
where the Laplace pressure in the liquid bridge between surface asperities is set
by the asperity height; and (3) the spherical regime (φ2 < φLB ), where the liquid
film thickness is larger than the size of the surface asperities and surface roughness
no longer significantly affects the cohesive force. As illustrated in Figure 4.21, the
cohesive force between the particles depends approximately on the cube root of the
liquid volume fraction φLB in the asperity regime, it increases approximately linearly
with φLB in the roughness regime, and it approaches a constant in the spherical
regime (Halsey and Levine, 1998). Scaling estimates by Halsey and Levine (1998)
indicate that in general, φ1 ≈ r 2d /Vp and φ2 ≈ 2r R/Vp ≈ ( r /R)2 , where Vp is the
volume of a single particle. A similar scaling was observed in an experimental study
of effect of a liquid film on a spouted bed by Zhu et al. (2013).
The different regimes of rough particle interaction result in distinct macroscopic
behaviors of the particles. For example, the effect of different liquid fractions on
the critical angle θC of a conical sand pile was examined theoretically by Halsey and
Levine (1998) and experimentally by Mason et al. (1999). Within the asperity regime
(φLB < φ1 ), the critical angle does not seem to be significantly affected by the liquid
surface tension σ , implying that the particle behavior is much the same as for a dry
granular system. In the spherical regime (φLB ≥ φ2 ), the cohesive forces between
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.5. Liquid Bridging Adhesion 117

sphere 1
Figure 4.22. Sketch illustrating the cylindrical region of
Z(r) b
radius b (shaded) used to approximate the liquid bridge for
calculation of the viscous force.
sphere 2

two particles are essentially the same as those between two smooth particles. In
the roughness regime (φ1 < φLB < φ2 ), the cohesive force due to the capillarity is
approximately linear in the volume of the liquid bridges, so that using (4.5.14) we
can write Fcap approximately as
φLB
Fcap = 4π Rσ G f cos θ . (4.5.17)
φ2

4.5.3. Viscous Force

In addition to the capillary force, there is an additional viscous force as two particles
move toward or away from each other due to the greater viscosity of the liquid
compared to the surrounding gas. Following the approach of Matthewson (1988),
the viscous force is evaluated using an approximation of the liquid bridge as a
cylindrical region with radius b shown in Figure 4.22. If Z(r) is the distance between
the sphere surfaces at radius r, the pressure within the liquid bridge is given by the
lubrication equation (2.4.1),
 
1 ∂ 3∂p
ḣ = rZ , (4.5.18)
12μL r ∂r ∂r
where ḣ is negative for approaching spheres. Integrating twice over r, the pressure
within the liquid bridge is obtained as
 r
ξ
p(r, t ) − p0 = 6μL ḣ 3 (ξ )
dξ . (4.5.19)
b Z
For small separation distances Z = h + r2 /r p , so performing the integration
(4.5.19) gives
 Z(r)  
dZ 3 1 1
p(r, t ) − p0 = 3μL r p ḣ = μL r p ḣ − . (4.5.20)
Z(b) Z
3 2 Z2 (b) Z2 (r)
Integrating the pressure over the region with area π b2 wetted by the liquid bridge
gives
 b
3π
Fvisc = 2π [p(r) − p0 ]rdr = − μL r2p ḣ[1 − h/Z(b)]2 (4.5.21)
0 2h
From (4.5.13) and (4.5.15), it follows that 1 − h/Z(b) = G f , so that (4.5.21) becomes

3 ḣ
Fvisc = − π μL r2p G2f . (4.5.22)
2 h
The result (4.5.22) is the same as the lubrication force expression (2.4.14) for
unbounded spaces with the exception of the factor G f . A similar derivation can be
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

118 Contact Mechanics with Adhesion Forces

done for two spheres of arbitrary size, for which case the sphere radius r p in (4.5.22)
is replaced by half of the equivalent radius R and (4.5.22) becomes

Fvisc = −6π μL R2 G2f ḣ/h. (4.5.23)

This is the same as the formula derived by Matthewson (1988) for collision of a
sphere with a flat plate, for which problem R is equal to the sphere radius.
The ratio of the viscous force to the capillary force is given by
Fvisc 3 R μL ḣ G f
= . (4.5.24)
Fcap 2 h σ cos θ

The ratio μL ḣ/σ is simply the capillary number Ca for the particle collision. The
viscous force becomes particularly important for high approach velocities (large Ca)
and small separation distance (small h/R).

4.5.4. Rupture Distance

The liquid bridge adhesion forces apply only up to the critical separation distance
hrupt at which the bridge ruptures. Lian et al. (1993) numerically computed the
rupture distance in a static case by minimization of the free energy subject to the
constraint of constant liquid volume. This paper notes that the computational results
are well approximated by a simple approximate expression of the form
 
θ  1/3
hrupt /r p = 1 + VL /r3p . (4.5.25)
2
In an experimental study, Pitois et al. (2001) modified this expression to include
effects of particle motion on rupture of the liquid bridge. Their experimental data
are reported as a function of the capillary number Ca ≡ μL |vR · n|/σ , where vR · n is
the normal component of the particle relative velocity and μL is the liquid viscosity.
Pitois et al. find that over the interval 0.001 ≤ Ca ≤ 0.1 the relative change in rupture
distance in the dynamic case compared with the static value varies in proportion to
Ca1/2 , which yields a modified rupture criterion for dynamic contact as
 
θ   1/3
hrupt /r p = 1 + 1 + Ca1/2 VL /r3p . (4.5.26)
2

4.5.5. Capillary Torque on a Rolling Particle

The presence of a liquid bridge might also introduce modifications to the resistance
forces and torques that occur during collisions with oblique particle motion, result-
ing in motions such as sliding, rolling, and twisting. Although most of the literature
on liquid bridge adhesion has been limited to normal particle interactions, there
have been two recent papers that experimentally examined the problem of a particle
rolling on a flat surface in the presence of a liquid film. The first of these papers, writ-
ten by Bico et al. (2009), considers the rolling motion of a particle down an inclined
plane that is covered with a thin film of transparent oil measuring 50–500 µm in
thickness. Experiments are conducted with different sphere materials and sizes,
lubricating oils with different viscosity and surface tension, and a range of slope
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.5. Liquid Bridging Adhesion 119

angles for the inclined plane. Some “overhang” experiments are also conducted,
where the sphere is underneath the plane and the gravitational force is countered
by the capillary force of the liquid bridge. Although the presence of the liquid film
both modifies the sliding friction coefficient of the particle on the planar surface and
introduces an enhanced viscous force within the fluid layer, the experimental results
also indicate a significant influence of surface tension on the particle rolling velocity.
In fact, the experiments indicate that the particle rolling speed varies approximately
as v ∝ σ −0.35 as the film surface tension is changed. The authors explain the effect
of surface tension in these experiments by noting that in the presence of rolling, the
meniscus surrounding the particle contact point becomes asymmetric, and they sug-
gest that this asymmetry may give rise to a tangential capillary force and torque that
opposes the rolling motion. Visualization of the wake structure in these experiments
exhibits three different regimes as a function of the capillary number Ca. For slowly
rolling particles with Ca < 1, the meniscus has an approximately circular shape. For
more rapidly rolling particles with Ca > 1, the meniscus breaks up to form two ridges
that trail behind the particle, leaving a tire-print type track in the liquid film. If the
particle is overhanging underneath the inclined plane, the observed track has only
one ridge.
Schade and Marshall (2011) report a series of experiments conducted at low cap-
illary numbers, where a particle mounted on a wire is allowed to roll on a translated
glass plate at a fixed speed in the presence of a liquid film. The liquid properties
and rolling velocity are varied. Because the centroid position of the particle was
fixed in these experiments, it was possible to perform detailed imaging of the liquid
film using both the laser-induced fluorescence (LIF) and particle-image velocimetry
(PIV) methods in order to examine the meniscus asymmetry in detail. Although
the meniscus profile in these experiments was approximately circular, distinct differ-
ences were observed in both the contact angles and radial locations of the contact
points between the leading and trailing sides of the particle. A sketch illustrating
the observed differences is given in Figure 1.12b. On the leading side of the particle
(the side toward which the particle is rolling), the contact angle is increased relative
to that in the static state and the contact point moves inward toward the contact
point. On the trailing side of the particle, the contact angle decreases slightly rel-
ative to the ambient and the contact point moves outward away from the contact
point.
The net effect of the contact point motion is that the meniscus slightly trails
behind the particle, so that the horizontal position of the center of the meniscus is
slightly behind the horizontal position of the particle centroid. As a consequence
of this shift, the line of force of the attractive capillary pressure force, which arises
from the meniscus curvature, no longer passes through the particle center, but rather
passes slightly behind the particle center by a shift distance ξ . This backward shift
gives rise to a torque that impedes the particle rolling motion, with magnitude equal
approximately to the magnitude of the capillary force Fcap , given by (4.5.14), times
the meniscus shift distance ξ . This mechanism for rolling resistance resulting from
a backward shift in the region over which the adhesion force acts is analogous to
that discussed with respect to the rolling resistance torque due to van der Waals
adhesion in Section 4.2.4, with the difference that in the current case it is not the
contact region itself that shifts, but the liquid meniscus.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

120 Contact Mechanics with Adhesion Forces

Figure 4.23. Plot showing the experimental data of Schade and Marshall (2011) for advancing
contact angle (open symbols) and receding contact angles (filled symbols) as a function of
capillary number, compared to the correlations of Hoffman (1975) for advancing contact
angle (solid line) and of de Gennes (1986) for receding contact angle (dashed line). Data is
obtained for different Reynolds and capillary numbers. Asymptotic values of the advancing
and receding contact angles are indicated by dashed-dotted lines.

A second mechanism causing an additional torque on a rolling particle in the

presence of a liquid film arises from the difference in contact angle between the par-
ticle leading and trailing sides. Measurements of these contact angles as a function
of capillary number are presented in Figure 4.23. The contact angle at the lead-
ing edge of the particle compares well with the correlation of Hoffman (1975) for
advancing contact angle on a flat surface for low capillary numbers (Ca < 0.01), but
then it asymptotes to an approximately constant value at higher capillary numbers.
Similarly, the contact angle at the trailing side compares well with the theoretical
expression of de Gennes (1986) for capillary number below about 0.01, but asymp-
totes to a constant value at higher capillary numbers. It therefore appears that the
asymmetry in the meniscus shape can be related to the classical contact angle hys-
teresis phenomenon, in which the contact angle of a meniscus on a moving plate can
vary between an advancing and a receding value, both of which are dependent on
the plate speed.

4.6. Sintering Force

Sintering is a thermal process for bonding two or more particles into a single coherent
solid structure via mass transport between the particles, some of which occurs at the
atomic scale. Sintering processes play an important role in the consolidation of metal
and ceramic powders and polymeric particles. These types of problems have taken on
an increasingly important role in recent years with the development and widespread
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.6. Sintering Force 121

Figure 4.24. Sketch illustrating the four

stages of sintering: (a) adhesion stage;
(b) initial sintering stage; (c) interme- (a) (b)
diate sintering stage; (d) final sintering
stage. Grain boundaries are indicated
using a gray line.

(c) (d)

application of 3D printing technologies. In the growing field of nanotechnology,

coalescence of nanoparticles via sintering governs growth of the particle size and
properties in certain flame synthesis and chemical-vapor decomposition synthesis
techniques for nanoparticle production.
The sintering of a particle packing proceeds through a series of stages, as shown
in Figure 4.24. The first image in this figure represents a collection of loose particles
in the adhesion stage, in which particles first attach to each other (e.g., via one of
the adhesion forces discussed previously in this chapter). Next, there is an initial
sintering stage in which particle bonding occurs over the small contact regions, but
the particles are still very distinct. As the bonds continue to grow, this progresses
into an intermediate sintering stage in which the volume of bonds grows and begins
to fill up the pore space. In the final sintering stage nearly all of the pores between
the particles are filled with material (German, 1996). In the current section, we are
concerned with initial-stage sintering, in which the material changes continuously
from a particulate state to a monolithic state but still may be reasonably described
as a discrete packing of particles. Three distinct types of sintering mechanisms have
been identified: (1) elastic adhesive contact (i.e., JKR contact); (2) “zipping” contact
growth driven by adhesive intersurface forces and accommodated by viscoelastic
deformation; and (3) “stretching” contact growth driven by surface tension and
accommodated by viscous flow (i.e., the classical viscous sintering).

4.6.1. Sintering Regime Map

A regime mapping showing regions where each sintering mechanism is dominant
was developed by Lin et al. (2001) for polymeric particles. The material viscoelastic
response is assumed to have a Maxwell form, which as shown in Figure 3.2 consists of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

122 Contact Mechanics with Adhesion Forces

a Hookian spring and a damper connected in series. The effective creep compliance
C(t ) for the collision is defined as

C(t ) = [D1 (t ) + D2 (t )](1 − ν 2 ), (4.6.1)

where the Poisson modulus ν is assumed to be the same for both materials and
D1 (t ) and D2 (t ) are the creep compliance for the two spheres. It is recalled that
creep compliance is defined as the instantaneous uniaxial strain divided by the
corresponding stress. For a generalized power-law Maxwell material, the effective
creep compliance C(t ) is assumed to vary with time as

C(t ) = C0 + C1t m (4.6.2)

where C0 = (1 − vi2 )/Ei is the instantaneous compliance for sphere i, C1 is a material

constant, and the exponent m is a material constant in the interval 0 < m ≤ 1. For
two identical particles, C0 = 1/2E, where E is the effect elastic modulus defined
in (3.1.1).
If the time after particle contact is sufficiently short that the material is unrelaxed
(t < t0 ), the dimensionless contact radius is given by the JKR model (4.2.34) as
 1/3  1/3
a0 9π γ 18πC0 γ
= = (4.6.3)
R RE R

where R is the effective particle radius given in (3.1.1). For intermediate times
in the interval t0 < t < tvis , the dimensionless contact radius can be estimated for
viscoelastic adhesive contact (Lin et al., 2001) as
 2/7  1/7
a δC C1 γ
= (63π 3 )1/7 t (4.6.4)
R R R

where δC is the critical normal overlap from JKR theory, given by (4.2.37). Finally,
if the time is large enough (t > tvis ), the contact radius can be obtained using the
approach taken by Frenkel (1945), who equated the rate of surface tension work to
the viscous flow energy dissipation rate to obtain
 1/2
a 2σ t
= . (4.6.5)
R μR

Replacing the viscosity in this equation by the effective particle creep viscosity,
μc,N = 1/(4C1 ) and the surface tension σ by the surface energy density γ , we can
write
 
a 8C1 γ 1/2
= t . (4.6.6)
R R

For instance, for fly-ash particles in a coal combustion boiler, typical values
for these parameters are C0 = 1 × 10−10 Pa−1 , C1 = 1 × 10−12 Pa−1 s−1 , m = 1,
γ = 0.5 J/m2 , and δc /R = 0.01. The change in dimensionless contact radius, a/R, is
sketched as a heavy solid line in Figure 4.25. If t < t0 (point A), the dimensionless
contact radius is a constant which depends on the dimensionless parameter C0 γ /R.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.6. Sintering Force 123

101

JKR elastic contact (C0γ/R=0.00025)

Vicoelastic contact (δc/R=0.01) ~t
1/2
Figure 4.25. A regime map for three dif- Viscous flow BB
=(tvis,(a/R)vis)
ferent sintering mechanisms, similar to 100 1/7
~t
that developed by Lin et al. (2001). Vari- A =(t
A 0,(a/R)JKR)

a/R
ation of the contact region radius for
a typical sintering case is indicated by
a solid line, which is nearly indepen- 10-1 JKR contact
dominated Viscous flow
dent of time in the JKR-contact regime, Viscoelastic contact dominated
increases in proportion to t 1/7 in the vis- dominated
coelastic regime, and increases in pro-
10-2
portion to t 1/2 in the viscous flow regime.
10-5 10-4 10-3 10-2 10-1 100

C1γt / R

For intermediate times (t0 < t < tvis ), the power law behavior for a/R exhibits only
very slow time dependence, to the 1/7 power. For t > tvis (point B), a/R depends on
the square root of the normalized time C1 σ t/R, indicative of viscous sintering.

4.6.2. Approximate Sintering Models

Several attempts have been made to apply the DEM approach to solve for sintering
of a realistic random packing of particles, including Jagota and Scherer (1993, 1995),
Luding et al. (2005), and Luding and Suiker (2008). In the model of Jagota and
Scherer, the viscous sintering regime is assumed (t > tvis in Figure 4.25) and the
normal contact force is modeled as
R
Fn = 3π σ a(t ) f (a/R) − 6π μc,N a(t ) v · n, (4.6.7)
|Xi j | R

where μc,N is the particle creep viscosity, σ is the surface tension, a(t ) is the contact
radius, |Xi j | is the distance between the centers of particles i and j, and f (a/R)
is a dimensionless function of contact radius which has order unity. The first and
second terms on the right-hand side of (4.6.7) represent the sintering rate driven by
surface tension and by viscous dissipation, respectively. Similarly, the model used
by Jagota and Scherer for shear force transmission through the contact is related to
both the difference in tangential velocity and the spin of the two particles through
the tangential contact viscosity.
During the sintering process, the contact forces are often driven by the cou-
pled grain-boundary and surface diffusion, instead of the viscous flows. A detailed
numerical simulation comparing the grain boundary and surface diffusion mecha-
nisms during sintering of two particles was reported by Bouvard and McMeeking
(1996). The results of this study were used in the context of DEM by Parhami and
McMeeking (1998) and by Martin et al. (2006) to perform simulations of diffusion-
controlled particle sintering, including problems with both free and pressure-assisted
sintering in three dimensions. In this work, the normal contact forces between two
particles with equivalent radius R are expressed as a sum of both sintering tensile
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

124 Contact Mechanics with Adhesion Forces

ψ Figure 4.26. Sketch showing the definition of the dihedral angle between
the particle surfaces at the neck.

force due to the surface energy and normal viscous force due to the grain boundary
and surface diffusion as

kB T dδ
Fn = π γ [8R(1 − cos(ψ/2)) + a(t ) sin(ψ/2)] − π a4 (t ) N , (4.6.8)
8Db δb dt

where a(t ) is the sintering contact radius, ψ is the dihedral angle between the particle
surfaces at the neck (as shown in Figure 4.26), γ is the surface energy density, 
is the atomic volume, and Db = D0b exp(−Qb /RT ) is the diffusion coefficient for
vacancy transport in the grain boundary with thickness δb and activation energy Qb .
The sintering contact radius grows according to Coble’s geometric model as (Coble,
1958)

da(t ) R dδN
= , (4.6.9)
dt a(t ) dt

with the initial value of a(t ) given by a2 = RδN for elastic deformation or by
a2 = BRδN for plastic deformation, where B is a constant varying between 1 for
linear hardening and 2.9 for perfectly plastic deformation. The normal contact
force is set to zero once the contact radius has grown to the equilibrium value
aeq = R sin(ψ/2).
DEM models based on the normal contact force expression in the presence of
sintering for the viscous flow regime, such as (4.6.7), and in the diffusion-dominated
regime, such as (4.6.8), are applicable to problems involving quasi-static packing
and consolidation of the sintering particles for cases without appreciable inertia.
In a general flow field, the particles collide with a nonzero impact velocity and the
elastic repulsive force between the particles must also be considered. For very small
particles, such that DMT theory applies, the Hertz elastic force can be added to
these two equations for a more complete description. In the general case of finite
Tabor number, the particle deformation under collision will influence the adhesive
force and so the two forces must be combined, such as was done in the JKR and
M-D models for nonsintering adhesion.

4.6.3. Hysteretic Sintering Contact Model

An alternative modeling approach for sintering phenomena was developed by Lud-
ing and coworkers (2005, 2008) using a hysteric contact model, which can be regarded
as an extension of the Walton and Braun (1986) model for elastic-plastic contact
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

4.6. Sintering Force 125

Figure 4.27. Load-displacement plot for

the hysteretic sintering contact model of
Luding et al. (2005).

force. The hysteretic sintering contact model is expressed using a force-displacement

relation given by
⎧
⎪
⎨ kL δN Loading
Fn = kU (δN − δres ) Un/reloading (4.6.10)
⎪
⎩
−kA δN Unloading
where kL and kU are the spring constants during loading and unloading processes,
respectively, such that kL ≤ kU . As seen in Figure 4.27, a contact begins at δN = 0,
such that during initial compressive loading, the contact repulsive force increases
with the overlap as Fn = kL δN until a maximum overlap δmax is attained. Beyond this
time the particles start to move apart, so that δN decreases in time. The repulsive
force between the particles decreases during this unloading process in proportion to
Fn = kU (δ − δres ). For subsequent loading with the same contact, the spring constant
remains equal to kU . The residual overlap δres = (1 − kL /kU )δmax is attained when
the repulsive force between the particles vanishes. As the particles continue to be
pulled apart, there is an attractive force between them due to the material bonding
at the particle contact. This attractive force increases with decrease in δN until a
maximum attractive force is attained, given by FA,max = −kA δA,max , where δA,max =
kU δres /(kU + kA ) and kA represents the effects of tensile softening. The magnitude
of the attractive contact force decreases as δN is decreased further according to Fn =
−kA δN until the load vanishes at δN = 0. In the sintering process, the values of kL , kU
and kA are strongly temperature-dependent, and some related empirical expressions
that can be used to estimate these parameters are given by Luding et al. (2005).
More extensive work on validation of DEM force models for sintering contact
between two spherical particles is needed. An experimental set-up used to measure
the critical sticking velocity of an incident particle in a vacuum column in which
particles impact on a flat surface was developed by van Beek (2001). Abd-Elhady
et al. (2005) extended this set-up for studying critical sticking velocity of an incident
particle with a sintered particulate fouling layer, which is exposed to a high gas-side
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

126 Contact Mechanics with Adhesion Forces

temperature ambience. Models for both normal and shear contact were proposed
and tested by Abd-Elhady for sintered particles. Nevertheless, DEM modeling of
sintered particles is still a relatively recent development.

REFERENCES

Abd-Elhady MS. Gas-side particulate fouling in biomass gasifiers. Ph.D. thesis, Eindhoven
University of Technology, The Netherlands (2005).
Agresar G. A Computational Environment for the Study of Circulating Cell Mechanics and
Adhesion. Ph.D. Dissertation, University of Michigan, Ann Arbor, 1996.
Bell GI. Models for the specific adhesion of cells to cells. Science 200, 618–627 (1978).
Bell GI, Dembo M, Bongrand P. Cell adhesion, competition between nonspecific repulsion
and specific bonding. Biophysics Journal 45, 1051–1064 (1984).
Bico J, Ashmore-Chakrabarty J, McKinley GH, Stone HA. Rolling stones: The motion of a
sphere down an inclined plane coated with a thin liquid film. Physics of Fluids 21, 082103
(2009).
Bouvard D, McMeeking RM. The deformation of interparticle necks by diffusion controlled
creep. Journal of the American Ceramic Society 79, 666–672 (1996).
Carpick RW, Ogletree DF, Salmeron M. A general equation for fitting contact area and
friction vs load measurements. Journal of Colloid and Interface Science 211, 395–400
(1999).
Cavallaro U, Christofori G. Cell adhesion in tumor invasion and metastasis: Loss of the glue
is not enough. Biochimica et Biophysica Acta 1552, 39–45 (2001).
Chesla SE, Selvaraj P, Zhu C. (1998) Measuring two-dimensional receptor-ligand binding
kinetics by micropipette. Biophysics Journal 75, 1553–1572 (1998).
Chesnutt JKW, Marshall JS. Blood cell transport and aggregation using discrete ellipsoidal
particles. Computers & Fluids 38, 1782–1794 (2009).
Chokshi A, Tielens AGGM, Hollenbach D. Dust coagulation. The Astrophysical Journal 407,
806–819 (1993).
Coble RL. Initial sintering of alumina and hematite. Journal of the American Ceramic Society
41, 55–62 (1958).
Coombs D, Dembo M, Wofsy C, Goldstein B. Equilibrium thermodynamics of cell-cell
adhesion mediated by multiple ligand-receptor pairs. Biophysics Journal 86, 1408–1423
(2004).
Cundall PA, Strack ODL. A discrete numerical model for granular assembles. Geotechnique
29(1), 47–65 (1979).
Dahneke B. Measurements of the bouncing of small latex spheres. Journal of Colloid and
Interface Science 45, 584–590 (1973).
de Gennes PG. Deposition of Langmuir-Blodgett layers. Colloid & Polymer Science 264,
463–465 (1986).
Dembo M, Torney DC, Saxman K, Hammer D. The reaction-limited kinetics of membrane-
to-surface adhesion and detachment. Proceedings of the Royal Society of London B 234,
55–83 (1988).
Derjaguin BV. Analysis of friction and adhesion. IV. The theory of the adhesion of small
particles. Kolloid Z. (in German) 69(2), 155–164 (1934).
Derjaguin BV, Landau L. Theory of the stability of strongly charged lyophobic sols and of the
adhesion of strongly charged particles in solutions of electrolytes. Acta Physico Chemica
URSS 14, 633–662 (1941).
Derjaguin BV, Muller VM, Toporov YP. Effect of contact deformations on the adhesion of
particles. Journal of Colloid and Interface Science 53(2), 314–326 (1975).
Ding W, Zhang H, Cetinkaya C. Rolling resistance moment-based adhesion characterization
of microspheres. Journal of Adhesion 84(12), 996–1006 (2008).
Dominik C, Tielens AGGM. Resistance to sliding on atomic scales in the adhesive contact of
two elastic spheres. Philosophical Magazine A 73(5), 1279–1302 (1996).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

References 127

Dominik C, Tielens AGGM. Resistance to rolling in the adhesive contact of two elastic
spheres. Philosophical Magazine A 72(3), 783–803 (1995).
Ennis BJ, Li J, Tardos GI, Pfeffer R. The influence of viscosity on the strength of an
axially strained pendular liquid bridge. Chemical Engineering Science 45(10), 3071–3088
(1990).
Fiscina JE, Lumay G, Ludewig F, Vandewalle N. Compaction dynamics of wet granular
assemblies. Physical Review Letters 105, 048001 (2010).
Frenkel J. Viscous flow of crystalline bodies under the action of surface tension. Journal of
Physics 9, 385–391 (1945).
German RM. Sintering Theory and Practice. John Wiley & Sons: Wiley-Interscience,
(1996).
Halsey TC, Levine AJ. How sandcastles fall. Physical Review Letters 80, 3141–3144 (1998).
Hoffman RL. A study of the advancing interface. I. Interface shape in liquid-gas systems.
Journal of Colloid and Interface Science 50(2), 228–241 (1975).
Hotta K, Takeda K, Iinoya K. The capillary binding force of a liquid bridge. Powder Technol-
ogy 10, 231–242 (1974).
Israelachvili J. Intermolecular and Surface Forces, Academic Press, 2nd edition, London
(1991).
Jagota A, Scherer GW. Viscosities and sintering rate of a two-dimensional granular composite.
Journal of the American Ceramic Society 76, 3123–3135 (1993).
Jagota A, Scherer GW. Viscosities and sintering rates of composite packings of spheres.
Journal of the American Ceramic Society 78, 521–528 (1995).
Johnson KL, Greenwood JA. An adhesion map for the contact of elastic spheres. Journal of
Colloid and Interface Science 192, 326–333 (1997).
Johnson KL, Kendall K, Roberts AD. Surface energy and the contact of elastic solids. Pro-
ceedings of the Royal Society of London A 324, 301–313 (1971).
Li S-Q, Marshall JS, Liu G, Yao Q. Adhesive particulate flow: The discrete element method
and its application in energy and environmental engineering. Progress in Energy and Com-
bustion Science 37(6), 633–668 (2011).
Lian G, Thornton C, Adams MJ. A theoretical study of the liquid bridge forces between two
rigid spherical bodies. Journal of Colloid and Interface Science 161, 138–147 (1993).
Lin YY, Hui CY, Jagota A. The role of viscoelastic adhesive contact in the sintering of
polymeric particles. Journal of Colloid and Interface Science 237, 267–282 (2001).
Liu GQ, Li SQ, Yao Q. A JKR-based dynamic model for the impact of micro-particle with a
flat surface. Powder Technology 207, 215–223 (2011).
Longest PW, Kleinstreuer C. Comparison of blood particle deposition models for non-parallel
flow domains. Journal of Biomechanics 36, 421–430 (2003).
Luding S, Suiker A. Self-healing of damaged particulate materials through sintering. Philo-
sophical Magazine 88, 3445–3457 (2008).
Luding S, Manetsberger K, Muellers J, A discrete model for long time sintering. Journal of
the Mechanics and Physics of Solids 53, 455–491 (2005).
Marshall JS. Particle aggregation and capture by walls in a particulate aerosol channel flow.
Journal of Aerosol Science 38, 333–351 (2007).
Marshall JS. Discrete-element modeling of particulate aerosol flows. Journal of Computational
Physics 228, 1541–1561 (2009).
Martin CL, Schneider LCR, Olmos L, Bouvard D. Discrete element modeling of metallic
powder sintering. Scripta Materialia 55, 425–428 (2006).
Masliyah JH, Bhattacharjee S. Electrokinetic and Colloid Transport Phenomena. John Wiley
and Sons, New Jersey (2006).
Mason T G, Levine A J, Ertas D, Halsey TC. Critical angle of wet sandpiles. Physical Review
E 60(5), 5044–5047 (1999).
Matthewson MJ. Adhesion of spheres by thin liquid films. Philosophical Magazine A 57(2),
207–216 (1988).
Maugis D. Adherence of elastomers: Fracture mechanics aspects. Journal of Adhesion Science
and Technology 1(2), 105–134 (1987).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

128 Contact Mechanics with Adhesion Forces

Maugis D. Adhesion of spheres: The JKR-DMT transition using a Dugdale model. Journal
of Colloid and Interface Science 150(1), 243–269 (1992).
Maugis D. Contact, Adhesion and Rupture of Elastic Solids, Springer-Verlag, Berlin (1999).
Mehrotra VP, Sastry KVS. Pendular bond strength between unequal-sized spherical particles.
Powder Technology 25, 203–214 (1980).
Muller VM, Yushenko VS, Derjaguin BV. On the influence of molecular forces on the
deformation of an elastic sphere and its sticking to a rigid plane. Journal of Colloid and
Interface Science 77(1), 91–101 (1980).
N’Dri NA, Shyy W, Tran-Son-Tay R. Computational modeling of cell adhesion and movement
using a continuum-kinetics approach. Biophysics Journal 85, 2273–2286 (2003).
Parhami F, McMeeking RM. A network model for initial stage sintering. Mechanics of
Materials 27, 111–124 (1998).
Peri MDM, Cetinkaya C. Rolling resistance moment of microspheres on surfaces. Philosoph-
ical Magazine 85(13), 1347–1357 (2005).
Pitois O, Moucheront P, Chateau X. Liquid bridge between two moving spheres: An exper-
imental study of viscosity effects. Journal of Colloid and Interface Science 231, 26–31
(2000).
Pitois O, Moucheront P, Chateau X. Rupture energy of a pendular bridge. European Physical
Journal B 23, 79–86 (2001).
Probstein RF. Physicochemical Hydrodynamics, An Introduction, 2nd ed.,Wiley Interscience,
New York (2003).
Russel WB, Saville DA, Schowalter W R. Colloidal Dispersions. Cambridge University Press,
Cambridge, UK (1989).
Savkoor AR, Briggs GAD. The effect of tangential force on the contact of elastic solids in
adhesion. Proceedings of the Royal Society of London A 356, 103–114 (1977).
Schade P, Marshall JS. Capillary effects on a particle rolling on a plane surface in the presence
of a thin liquid film. Experiments in Fluids 51(6), 1645–1655 (2011).
Smith AW. Biofilms and antibiotic therapy: Is there a role for combating bacterial resistance
by the use of novel drug delivery systems? Advanced Drug Delivery Reviews 57, 1539–1550
(2005).
Tabor D. Surface forces and surface interactions. Journal of Colloid and Interface Science
58(1), 2–13 (1977).
Tachev KD, Angarska JK, Danov KD, Kralchevsky PA. Erythrocyte attachment to sub-
strates: determination of membrane tension and adhesion energy. Colloids and Surfaces B:
Biointerfaces 19, 61–80 (2000).
Thornton C. Interparticle sliding in the presence of adhesion. Journal of Physics D-Applied
Physics 24, 1942–1946 (1991).
Thornton C, Ning Z. A theoretical model for the stick/bounce behaviour of adhesive, elastic-
plastic spheres. Powder Technology 99, 154–162 (1998).
Thornton C, Yin KK. Impact of elastic spheres with and without adhesion. Powder Technology
65, 153–166 (1991).
Tozeren A, Sung KLP, Chien S. Theoretical and experimental studies on cross-bridge migra-
tion during cell disaggregation. Biophysics Journal 55, 479–487 (1989).
Tsuji Y, Tanaka T, Ishida T. Lagrangian numerical simulation of plug flow of cohesionless
particles in a horizontal pipe. Powder Technology 71, 239–250 (1992).
van Beek MC. Gas-side fouling in heat-recovery boilers. Ph.D. Thesis, Eindhoven University
of Technology, The Netherlands (2001).
Verwey EJW, Overbeek JTG. Theory of the Stability of Lyophobic Colloids. Elsevier,
Amsterdam (1948).
Wall S, John W, Wang H, Goren SL. Measurements of kinetic energy loss for particles
impacting surfaces. Aerosol Science and Technology 12, 926–946 (1990).
Walton OR, Braun RL. Viscosity, granular temperature, and stress calculations for shearing
assemblies of inelastic, frictional disks. Journal of Rheology 30, 949–980 (1986).
Willett CD, Adams MJ, Johnson SA, Seville JPK. Capillary bridges between two spherical
bodies. Langmuir 16(24), 9396–9405 (2000).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-04 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:23

References 129

Zhu C. Kinetics and mechanics of cell adhesion. Journal of Biomechanics 33, 23–33 (2000).
Zhu C, Williams TE. Modeling concurrent binding of multiple molecular species in cell
adhesion. Biophysics Journal 79, 1850–1857 (2000).
Zhu RR, Li SQ, Yao Q. Effects of cohesion on the flow patterns of granular materials in
spouted beds. Physical Review E 87, 022206 (2013).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5 Fluid Forces on Particles

It is necessary in any simulation of a particulate flow to utilize a model for the forces
exerted by the fluid on the particles. In development of such a model, it is usually
assumed that the particles are immersed in a fluid flow that has characteristic length
scales that are much greater than the particle diameter. If this is the case, then the
fluid flow can be approximated from the perspective of a particle as a uniform flow
plus a linear velocity gradient, which greatly simplifies the task of model development
for the fluid forces on the particle. Even with this simplification, there is still a vast
literature devoted to fluid forces on bodies of various shapes and sizes in different
Reynolds number and Mach number regimes. The current chapter focuses on forces
on spherical particles in an incompressible flow with particle Reynolds numbers less
than about 10, and in some cases with particle Reynolds numbers less than unity.
The problem of fluid flows with ellipsoidal particles is taken up in Chapter 7.
The first five sections of this chapter discuss various forces acting on a particle
immersed in a fluid shear flow in isolation, without the presence of walls or other
particles. The first two of these sections discuss the drag and lift forces that act on par-
ticles immersed in a steady velocity field, whereas the third section concerns forces
that act in unsteady flows, including added mass force, inertial pressure-gradient
force, and the so-called history force. The fourth section introduces the force due
to Brownian motion. A scaling analysis examining the order of magnitude of these
different forces relative to the dominant drag force is given in the fifth section.
The remaining three sections of the chapter examine the effect of container walls
and of other particles on the particle forces. The effect of a nearby plane wall,
discussed in the sixth section, is a particularly important problem for adhesive par-
ticles, which tend to stick to walls immersed in the flow field. Surrounding particles
also have an important impact on particle fluid forces as the particle concentration
becomes large. The seventh section discusses the influence of surrounding particles
on the drag force with a specified mean fluid flow. The eighth section describes
a method called Stokesian dynamics, which is used to provide a rapid computa-
tion of the hydrodynamic interaction between particles in a flow, accounting for
the modification of the fluid velocity field by the presence of surrounding parti-
cles. The final section discusses different types of particle interactions with acoustic
fields.

130
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.1. Drag Force and Viscous Torque 131

5.1. Drag Force and Viscous Torque

In most particulate flows, there are two primary forces that are most important for
controlling the particle dynamics, and these two forces are in approximate balance
with each other as the particles move about. Additionally, there will be a number of
secondary forces that are less important, although perhaps not completely negligible.
Identification and accurate modeling of the primary forces is critical for solution of
the particle transport problem, although secondary forces are often approximated
with a less rigorous model or even neglected. For problems involving neutrally
buoyant particles with diameter greater than about 1 µm, the primary forces are
usually drag and particle inertia. This statement remains valid even for cases where
the particle Reynolds number is small, for which case both the particle inertia and
drag force will be small, but other forces will usually be even smaller. For nanometer-
size particles, Brownian motion can become sufficiently large so as to exceed the
particle inertia, in which case the particle drag is approximately balanced by the
Brownian force. This is the assumption typically used in the Brownian dynamics
method discussed in Section 2.3.2. For particles that are not neutrally buoyant,
the reduced gravity force (which combines gravity and buoyancy) can also become
important in certain particle size ranges.
The classical solution for drag force on a sphere was obtained by Stokes, given by
(1.2.1). This solution assumes that the particle Reynolds number is small compared to
unity, that the fluid flow relative to the particle is uniform, that the no-slip condition
is satisfied on the surface of the particle, and that the space is unbounded with
no solid surfaces present except for the particle itself. All of these approximations
break down at various instances in particulate flow problems. A modified form of
the Stokes expression for drag force Fd on a spherical particle can be written as

Fd = −3π dμ(v − u) f, (5.1.1)

where f is called the friction factor, v is the particle velocity, and u is the velocity
that the fluid would have had at the particle centroid had the particle not been
present. The difference v − u is known as the slip velocity. The friction factor is
used to account for breakdown of some of the approximations noted earlier, so
that the original Stokes solution is recovered when f = 1. The friction factor can be
decomposed as the product of various correction factors, given by

f = CI CCCF , (5.1.2)

where CI accounts for effect of fluid inertia, CF accounts for effects of crowding by
surrounding particles, and CC accounts for the effect of nonzero slip on the particle
surface. The inertia and slip correction factors are discussed in the following, and
the effect of surrounding particles is discussed in Section 5.7.

5.1.1. Effect of Flow Nonuniformity

Perhaps the first objection that one might make to application of the Stokes drag
expression for a particle in a general fluid flow is that the incident flow field relative
to the particle is not actually uniform, but rather changes with some characteristic
length scale L. How important is nonuniformity of the incident flow to the particle
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

132 Fluid Forces on Particles

drag force? This question can be addressed with reference to a more general solution
due to Faxén (1922), who found that the force on a sphere of diameter d immersed
in a flow field u(x) with small particle Reynolds number and small value of the
parameter d/L is given by
μπ 3 2
Fd = −3π dμ(v − u) + d ∇ u, (5.1.3)
8
where both u and ∇ 2 u are evaluated at the particle centroid. The second term in
(5.1.3) is similar to a buoyancy force. To see this, we recall that for a flow with
small flow Reynolds number the fluid inertia term is negligible, so the Navier-Stokes
equation reduces to the Stokes equation

0 = −∇ p + μ∇ 2 u, (5.1.4)

where p is fluid pressure and u is the fluid velocity field. Solving for ∇ 2 u = (1/μ)∇ p
from (5.1.4) and substituting into (5.1.3) gives
π 3
Fd = −3π dμ(v − u) + d ∇ p. (5.1.5)
8
The last term in (5.1.5) is proportional to the sphere volume times the ambient fluid
pressure gradient to which the sphere is subjected, which is equal to the buoyancy
force. We note that this analysis requires that the flow Reynolds number Re f is
small, which is substantially more restrictive than the requirement that the particle
Reynolds number be small. A scaling analysis indicates that the ratio of the second
term on the right-hand side of (5.1.3) to the first (Stokes) term is of order (d/L)2 .
Applicability of the discrete-element method usually requires that d/L is small, so
the second term in (5.1.3) is negligible. An exception might be flow near walls, which
is discussed further in Section 5.6.
Nonuniformity of the flow field also introduces the possibility that nonzero
rotation rate of the material elements in the flow surrounding the particle might
induce a torque on the particle. This issue was also examined by Faxén (1922), who
concluded that a viscous torque MF acts on the particle that is proportional to the
difference between the local fluid rotation rate ω/2, where ω = ∇ × u is the fluid
vorticity evaluated at the particle centroid, and the particle rotation rate . The
viscous torque is on the sphere is given by the formula
 
3 1
MF = π μd ω− . (5.1.6)
2

5.1.2. Effect of Fluid Inertia

For small particle Reynolds numbers (Re p = ρ f d |v − u| /μ 1), one-third of the
total drag force acting on a sphere is due to asymmetry in the pressure distribu-
tion and two-thirds is due to viscous shear stress on the sphere surface. As shown
in Figure 5.1, the particle wake becomes increasingly pronounced as the particle
Reynolds number approaches and exceeds unity. At a Reynolds number of about
24, a standing vortex ring is observed to form behind the sphere (Figure 5.1c).
This vortex ring moves backward gradually as the Reynolds number is increased
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.1. Drag Force and Viscous Torque 133

Figure 5.1. Flow visualization showing the wake behind a sphere at Reynolds numbers
(a) 9.15; (b) 17.9; (c) 37.7; (d) 73.6; (e) 118; and (f) 550. Images (a)–(e) are the wakes of
solid spheres visualized by suspended aluminum powder. Image (f) is the wake of an oil drop
of carbon tetrachloride in water colored by fluorescein. [Reprinted with permission from
Taneda (1956).]

(Figure 5.1d, e). At a Reynolds number of about 130, oscillations are first observed
in the wake at the rear of the vortex ring, which at higher Reynolds numbers leads
to periodic shedding of horseshoe-type vortex structures on alternating sides of the
sphere (Figure 5.1f).
The wake development leads to increase in the rear-aft pressure asymmetry,
causing the pressure drag on the particle to increase relative to the viscous drag.
This increase in pressure drag leads to the so-called inertial correction of the Stokes
drag expression, valid for particle Reynolds numbers that are not sufficiently small
that the fluid inertia is negligible but also not sufficiently large that the pressure
force completely dominates the drag force. The original Stokes expression (5.1.4)
is obtained by neglecting the fluid inertia term in the Navier-Stokes equations, and
this equation is valid in the region near the sphere for particle Reynolds numbers
up to about unity. Oseen (1910) developed a zeroth-order correction of the Stokes
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

134 Fluid Forces on Particles

equation valid in the region far away from the sphere by linearizing the inertia term
in the Navier-Stokes equation, resulting in the Oseen equation

(U · ∇ )u = −∇ p + μ∇ 2 u, (5.1.7)

where U is the uniform fluid velocity at infinity and the sphere is assumed to be sta-
tionary. Matching the far-field solution of (5.1.7) to the near-field solution of (5.1.4)
gives a uniformly valid asymptotic solution of the velocity field for low Reynolds
number flow. This procedure can be used to obtain a correction for the Stokes drag
coefficient formula, resulting in an expression for the inertia correction factor as

CI = 1 + (3/16)Re p . (5.1.8)

Proudman and Pearson (1957) use a singular perturbation scheme that matches
solutions near and far from the sphere to obtain a first-order correction for the
sphere flow field valid at finite Reynolds numbers. This perturbation scheme results
in an expression for the inertia correction factor as

3 9
CI = 1 + Re p + Re2 ln(Re p /2). (5.1.9)
16 160 p
Increase in particle Reynolds number up to about 100 is accompanied by formation
of a standing vortex ring within the wake, resulting in formation of an elongated wake
“bubble” (Figure 5.1c). A popular empirical expression for the inertial correction
factor is that of Schiller and Naumann (1933),

CI = 1 + 0.15Re0.687
p . (5.1.10)

This expression remains valid to within 5% of the experimental data for particle
Reynolds numbers up to about 800. There are many other empirical expressions for
drag coefficient that are valid for even higher Reynolds numbers, but for adhesive
particle flows the particle Reynolds numbers are typically small and so the expres-
sions given here are usually adequate.
A plot comparing the predictions of expressions (5.1.8), (5.1.9), and (5.1.10)
for the inertial correction factor is given in Figure 5.2a. The empirical Schiller-
Naumann expression is a close fit to the experimental data. This correlation indi-
cates that the inertial correction factor increases the drag by a factor of about 0.2
for Re p = 1, 0.5 for Re p = 5, and 1.7 for Re p = 10. The Oseen equation (5.1.8)
follows reasonably close to the Schiller-Naumann expression for Re p < 2, but for
larger Reynolds number it gives excessively high estimates for the sphere drag. The
Proudman-Pearson expression (5.1.9) is closer to the Schiller-Naumann expression
for Re p < 2 than is the Oseen expression, but at higher Reynolds numbers it gives
excessively high estimates for sphere drag. This deviation is not surprising because
the Proudman-Pearson expression is derived from an asymptotic theory assuming
a small particle Reynolds number. The drag coefficient Cd = Fd / π8 ρ |v − u|2 d2 is
plotted as a function of particle Reynolds number in Figure 5.2b, comparing exper-
imental data with the Stokes, Oseen, and Schiller-Naumann expressions for a wide
range of Reynolds numbers. For DEM calculations, we recommend using the orig-
inal Stokes expression (with CI = 1) for particle Reynolds numbers up to about
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.1. Drag Force and Viscous Torque 135

5 3
10
Laminar attached Separated Experiments
4.5
flow flow
Stokes
4 2
10 Oseen
3.5
PP SN
1
3 10
CI 2.5
Cd
Oseen Laminar separation &
2 0 turbulent wake
10

1.5 SN
1 -1
Stokes 10

0.5

0 -2
10 -1
0 2 4 6 8 10 0 1
10
2 3
10 10
4
10
5 6
10 10 10 10
ReP Re p
(a) (b)
Figure 5.2. (a) Comparison of different expressions for the inertia correction factor CI as a
function of particle Reynolds number ReP , including the Stokes expression (CI = 1) (dashed
line), the Oseen expression (5.1.8) (dashed-dotted line), the Proudman-Pearson expression
(5.1.9) (long dashed line), and the Schiller-Naumann expression (5.1.10) (solid line). (b) Plot
of drag coefficient for flow past a sphere including Stokes, Oseen, and Schiller-Naumann
expressions and experimental data reported by White (2006).

Re p = 1, since this linear expression is most accommodating of an implicit numeri-

cal solution of the particle momentum equation (see further comments on numerical
instability of the particle momentum equation in Chapter 11). For higher values of
the particle Reynolds number, the Schiller-Naumann expression (5.1.10) is the best
choice.

5.1.3. Effect of Surface Slip

Because particles of interest in adhesive particle flows are often quite small, it can
sometimes happen in gas-particle flows that the particle diameter approaches the
order of magnitude of the mean free path λ of the gas molecules. Mean free path is
the average distance traveled by a molecule of the gas before collision with another
gas molecule. Typical values of λ are about 68 nm for gases such as O2 or N2 at
standard temperature and pressure. Kinetic theory can be used to relate the mean
free path to the temperature and pressure of an ideal gas as

k T
λ= √ B , (5.1.11)
2π σM
2 p

where the Boltzmann constant kB = 1.380648813 × 10−23 J/K is equal to the ratio
of the universal gas constant to Avogadro’s number, T and p are the absolute
temperature and pressure, and σM is the molecular diameter. Kinetic theory also
provides a relationship between molecular diameter and the fluid viscosity μ as
 1/2
5 MgkB T
μ= , (5.1.12)
16σM2 π
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

136 Fluid Forces on Particles

where Mg is the gaseous molecular mass. Solving for σM 2

in (5.1.11) and using the
ideal gas law, the expression (5.1.12) can be written as
√
5 2π ρ f λ
μ=  (kB T )1/2 , (5.1.13)
16 Mg

√
where ρ f is the gas density. Recalling the expression c = κRT for the speed of
sound in an ideal gas, where κ = c p /cv is the ratio of specific heats (∼
= 1.4 for diatomic
molecules) and R = kB /Mg is the specific gas constant (the ratio of the universal gas
constant to the gas molar mass), the result (5.1.13) can be used to obtain
√
5 2π
μ = √ ρ f λc. (5.1.14)
16 κ

Hence, the gas viscosity is directly proportional to the mean-free-path of the gas
molecules. Equation (5.1.14) provides a useful formula for estimating mean-free-
path based on macroscopic quantities such as gas viscosity, density, and speed of
sound.
The no-slip condition on the particle surface begins to break down when the
ratio of mean-free-path to particle diameter approaches about 1,000. As discussed
in Section 1.2, the ratio of mean-free-path to particle diameter is called the Knudsen
number, given by

λ
Kn ≡ . (5.1.15)
d

From (5.1.14), λ ∼ μ/ρ f c, so (5.1.15) becomes

μ Ma
Kn ∼ = , (5.1.16)
ρ f cd Re p

where the Mach number Ma = U/c is the ratio of the characteristic fluid velocity
scale U and the speed of sound in the fluid. From (5.1.16), it follows that particles with
very small Reynolds numbers may exhibit significant effects due to flow rarefaction
even in flows with small Mach numbers.
A solution for the effect of surface slip on the drag on a sphere at low Reynolds
number was obtained by Basset (1888) using the common assumption that the slip
velocity uslip is proportional to the wall shear stress τwall , or

Guslip = τwall , (5.1.17)

where G is the coefficient of sliding friction. Basset considered the case where the
Knudsen number is sufficiently small that the continuum equation can still be used
(approximately Kn < 0.1). He solved the Stokes equation (5.1.4) with the boundary
condition (5.1.17) to obtain an expression for the slip correction factor as

Gd + 4μ
CC = . (5.1.18)
Gd + 6μ
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.1. Drag Force and Viscous Torque 137

0.8

Figure 5.3. Plot comparing different expres- Bassett

0.6
sions for the slip correction factor as a func-
tion of Knudsen number, including that of
CC
Basset (5.1.18) (dashed-dotted line), Epstein
0.4
(5.1.19) (long dashed line), Cunningham
(5.1.21) (short dashed line), and Millikan
(1923) (5.1.22) (solid line). 0.2
Epstein Millikan
Cunningham
0
0 2 4 6 8
Kn

For no slip G → ∞ and CC → 1, whereas for perfect slip G = 0 and CC = 2/3.

Epstein (1924) used kinetic theory to derive an expression for slip correction factor
that is valid over a broader range of Knudsen numbers, given by
2μ
CC = 1 − . (5.1.19)
Gd
Schmitt (1959) argued that the ratio μ/G should be proportional to the mean-free-
path, and proposed

μ/G = Aλ, (5.1.20)

where A ∼ = 1.3. Using (5.1.20), the various expressions for CC can be written as
functions of the Knudsen number. A similar expression was derived earlier by Cun-
ningham (1910), also using kinetic theory, as
1
CC = . (5.1.21)
1 + 3.26Kn
In recognition of this work, the slip correct factor is often called the Cunningham
correction factor.
An empirical expression for the slip correction factor was obtained by Millikan
(1923), based on his famous oil-drop experiments, as
1
CC = . (5.1.22)
1 + Kn[A1 + A2 exp(−A3 /Kn)]
For air at standard temperature and pressure, A1 = 2.49, A2 = 0.84, and A3 = 1.74.
Using (5.1.22), CC ∼
= 0.99 for 15 µm diameter particles in air at standard temperature
and pressure, and CC ∼ = 0.87 for 1 µm diameter particles. A plot comparing the
different expressions for slip correction factor is shown in Figure 5.3. The Basset
expression (5.1.18) gives values for CC that are much greater than Millikan’s empirical
fit, where the latter is in good agreement with the experimental data. Epstein’s
expression is accurate for very small Knudsen numbers but quickly deviates from
the experimental data as Kn increases above about 0.02. The simple Cunningham
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

138 Fluid Forces on Particles

relationship (5.1.21) is very close to the Millikan correlation throughout the range
of Kn values considered.

5.2. Lift Force

Particles placed in a unidirectional flow with a velocity gradient normal to the flow
direction are observed to drift in the direction of the velocity gradient. A well-known
demonstration of this drift was noted by Segré and Silberberg (1962), who observed
that neutrally buoyant particles immersed in a fluid flowing through a circular tube
tend to drift over time to a radial location that is midway between the tube wall and
the tube axis. It was shown by Saffman (1956) and Bretherton (1962) that a lateral
(lift) force is not possible for the problem of unidirectional flow past an isolated, rigid
spheroidal body in a flow governed by the Stokes equations (5.1.4), whatever the
velocity profile. Consequently, the lateral motion observed by Segré and Silberberg
must involve the effect of fluid inertia, even though the inertia associated with the
small particle sizes involved may be quite small compared with the other terms
effecting particle motion.
The lateral force that acts on the particle is observed to depend on both
the particle rotation rate and the velocity gradient of the flow in the vicinity of the
particle. The problem is therefore usually divided into two separate problems. The
first problem concerns a particle moving at a velocity v in a uniform shear flow with
velocity u = u(y)ex , where v and u are in the same direction. The second problem is
that of a particle traveling at a velocity v and spinning with rotation rate  through a
fluid with uniform velocity u = U ex u. Both of these problems result in lateral forces
on the particle, often called the Saffman force and the Magnus force, respectively,
where these forces can be added together to obtain the total lateral force.

5.2.1. Saffman Lift Force

For the case of a particle traveling in the same direction as a plane shear flow with
ambient velocity gradient G = du/dy, Saffman obtained an analytic solution for the
magnitude F of the lift force of the form
F = 6.46μ(d/2)2 vs (G/ν)1/2 , (5.2.1)
where vs = |v − u| is the magnitude of the particle slip velocity and ν is the fluid
kinematic viscosity. This asymptotic solution is valid under the assumptions that the
particle Reynolds number Re p = ρ f vs d/μ is much less than the square root of the
“shear Reynolds number” ReG ≡ ρ f d2 G/μ, and that both Re p and ReG are much
less than unity. It is noted that this leading-order solution is independent of the
particle rotation rate. The lift force is oriented in a direction t that is parallel to the
fluid velocity gradient, pointing from the low-velocity to the high-velocity side of
the sphere.
The Saffman
 expression was extended by McLaughlin (1991) for cases where
Re p = O( ReG ), but where both Re p and ReG are still assumed to be small, to
obtain an analytical expression for lift force magnitude as
9J
F = μ(d/2)2 vs (G/ν)1/2 , (5.2.2)
π
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.2. Lift Force 139

where J is a function of the dimensionless parameter κ ≡ Re1/2

G /Re p . McLaughlin
provided numerical computations for J as well as asymptotic solutions for large and
small values of κ, given by

2.255 − 0.6463/κ 2 + O(1/κ 4 ) for κ 1

J= . (5.2.3)
−32π κ ln(1/κ ) + O(κ )
2 5 2 5
for κ 1

As κ → ∞, J → 2.255 and the Saffman result (5.2.1) is recovered. A curve fit to the
numerical values of J(κ ) for 0.1 ≤ κ ≤ 20 was reported by Mei (1992) as

J = 0.6765{1 + tanh[2.59 log10 κ + 0.191)]}{0.667 + tanh[6(κ − 0.32)]}. (5.2.4)

Numerical simulations of lift force on a spherical particle in a linear shear flow were
reported by Dandy and Dwyer (1990) for a fixed particle over Reynolds number
range 0.1 ≤ Re p ≤ 100. Mei (1992) developed a correlation for lift force based on
the results of Saffman (1965) and the numerical results of Dandy and Dwyer (1990)
which gives the ratio of the lift force at finite Reynolds number to the Saffman lift
force (5.2.1) as

F (1 − 0.3314β 1/2 ) exp(−Re p /10) + 0.3314β 1/2 for Re p < 40

= (5.2.5)
F ,SAF F 0.0524(βRe p ) 1/2
for Re p > 40

where β ≡ Gd/vs = ReG /Re p . Kurose and Komori (1999) and Bagchi and Balachan-
dar (2002a) simulated flow past both fixed and rotating spheres in a linear shear flow
for Reynolds numbers ranging from order unity to several hundred. Kurose and
Komori observed that the sum of the lift force for a fixed particle in a linear shear
flow and that for a rotating particle in a uniform flow is nearly equal to the lift force
for a rotating particle in a linear shear flow provided that the particle Reynolds
number is not too large compared to unity. This conclusion is also supported by the
asymptotic study of Saffman (1965) at small particle Reynolds numbers.
A comparison of the numerical results of Dandy and Dwyer (1990), Kurose
and Komori (1999), and Bagchi and Balachandar (2002a) for lift coefficient CL =
8F /ρ f vs2 πd2 with β = 0.4 and both fixed spheres and freely rotating spheres is given
in Figure 5.4, in comparison to the Saffman and McLaughlin analytic formulas and
to the correlation of Mei (1992). For small particle Reynolds numbers, of order unity
or less, the particle rotation rate appears to have little influence on the lift force and
the data are well predicted by both the McLaughlin solution. The lift coefficient
decreases rapidly for both fixed and rotating sphere conditions as the Reynolds
number increases above order unity. The lift coefficient data at large Reynolds
numbers generally fall between the predictions of the Saffman and McLaughlin
formulas.
All of the developments for lift force discussed earlier are concerned with the
problem of a particle in a linear shear flow. Shear flows can be decomposed into a
local rotation plus a straining flow. It is of interest to examine the lift force for other
flow fields and to see to what extent the lift force in such flows differs from that for
a linear shear flow. The problem of a particle immersed in a rigid-body rotation,
such as would occur within the core of a vortex or in a centrifuge, has been a focus
of particular interest. This problem was examined using an asymptotic analysis by
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

140 Fluid Forces on Particles

Figure 5.4. Plot of lift force coefficient

10
0
Saffm
an versus particle Reynolds number com-
paring predictions of Saffman (solid
Me line), McLaughlin (dashed-dotted line),
i
and Mei (dashed line) for β = 0.4.

Mc
CL
Lau
-1
10 The theoretical results are compared

ghl
to numerical predictions with the same
in value of β from Bagchi and Balachan-
dar (2002a) for spheres rotating with
10
-2 no net torque (filled squares) and non-
rotating spheres (open squares), from
Dandy and Dwyer (1990) for nonrotat-
ing spheres (delta), and from Kurose
-3
and Komori (1999) for nonrotating
10
10
0
10
1
10
2 spheres (circles).

Rep

Herron et al. (1975) for small particle Reynolds numbers. Experimental results for
the lift force on contaminated microbubbles, which behave similarly to solid spheres,
immersed in vortices are reported by Sridhar and Katz (1995), who observe lift force
between one and two orders of magnitude higher than predicted by expressions
for linear shear flow. Bagchi and Balachandar (2002b) compare lift force for both
fixed and freely rotating particles immersed in a vortex flow with that for particles
immersed in a linear shear flow, both flows having the same vorticity, with Reynolds
numbers in the range 10 ≤ Re p ≤ 100 based on numerical simulations using a high-
resolution pseudospectral method. They also observe significantly higher lift force
for the vortex flow at moderate values of the Reynolds number.

5.2.2. Magnus Lift Force

Problems involving colliding particles often exhibit large particle rotation rates due
to the sliding resistance torque exerted on the particles during collisions. A spin-
ning particle moving through a fluid is observed to experience a lateral force. This
phenomenon, called the Magnus effect, is commonly observed in the motion of a
curveball in baseball or in the various spin shots used in tennis. For flow with small
particle Reynolds numbers, Rubinow and Keller (1961) showed that a sphere spin-
ning with rotation rate  and translating at a velocity v through an otherwise still
fluid experiences a lift force Fm given by

Fm = π ρ f (d/2)3  × v. (5.2.6)

Experimental results for lift force on a spinning sphere traveling in an otherwise

stationary fluid are reported by Oesterlé and Dinh (1998), Tsuji et al. (1985), and Tri
et al. (1990), and numerical results are given by Bagchi and Balachandar (2002a).
These studies have found that although (5.2.6) gives the correct trends for the Magnus
force even for Reynolds numbers as high as 100, experimental and numerical data
are better fit if the expression in (5.2.6) is multiplied by a factor CS ≤ 1. Values of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.3. Forces in Unsteady Flows 141

CS proposed by different investigators include CS = 0.25 (Tanaka et al., 1990), CS =

0.4 (Tri et al., 1990), and CS = 0.55 (Bagchi and Balachandar, 2002a).
The problem of a rotating particle immersed in a linear shear flow was con-
sidered analytically by Saffman (1965) and via numerical computation by Kurose
and Komori (1999) and Bagchi and Balachandar (2002a). Saffman employed an
asymptotic analysis to show that the total lift is given by
 
11
F ,T OT = 6.46μ(d/2)2 vs (G/ν)1/2 + π ρ f (d/2)3 vs  − G , (5.2.7)
8π

where  is the magnitude of the particle rotation rate. The first term in (5.2.7) is
the same as given in (5.2.1) for a fixed particle in a linear shear flow, and the part
proportional to  in the second term is the same as the Rubinow-Keller expression
(5.2.6) for the Magnus force. The factor 11G/8π ∼ = 0.4377G in the last term in (5.2.7)
is close to the value G2 for the rotation rate of fluid elements in a flow with shear
rate G.
It has been noted by a number of investigators that a spherical particle in a linear
shear flow develops a spin that asymptotes at a torque-free rotation rate st which
is different than the rotation rate  f = G/2 of fluid elements in the shear flow. Lin
et al. (1970) employ an asymptotic method for small Reynolds numbers to show
that
st
= 1 − 0.0385Re3/2
G . (5.2.8)
f

Based on numerical simulations at higher Reynolds numbers, Bagchi and Balachan-

dar (2002a) develop a correlation for this ratio as
⎧
st ⎨ 1 − 0.03464Re0.95p for 0.5 < Re p ≤ 5
= . (5.2.9)
f ⎩ 1 − 0.0755Re0.455 for 5 ≤ Re ≤ 200
p p

In the vanishing Reynolds number limit, st →  f , in agreement with the viscous
torque expression (5.1.6).

5.3. Forces in Unsteady Flows

Several additional types of forces act on a particle if the fluid velocity at the particle
centroid location varies in time. This can occur either due to an unsteady velocity
field or due to movement of the particle in a nonhomogeneous velocity field.

5.3.1. Pressure-Gradient (Buoyancy) Force

Pressure-gradient force exists whenever a body is placed in an ambient pressure
gradient. The general expression for the force Fp due to a pressure field p acting on
the surface S of a body with outward unit normal n is

F p = − pn da. (5.3.1)
S
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

142 Fluid Forces on Particles

Using the divergence theorem and assuming a constant pressure gradient ∇ p, this
expression reduces to

F p = − ∇ pdv = −V ∇ p. (5.3.2)
V

Pressure-gradient force can arise from any type of pressure gradient. For instance,
in a gravitational field the hydrostatic pressure variation is given by ∇ p = −ρ f gez ,
where g is the gravitational acceleration. Substituting this expression into (5.3.2)
gives

F p = ρ f gV ez , (5.3.3)

which is the buoyancy force predicted by Archimedes’ law. Combining (5.3.3) with
the gravitational force acting on the particle yields the reduced gravity force

Fg = (ρ f − ρ p )gV ez = −ρ p gRV ez , (5.3.4)

where gR = (1 − χ )g is called the reduced gravitational acceleration and χ = ρ f /ρ p .

In many fluid flows of interest, the flow Reynolds number is large and, outside of
boundary layers, the viscous term is small compared with the inertial term. In such
cases, the pressure gradient associated with fluid acceleration is ∇ p = −ρ f Du/Dt,
where DDt denotes the material derivative following a fluid material element.
Substituting this expression into (5.3.2) gives the pressure-gradient force as
π Du
Fp = ρ f d3 . (5.3.5)
6 Dt
This expression is often known as the inertial pressure-gradient force to distinguish
it from the buoyancy force due to the hydrostatic pressure gradient. An interesting
and important example of the inertial pressure-gradient force occurs when a particle
is exposed to a vortex flow field, as was discussed in Section 1.2.

5.3.2. Added Mass Force

The added mass force is a term used to describe the additional force that is required
to accelerate a body immersed in a dense fluid compared to that required when the
body is in a vacuum. This additional force makes the body respond as if it has a
greater mass than it actually does, hence the term “added mass.” The added mass
arises from the fact that an object accelerating relative to the surrounding fluid must
also increase the kinetic energy associated with motion of the surrounding fluid. This
increased kinetic energy is derived from work that is performed by the “added mass
force” as the body is accelerated.
The classical expression for added mass force Fa of a body traveling with speed
v in an inviscid fluid with uniform velocity field u(t ) is (Batchelor, 1967)
 
π 3 dv du
Fa = −c M ρ d − , (5.3.6)
6 f dt dt
where for a sphere the added mass coefficient is cM = 1/2. The notation ddt is
the time derivative following the particle. In the case that the ambient fluid flow is
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.3. Forces in Unsteady Flows 143

not uniform, Auton et al. (1988) showed that the correct expression for force on a
particle in an inviscid flow has the form
   
Du dv Du
F = ρfV − cM − + cL (u − v) × ω , (5.3.7)
Dt dt Dt

where V is the particle volume, cL is a lift coefficient, ω is the ambient fluid vorticity,
and D/Dt denotes the material derivative following a fluid element. For a spherical
particle, cM = cL = 1/2. We recognize several different forces within this expression.
For instance, the first term within the brackets is simply the inertial pressure-gradient
force given by (5.3.5). The third term is an inviscid lift force that acts on a particle
in the presence of nonzero fluid vorticity. The ratio of the magnitude of this inviscid
lift term and the magnitude of the Saffman lift (5.2.1) for a linear shear flow gives

F ,INV
∼ 1/2
= 0.16ReG . (5.3.8)
F ,SAF F

Hence, for small values of ReG the Saffman lift is substantially larger than the inviscid
lift, and the latter may be safely neglected, at least for a solid particle. On the other
hand, numerical simulations by Legendre and Magnaudet (1998) for a clean bubble
at a moderate particle Reynolds number (Re p = 500) report a value of lift coefficient
of cL = 0.49, which is quite close to the inviscid result.
The second term in (5.3.7) provides an expression for added mass force that is
valid for general ambient flow fields as
 
π dv Du
Fa = −cM ρ f d3 − . (5.3.9)
6 dt Dt

The only difference between the expression (5.3.6) for uniform flow and the expres-
sion (5.3.9) for general flows is the type of derivative acting on the fluid velocity u.
The difference between these two derivatives is given by

Du du
− = (u − v) · ∇u. (5.3.10)
Dt dt

When this is substituted into (5.3.9), the difference between the expression (5.3.9)
and the standard expression (5.3.6) is of the same order of magnitude as the inviscid
lift force. While the expression (5.3.9) for added mass force was derived for inviscid
flow, it was demonstrated in the numerical simulations by Rivero et al. (1991) that the
expression remains accurate with cM = 1/2 even in viscous flows over a wide range of
Reynolds number and fluid accelerations (see also discussion in Magnaudet, 1997).

5.3.3. History Force

The Basset history force is associated with the response of the particle viscous shear
force to acceleration of the flow relative to the particle. The basic effect can be
interpreted in part by consideration of the classic Stokes problem of a flat plate lying
along the x-axis in an unbounded fluid domain. At time t = t  the plate is accelerated
in the x-direction with a small velocity increment dU . The fluid above the plate
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

144 Fluid Forces on Particles

is driven into motion in order to satisfy the no-slip condition. The Navier-Stokes
equation for the fluid velocity u(y, t ) reduces to

∂u ∂ 2u
= ν 2. (5.3.11)
∂t ∂y
This problem admits a similarity solution of the form
 
y
u(y, t ) = dU erf √ , (5.3.12)
2 ν(t − t  )
where erf(·) is the error function. The shear stress at the plate surface (y = 0) varies
as a function of time as
  
∂u  ρ f μ 1/2 dU
τw = μ y=0 = √ . (5.3.13)
∂y  π t − t

Because (5.3.11) is a linear equation, we can decompose a gradual velocity variation

U (t ) into a series of small discrete jumps and write the total velocity as the sum of
that due to each of these jumps. The wall shear stress at time t due to this series of
discrete jumps can then be obtained from (5.3.13) as
   
∂u  ρ f μ 1/2 t dU/dt  
τw = μ y=0 = √ dt . (5.3.14)
∂y  π 0 t − t

This example illustrates some of the physics associated with how acceleration of the
particle relative to the velocity u of the surrounding fluid can lead to a time-varying
viscous force that augments the steady-state force on the particle. For a spherical
particle at low particle Reynolds number, an expression for this additional “history”
force was obtained by Boussinesq (1885) and Basset (1888) as
 t  
du dv
Fh = μd K(t − t  ) − dt  , (5.3.15)
−∞ dt  dt 
where the kernel function K(t − t  ) is given by
1/2
 3 π ρ f d2
K(t − t ) = . (5.3.16)
2 μ(t − t  )

This force is often referred to as the Boussinesq-Basset history force, or simply the
Basset force. Druzhinin and Ostrovsky (1994) used this expression to study the effect
of the Basset force for particle transport in both a vortex and a cellular flow field at
low Stokes numbers. For the problems of drift of a heavy particle out of a vortex
or drift of a light particle toward a vortex, Druzhinin and Ostrovsky found that the
Basset force acts in the direction opposite to that of the particle drift, significantly
decreasing the rate of drift compared with simulations with no Basset force. A later
experimental and numerical study of this problem is reported by Candelier et al.
(2004), who find that the difference in drift rate with and without the Basset force
varies with the square root of the Stokes number. Although the steady-state viscous
drag is normally the dominant force on the particle, particularly at low Reynolds
numbers, Thomas (1992) observed in an investigation of particle motion through
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.4. Brownian Motion 145

a shock wave that within the shock, the Basset force can be of a similar order of
magnitude to the viscous drag.
Numerical studies conducted by Mei et al. (1991) and Rivero et al. (1991) for a
flow with small-amplitude oscillations concluded that while the kernel (5.3.16) yields
reasonable results at short times, it predicts values for the Basset force at long times
that are too large. They trace this problem to the fact that in the far-field (Oseen)
region of the sphere wake, advection becomes a more effective method for vorticity
transport than diffusion due to the small velocity gradients. The Basset-Boussinesq
expression (5.3.15), (5.3.16) was derived using the Stokes equation (5.1.4), with
assumes purely diffusion-controlled vorticity transport. Alternative expressions for
history force are proposed by Mei and Adrian (1992) and Lovalenti and Brady
(1993), giving an integration kernel that decays as t −2 at long time.

5.4. Brownian Motion

Observe what happens when sunbeams are admitted into a building and shed light on
its shadowy places. You will see a multitude of tiny particles mingling in a multitude of
ways . . . their dancing is an actual indication of underlying movements of matter that are
hidden from our sight . . . . It originates with the atoms which move of themselves. Then
those small compound bodies that are least removed from the impetus of the atoms are
set in motion by the impact of their invisible blows and in turn cannon against slightly
larger bodies. So the movement mounts up from the atoms and gradually emerges to the
level of our senses, so that those bodies are in motion that we see in sunbeams, moved
by blows that remain invisible (Lucretius, “On the Nature of Things,” c. 60 BC).

It has been known since ancient times that very small bodies execute strange,
random motions when suspended in a gas. The verse by the Roman poet and philoso-
pher Lucretius presented here is remarkable not only for its understanding of the
particulate nature of the gas but also for its interpretation of this random motion as
being due to the collision of dust particles with the elementary particles that make
up the fluid. A similar random motion was observed in the 1827 by the botanist
Robert Brown (Brown, 1828), who attributed random motions of pollen particles
suspended in water to the collisions of the particles with the water molecules. These
random motions have played an important role not only in the understanding of the
molecular nature of matter, but also as an important example in the development
of the field of stochastic processes, with key early contributions on development
of a mathematical theory of Brownian motion of particles by Einstein (1906) and
Smoluchowski (1906).
The random force acting on a small particle immersed in a fluid that gives rise
to Brownian motion is known as the Brownian force. In order to examine Brownian
force, we consider a simplified model in which the particle momentum equation is
reduced to three terms – the particle inertia, the Stokes drag, and the Brownian
force. The resulting equation, given by
d2 x dx
m + 3π dμ = Fb (t ), (5.4.1)
dt 2 dt
is a form of the Langevin equation discussed in Chapter 2. In this equation, x(t ) is
the position vector of the particle centroid and Fb (t ) denotes the Brownian force.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

146 Fluid Forces on Particles

For very small particles, it might be necessary to modify the drag term in (5.4.1) to
include the slip correction factor CC discussed in Section 5.1, but we will omit the
effect of slip for the present discussion.
The Brownian force can be characterized on the basis of two assumptions. The
first assumption is that the Brownian force is an uncorrelated random function that
can be represented by a Gaussian probability distribution with mean and covariance
given by

Fb (t ) = 0 (5.4.2)

Fb (t )FTb (t + τ ) = Bδ(τ ). (5.4.3)

The notation · denotes the ensemble average, defined by

1  (i)
N
Fb (t ) = Fb (t ). (5.4.4)
N
i=1

Here Fb(i) (t ) is the ith realization of a stochastic process having the same initial
conditions. The ! ∞notation δ(t ) denotes the Dirac delta, defined such that δ(t ) = 0
for t = 0 and −∞ δ(t )dt = 1, and B is a constant, second-order tensor. The second
assumption is that when the particle is in an equilibrium state (such that there is no
mean motion), the particle kinetic energy is partitioned as
1 1
m vvT = kB T I, (5.4.5)
2 2
where kB is the Boltzmann constant, T is absolute temperature, m is the particle
mass, and I is the identity tensor.
The Langevin equation (5.4.1) is a first-order linear ordinary differential equa-
tion for the particle velocity dx/dt, which admits the solution

dx 1 t
= F (t  ) exp[−3πdμ(t − t  )/m]dt  . (5.4.6)
dt m −∞ b
From (5.4.6) and (5.4.3), the velocity autocorrelation is obtained as
$ %  
dx dxT B 3π μd
R(τ ) ≡ (t ) (t + τ ) = exp − τ . (5.4.7)
dt dt 6π μdm m
Because R(0) = vv, (5.4.7) with τ = 0 and (5.4.5) yield an expression for the con-
stant tensor B as

B = 6π μdkB T I. (5.4.8)

The ensemble average of the square of the particle displacement xxT is called the
mean-square displacement (MSD). The autocorrelation function (5.4.7) can be used
to show that the MSD over long time (t d2 /ν) approaches
2kB T
xxT = tI, (5.4.9)
3π μd
from which it follows that the averaged position of a particle increases in proportion
to the square root of time.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.5. Scaling Analysis 147

Einstein (1906) and Smoluchowski (1906) showed that over long times, a set of
particles undergoing Brownian motion diffuse outward from regions of high particle
concentration in a manner as if the concentration φ(x, t ) obeyed a diffusion equation
of the form
∂φ
= Db ∇ 2 φ. (5.4.10)
∂t
The diffusion coefficient Db is equal to the coefficient on the right-hand side of
(5.4.9), or
kB T
Db = , (5.4.11)
3π μd
which is known as the Stokes-Einstein relation. In terms of the diffusion coefficient,
the MSD in (5.4.9) can be written as

xxT = 2DbtI. (5.4.12)

Equation (5.4.12) is valid for long times, satisfying t τ p where τ p = m/3πdμ is the
particle response time. In the opposite extreme of very short times, such that t τ p,
the particle dynamics are dominated by the particle inertia, and the corresponding
particle motion changes from a diffusive nature with MSD proportional to t to a
ballistic nature with MSD proportional to t 2 . The short- and long-time behavior of
particles under Brownian motion was examined in a recent experimental study by
Li et al. (2010), which used a dual-beam optical tweezer to measure the MSD of
a 3 µm silica bead as a function of time. The MSDs over short time intervals are
found to be significantly different from the predictions from Einstein’s theory of
Brownian motion in a diffusive regime, but agree well with the predictions of the
ballistic Brownian motion theory that the MSD is proportional to t2 .
For purposes of numerical solution, it is convenient to write (5.4.1) in terms of
the velocity differential dv as

mdv + 3π dμvdt = B1/2 dW. (5.4.13)

In this equation, the scalar B is defined such that B = BI, so from (5.4.8)

B = 6π μdkB T. (5.4.14)

The vector dW can be written in component form as dW = dWi ei , where the compo-
nents dWi are independent random variables with Gaussian probability distribution
having zero mean and variance dt. Comparing (5.4.13) with (5.4.1), the Brownian
force Fb is given by
dW
Fb = B1/2 . (5.4.15)
dt

5.5. Scaling Analysis

The relative importance of the different forces discussed in this chapter can be exam-
ined using a scaling analysis, which we present for the case of a particle of diameter
d moving in a linear shear flow with shear rate G. The major parameters used in this
scaling analysis are the dimensionless particle diameter εd ≡ d/L, the flow Reynolds
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

148 Fluid Forces on Particles

Table 5.1. Order of magnitude estimates for the different fluid

forces on the particles, compared with the particle inertia or the
particle drag, which are assumed to be of the same order of
magnitude

Order of magnitude relative

Force to particle inertia/drag

Drag force 1
Saffman lift force Re1/2
G
Magnus lift force ReG
Added mass force χ
Pressure-gradient force χ
Basset history force (Re f εSt)1/2 = Re1/2
p
Gravitational force (εFr2 )−1
Brownian force (1/εSt Pe f )1/2 = Pe−1/2
p

number Re f = ρ f U L/μ, the shear Reynolds number ReG = ρ f Gd2 /μ, the density
ratio χ = ρ f /ρ p , the Stokes number St ≡ ρ p d2U /18μL, the particle Froude num-

ber Fr = U/ gR d, and the flow Peclet number associated with Brownian motion
Pe f = LU/Db . Here L and U are length and velocity scales associated with the flow
field and Db is the Brownian diffusion coefficient defined in (5.4.11). As is typical of
a particle in the micrometer size range, it is assumed that fluid drag is the dominant
force balanced by the particle inertia. For particle flows at low Stokes numbers,
balancing the order of magnitude of particle inertia and drag results in the scaling
|v − u| = O(StU ) for the particle slip velocity. The scaling estimate also neglects
particle collisions, instead concentrating on the forces acting on an isolated particle
immersed in the fluid, and it assumes that the particle Reynolds number is small.
A listing of the order of magnitude of the different forces discussed in the chapter
is given in Table 5.1. This table lists the ratio of the order of magnitude of each force
divided by the order of magnitude of the particle inertia. The entry next to drag is
unity, as drag is of the same order of magnitude as inertia. The ratio of the Saffman
lift force (5.2.1) to the Stokes drag force gives
 
F O μd2 vS (G/ν)1/2    
= = O d(G/ν)1/2 = O Re1/2 G . (5.5.1)
Fd O(μdvS )

Taking the ratio of the Magnus force expression (5.2.6) to the Stokes drag force and
assuming that the particle rotation rate  is of the same order of magnitude as the
shear rate G gives

Fm O[ρ f d3 vS ]
= = O[ρ f d2 G/μ] = O(ReG ) (5.5.2)
Fd O(μdvS )

The ratio of the reduced gravity force (5.3.4) to the Stokes drag, using the observation
that vS = O(StU ), is obtained as
 
Fg O(ρ p gR d3 ) εd Re f
= =O . (5.5.3)
Fd O(μdvS ) χ St Fr2
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.5. Scaling Analysis 149

Using the relationship (1.2.14) between the Stokes number and the flow Reynolds
number, this expression becomes
 
Fg 1
=O . (5.5.4)
Fd εd Fr2
The inertial pressure-gradient force and the added mass force have the same order
of magnitude. In this case it is more convenient to divide the force by the order
of magnitude of the particle inertia Fi = O(mdv/dt ) = O(ρ p d3U/τ p ), where τ p is a
particle time scale, to get
 
Fp F O(ρ f d3U/τ p )
=O a = = O(χ ). (5.5.5)
Fi Fi O(ρ p d3U/τ p )
The ratio of the history force (5.3.15) to the drag force gives
√
Fh O(μd2 vS / ντ )
= . (5.5.6)
Fd O(μdvS )
The time scale most suited for the viscous stresses on the particle surface is given by
the time associated with the Stokes drag, or τS = d/vS . Substituting this time scale
into (5.4.21) and using vS = O(StU ) for the slip velocity gives
 
Fh dvS 1/2
=O = O(εd Re f St)1/2 . (5.5.7)
Fd ν
We recall from (1.2.13) that the combination εd Re f St in (5.5.7) has the same order
of magnitude as the particle Reynolds number Re p .
From (5.4.15), the Brownian force has order of magnitude (B/τ )1/2 , where τ
is an appropriate particle time scale. If we again use the Stokes drag time scale
τS = d/vS , the ratio of the Brownian force to the drag force is given by
 
Fb O(μdkB T/τ ) kB T 1/2
= =O . (5.5.8)
Fd O(μdvS ) μd2 vS
Recalling the estimate vS = O(StU ), (5.5.8) can be written as
 1/2
Fb 1
=O . (5.5.9)
Fd εd St Pe f

A particle Peclet number can alternatively be defined using the particle diameter d
and slip velocity vS as length and velocity scales to write

Pe p ≡ vS d/Db = O(εd St Pe f ), (5.5.10)

in terms of which (5.5.9) becomes

Fb & '
= O Pe−1/2
p . (5.5.11)
Fd
Brownian force starts to become important when Pe p = O(1). When Pe p 1, the
Brownian force will become one of the dominant forces, in which case the particle
motion is governed primarily by a balance between the Stokes drag and the random
Brownian force.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

150 Fluid Forces on Particles

As a first example, we consider a dust particle with diameter d = 10 µm in

an air flow with characteristic length and velocity scales L = 1 cm and U = 1 m/s.
Assuming that the dust is made of quartz, the fluid and particle densities are ρ f = 1.2
kg/m3 and ρ p = 2650 kg/m3 . The shear rate is estimated as G ∼ = U/L = 100 s−1 . The
Brownian diffusion coefficient at room temperature (T = 300 K) is given by (5.4.11)
as Db = 2.44 × 10−12 m2 /s. For this example, the dimensionless parameters listed at
the start of this section are given by
χ = 0.000453, Re f = 667, ReG = 6.7 × 10−4 , St = 0.082,
εd = 0.001, Pe f = 4.1 × 10 ,9
Fr = 101.
Using the order of magnitude estimates stated earlier, the typical particle Reynolds
number and Peclet number can be estimated as
Re p ∼
= 0.055, and Pe p = 3.4 × 105 .
From the estimates in Table 5.1, for this example the force ratios are given approxi-
mately by
F Fm Fa F
≈ 0.026, ≈ 0.0007, ≈ a ≈ 0.0004,
Fd Fd Fd Fi
Fh Fg Fb
≈ 0.23, ≈ 0.1, ≈ 0.0017.
Fd Fd Fd
We conclude from these estimates that Magnus lift, added mass force, pressure-
gradient force, and Brownian force are all negligible. The Saffman lift force is also
small and can probably be neglected. The history and gravity forces are significant
enough that they will likely influence the analysis; however, both remain minor forces
compared to drag and inertia.
In a second example, we consider a neutrally buoyant bacterial cell (with no flag-
ella) having diameter d = 1 µm in a microfluidic flow of water through a microchannel
with L = 1 mm and U = 10 cm/s. The shear rate is estimated as G ∼ = U/L = 100 s−1 ,
and the Brownian diffusion coefficient is Db = 4.4 × 10−13 m2 /s. For this example,
the dimensionless parameters listed at the start of this section are given by
χ = 1, Re f = 100, ReG = 1.0 × 10−7 , St = 5.6 × 10−6 ,
εd = 0.001, Pe f = 2.3 × 108 , Fr = 31.9.
The typical particle Reynolds number and particle Peclet number are given by
Re p ∼
= 5.6 × 10−7 and Pe p = 1.27.
The estimates in Table 5.1 give the force ratios for this example as
F Fm Fa F
≈ 3.2 × 10−4 , ≈ 10−7 , ≈ a ≈ 1,
Fd Fd Fd Fi
Fh Fg Fb
≈ 7.5 × 10−4 , ≈ 0, ≈ 0.89.
Fd Fd Fd
In this example, the reduced gravity force is identically zero, and the Saffman and
Magnus lift forces and the history force are all negligible. However, the added mass
and pressure-gradient forces and the Brownian force are all of about the same order
of magnitude as the particle inertia and drag force.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.6. Near-Wall Effects 151

G
Ω v
Figure 5.5. Schematic diagram of a particle trav- vy
eling in a linear shear flow near a plane wall.
y vx
x ys
0

5.6. Near-Wall Effects

The presence of a wall in the flow field can significantly modify the hydrodynamics
forces acting on particles close to the wall. We consider a particle of diameter d
with centroid at a height d/2 + ys above a wall, where the offset distance ys is on
the order of the particle diameter or smaller. Assuming that the particle diameter is
much less than the characteristic fluid length scale, the influence of the wall on the
particle forces can be examined by considering the simplified problem of a particle
above a flat, infinite wall subject to a linear shear flow in the x-direction with shear
rate G. The particle centroid has translational velocity v = vx ex + vy ey and the par-
ticle rotates with rotation rate  = ez , where  is positive for a particle rotating
in a counter-clockwise direction. A schematic diagram of this flow field is given in
Figure 5.5.
The limiting case where the particle touches the plate (ys → 0) is of particular
interest for adhesive particles, which are commonly drawn to surfaces in the flow
field. Included in this limit is the important problem of a particle rolling along a
surface, for which  = −2vx /d and vy = 0. As pointed out by Goldman et al. (1967a)
and O’Neill and Stewartson (1967), the force and torque acting on a particle exhibit a
logarithmic singularity as the separation distance ys approaches zero for any problem
with nonzero particle translation velocity or rotation rate. Hence, a perfectly smooth
spherical particle in contact with a perfectly smooth, inclined wall cannot roll down
the wall under any finite body force, according to the restrictions imposed by the
Navier-Stokes equations and the no-slip boundary condition. This statement is, of
course, at odds with the everyday observation that apparently smooth particles do in
fact roll down apparently smooth walls. Because this force singularity is proportional
to log(ys ), even very small values of the separation distance ys are sufficient to allow
particle motion with reasonable force magnitudes. It was demonstrated by Smart
et al. (1993) that this singularity can be effectively resolved by accounting for the
effects of surface roughness, such that the separation distance ys is not allowed to
decrease below the approximate height of the particle surface roughness elements.

5.6.1. Drag Force

The drag force on a spherical particle at low particle Reynolds number is assumed
to vary linearly with the shear rate G, the particle translation velocities in the x- and
y-directions, and the particle rotation rate, so that
FD,x = 32 π μd2 GCG + 3π μdvxCU + 32 π μd2 CR (5.6.1)
FD,y = 3π μdvyCV . (5.6.2)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

152 Fluid Forces on Particles

Similarly, the torque on the particle can be expressed as

MF = −π μd3 GDG − π μd3 DR − 2π μd2 vx DU . (5.6.3)

We define the dimensionless separation distance ζ ≡ 2ys /d and the relative particle
velocity in the x-direction, vS = vx − G(1 + ζ )d/2. A variety of studies have been
performed in order to obtain expressions for the coefficients CG , CU , CR , CV , DR , DG ,
and DU in (5.6.1), (5.6.2), and (5.6.3) as functions of the dimensionless separation
distance ζ . In the limit of a particle far away from the wall (ζ → ∞), the Stokes drag
result is obtained for CG = −CU = −CV = DR = −2DG = 1 and CR = DU = 0. This
provides a limit that expressions for these coefficients with finite ζ must satisfy.
Most analyses examining particle forces and torques near a wall adopt one of
two approaches. The first approach, valid for large separation distances (ζ 1), is
to use some variation of the method of images to account for the reflection of the
sphere over the wall surface, where the sphere is modeled by a superposition of a
stokeslet, a doublet, and a uniform flow. For instance, Blake (1971) derived the image
system of a stokeslet over a no-slip wall, and Happel and Brenner (1983, 288–297)
utilized this technique to examine particle interaction with a plane wall as well as with
other particles based on the earlier work of Brenner. The second approach, valid for
small separation distances (ζ 1), involves the use of the lubrication approximation
for the gap flow between the particles and the wall, which is matched to an outer
solution for the flow in the region away from the wall. This approach was used by
Goldman et al. (1967a) and O’Neill and Stewartson (1967), as well as numerous later
investigators, to investigate the forces on a particle that is nearly in contact with a
wall.
Goldman et al. (1967a) examined a number of scenarios using lubrication theory
that give limiting expressions for these coefficients valid for ζ 1. Starting with the
problem of a nonrotating sphere ( = 0) translated over a wall with no shear flow
(G = 0), they derived limiting expressions for CU and DU as

CU = 8
15
ln(ζ ) − 0.9588, (5.6.4a)

DU = − 10
1
ln(ζ ) − 0.1895. (5.6.4b)

Goldman et al. (1967a) then considered the problem of a rotating particle with no
translation (vx = 0) near a wall, finding that for ζ 1 the coefficients can be written
as

CR = − 15
2
ln(ζ ) − 0.2526, (5.6.4c)

DR = − 25 ln(ζ ) + 0.3817. (5.6.4d)

In a follow-up paper examining a stationary, nonrotating particle in a shear flow,

Goldman et al. (1967b) used lubrication theory to show that the coefficients CG and
DG approach the constant values

CG = 1.7005 + O(ζ ) (5.6.4e)

DG = 0.9440 + O(ζ ) (5.6.4f)

as the particle approaches the wall.

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.6. Near-Wall Effects 153

The lubrication theory results in (5.6.4a–f) are complemented by results using

the method of reflections, which are valid for ζ 1, obtained by a variety of different
investigators. For instance, Goldman et al. (1967b) show that for large separation
distances, CG approaches
 
9 1
CG = 1 + . (5.6.5)
16 1 + ζ
This result in fact provides a reasonable approximation for arbitrary values of ζ ,
with a maximum error of about 10% as ζ → 0. Empirical fits for CG and CU that are
valid for arbitrary ζ are reported by Zeng et al. (2009) as
 
9 1
CG = 1 + 0.138 exp(−ζ ) + , (5.6.6)
16 1 + ζ
 
0.07 8 ζ
CU = −1.028 + + ln . (5.6.7)
1 + ζ2 15 1 + 0.948ζ
The problem of a spherical particle moving toward or away from a wall in an oth-
erwise quiescent fluid was examined by Brenner (1961), who derived an expression
for CV valid for arbitrary ζ using the method of reflections as
∞  
4 n(n + 1) 2 sinh[(2n + 1)α] + (2n + 1) sinh(2α)
CV = − sinh(α) −1 ,
3 (2n − 1)(2n + 3) 4 sinh2 [(n + 1/2)α] − (2n + 1)2 sinh2 α
n=1
(5.6.8)

where α ≡ cosh−1 (1 + ζ ). In the limit of small ζ , Cox and Brenner (1967) and Cooley
and O’Neill (1968) derived the lubrication-theory result
1 1
CV = − + ln(ζ ) − 0.97128. (5.6.9)
ζ 5
Wakiya (1960) gives a solution valid for large ζ as
1
CV = − . (5.6.10)
1 − (9/8)(1 + ζ )−1 + (1/2)(1 + ζ )−3
Ziskind et al. (1998) show that a close approximation to the results of the exact
solution (5.6.8) is obtained by using (5.6.9) for ζ < 1 and (5.6.10) for ζ > 1.
Modification of the results presented here at finite particle Reynolds numbers
has been examined using numerical simulation in a number of studies. The problem
of a stationary particle attached to a wall (ζ = 0) is of particular importance in adhe-
sive particle flows in which particles adhere to a wall. This problem was examined by
Sweeney and Finlay (2007) for a particle immersed in a Blasius boundary layer with
Rex = U x/ν = 1000 and several different values of particle Reynolds number Re p .
An empirical solution to the drag force obtained in this study is reported, giving
−1
0.2817 −1
CG = 1.7005 1 − sin (0.238Re p ) , (5.6.11)
Re0.0826
p

where for this study G in (5.6.1) is replaced by the ambient velocity within the Blasius
boundary layer at the particle centroid divided by the particle radius d2. This result
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

154 Fluid Forces on Particles

reduces to (5.6.4e) as Re p → 0. Zeng et al. (2005, 2009) performed an extensive

numerical study of finite Reynolds number effects for a particle in a linear shear
flow. These studies examine both the problem of a stationary particle at different
separation distances above the wall as well as a translating particle in a quiescent flow.
The latter problem was also examined experimentally by Takemura and Magnaudet
(2003) using clean and contaminated bubbles rising parallel to a plane wall and
theoretically by Vasseur and Cox (1977) for small, but finite, particle Reynolds
numbers. If the low-Reynolds number results in (5.6.6) and (5.6.7) are denoted by
CG,0 and CU,0 , respectively, Zeng et al. (2009) report a correlation based on their
numerical results valid for finite Reynolds numbers in the range 2 ≤ Re p ≤ 250
as
 β   β 
CG = CG,0 1 + α1 Re p1 , CU = CU,0 1 + α2 Re p2 , (5.6.12)

where

α1 = 0.15 − 0.046(1 − 0.04ζ 2 ) exp(−0.35ζ )

β1 = 0.687 − 0.066(1 − 0.19ζ 2 ) exp(−0.535ζ 0.9 )

α2 = 0.15[1 − exp(− ζ /2)]

β2 = 0.687 + 0.313 exp(− 2ζ ).

[These expressions have been modified to account for different definitions of the
parameter ζ between the current chapter and the paper of Zeng et al. (2009).]
Equation (5.6.12) reduces to the Schiller and Naumann (1933) expression (5.1.10)
for drag in an unbounded fluid as ζ → ∞.
The importance of particle rotation on the drag force at finite Reynolds numbers
was examined using numerical simulations by Zeng et al. (2009), comparing drag on
a stationary particle in a shear flow with that for a freely rotating particle. A similar
numerical study was performed by Stewart et al. (2010) for a spherical particle
rolling on a plane wall with different rotation rates, the results of which are shown in
Figure 5.6. In both cases, it is found that particle rotation has negligible influence on
the drag force on the particle.

5.6.2. Lift Force

The lift force is an inherently nonlinear function of shear rate, rotation rate, and
particle translation velocity, as was also observed in Section 5.2 for a particle in an
unbounded flow. Cherukat and McLaughlin (1994) solved both the problem of lift on
a nonrotating sphere and a freely rotating sphere at low particle Reynolds numbers.
Validity of their solution requires that the distance = d(1 + ζ )/2 between the
particle centroid and the wall is less than both the characteristic length scale S = ν/vS
associated with the sphere motion relative to the flow and the length scale G =
(ν/G)1/2 associated with the shear flow. Defining κ ≡ 1/(1 + ζ ), the condition
min( S , G ) is equivalent to the requirement that Re p κ and Re1/2
G κ. Cherukat
and McLaughlin (1994) find that the lift force for this problem can be written as

F = ρIvS2 d2 /4, (5.6.13)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.6. Near-Wall Effects 155

CD

Figure 5.6. Numerical results of Stewart et al. (2010) (symbols) for drag on a particle rolling
on a plane wall for different rotation rates (corresponding to different values of the parameter
α = d/2vx ), in comparison to the correlation of Zeng et al. (2009) (solid line). [Reprinted
with permission from Stewart et al. (2010).]

where vS = vx − (1 + ζ )dG/2 is the particle velocity relative to the shear flow veloc-
ity at the particle centroid. Defining S ≡ Gd/2vs , polynomial approximations for
the coefficient I in (5.6.13) are reported by Cherukat and McLaughlin as

I = [1.7716 + 0.2160κ − 0.7292κ 2 + 0.4854κ 3 ]

 
3.2397
− + 1.1450 + 2.0840κ − 0.9059κ 2 G (5.6.14a)
κ
+ [2.0069 + 1.0575κ − 2.4007κ 2 + 1.3174κ 3 ]2G

for a nonrotating particle and as

I = [1.7631 + 0.3561κ − 1.1837κ 2 + 0.845163κ 3 ]

 
3.2139
− + 2.6760 + 0.8248κ − 0.4616κ 2 G 5.6.14b)
κ
+ [1.8081 + 0.879585κ − 1.9009κ 2 + 0.98149κ 3 ]2G

for a freely rotating particle. The Cherukat-McLaughlin lift solutions for fixed and
freely rotating spheres are also found to be quite close to each other, suggesting that
sphere rotation rate has only a small effect on the lift force for a particle traveling
in a shear flow near a wall. For instance, in Figure 5.7 the coefficient I is plotted as
a function of G for two values of ζ using (5.6.14a) and (5.6.14b). In both cases, the
solutions of I are quite close to each other.
The limit of a stationary particle resting on the plane wall was examined by
Leighton and Acrivos (1985), for which the lift is given by F = 9.22ρ(d/2)4 G2 . The
Cherukat and McLaughlin solution in this limit gives F = Iρ(d/2)4 G2 , where for
κ = 1 and G = −1, (5.6.14a) gives I = 9.28. This result differs from the Leighton
and Acrivos solution by 0.65%. The problem of a sphere traveling in contact with
a plane wall is examined both theoretically and experimentally for cases with low
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

156 Fluid Forces on Particles

80 160

140

120
60
100

80
I 40
I 60

40

20
20
0

-20
0 -40

-60
-4 -2 0 2 4 -4 -2 0 2 4

ΛG ΛG
(a) (b)
Figure 5.7. Plots of the Cherukat-McLaughlin coefficient I in the lift force for a particle in a
shear flow near a wall as a function of G for cases with dimensionless separation distance
(a) δ = 0.1 and (b) δ = 4 with both a nonrotating sphere (solid line) and a freely rotating
sphere (dashed line).

particle Reynolds number by Krishnan and Leighton (1995) and King and Leighton
(1997), including effects of surface roughness.
A number of investigators have examined the effect of finite Reynolds numbers
on the lift force, including experiments by Takemura and Magnaudet (2003) for
rising bubbles near a wall, calculations for a particle in a Blasius boundary layer
by Sweeney and Finlay (2007), and numerical computations and experiments of a
sphere rolling on a plane by Stewart et al. (2010), among others. Zeng et al. (2009)
summarizes these various correlations and evaluates these expressions based on their
numerical solutions. However, as the correlations are numerous, fairly complex, and
tend to produce widely scattered results depending on the specific conditions used,
we defer to the original literature for further information on finite Reynolds number
effects.

5.7. Effect of Surrounding Particles

Each particle in a particulate flow is surrounded by other particles. Unless the particle
concentration is dilute, the presence of surrounding particles will influence the drag
force for any given particle. A number of investigators have examined this “particle
crowding” effect in different regimes. Some investigators have taken an empirical
approach and have developed correlations for drag correction based on pressure
drop data for flow through a fixed packed bed (Ergun, 1952) or a fluidized bed (Wen
and Yu, 1966), or by measurement of particle terminal velocity during sedimentation
(Richardson and Zaki, 1954). Other investigators develop the drag correlations by
employing microscale simulations such as direct numerical simulation (DNS) or
Lattice-Boltzmann method (LBM) simulations (Hill et al., 2001a,b; van der Hoef
et al., 2005).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.7. Effect of Surrounding Particles 157

Atot

g
Figure 5.8. Schematic illustrating flow through a fluidized bed in a
cylinder. L

us

For definiteness, let us consider the problem of flow through a bed of particles
contained in a cylinder with uniform cross-sectional area Atot , through which a fluid
flows with flow rate Q in the upward (positive x) direction, as shown in Figure 5.8.
The fluid superficial velocity us in the x-direction is defined by the ratio of the flow
rate to the total cross-sectional area of the flow, or us = Q/Atot . If ε ≡ 1 − φ is the
void fraction, the part of the cylinder cross-sectional area occupied by the fluid is
εAtot . The average velocity of the fluid phase within the flow is therefore given by
u f = Q/εAtot , so we can write

u f = us /ε. (5.7.1)

The total fluid force plus the gravitational force acting on an individual particle
within the bed is given by

F = Fd − ρPVp gex − Vp ∇ p, (5.7.2)

where Fd is the drag force on an individual particle and Vp = π d3 /6 is the particle

volume. The pressure gradient ∇ p in (5.7.2) can arise from a number of mechanisms,
including a hydrostatic contribution due to gravity, a contribution due to acceleration
a0 = Dus /Dt of the fluid superficial velocity, and an additional contribution due to
the flow of fluid through the bed. The first and last of these mechanisms lead to
negative values of pressure gradient for upward flow in the bed. Separating out the
first two of these mechanisms, we can decompose the pressure gradient as

∇ p = −ρ f (g − a0 )ex + ∇ pd , (5.7.3)

where ∇ pd is the part of the pressure gradient due to the drag force between the
particles and the fluid. Substituting (5.7.3) into (5.7.2) gives

F = (Fd − Vp ∇ pd ) + (ρ f − ρP )Vp gex − ρ f a0Vp . (5.7.4)

The first term on the right-hand side of (5.7.4) represents the total force on the
particle that arises from interaction with the fluid, and this combination is denoted
by

F̂d ≡ Fd − Vp ∇ pd . (5.7.5)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

158 Fluid Forces on Particles

The pressure gradient ∇ pd can be determined using the control volume V for a
particle bed of length L shown in Figure 5.8. Equating the pressure force −pd Atot
on the control volume to the volumetric force n p F̂d , where n p is the number of
particles in the control volume, gives

pd = −n p F̂d /Atot . (5.7.6)

The number of particles per unit volume within the particle bed is given by φ/Vp , so
that

n p = Atot Lφ/Vp . (5.7.7)

Substituting (5.7.7) into (5.7.6) gives

∇ pd = pd /L = −φ F̂d /Vp . (5.7.8)

Substituting (5.7.8) into (5.7.5) gives a relationship between Fd and F̂d as

F̂d = Fd /ε. (5.7.9)

The literature on fluid flow through particulate beds contains a number of apparent
inconsistencies in the various correlations used for the friction factor f defined in
(5.1.1). These differences can generally be traced back to whether the average fluid
velocity u f or the superficial fluid velocity us is used, as related by (5.7.1), or whether
the form Fd or F̂d of the drag force is reported. The force on the particles depends
in general on the difference between the average fluid velocity u f and the particle
velocity, where we denote the average fluid velocity in vector form simply as u.
We note that although u is defined as the fluid velocity at a particle centroid in the
absence of the particle, it does include the effects of surrounding particles on the
fluid velocity field.
In continuum models for particulate flows, the fluid-solid interaction force per
unit volume, fgs , is commonly assumed to be proportional to the difference between
the volume-averaged particle velocity v and the average fluid velocity u, so we can
write

fgs = −β(v − u). (5.7.10)

The coefficient β in this equation is called the interphase momentum transfer coef-
ficient. Discrete models, such as DEM, require an expression for drag force Fd on
individual particles as a function of v − u. Dividing fgs by the number of particles per
unit volume, φ/Vp , gives
βVp βVp
Fd = − (v − u) = − (v − u). (5.7.11)
φ 1−ε
Comparing (5.7.11) with (5.1.1), the friction factor f for the drag force can be written
as
βd2
f = . (5.7.12)
18μ(1−ε)
We write an expression for the entire friction factor f, rather than just the crowding
correction factor CF , in this section because the inertial tern is generally included
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.7. Effect of Surrounding Particles 159

in these correlations. Also, the conditions considered in this section generally have
very little slip, so we can take CC ∼
= 1.

5.7.1. Flow through Packed Beds

A well-known correlation for flow through a packed bed of particles was proposed
by Ergun (1952), which reduces to the classic Kozeny-Carman relationship (Kozeny,
1927; Carman, 1937) for small particle Reynolds numbers and to the Burke-Plummer
expression (Burke and Plummer, 1928) for high Reynolds numbers. The Ergun
correlation gives the pressure gradient ∇ pd due to drag force in the particle bed as

1 − ε ρ f us
2
pd (1 − ε)2 μus
∇ pd = = −150 − 1.75 . (5.7.13)
L ε3 d2 ε3 d
From (5.7.8), (5.7.9), and (5.7.11), we have
εVp βVp
Fd = − ∇ pd = uf. (5.7.14)
1−ε 1−ε
Substituting (5.7.13) into (5.7.14) and solving for β gives
ε ε2 (1 − ε)2 μ ρ f |u − v|
β=− ∇ pd = − ∇ pd = 150 + 1.75(1 − ε) , (5.7.15)
uf us ε d2 d
where we have used (5.7.1) to write the superficial velocity as us = εu f . Substituting
into (5.7.12) gives the friction factor as
βd2 (1 − ε)
f = = 8.33 + 0.097Re p , (5.7.16)
18μ(1−ε) ε
where the particle Reynolds number is defined as Re p = ρ f d |u − v| /μ. The crowd-
ing correction factor given in (5.7.16) includes finite Reynolds number effects, so
an additional inertial correction factor should not be used. The Ergun correlation
fits the packed bed experimental data well over a large range of Reynolds numbers,
but it is based on data obtained for a relatively small range of void fraction values,
ranging between approximately 0.44 and 0.53.

5.7.2. Flow through Fluidized Beds

Much greater variation in void fraction values can be achieved using fluidized particle
beds. For an equilibrium fluidized bed, both the particle velocity v and the total force
F vanish and the pressure gradient is given by
dp
= −[(1 − ε)ρ p + ερ f ]g, (5.7.17)
dz
where the term in brackets is the averaged density of the mixture. Substituting
(5.7.17) into (5.7.2) and setting the total force equal to zero gives
Fd = (ρP − ρ f )Vp gεex . (5.7.18)
Measurement of void fraction of the fluidized bed for different fluid flow rates Q
can thus be used with (5.7.18) to obtain data for particle drag force as a function of
average fluid velocity.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

160 Fluid Forces on Particles

The term multiplying ε on the right-hand side of (5.7.18) is the drag force Fdt
acting on an isolated sphere of diameter d falling at its terminal velocity ut in an
otherwise stationary fluid. Consequently, we can write (5.7.18) as

Fd = εFdt . (5.7.19)

The result (5.7.19) relates the magnitude of the drag force on a particle within the
fluidized bed subject to an upward superficial velocity us to that acting on an isolated
particle falling at a terminal velocity ut in a stationary fluid. Using the Stokes formula
(5.1.1), we can therefore write

Fd = 3π dμu f f , Fdt = 3π dμut , (5.7.20)

where u f = us /ε is the mean fluid velocity in the particle bed. Substituting the two
equations in (5.7.20) into (5.7.19) gives
   
ut ut
f = ε= ε2 . (5.7.21)
uf us

Richardson and Zaki (1954) showed experimentally that the ratio us /ut can be well
fit by an expression of the form
us
= φvn , (5.7.22)
ut

where the exponent n is a function of the particle Reynolds number Re p evaluated

using velocity ut . For small values of Re p this exponent approaches a value n = 4.65.
Substituting (5.7.22) into (5.7.21) with n = 4.65 gives the friction factor as

f = ε−2.65CI , (5.7.23)

where CI is an appropriate inertial correction factor.

The data from experimental studies spanning a large range of particle Reynolds
numbers were collected by Di Felice (1994), who noted that although (5.7.23) per-
forms well for both large and small particle Reynolds numbers, it exhibits significant
errors at intermediate Reynolds numbers. Di Felice fit available data over a particle
Reynolds number range from 0.01 to 104 by an empirical expression of the form
& '
f = ε1−ζ CI , ζ = 3.7 − 0.65 exp − 12 [1.5 − ln(Re p )]2 . (5.7.24)

This expression approaches the Wen-Yu expression for low particle Reynolds num-
ber, if the exponent 2.65 in (5.7.23) is rounded to 2.7.
While the Wen-Yu expression (5.7.23) has been validated over a wide range of
values of the void fraction ɛ, the large value of the exponent makes the expression
for CF sensitive to the value of ɛ as void fraction decreases to significantly less than
unity. The Ergun correlation gives accurate predictions for low void fractions, but
it does not approach the correct limit f → 1 as ε → 1. Gidaspow (1994) therefore
proposed combining these two correlations as

f = ξ fF,E + (1 − ξ ) fF,WY , (5.7.25)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.7. Effect of Surrounding Particles 161

Figure 5.9. Fluid streamlines for flow around parti-

cles in a face-centered cubic array computed using
the Lattice-Boltzmann method. [Reprinted with per-
mission by Hill et al. (2001b).]

where ξ is a stitching function. Here fF,E denotes the Ergun correlation (5.7.16) and
fF,WY denotes the Wen-Yu correlation (5.7.23). Gidaspow used a stitching function
of the form
0 for ε > 0.8
ξ= . (5.7.26)
1 for ε ≤ 0.8
This choice leads to a discontinuity in the particle drag expression at ε = 0.8, the
magnitude of which becomes increasingly large as the particle Reynolds number
increases. Later authors proposed continuous stitching functions that join these two
expressions over an interval in ε. For instance, Huilin et al. (2003) use an inverse
tangent function
tan−1 [262.5(0.8 − ε)]
ξ= + 0.5, (5.7.27)
π
and Dahl and Hrenya (2005) use a linear interpolation
⎧
⎨0 ε ≥ 0.8
ξ = 8 − 10ε, 0.7 ≤ ε < 0.8 . (5.7.28)
⎩
1 ε < 0.7
One problem with this combined approach is that it results in an abrupt change in
the value of f in the region 0.7 ≤ ε ≤ 0.8. The magnitude of this change becomes
progressively larger as the particle Reynolds number increases.

5.7.3. Simulations
Simulations using the Lattice-Boltzmann method (LBM) for flow through random
and ordered arrays of stationary spheres are reported by Hill et al. (2001a,b), from
which correlations for friction factor are reported for small and moderate particle
Reynolds numbers. Examples of streamlines of the flow around the particles in a
face-centered cubic array are shown in Figure 5.9. Extensions to the Hill et al. drag
correlations to apply to a broader range of values of Rep and ε were proposed by
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

162 Fluid Forces on Particles

Benyahia et al. (2006) and Beetstra et al. (2007). A review of various expressions for
particle friction factor is given by Deen et al. (2007).
The LBM simulation study by van der Hoef et al. (2005) for a wide range of
void fractions with low particle Reynolds numbers, satisfying Re p < 0.2, is reported
for both arrays formed of uniform-size particles and arrays with particles of two
different sizes. For beds with uniform particle size, van der Hoef et al. (2005) shows
that their own data as well as the low Reynolds number simulation data of Ladd
(1990) obtained with a particle multipole approach and that of Hill et al. (2001a)
obtained with LBM can be accurately fit using a simple expression for friction factor
of the form
1−ε √
f = 10 + ε3 (1 + 1.5 1 − ε). (5.7.29)
ε
The first term in this expression is the same as that proposed by Carman (1937)
based on experimental data for packed beds.
The interphase momentum transfer coefficient β can be expressed in terms of a
dimensionless drag F * as

μ(1 − ε) 2 ∗
β = 18 ε F , (5.7.30)
d2
where F * is related to the friction coefficient f by f = F ∗ ε2 . The expressions pro-
posed by Benyahia et al. (2006), based on the simulation data of Hill et al. (2001a,b)
as well as known limiting forms of the drag function, give F * for a wide range of
Reynolds number and void fraction values as follows:
Low Reynolds number region (Re p < ReTRANS )
⎧
⎨ F + F1 ε2 Re2 , when ε < 0.99
∗ 0 p
F = 4
(5.7.31a)
⎩ 1 + 3ε Re , when ε ≥ 0.99
16 p

High Reynolds number region (Re p ≥ ReTRANS )

F3 ε
F ∗ = F2 + Re p . (5.7.31b)
2
The coefficients F0 , F1 , F2 , and F3 in these expressions are functions of void fraction
ε and volumetric concentration φ = 1 − ε, and are given by
⎧ 
⎪
⎪ 1 + 3 φ/2 + 2.109φ ln φ + 17.14φ 10φ
⎪
⎨ (1−w) 1 + 0.681φ − 8.48φ 2 + 8.16φ 3 + w ε3 , when 0.6 < ε < 0.99
F0 =
⎪
⎪ 10φ
⎪
⎩ 3 , when ε ≤ 0.6
ε
(5.7.32a)
⎧ (
⎪
⎪
⎨ 1 2, when 0.9 ≤ ε < 0.99
F1 = 40 φ (5.7.32b)
⎪
⎪
⎩
0.11 + 0.000051 exp(11.6φ), when ε < 0.9
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.7. Effect of Surrounding Particles 163

⎧ 
⎪
⎪ 1 + 3 φ/2 + 2.109φ ln φ + 17.89φ 10φ
⎨ (1−w) + w 3 , when ε > 0.6
F2 = 1 + 0.681φ − 11.03φ + 15.41φ
2 3 ε
⎪
⎪ 10φ
⎩ 3 , when ε ≤ 0.6
ε
(5.7.32c)
⎧
⎪
⎨ 0.9351φ + 0.03667, when ε > 0.9047
F3 = 0.0232 (5.7.32d)
⎪
⎩ 0.0637 + 0.212φ + , when ε ≤ 0.9047
ε5
where

w ≡ exp[−10(0.4 − φ)/φ] (5.7.32e)

The transition point between the low and high Reynolds number regions in the drag
laws presented here can be determined by equating F * at the transition point, giving

F3 + F32 − 4F1 (F0 − F2 )
ReTRANS = . (5.7.33)
εF1
We note that these expressions differ slightly from those given by Benyahia et al.
(2006) due to a difference in the definition of the particle Reynolds number, where
in the current book we use Re p = ρ f d |u − v| /μ.
Figure 5.10 shows a comparison of the friction factor computed using different
correlations plotted as a function of void fraction, for values of particle Reynolds
numbers of Re p = 0.1, 1, 10, and 100. For example, a 100 µm diameter dust particle
in air with a slip velocity of |v − u| = 1 m/s has a Reynolds number of Re p = 6.7,
whereas a 10 µm diameter red blood cell in blood plasma with a slip velocity of
1 mm/s has a Reynolds number of Re p = 0.003. The void fraction is plotted from
the value for a dense packing (approximately 0.4) to unity. The correlations plotted
include (1) the Gidaspow combined Ergun-Wen-Yu correlation (5.7.25) with the
Huilin-Gidaspow stitching function (5.7.27), (2) the Di Felice correlation (5.7.24),
(3) the low Reynolds number van der Hoef et al. correlation (5.7.29), and (4) the
Benyahia et al. modification of the Hill-Koch-Ladd correlation (5.7.31) and (5.7.32).
The Schiller-Naumann inertial correction factor (5.1.10) is used together with the
Wen-Yu and Di Felice correlations to obtain the combined friction factors. The
Gidaspow correlation is found to smoothly bridge between the Ergun and Wen-Yu
correlations for the three lower values of Reynolds number, but for Re p = 100 there
is a large jump in friction factor value at the joining point. The Di Felice correlation
agrees well with the Wen-Yu correlation for all Reynolds numbers examined. The
van der Hoef and Hill-Koch-Ladd correlations agree closely with each other for the
three lower Reynolds numbers, indicating that the particle Reynolds number does
not have a major effect on the friction factor for Re p ≤ 10.
Because the crowding correction factor is based on the particle concentration,
it is valid only in a time-averaged sense. An obvious shortcoming of (5.1.30) or
(5.1.31) is that these equations take no account of the effect of a sharp gradient in
the local concentration, such as might occur near the top part of a packed bed or
along the sides of a particle agglomerate. Furthermore, the discrete element method
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

164 Fluid Forces on Particles

(a) (b)

(c) (d)
Figure 5.10. Comparison of friction factor predictions from different correlations as a function
of void fraction ε for particle Reynolds numbers (a) 0.1; (b) 1; (c) 10; and (d) 100. Correlations
used include the Gidaspow combined correlation (5.7.25) with the Huilin-Gidaspow stitching
function (dashed line), Di Felice (open squares), van der Hoef et al. (2007) low Reynolds
number (open circles), and Benyahia et al.’s (2006) modification of the Hill et al. (2001b)
correlation (filled circles).

is based on evolution of the instantaneous particle velocity, so use of a time- or space-

averaged correlation for the drag correction in the particle evolution equations will
smear over important instantaneous effects. The problem of developing a simple
analytical expression that yields a correction for the instantaneous particle drag force,
accounting for the time-varying positions and velocities of surrounding particles, is
an outstanding and important research challenge.

5.7.4. Effect of Particle Polydispersity

In practice, most particulate flows involve particles with a range of diameters, rather
than a single diameter. Such systems are said to be polydisperse. In polydisperse
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.8. Stokesian Dynamics 165

systems, the number of particles ni per unit volume of species i is given by φi /Vp,i ,
where φi is the volume concentration of particle diameter di and where Vp,i = π di3 /6.
An interphase momentum transfer coefficient βi can be defined such that the drag
force on a particle with size di is given by
βiVp,i
Fd,i = − (v − u), (5.7.34)
φi
)
where the void fraction satisfies ε = 1 − φi . An expression for the βi coefficient
for particle species i is obtained from the Ergun equation (5.7.15) as
μφi
βi = [150(1 − ε) + 1.75εRe p,i ], (5.7.35)
εdi2
where Re p ≡ ρ f di |vi − u|/μ. The Wen-Yu friction factor depends only on the void
fraction, and so it is not changed for a polydispersed system.
For polydisperse systems, it is convenient to define a diameter ratio yi by
di
yi = . (5.7.36)
d
Here, d denotes the Sauter diameter, which is defined in terms of the number Ni
of particles with diameter di as
)
N d3
d = ) i i2 . (5.7.37)
Ni di
A computational study using LBM for bidisperse systems at low Reynolds numbers
is reported by van der Hoef et al. (2005). Based on the assumption that the drag
force Fd,i acting on particles of species i is proportional to the particle diameter di ,
these authors argue that the friction factor fi for particles of species i can be written
using a weighted average as
 
fi = εyi + (1 − ε)y2i + 0.064εy3i fMD , (5.7.38)

where fMD is the friction factor for a monodisperse system with the same void
fraction, given for the van der Hoef et al. (2007) study by (5.7.29).

5.8. Stokesian Dynamics

As forces and torques are exerted on particles as the particles move relative to a fluid,
so too are forces and torques exerted from the particles back onto the fluid. In the
case of dense particle flows, in which the momentum coupling parameter defined in
(1.2.26) is not small, the effects of the particles on the fluid can lead to significant flow
modification. In some cases this flow modification influences the entire flow field,
whereas in other cases this modification may be confined to the region surrounding
an agglomerate or a cluster of particles and not significantly influence the rest of the
flow.

5.8.1. Example for Falling Cluster of Particles

A series of experiments and computations investigating the effect of two-way
coupling on the motion of a cluster of particles was performed by Nitsche and
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

166 Fluid Forces on Particles

Batchelor (1997), who studied the problem of a “droplet” (or cluster) of parti-
cles falling in an otherwise stationary fluid. The particles were 0.9 mm diameter glass
beads immersed in glycerine, which were initially grouped into clusters with diameter
of approximately 0.7 cm and allowed to settle under gravitational force. The settling
velocity was measured and compared against computations, as well as against the
predicted settling velocity for isolated particles. The computations were performed
by balancing the drag force on each particle with the gravitational force, so that each
particle travels at its terminal velocity. By using the Stokes drag expression (5.1.1),
with f = 1, the relative velocity between the fluid and a particle is obtained as
mgR
v−u=− e , (5.8.1)
3π μd z
where gR is the reduced gravitational acceleration. However, instead of setting the
fluid velocity u = 0 at the particle centroids, Nitsche and Batchelor accounted for the
induced velocity by the other particles on the fluid velocity. This was done by writing
the fluid velocity at any point x in the flow field as a sum of the velocity induced
by the different particles, such that the fluid velocity ui evaluated at the centroid of
particle i can be written as

ui = W(xi , x j ) · Fd, j , (5.8.2)
j=i

where Fd, j is the Stokes drag force on particle j and W(xi , x j ) is the Oseen tensor,
defined by
 
1 I (x − x0 )(x − x0 )T
W(x, x0 ) = + . (5.8.3)
8π μ r r3
In (5.8.3), I is the identity tensor and r ≡ |x − x0 | is the distance between points x
and x0 .
The experimental results of Nitsche and Batchelor show that the particle cluster
falls as a nearly spherical “blob,” which leaves behind a trail of particles (Fig-
ure 5.11a). The motion of particles backward into the tail region is due to the fact
that the fluid streamlines penetrate into the particle blob, as shown in Figure 5.11b.
These outer streamlines carry particles backward toward the rear of the blob, feeding
the particles that are entrained into the tail region. Nitsche and Batchelor found that
the blob of particles settles downward at a faster rate than would an individual
particle. In fact, if v0 = mgR /3π μd is the settling velocity under Stokes flow of an
isolated particle, Nitsche and Batchelor showed that an approximate expression for
the settling velocity vb of a blob of Nb particles with particle diameter d and blob
radius Rb is approximately given by
vb ∼ 3 d
= 1 + Nb . (5.8.4)
v0 5 Rb
This approximation agrees closely with the computational data. The increased set-
tling velocity of the particle blob is due to the effect of surrounding particles in
dragging the fluid downward, so as to produce an average downward fluid velocity
u in (5.8.1).
Additional experimental studies of falling particle clouds by Noh and Fernando
(1993) and Metzger et al. (2007) report the presence of a transition point in the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.8. Stokesian Dynamics 167

Boundary of region
of closed fluid
streamlines

Blob boundary

(a) (b)
Figure 5.11. Results showing the sedimentation of a “blob” of particles under gravity: (a)
photograph of particle blob with trailing particles; (b) diagram showing fluid streamlines
relative to the falling particle blob. [Reprinted with permission from Nitsche and Batchelor
(1997).]

cloud structure as it falls. For falling clouds under low flow Reynolds number con-
ditions, Metzger et al. describe a process in which the cloud transitions from an
initially spherical shape into a torus, which eventually breaks up into two or more
subclouds in a repeating cascade process (Figure 5.12). This process occurs over a

Figure 5.12. Results showing deformation and break-up of

an initially spherical particle cloud: (a) simulation using the
Stokesian dynamics method and (b) experiment. [Reprinted
with permission from Metzger et al. (2007).]

3 mm

(a) (b)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

168 Fluid Forces on Particles

Oseen wake
(1/r decay)

Figure 5.13. Plot of velocity field due to the Oseen solution about a
spherical particle traveling with velocity v, drawn in a frame traveling
Stokes region with the particle. The velocity field reduces to the Stokes flow solution
near the particle and has the form of a potential source flow with
rp / Re
v
O(1/r2 ) decay and a thin wake region with O(1/r) decay far away
from the particle (at distances greater than the inertial screening
length r p /Re p ).

Oseen region
(1/r 2 decay)

long fall length scale and requires that the system contain a large number of particles.
Numerical simulations of the transitions in a falling particle cloud were performed
by Subramanian and Koch (2008) and Pignatel et al. (2011) using an Oseen dynamics
method, which is similar to the Stokesian dynamics approach employed by Nitsche
and Batchelor (1997) but uses the full Oseen solution for hydrodynamic interaction
of the particles (Proudman and Pearson, 1957).
The Oseen solution for the flow field generated by a spherical particle with
radius r p translating with a velocity U ex relative to the surrounding fluid at low
particle Reynolds number Re is given in spherical coordinates, with the polar axis
(θ = 0) coincident with the direction of the ambient flow, by

U r2p rp 3(1 − cos θ )r Rer(1 + cos θ )

ur = − cos θ − exp −
r2 2r 4r p 2r p
 *
3 Rer(1 + cos θ )
+ 1 − exp −
2Re 2r p
 
r p sin θ 3 sin θ Rer(1 + cos θ )
uθ = U r p
2
− exp − . (5.8.5)
4r3 4r p r 2r p

This solution approaches the Stokes solution for flow past a sphere for distances r
much smaller than the inertial screening length r p /Re. At very large distances from
the particle, the velocity field approximates a point source with velocity magnitude
decaying as O(1/r2 ), except within a thin wake region near θ = π . Within this wake
region the velocity is directed inward toward the sphere and the velocity magnitude
decays with distance as O(1/r). A sketch showing the different features of the Oseen
flow is given in Figure 5.13. Unlike the Stokes solution, the Oseen flow solution is
uniformly valid throughout the flow domain. Use of the oseenlet to model particle
hydrodynamic interactions, instead of the sum of the stokeslet and the potential
doublet, allows us to extend the Stokesian dynamics method to flows in which the
flow Reynolds number is not small, provided that the particle Reynolds number
remains small.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.8. Stokesian Dynamics 169

5.8.2. General Theory

In general, the velocity induced by a stress distribution t(x) defined on the surface
S of the particle at a point x located outside of the particle is given for a Stokes flow
by

1
ui (x) = − G (x, x )t j (x )da . (5.8.6)
8π μ S i j

Here we use tensor indices for the components of the second-order tensor G and
the vectors u and t. The integration in (5.8.6) is with respect to the primed position
vector. The tensor G is related to the Oseen tensor W by Gi j = 8π μWi j . The stress
vector for a Newtonian fluid is given by

t = −pn + 2μD · n, (5.8.7)

where n is the outward unit normal of S, p is pressure, and D is the rate of deformation
tensor, defined as the symmetric part of the velocity gradient tensor.
The basic idea of the Stokesian dynamics method is to expand the integral in
(5.8.6) in terms of a multipole expansion about the centroid ξ of the particle, so as
to write fluid velocity at a point x as (Pozrikidis, 1992, 45)
   
1 ∂Gi j
ui (x) = − Gi j (x, ξ ) t j (x )da + (x, ξ ) (xk − ξk )t j (x )da + · · · .
8π μ S ∂ξk S
(5.8.8)

Usually in Stokesian dynamics this expansion is only used for points located suffi-
ciently distant from the particle that the expansion can be truncated after two terms.
Of course, in this case the multipole expansion only accounts for the far-field flow
and as particles move toward each other we must also include lubrication forces,
such as those discussed in Section 2.4. The first integral in the multipole expansion
(the zeroth moment) is simply the net force exerted on the particle. This term is the
same as the stokeslet term (or point force) that was used in the example calculation
discussed in Section 5.8.1. The second integral in the multipole expansion (the first
moment) is the flow due to a force doublet. If the force doublet is decomposed into
symmetric and antisymmetric parts, the symmetric part can be written in terms of
a quantity called a stresslet, and the antisymmetric part can be written in terms of
a point torque, or a rotlet. The advantage of the multipole expansion is that the
integrals in (5.8.8) depend only on the stress and location on the particle surface
(relative to the centroid position) and not on the position of the point x at which
we seek to know the velocity. For this reason, these integrals can be evaluated only
once each time step and used to find the velocity anywhere in the flow field. A
summary of the basic theory of Stokesian dynamics is given in the review article by
Brady and Bossis (1988).
Because Stokesian dynamics calculations usually do not account for contact
forces between the particles, it can sometimes happen that particles come too close
together. In this case, because the stokeslet and higher order terms (such as stresslet
and rotlet) are singular at the particle centroid, the computations can quickly break
down. To address this problem investigators have developed regularized versions of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

170 Fluid Forces on Particles

the stokeslet that are not singular (e.g., Cortez, 2001) by distributing the force over a
finite region. If used together with a DEM computation, however, these singularities
would not be a problem since the contact forces used in the DEM approach keep
the particles sufficiently far away from each other.
The Stokesian dynamics method requires solution of a matrix equation for
the fluid velocity at each particle centroid. This can be seen by considering the
expression (5.8.2) for the fluid velocity and noting that the Stokes equation for
Fd is proportional to the fluid velocity u evaluated at the center of each particle.
For a system of N particles, direct solution of this matrix problem requires O(N 3 )
calculations per time step, and an efficient iterative matrix solver still requires O(N 2 )
calculations. To address these computational requirements, an accelerated approach
using Fourier transforms has been proposed by Sierou and Brady (2001) which
reduces the computational cost per time step to O(N ln N).
A significant limitation of the Stokesian dynamics method is that the flow
Reynolds number must be sufficiently small that the Stokes equation (5.1.4) can
be used to solve for the flow field through the entire flow domain. This is obviously a
much more restrictive requirement than simply requiring that the particle Reynolds
number be small, and it limits the Stokesian dynamics method to problems such
as microfluidics or similar flows in small-scale, confined regions. The force-coupling
method proposed by Lomholt and Maxey (2003) seeks to ease this restriction by
distributing the particle force to a grid and then solving the flow field on the grid
using the full Navier-Stokes equation. This approach is more time consuming than
Stokesian dynamics, but it considerably extends the types of problems for which such
an approach can be used and still retains many of the advantages of the Stokesian
dynamics method. Alternatively, the Oseen dynamics method used by Subramanian
and Koch (2008) for a falling particle cloud is valid for arbitrary flow Reynolds
numbers, assuming only small particle Reynolds numbers.
The presence of walls in the flow domain introduces another complicating fea-
ture of the Stokesian dynamics method. For simple flow geometries, such as a flat
surface or a sphere, analytical expressions can be used for the image of a stokeslet
over the wall (Blake, 1971; Bossis et al., 1991). For more complex geometries, a
boundary element method can be used to compute the Stokes flow within the domain
(Pozrikidis, 1992). Complications can arise when using a boundary element method
if particles move close to the wall, for which case it might be difficult to resolve the
particle image set when the particle is much smaller than the panel used to discretize
the flow domain. In such cases the pseudoimage method, described in Section 8.4
with reference to particle-induced electric fields, might be of use.

5.9. Particle Interactions with Acoustic Fields

Acoustic fields are used in a number of applications to accelerate the rate of parti-
cle agglomeration. For instance, in coal power plants it has been found effective to
expose the exhaust stream to high intensity acoustic excitation just downstream of the
combustion chamber in order to increase formation of agglomerations of micron- and
submicron-size particles upstream of the electrostatic precipitators (Gallego-Juárez
et al., 1999; Liu et al., 2009). This acoustic pretreatment significantly enhances per-
formance of the precipitators due to the fact that the net fluid drag on an agglomerate
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.9. Particle Interactions with Acoustic Fields 171

is substantially less than the drag on the sum of the individual particles in isolation,
as discussed in Section 5.7. In a related application, de Sarabia et al. (2003) exam-
ine the effect of ultrasonic forcing on agglomeration due to liquid bridge forces for
particulates in diesel exhaust streams. Very intense acoustic excitation has also been
used as a method to break up particle agglomerates. For instance, Ding and Pacek
(2008) examined use of ultrasound to break up nanoparticle colloidal agglomerates
with different ionic strengths. Acoustic excitation has been used by numerous inves-
tigators for cleaning of particles from a planar surface. For instance, Chen and Wu
(2010) and Furhmann et al. (2013) showed that exposure of a flat plate to acoustic
waves in air significantly enhances the effectiveness of different aerodynamic parti-
cle removal approaches. Ultrasonic excitation is a common method used in cleaning
silicon wafers, ceramic membranes, and other sensitive surfaces in liquids (Brereton
and Bruno, 1994; Kuehn et al., 1996; Lamminen et al., 2004). As noted by Kim et al.
(2009) and Keswani et al. (2009), ultrasonic cleaning not only moves the attached
particles directly, but it leads to formation of cavitation bubbles on the surface which
dislodge nearby particles.
Acoustic waves are also used to manipulate particles on a surface. For instance,
an ultrasound trap is a device that focuses particles onto the nodal points of an
acoustic pressure field as part of a biological assay process (Kuznetsova et al., 2007;
Sobanski et al., 2001). Acoustic radiation force produces a particle drift at a rate that
varies with particle size. This observation has been used by a number of investiga-
tors for development of acoustic particle separation devices in microfluidic systems
(Kapishnikov et al., 2006; Laurell et al., 2007; Nam et al., 2011).
Acoustic-particle interaction effects are usually categorized as first-order or
second-order, with the former depending linearly on the imposed acoustic velocity
field and the latter depending nonlinearly on the acoustic velocity field. The dominant
first-order effects leading to particle agglomeration are orthokinetic motion and the
acoustic wake effect (or hydrodynamic interaction) (Dong et al., 2006; Song et al.,
1994; Sheng and Shen, 2006). Orthokinetic motion arises from the oscillatory motion
of particles of different sizes induced by the acoustic wave motion. The acoustic
wake effect arises from the hydrodynamic interaction of one particle with the wake
of another particle.
The second-order acoustic-particle interactions give rise to a steady force on the
particle called the acoustic radiation pressure, which can be used to trap particles
into velocity nodes or antinodes of the acoustic wave field. The seminal theoretical
work on acoustic radiation pressure was the paper by King (1934), who derived an
expression for radiation pressure on a rigid particle suspended freely in an inviscid
fluid with arbitrary particle diameter relative to the acoustic wavelength. Most sub-
sequent work builds on King’s theory by releasing various of his assumptions. The
effect of sphere elasticity was examined by Hasegawa and Yosioka (1969), as well as
in subsequent papers by the Hasegawa and colleagues. Effects of fluid viscosity are
particularly important both because small suspended particles typically have small
particles Reynolds numbers and because the presence of viscosity introduces new
physical effects, such as acoustic streaming around the particle, which can have a
substantial influence on the radiation force. Theoretical studies of radiation pressure
on a particle in a viscous fluid in various limits are reported in a series of papers by
Danilov (Danilov, 1985; Danilov and Mironov, 1984; 2000) and in a series of more
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

172 Fluid Forces on Particles

comprehensive papers by Doinikov (Doinikov, 1994, 1997). The radiation pressure

is often neglected for studies of particle agglomeration because in many cases the
time scale required for particle drift under this force is long compared to the parti-
cle agglomeration time scale associated with the two first-order forces discussed in
the previous paragraph (Song, 1990).

5.9.1. Orthokinetic Motion

In the presence of an acoustic wave, particles oscillate back and forth at the same
frequency as the acoustic wave velocity. The particle phase is different than that of
the acoustic wave velocity, however, due to effect of particle inertia. To see this,
consider a simplified problem introduced by Tiwary and Reethof (1986) in which
the fluid velocity oscillates as
u = U0 sin(ωt ), (5.9.1)
where ω is the acoustic angular frequency. For sufficiently small particles in an
aerosol, it is reasonable to neglect effects such as added mass and Basset force to
write the particle moment equation simply in terms of the particle inertia and fluid
drag, giving
dv
m = 3π μd(u − v), (5.9.2)
dt
where we have used scalars for the fluid velocity u(t) and particle velocity v(t),
because the motion in this example is assumed to be one-dimensional. In this equa-
tion it is assumed that the particle Reynolds number is sufficiently small that the
Stokes drag expression applies and that the particle diameter is much smaller than
the acoustic wave length. The particle velocity oscillates with the same frequency as
the acoustic wave, but with a different phase and amplitude, so we can write
v = η pU0 sin(ωt − φ p ), (5.9.3)
where the particle entrainment coefficient η p is the ratio of the particle velocity
amplitude to the fluid velocity amplitude and the phase factor φ p is the phase delay
between the particle and fluid oscillations. Substituting (5.9.3) into (5.9.2) and solving
for η p and φ p gives

η p = 1/ 1 + (ωτ p )2 , φ p = tan−1 (ωτ p ), (5.9.4)

where τ p = m/3π μd is the particle time scale. The product ωτ p can be regarded as an
acoustic Stokes number, where we see that in the limit ωτ p→0 the particle amplitude
and phase approach those of the fluid.
Orthokinetic particle collisions occur because the amplitude and phase of the
particle motion in response to acoustic excitation depend on the particle diameter,
as ωτ p varies with diameter in proportion to d2 . In general, the amplitude of particle
oscillation decreases as ωτ p increases. The relative velocity between two nearby par-
ticles (i.e., particles that are sufficiently close to each other to experience essentially
the same fluid velocity) with diameters d1 and d2 is given by
Urel,12 = [η1 cos(ωt − φ1 ) − η2 cos(ωt − φ2 )]U0 , (5.9.5)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

5.9. Particle Interactions with Acoustic Fields 173

Boundary of agglomeration volume

Small particle
Figure 5.14. Schematic diagram showing the acoustic
agglomeration volume surrounding a large particle
oscillating under an acoustic wave in the presence of
a smaller particle. [Reprinted with permission from Li
et al. (2011).]

Particle oscillation path Large particle

where ηi and φi are the entrainment coefficient and phase for particle i. In the
frame of one particle, any second particle located within an agglomeration volume
surrounding the first particle, as shown in Figure 5.14, would collide with the first
particle, assuming that there is no deflection of the particle path due to hydrodynamic
interaction. In a typical DEM simulation of acoustic particle interaction, the fluid
time step is chosen to be sufficiently small to resolve the acoustic wave oscillation.
In this case, the motion of individual particles is accurately resolved in time, such
that orthokinetic collisions are obtained as a result of the simulation and no special
modeling is required. However, this will often require that the time scale used for
the fluid transport must be set extremely small.
Use of a time-averaged modeling approach allows much larger fluid time steps
to be used which do not capture the acoustic wave motion; instead, the effect of
acoustic waves is incorporated in the particle transport model. In a population bal-
ance approach, the effect of acoustic waves is usually accounted for by modifying
the collision kernel so that particle collision occurs in a certain fraction fC of cases
where two particles enter into their respective agglomeration volumes. The fraction
fC is called the collision efficiency, and it can be written as a function of the relative
acoustic Stokes number St12 as (Dong et al., 2006)
 B
St12
fC = , (5.9.6)
St12 + A
where A and B are constants, given empirically by A = 0.65 and B = 3.7. For two
particles with diameters d1 and d2 , such that d1 > d2 , the relative acoustic Stokes
number is defined by
ρ p η12U0 d22
St12 ≡ , (5.9.7)
18μd1
where
ω(τ1 − τ2 )
η12 ≡   (5.9.8)
1 + (ωτ1 )2 1 + (ωτ2 )2
is the relative entrainment between the two particles.

5.9.2. Acoustic Wake Effect

Two idealized types of hydrodynamic interactions between particles under an
imposed acoustic field are illustrated in Figure 5.15. In the first type, the line connect-
ing the centroids of two interacting particles is perpendicular to the wave propagation
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

174 Fluid Forces on Particles

Figure 5.15. Configurations of particles

and acoustic wave field for hydro-
dynamic interaction between particles
with connecting line (a) orthogonal to
the wave direction and (b) parallel to
the wave direction.
(a) (b)

direction (Figure 5.15a), whereas in the second type this line is parallel to the wave
propagation direction (Figure 5.15b). For the first type of interaction, the velocity
relative to the particles induced by the acoustic wave field is squeezed between the
two particles, leading to a reduction in pressure in the region between the parti-
cles. This pressure reduction results in an agglomeration force that acts to draw the
particles toward one another. However, Shaw (1978) showed that this type of hydro-
dynamic interaction generally has a small influence on the particles. The second type
of hydrodynamic interaction occurs when one particle travels into the wake of an
upstream particle. The viscous wake of the leading particle causes the downstream
particle to be exposed to a decreased relative fluid velocity compared with that of
the leading particle, allowing the downstream particle to travel slightly faster than
the leading particle. In the second half of the acoustic wave cycle the roles of the two
particles are reversed, but still with the effect that the particles are drawn toward
each other.
The tendency of particles that lie in each other’s wakes in the presence of an
acoustic wave to be attracted to each other is called the acoustic wake effect. An
interesting experiment that visualizes the acoustic wake effect was developed by
Hoffman and Koopmann (1997) in which two particles suspended in a fluid settle
side-by-side under gravity. At a given time, the particles are subjected to a horizontal
acoustic wave. The particles continue to settle downward, but the acoustic wake
effect makes them very gradually become attracted to each other as they oscillate
back and forth in the horizontal direction, so that the particle path forms a tuning-
fork pattern such as that shown in Figure 5.16.
In general, orthokinetic motion is the dominant cause of particle collisions at
low acoustic frequencies and for particles with significantly different diameters. The
acoustic wake effect dominates at high acoustic frequencies and for particles of
similar size. A model that accounts for the combined action of orthokinetic motion
and acoustic wake effect on the collision kernel used in the population balance model

Direction of
acoustic wave
oscillation g

Figure 5.16. DEM simulation of the tuning-fork pattern observed for a pair
of microspheres settling vertically under gravity and subject to a horizontal
planar acoustic wave field.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

References 175

is reported by Song et al. (1994). The acoustic wake effect can be implemented very
simply in a DEM approach using an expression by Dianov et al. (1968) for the
velocity reduction on downstream particles due to the acoustic wake of an upstream
particle, which is derived based on the Oseen linearization of the convective term
in the Navier-Stokes equation for low particle Reynolds numbers. The derivation
makes the additional assumption that the acoustic frequency is sufficiently low that
| ∂u
∂t
| |ν∇ 2 u|, which is valid provided that the acoustic Reynolds number satisfies
ωr p /ν
2
1. The Dianov et al. theory, modified by González-Gómez et al. (2000),
gives an expression for the change in fluid velocity in spherical coordinates at a
location (r, θ ) due to a particle located at the origin subject to a relative “slip”
velocity w = u − v in the x-direction as
A A " r # " r #
ur = 20 − 20 1 + (|w| + w cos θ ) exp − (|w| − w cos θ ) , (5.9.9)
r r 2ν 2ν

A0 w " r #
uθ = − exp − (|w| − w cos θ ) sin θ , (5.9.10)
2rν 2ν
where
 
3νr p 3r p |w|
A0 = 1+ . (5.9.11)
2 8ν
The technique is the same as the Oseen dynamics approach discussed in the previous
section. By determining the relative velocity of all particles surrounding a given
particle, the equations given here can be used to determine the effect of nearby
particle wakes on the fluid velocity incident to the particle, which when summed
yield the net fluid velocity relative to the give particle. Details on implementation
of this approach in the context of a DEM simulation are given by González-Gómez
et al. (2000). This modified DEM approach can be used to predict acoustic attraction
and agglomeration of particles, and predictions obtained with this approach (e.g.,
Figure 5.16) compare well with experimental data by Hoffmann and Koopmann
(1996) and González et al. (2002).

REFERENCES

Auton TR, Hunt JCR, Prud’homme M. The force on a body in inviscid unsteady non-uniform
rotational flow. Journal of Fluid Mechanics 197, 241–257 (1988).
Bagchi P, Balachandar S. Effect of free rotation on the motion of a solid sphere in linear shear
flow at moderate Re. Physics of Fluids 14(8), 2719–2737 (2002a).
Bagchi P, Balachandar S. Shear versus vortex-induced lift force on a rigid sphere at moderate
Re. Journal of Fluid Mechanics 473, 379–388 (2002b).
Basset AB. A Treatise on Hydrodynamics, Vol. 2, Deighton, Bell and Co., Cambridge (1888)
[reprinted by Dover Publ., New York, 1961].
Batchelor GK. An Introduction to Fluid Dynamics, Cambridge University Press, Cambridge
(1967).
Beetstra R, van der Hoef MA, Kuipers JAM. Drag force of intermediate Reynolds number
flow past mono- and bidisperse arrays of spheres. AIChE Journal 53, 489–501 (2007).
Benyahia S, Syamlal M, O’Brien TJ. Extension of Hill-Koch-Ladd drag correlation over all
ranges of Reynolds number and solids volume fraction. Powder Technology 162, 166–174
(2006).
Blake JR. A note on the image system for a stokeslet in a no-slip boundary. Mathematical
Proceedings of the Cambridge Philosophical Society 70, 303–310 (1971).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

176 Fluid Forces on Particles

Bossis G, Meunier A, Sherwood JD. Stokesian dynamics simulations of particle trajectories

near a plane. Physics of Fluids A 3(8), 1853–1858 (1991).
Boussinesq JV. Sur la résistance qu’oppose un fluide indéfini au repos, sans pesanteur, au
mouvement varié d’une sphère solide qu’il mouille sur toute sa surface, quand les vitesses
restent bien continues et assez faibles pour que leurs carrés et produits soient négligeables.
Comptes Rendu de l’Academie des Sciences 100, 935–937 (1885).
Brady J, Bossis G. Stokesian dynamics. Annual Review of Fluid Mechanics 20, 111–157 (1988).
Brenner H. The slow motion of a sphere through a viscous fluid towards a plane surface.
Chemical Engineering Science 16, 242–251 (1961).
Brereton GJ, Bruno BA. Particle removal by focused ultrasound. Journal of Sound and
Vibration 173(5), 683–698 (1994).
Bretherton FP. The motion of rigid particles in a shear flow at low Reynolds number. Journal
of Fluid Mechanics 14, 284–304 (1962).
Brown R. A brief account of microscopical observations made in the months of June, July
and August, 1827, on the particles contained in the pollen of plants; and on the general
existence of active molecules in organic and inorganic bodies. Philosophical Magazine 4,
161–173 (1828).
Burke SP, Plummer WB. Gas flow through packed columns. Ind. Eng. Chem. 20, 1196–1200
(1928).
Candelier F, Angilella JR, Souhar M. On the effect of the Boussinesq-Basset force on the
radial migration of a Stokes particle in a vortex. Physics of Fluids 16(5), 1765–1776 (2004).
Carman PC. Fluid flow through granular beds. Transactions of the Institute of Chemical
Engineers 15, 150–166 (1937).
Chen D, Wu J-W. Dislodgement and removal of dust-particles from a surface by a technique
combining acoustic standing wave and airflow. Journal of the Acoustical Society of America
127, 45–50 (2010).
Cherukat P, McLaughlin JB. The inertial lift on a rigid sphere in a linear shear flow field near
a flat wall. Journal of Fluid Mechanics 263, 1–18 (1994).
Cooley MDA, O’Neill ME. On the slow rotation of a sphere about a diameter parallel to a
nearby plane wall. Journal of the Institute of Mathematics and its Applications 4, 163–173
(1968).
Cortez R. The method of regularized stokeslets. SIAM Journal of Scientific Computing 23(4),
1204–1225 (2001).
Cox RG, Brenner H. The slow motion of a sphere through a viscous fluid towards a plane
surface. II Small gap widths, including inertial effects. Chemical Engineering Science 22(12),
1753–1777 (1967).
Cunningham E. On the velocity of steady fall of spherical particles through fluid medium.
Proceedings of the Royal Society of London A 83, 357–365 (1910).
Dahl SR, Hrenya CM. Size segregation in gas-solid fluidized beds with continuous size distri-
butions. Chemical Engineering Science 60, 6658–6673 (2005).
Dandy DS, Dwyer HA. A sphere in shear flow at finite Reynolds number: effect of shear on
particle lift, drag, and heat transfer. Journal of Fluid Mechanics 216, 381–410 (1990).
Danilov SD. Average force acting on a small sphere in a travelling-wave field in a viscous
fluid. Soviet Physics Acoustics 31(1), 26–28 (1985).
Danilov SD, Mironov MA. Radiation pressure force acting on a small particle in a sound field.
Soviet Physics Acoustics 30(4), 280–283 (1984).
Danilov SD, Mironov MA. Mean force on a small sphere in a sound field in a viscous fluid.
Journal of the Acoustical Society of America 107(1), 143–153 (2000).
Deen NG, Van Sint Annaland M, Van der Hoef MA, Kuipers JAM. Review of discrete
particle modeling of fluidized beds. Chemical Engineering Science 62, 28–44 (2007).
Dianov DB, Podol’skii AA, Turubarov VI. Calculation of the hydrodynamic interactions of
aerosol particles in a sound field under Oseen flow conditions. Soviet Physics Acoustics
13(3), 314–319 (1968).
Di Felice R. The voidage function for fluid-particle interaction systems. International Journal
of Multiphase Flow 20, 153–159 (1994).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

References 177

Ding P, Pacek AW. De-agglomeration of goethite nano-particles using ultrasonic comminu-

tion device. Powder Technology 187(1), 1–10 (2008).
Doinikov AA. Acoustic radiation pressure on a rigid sphere in a viscous fluid. Proceedings of
the Royal Society of London A 447, 447–466 (1994).
Doinikov AA. Acoustic radiation force on a spherical particle in a viscous heat-conducting
fluid. I. General formula. Journal of the Acoustical Society of America 101(2), 713–721
(1997).
Dong S, Lipkens B, Cameron TM. The effects of orthokinetic collision, acoustic wake, and
gravity on acoustic agglomeration of polydisperse aerosols. Journal of Aerosol Science 37,
540–553 (2006).
Druzhinin OA, Ostrovsky LA. The influence of Basset force on particle dynamics in two-
dimensional flows. Physica D 76, 34–43 (1994).
Einstein A. Über die von der molekularkinetischen Theorie der Wärme geforderte Bewegung
von in ruhenden Flüssigkeiten suspendierten Teilchen. Annalen der Physik 17, 549–560
(1906).
Epstein PS. On the resistance experienced by spheres in their motion through gases. Physical
Review 23, 710–733 (1924).
Ergun S. Fluid flow through packed columns. Chemical Engineering Progress 48(2), 89–94
(1952).
Faxén H. Der Widerstand gegen die Bewegung einer starren Kugel in einer zähen Flüssigkeit,
die zwischen zwei parallelen ebenen Wänden eingeschlossen ist. Annalen der Physik
373(10), 89–119 (1922).
Furhmann A, Marshall JS, Wu J-R. Effect of acoustic levitation force on aerodynamic particle
removal from a surface. Applied Acoustics 74, 535–543 (2013).
Gallego-Juárez JA, de Sarabia ERF, Rodrı́guez-Corral G, Hoffman TL, Gálvez-Moraleda
JC, Rodrı́guez-Maroto JJ, Gómez-Moreno FJ, Bahillo-Ruiz A, Martı́n-Espigares M, Acha
M. Application of acoustic agglomeration to reduce fine particle emissions from coal com-
bustion plants. Environmental Science and Technology 33, 3843–3849 (1999).
Gidaspow D. Multiphase Flow and Fluidization. Academic Press, San Diego (1994).
Goldman AJ, Cox RG, Brenner H. Slow viscous motion of a sphere parallel to a plane wall.
I. Motion through a quiescent fluid. Chemical Engineering Science 22(4), 637–651 (1967a).
Goldman AJ, Cox RG, Brenner H. Slow viscous motion of a sphere parallel to a plane wall.
II. Couette flow. Chemical Engineering Science 22(4), 653–660 (1967b).
González I, Hoffmann TL, Gallego JA. Visualization of hydrodynamic particle interactions:
validation of a numerical model. Acta Acustica 88, 19–26 (2002).
González-Gómez I, Hoffman TL, Gallego-Juárez JA. Theory and calculation of sound
induced particle interactions of viscous origin. Acta Acustica 86, 784–797 (2000).
Happel J, Brenner H. Low Reynolds Number Hydrodynamics. Martinus Nijhoff Publ., The
Hague, Netherlands, (1983).
Hasegawa T, Yosioka K. Acoustic-radiation force on a solid elastic sphere. Journal of the
Acoustical Society of America 46, 1139–1143 (1969).
Herron IH, Davis SH, Bretherton FP. On the sedimentation of a sphere in a centrifuge.
Journal of Fluid Mechanics 68(2), 209–234 (1975).
Hill RJ, Koch DL, Ladd AJC. The first effects of fluid inertia on flows in ordered and random
arrays of spheres. Journal of Fluid Mechanics 448, 213–241 (2001a).
Hill RJ, Koch DL, Ladd AJC. Moderate-Reynolds-number flows in ordered and random
arrays of spheres. Journal of Fluid Mechanics 448, 243–278 (2001b).
Hoffman TL, Koopmann GH. Visualization of acoustic particle interaction and agglomera-
tion: theory and experiments. Journal of the Acoustical Society of America 99(4), 2130–2141
(1996).
Hoffman TL, Koopmann GH. Visualization of acoustic particle interaction and agglomer-
ation: theory evaluation. Journal of the Acoustical Society of America 101(6), 3421–3429
(1997).
Huilin L, Gidaspow D, Bouillard J, Wentie L. Hydrodynamic simulation of gas-solid flow in a
riser using kinetic theory of granular flow. Chemical Engineering Journal 95, 1–13 (2003).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

178 Fluid Forces on Particles

Kaftori D, Hetsroni G, Banerjee S. Particle behavior in the turbulent boundary layer. I.

Motion, deposition and entrainment. Physics of Fluids 7(5), 1095–1106 (1995).
Kafui KD, Thornton C, Adams MJ. Discrete particle-continuum fluid modelling of gas-solid
fluidized beds. Chemical Engineering Science 57, 2395–2410 (2002).
Kapishnikov S, Kantsler V, Steinberg V. Continuous particle size separation and size sorting
using ultrasound in a microchannel. Journal of Statistical Mechanics: Theory and Experiment
P01012 (2006).
Keswani M, Raghavan S, Deymier P, Verhaverbeke S. Megasonic cleaning of wafers in
electrolytic solutions: possible role of electro-acoustic and cavitation effects. Microelectronic
Engineering 86, 132–139 (2009).
Kim W, Kim T-H, Choi J, Kim H-Y. Mechanism of particle removal by megasonic waves.
Applied Physics Letters 94, 081908 (2009).
King LV. On the acoustic radiation pressure on spheres. Proceedings of the Royal Society of
London A 147(861), 212–240 (1934).
King MR, Leighton DT, Measurement of the inertial lift on a moving sphere on contact with
a plane wall in a shear flow. Physics of Fluids 9(5), 1248–1255 (1997).
Kozeny J. Ueber kapillare Leitung des Wassers im Boden. Sitzber. Akad. Wiss. Wien 136,
271–306 (1927).
Krishnan GP, Leighton DT. Inertial lift on a moving sphere in contact with a plane wall in a
shear flow. Physics of Fluids 7(11), 2538–2545 (1995).
Kuehn TH, Kittelson TB, Wu Y, Gouk R. Particle removal from semiconductor wafers by
megasonic cleaning. Journal of Aerosol Science 47, S427–S428 (1996).
Kurose R, Komori S. Drag and lift forces on a rotating sphere in a linear shear flow. Journal
of Fluid Mechanics 384, 183–206 (1999).
Kuznetsova LA, Bazou D, Coakley WT. Stability of 2-D colloidal particle aggregates held
against flow stress in an ultrasound trap. Langmuir 23, 3009–3016 (2007).
Ladd AJC. Hydrodynamic transport coefficients for random dispersions of hard spheres.
Journal of Chemical Physics 93, 3484–3494 (1990).
Lamminen MO, Walker HW, Weavers LK. Mechanisms and factors influencing the ultrasonic
cleaning of particle-fouled ceramic membranes. Journal of Membrane Science 237, 213–223
(2004).
Laurell T, Petersson F, Nilsson A. Chip integrated strategies for acoustic separation and
manipulation of cells and particles. Chemical Society Reviews 36, 492–506 (2007).
Legendre D, Magnaudet J. The lift force on a spherical bubble in a viscous linear shear flow.
Journal of Fluid Mechanics 368, 81–126 (1998).
Leighton D, Acrivos A. The lift on a small sphere touching a plane in the presence of a simple
shear flow. Journal of Applied Mathematics and Physics (ZAMP) 36, 174–178 (1985).
Li S-Q, Marshall JS, Liu G, Yao Q. Adhesive particulate flow: The discrete element method
and its application in energy and environmental engineering. Progress in Energy and Com-
bustion Science 37(6), 633–668 (2011).
Li T, Kheifets S, Medellin D, Raizen MG. Measurement of the instantaneous velocity of a
Brownian particle. Science 328, 1673–1675 (2010).
Lin JC, Perry JH, Schowalter WR. Simple shear flow round a rigid sphere: Inertial effects and
suspension rheology. Journal of Fluid Mechanics 44, 1–17 (1970).
Liu J, Zhang G, Zhou J, Wang J, Zhao W, Cen K. Experimental study of acoustic agglom-
eration of coal-fired fly ash particles at low frequencies. Powder Technology 193(1), 20–25
(2009).
Lomholt S, Maxey MR. Force-coupling method for particulate two-phase flow: Stokes flow.
Journal of Computational Physics 184, 381–405 (2003).
Lovalenti PM, Brady JF. The hbydrodynamic force on a rigid particle undergoing arbitrary
time-dependent motion at small Reynolds number. Journal of Fluid Mechanics 256, 561–605
(1993).
Lucretius, “On The Nature of Things,” translated by WE Leonard, reprinted by Dover, New
York (2004).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

References 179

Magnaudet JM. The forces acting on bubbles and rigid particles. Proceedings of the Fluids
Engineering Division Annual Meeting, ASME (FEDSM97-3522), Vancouver, June 22–26
(1997).
McLaughlin JB. Inertial migration of a small sphere in linear shear flows. Journal of Fluid
Mechanics 224, 261–274 (1991).
Mei R. An approximate expression for the shear lift force on spherical particles at finite
particle Reynolds number. International Journal of Multiphase Flow 18(1), 145–147 (1992).
Mei R, Adrian RJ. Flow past a sphere with an oscillation in the free-stream velocity and
unsteady drag at finite Reynolds number. Journal of Fluid Mechanics 237, 323–341 (1992).
Mei R, Lawrence CJ, Adrian RJ. Unsteady drag on a sphere at finite Reynolds number with
small-amplitude fluctuations in the free-stream velocity. Journal of Fluid Mechanics 233,
613–631 (1991).
Metzger B, Nicolas M, Guazzelli E. Falling clouds of particles in viscous fluids. Journal of
Fluid Mechanics 580, 283–301 (2007).
Millikan RA. The general law of fall of small spherical body through a gas, and its bearing
upon the nature of molecular reflection from surfaces. Physical Review 22, 1–23 (1923).
Nam J, Lee Y, Shin S. Size-dependent microparticles separation through standing surface
acoustic waves. Microfluidics and Nanofluidics 11, 317–326 (2011).
Nitsche JM, Batchelor GK. Break-up of a falling drop containing dispersed particles. Journal
of Fluid Mechanics 340, 161–175 (1997).
Noh Y, Fernando HJS. The transition in the sedimentation pattern of a particle cloud. Physics
of Fluids A 5(12), 3049–3055 (1993).
Oesterlé B, Dinh TB. Experiments on the lift of a spinning sphere in a range of intermediate
Reynolds numbers. Experiments in Fluids 25, 16–22 (1998).
O’Neill ME, Stewartson K. On the slow motion of a sphere parallel to a nearby plane wall.
Journal of Fluid Mechanics 27(4), 705–724 (1967).
Oseen CW. Über die Stokes’sche Formel und über eine verwandte Aufgabe in der Hydrody-
namik. Arkiv för Matematik Astronomi och Fysik 6, 1–20 (1910).
Pignatel F, Nicolas M, Guazzelli E. A falling cloud of particles at a small but finite Reynolds
number. Journal of Fluid Mechanics 671, 34–51 (2011).
Pozrikidis C. Boundary Integral and Singularity Methods for Linearized Viscous Flows. Cam-
bridge University Press, Cambridge UK (1992).
Proudman I, Pearson JRA. Expansions at small Reynolds numbers for the flow past a sphere
and a circular cylinder. Journal of Fluid Mechanics 2(3), 237–262 (1957).
Richardson JF, Zaki WN. Sedimentation and fluidization. Part I. Transactions of the Institution
of Chemical Engineers 32, 35–53 (1954).
Rivero M, Magnaudet J, Fabre J. Quelques résultats nouveaux concernant les forces exercées
sur une inclusion sphérique par un écoulement accéléré. C.R. Acad. Sci. Paris Série II 312,
1499–1506 (1991).
Rubinow SI, Keller JB. The transverse force on a spinning sphere moving in a viscous fluid.
Journal of Fluid Mechanics 11, 447–459 (1961).
Saffman PG. On the motion of small spheroidal particles in a viscous liquid. Journal of Fluid
Mechanics 1, 540–553 (1956).
Saffman PG. The lift on a small sphere in a slow shear flow. Journal of Fluid Mechanics 22,
385–400 (1965).
Saffman PG. Corrigendum to “The lift force on a small sphere in a slow shear flow.” Journal
of Fluid Mechanics 31, 624 (1968).
de Sarabia ERF, Elvira-Segura L, González-Gómez I, Rodrı́guez-Maroto JJ, Muñoz-Bueno R,
Dorronsoro-Areal JL. Investigation of the influence of humidity on the ultrasonic agglom-
eration of submicron particles in diesel exhausts. Ultrasonics 41, 277–281 (2003).
Schiller L, Naumann A. Über die gundlegenden Berechungen bei der Schwerkraftaufbere-
itung. Zeitschrift des Vereines Deutscher Ingenieure 77, 318–320 (1933).
Schmitt K. Grundlegende Untersuchungen zum Thermalpräzipitator. Staub 19, 416–421
(1959).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

180 Fluid Forces on Particles

Segré G, Silberberg A. Behavior of macroscopic rigid spheres in Poiseuille flow. Journal of

Fluid Mechanics 14, 136–157 (1962).
Shaw DT. Acoustic agglomeration of aerosol. In Recent Developments in Aerosol Science, DT
Shaw, editor, John Wiley & Sons, New York, pp. 279–319 (1978).
Sheng C, Shen X. Modelling of acoustic agglomeration processes using the direct simulation
Monte Carlo method. Journal of Aerosol Science 37, 16–36 (2006).
Sierou A, Brady JF. Accelerated stokesian dynamics simulations. Journal of Fluid Mechanics
448, 115–146 (2001).
Smart JR, Beimfohr S, Leighton DT. Measurement of the translational and rotational veloci-
ties of a non-colloidal sphere rolling down a smooth inclined plane at low Reynolds number.
Physics of Fluids A 5(1), 13–24 (1993).
Smoluchowski M. Zur kinetischen Theorie der Brownschen Molekularbewegung und der
Suspensionen. Annalen der Physik 21, 756–780 (1906).
Sobanski MA, Tucker CR, Thomas NE, Coakley WT. Sub-micron particle manipulation in
an ultrasonic standing wave: applications in detection of clinically important biomolecules.
Bioseparation 9, 351–357 (2001).
Song L. Modeling of acoustic agglomeration of fine aerosol particles. Ph.D. dissertation,
Pennsylvania State University, College Station, Pennsylvania (1990).
Song L, Kooperman GH, Hoffman TL. An improved theoretical model of acoustic agglom-
eration. Journal of Vibration and Acoustics 116, 208–214 (1994).
Sridhar G, Katz J. Drag and lift forces on microscopic bubbles entrained by a vortex. Physics
of Fluids 7(2), 389–399 (1995).
Stewart BE, Thompson MC, Leweke T, Hourigan K. Numerical and experimental studies of
the rolling sphere wake. Journal of Fluid Mechanics 643, 137–162 (2010).
Subramanian G, Koch DL. Evoluation of clusters of sedimenting low-Reynolds-number par-
ticles with Oseen interactions. Journal of Fluid Mechanics 603, 63–100 (2008).
Sweeney LG, Finlay WH. Lift and drag forces on a sphere attached to a wall in a Blasius
boundary layer. Journal of Aerosol Science 38, 131–135 (2007).
Takemura F, Magnaudet J. The transverse force on clean and contaminated bubbles rising
near a vertical wall at moderate Reynolds number. Journal of Fluid Mechanics 495, 235–253
(2003).
Tanaka T, Yamagata K, Tsuji Y. Experiment on fluid forces on a rotating sphere and spheroid.
Proceedings of the 2nd KSME-JSME Fluids Engineering Conference, Vol. 1, p. 366 (1990).
Taneda S. Experimental investigation of the wake behind a sphere at low Reynolds numbers.
Journal of the Physical Society of Japan 11(10), 1104–1118 (1956).
Thomas PJ. On the influence of the Basset history force on the motion of a particle through
a fluid. Physics of Fluids A 4(9), 2090–2093 (1992).
Tiwary R, Reethof G. Hydrodynamic interaction of spherical aerosol particles in a high
intensity acoustic field. Journal of Sound and Vibration 108(1), 33–49 (1986).
Tri BD, Oesterle B, Deneu F. Premiers resulats sur la portance d’une sphere en rotation aux
nombres de Reynolds intermediaies. C.R. Acad. Sci. Ser. H: Mec. Phys. Chim. Sci. Terre
Univers 311, 27 (1990).
Tsuji Y, Morikawa Y, Mizuno O. Experimental measurement of the Magnus force on a
rotating sphere at low Reynolds numbers. Journal of Fluids Engineering 107, 484–488
(1985).
Van der Hoef MA, Beetstra R, Kuipers JAM. Lattice-Boltzmann simulations of low-
Reynolds-number flow past mono- and bidisperse arrays of spheres: Results for the perme-
ability and drag force. Journal of Fluid Mechanics 528, 233–254 (2005).
Vasseur P, Cox RG. The lateral migration of spherical particles sedimenting in a stagnant
bounded fluid. Journal of Fluid Mechanics 80(3), 561–591 (1977).
Wakiya S. Research report 9. Faculty of Engineering, Niigata University, Japan (1960).
Wen CY, Yu YH. Mechanics of fluidization. Chemical Engineering Progress Symposium Series
62(62), 100–111 (1966).
White FM. Viscous Fluid Flow. 3rd ed, McGraw-Hill, New York (2006).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-05 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:35

References 181

Zeng L, Balachandar S, Fischer P. Wall-induced forces on a rigid sphere at finite Reynolds

number. Journal of Fluid Mechanics 536, 1–25 (2005).
Zeng L, Najjar F, Balachandar S, Fischer P. Forces on a finite-sized particle located close to
a wall in a linear shear flow. Physics of Fluids 21, 0.33302 (2009).
Ziskind G, Fichman M, Gutfinger C. Effects of shear on particle motion near a surface:
application to resuspension. Journal of Aerosol Science 29(3), 323–338 (1998)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6 Particle Dispersion in Turbulent Flows

Turbulent particle dispersion is a key aspect of a large number of important pro-

cesses, including droplet and smog transport in the atmosphere, dust contamination
in electronics manufacturing, particulate combustion processes, and automotive and
truck exhaust. In some of these processes the particles have very little interaction
with each other, for which case our interest primarily focuses on modeling the time
variation of the bulk concentration field, measured on a scale much larger than
the turbulent flow eddies. On the other hand, certain particle transport problems
involve particles that do interact with each other. For instance, droplet collision and
merger in a cloud is a basic process that governs the size of droplets, and eventually
determines onset of precipitation. Collision of particles in a dust cloud can cause
electrons or ions to be stripped from one or the other particle, causing charging
of the two colliding particles via the mechanism of tribocharging. For instance, the
frequent large-scale dust storms on Mars are a critical component of the planetary
weather system, causing the fine dust particles that make up the Martian soil to
become highly charged. In both of these examples, the rate of particle tribocharging
and the rate of droplet merger are dependent on the collision rate of particles in a
turbulent flow. Such problems depend critically on the local particle concentration,
and are sensitive to variation of the concentration field on length scales on the order
of the turbulent eddies throughout the inertial range.
Development of agglomerates of adhesive particles in turbulent flows is sensitive
to the turbulent flow in two different ways. The inertia imparted on the particles by
the turbulent fluctuations leads to enhanced collision rate, and hence increases the
rate of agglomerate formation. On the other hand, the shear stress fluctuations
associated with the turbulent flow can tear apart the agglomerates if sufficiently
strong. Turbulent flow with adhesive particles is therefore particularly sensitive to
local fluctuations in particle concentration and to the distribution of energy with
length scale within the turbulent flow.

6.1. Particle Motion in Turbulent Flows

Because it is not possible to directly simulate all scales of most turbulent flows,
it is necessary to introduce some model for the fluctuating fluid velocity field of
a turbulent flow. For many engineering applications, the turbulent flow is modeled
182
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.1. Particle Motion in Turbulent Flows 183

using a Reynolds-averaged Navier-Stokes (RANS) model, which solves for the mean
velocity field and certain measures that characterize the turbulent fluctuations. For
instance, in the well-known k-ɛ model (Launder and Sharma, 1974), the spatial
variations of the turbulent kinetic energy (which we will denote by q) and the
dissipation rate ɛ are obtained from two additional partial differential equations.
From these two parameters, estimates of the Lagrangian integral length scale L and
time scale τL can be obtained as

L = u30 /2ε, τL = 4q/3C0 ε, (6.1.1)

where u0 is a characteristic velocity scale of the energy-containing motion that is
related to the turbulent kinetic energy q by u0 = (2q/3)1/2 and C0 is the Kolmogorov
constant with value C0 ∼ = 6.5. The turbulent eddies range in size from this integral
scale, associated with the most energetic eddies, to the Kolmogorov length scale
η = (ν 3 /ε)1/4 , where ν is the fluid kinematic viscosity, which is the scale associated
with energy dissipation. Eddy length scales significantly larger than the Kolmogorov
scale are said to lie in the inertial range, for which the fluid viscosity is not an important
parameter and the flow can be scaled as if it were inviscid. Throughout this inertial
range, the eddy velocity scale u can be related to its length scale by u ∼ (ε )1/3 .
While RANS models are still prevalent for applications, large eddy simulation
(LES) approaches are becoming more common, particularly for relatively simple
engineering flow problems. In LES, the larger scales of the flow are directly com-
puted, ranging from the length scale of the computational domain to a scale GRID
associated with the grid resolution, which ideally is selected such that GRID < L .
Particle transport by the resolved eddies in the LES computation can be directly
computed using the DEM method discussed in Chapter 2, with the LES velocity
field used to obtain the fluid velocity u in the force expressions in Chapter 5. Nev-
ertheless, there is a significant amount of particle transport at scales smaller than
GRID in most real turbulent flows, which cannot be simulated by the LES method. It
is thus necessary for both RANS and LES models to introduce an additional model
that describes necessary statistical aspects of the subgrid-scale turbulent fluctuations
required to obtain a reasonable prediction for particle transport by the unresolved
turbulence.
Particles are transported relative to the mean-flow streamlines in turbulent flows
by two very different mechanisms. The first mechanism (usually called particle dis-
persion) occurs when material regions of the particulate fluid are exchanged between
different turbulent eddies. This is the same process by which turbulence enhances
diffusion of heat, mass, and momentum, and it is well known to cause an enhance-
ment in the rate at which particles are spread outward and mixed by the fluid flow.
The second mechanism (called particle clustering) is related to the drift of particles
across streamlines of the instantaneous flow field in regions of high streamline cur-
vature. A principal cause of this drift in turbulent flows is the action of turbulent
eddies in excluding particles that are more dense than the surrounding fluid via cen-
trifugal force, causing the particles to cluster in the interstitial region between the
eddies (Bec et al., 2005, 2007; Falkovich and Pumir, 2004; Grits et al., 2006). The
equations governing this mechanism are discussed in Section 1.2.2. The local parti-
cle concentration can fluctuate via this mechanism by large amounts, with reports
of local concentration values as large as 25 times the mean concentration (Squires
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

184 Particle Dispersion in Turbulent Flows

0.003

vorticity
0.002 300000
280000
260000
240000
0.001 220000
200000
180000
160000
0
Y

140000
120000
100000
80000
–0.001 60000
40000
20000

–0.002

–0.003
–0.003 –0.002 –0.001 0 0.001 0.002 0.003
X
Figure 6.1. Small-scale clustering of particles in interstitial space between vortices in homo-
geneous turbulence. [Reprinted with permission from Garcia (2009).]

and Eaton, 1991). Figure 6.1 shows predictions of a DNS study of homogeneous
turbulence by Garcı́a (2009) in which heavy particles are thrown out of the turbu-
lent eddies and collect in high-concentration regions between the eddies. This figure
illustrates the antidiffusive characteristic of the particle clustering mechanism, which
leads to formation of regions with highly heterogeneous concentration values.
The nature of particle transport by a given turbulent eddy varies significantly
with the length scale and velocity scale u of the eddy. The effect of the eddy on
nearby particles can be characterized by the eddy Stokes number, defined by

St ≡ mu/3π μ d , (6.1.2)

where d and m are the particle diameter and mass, respectively, and μ is the fluid
viscosity. The eddy Stokes number, which is a function of eddy size, differs from
the more general Stokes number St for a turbulent flow, which is usually based on
the integral size and velocity scale. Using the inertial-range scaling u ∼ (ε )1/3 and
recalling that a basic characteristic of inertial range turbulence is that the dissipation
rate is independent of scale (Tennekes and Lumley, 1972), the eddy Stokes number
is found to vary with approximately as St ∼ −2/3 . Consequently, larger eddies
have smaller eddy Stokes numbers, and vice versa, for a given particle size.
Let us suppose that a critical value of eddy size crit exists within the inertial range
for which St ( crit ) = 1. For eddy sizes crit , the eddy Stokes number is small and
the particles are simply carried with the fluid flow. On the other hand, small-scale
fluctuations, with crit , do not have much influence on the particle motion, but
instead the force on the particles induced by these small velocity fluctuations is
filtered out by the particle inertia (Ayyalasomayajula et al., 2008). To illustrate these
limits, consider a simple velocity field that varies sinusoidally in time as

u = ux ex = U sin(ωt )ex , (6.1.3)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.2. Particle Drift Measure 185

where U is the fluid velocity scale and the angular frequency ω induced by an eddy is
set equal to the ratio /U of eddy size to eddy velocity scale. If the particle motion in
this example is controlled by the combination of the particle inertia and drag, then
dv
m = −3π μ d(v − u). (6.1.4)
dt
Nondimensionalizing velocity and time as vx = vx /U and t  = tω, (6.1.4) becomes

dvx 1  1

+ vx = sin(t  ), (6.1.5)
dt Stω Stω
where Stω is the Stokes number computed using (6.1.2) with = U/ω. Integrating
this equation in time and letting t  be sufficiently large that exp(−t  /Stω ) ∼
= 0, the
particle dimensionless velocity and acceleration are obtained as
sin(t  ) − Stω cos(t  ) cos(t  ) + Stω sin(t  )
vx (t  ) = , ap,x (t  ) = . (6.1.6)
1 + St2ω 1 + St2ω
For small Stokes numbers (Stω 1), the leading order solution is simply vx ∼
 
sin(t ) = ux , indicating that the particle moves with the local velocity field. Taking the
square of the velocity and acceleration in (6.1.6) and averaging over the oscillation
period gives
 
vx2 2
a p,x 1
=  = , (6.1.7)
ux 2 2
a f,x 1 + St2ω
where the brackets denote the average value over the oscillation period and ux =
sin(t  ) and af,x = cos(t  ) are the dimensionless fluid velocity and acceleration. The
result (6.1.7) shows that the kinetic energy of the particle per unit mass is damped
by a factor 1/St2ω for large Stokes numbers relative to that of the fluid.

6.2. Particle Drift Measure

It has long been recognized that fluid acceleration plays a key role in determination
of particle drift relative to the surrounding fluid flow. To understand why this is so,
we return to the example of a particle governed by its own inertia and the Stokes
drag force, for which the momentum equation is given by (6.1.4). A particle drift
velocity vd is defined by vd ≡ v − u, which from (6.1.4) is given by
m dv
vd = − . (6.2.1)
3π μ d dt
Nondimensionalizing velocity by the root-mean square fluctuation velocity scale
u0 and time by the ratio L /u0 , where L is the Lagrangian integral length scale
defined in (6.1.1), the drift velocity expression in (6.2.1) admits the dimensionless
form
dv
vd = −StL , (6.2.2)
dt 
where primes denote dimensionless variables and StL is the Stokes number based on
Lagrangian integral scaling. For small values of the Stokes number, it was discussed
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

186 Particle Dispersion in Turbulent Flows

in Section 1.2.4 that the particle nearly follows the fluid streamlines, with a drift
velocity vd that scales as O(StL u0 ). The particle acceleration term can be expanded
using the definition of the drift velocity as

dv Du dvd

= + + vd · ∇  u , (6.2.3)
dt  Dt  dt 
where D/Dt and ddt denote material time derivative following a fluid element and
following a particle, respectively. The first term on the right-hand side of (6.2.3)
is O(1), and the second and third terms are both O(StL ). Substituting (6.2.3) into
(6.2.2) and retaining only leading-order terms in the Stokes number, an approximate
equation for the particle drift velocity is obtained as

vd = −St a + O(St2 ), (6.2.4)

where a ≡ Du /Dt  is the dimensionless fluid acceleration. This result is the basis of
the “fast-Eulerian method” proposed by Ferry and Balachandar (2001).
The result (6.2.4) implies that although a particle at low Stokes number may
travel predominantly with the fluid flow, it also drifts relative to the fluid in a direction
opposite to the direction of the fluid acceleration vector and at a rate proportional to
the product of the Stokes number and the acceleration. An example illustrating this
result was discussed in Section 1.2.2 with respect to the problem of particles thrown
out of a vortex by the centrifugal force. Of course, the notion of centrifugal force
is tied to solution of the governing equations in a rotating coordinate frame. In an
inertial frame, as used in obtaining (6.2.4), the fluid rotating around the vortex has a
centripetal acceleration a pointing inwards toward the vortex center. According to
(6.2.4), the particle will drift in a direction opposite to the fluid acceleration, moving
the particle radially outward from the vortex center.
Because of the important role that acceleration plays on particle dispersion, the
statistics of the turbulence acceleration field is a matter of considerable interest.
The velocity field in a turbulent flow is well known to exhibit a nearly Gaussian
probability distribution; however, a wide range of experimental investigations (Voth
et al., 1998; La Porta et al., 2001; Mordant et al., 2004; Ayyalasomayajula et al.,
2006) and computational studies (Yeung et al., 2007; Biferale et al., 2004; Sawford
et al., 2003) have demonstrated that the acceleration field exhibits a so-called super-
statistical distribution characterized by fat tails (Reynolds, 2003b; Beck, 2008). Data
for the acceleration probability density function (PDF) from La Porta et al. (2001)
is shown in Figure 6.2, along with a Gaussian distribution having the same variance.
For accelerations more than one order of magnitude smaller than the most probable
acceleration, there is a strong deviation between these two distributions. A best-fit
curve to the data is also plotted, which gives the probability distribution function
P(a) as

P(a) = C exp[−a2 /(1 + |aβ/σ |γ σ 2 )], (6.2.5)

where for these data β = 0.539, γ = 1.588, and σ = 0.508. The constant C is selected
such that the integral over the probability distribution equals unity, which for these
data gives C = 0.786.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.2. Particle Drift Measure 187

100
Rλ = 200
10–1 Rλ = 690
Rλ = 970
Figure 6.2. Plot showing the probability
10–2
density function of the turbulence accel-

Probability
eration, comparing experimental mea- 10–3
surements of La Porta et al. (2001) at
different microscale Reynolds numbers
10–4
to a best fit curve (solid line) and to a
Gaussian distribution with the same vari- 10–5
ance (dashed line). [Reprinted with per-
mission from La Porta et al. (2001).] 10–6

–20 0 20
a / <a2>1/2

The relationship between the turbulence acceleration field and the turbulent
eddy structures has been investigated by Lee and Lee (2005), Reynolds et al. (2005),
and Mordant et al. (2004). These studies have demonstrated clearly that the magni-
tude of the turbulent acceleration field is dominated by the centripetal acceleration
surrounding coherent vortex structures. An example showing this dominant feature
of the acceleration field is given in Figure 6.3, from Lee and Lee (2005), in which
contours of the enstrophy field (Es = ω · ω/2) are plotted for a turbulent flow over
a cross-section of a coherent vortex structure. The pressure gradient vector is plotted

Figure 6.3. Plot from a direct numerical simulation of homogeneous turbulence, showing a
vortex structure detected by iso-surface of enstrophy (E = ω · ω/2) and enstrophy contour
lines in its cross-section. The pressure gradient is indicated by arrows, which for inviscid flow is
proportional to the acceleration vector. [Reprinted with permission from Lee and Lee (2005).]
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

188 Particle Dispersion in Turbulent Flows

using arrows in this figure, where it is recalled that for an inviscid flow the pressure
gradient is related to the acceleration by ρ a = −∇ p.
Sala and Marshall (2013) pointed out that particle motion relative to a fluid
streamline is related only to the component of the particle drift that is normal to
the fluid streamline. Using (6.2.4) and the unit vector along the streamline, given
by s ≡ u/ |u|, the rate of particle drift relative to the streamline is found to be
proportional to St a × s. Taking the ratio of the magnitude of this vector with the
fluid velocity magnitude gives a dimensionless particle drift measure ϕs as
|a × u |
ϕs (x, t ) = StL , (6.2.6)
|u |2
where the primes denote dimensionless variables.

6.3. Particle Collision Models

Many processes in which particles interact with each other start with particle collision.
Whether we are concerned with droplet merger, particle agglomeration, particle
sintering, or particle tribocharging, each of these processes is initiated by a particle
collision event. Consequently, the ability to accurately predict particle collision rate
in turbulent flows is of great importance.

6.3.1. Collision Mechanisms

As noted in Section 1.3 with reference to filtration applications, there are several
different physical mechanisms that give rise to particle collisions. A summary and
a classification of these different mechanisms were developed in the review article
by Meyer and Deglon (2011). Two of the most important mechanisms have already
been discussed in earlier sections of this book with respect to collisions between
particles and larger bodies. For instance, in the orthokinetic (shear) mechanism
illustrated in Figure 6.4a, the particles follow the fluid streamlines and collisions
occur due to the finite particle size as particles are carried past each other by the
local fluid shear. In the accelerative-correlated mechanism illustrated in Figure 6.4b,
particles are mostly advected by the local fluid flow, but drift of particles across fluid
streamlines in regions of strong streamline curvature leads to collisions with other
particles. The accelerative-independent mechanism illustrated in Figure 6.4c occurs at
large Stokes numbers, in which particles that receive impulse from distant turbulent
eddies can travel substantial distances before collision, and as a result at the collision
location the particle velocities are essentially uncorrelated with the local fluid velocity
field. The differential settling mechanism illustrated in Figure 6.4d occurs due to
the fact that larger, heavier particles have a higher terminal settling velocity than
smaller, lighter particles. In a flow dominated by particle settling, the heavier particles
can overtake the slower-moving lighter particles, leading to collisions. For very
small particles, with diameters less than about 1 µm, the perikinetic (or Brownian
motion) mechanism illustrated in Figure 6.4e can also be significant, in which particles
move in a zig-zagging manner under the random forces caused by collisions with
individual fluid molecules. Since the random paths of the particles are uncorrelated,
they sometimes cross or come sufficiently close to each other that the particles
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.3. Particle Collision Models 189

Fluid streamline

Time 1 Time 2
Time 1 Time 2 Time 1
(a)
(b)

Time 2

Time 1

Time 1
g
Time 2
Time 2

Time 1

(c) (d) (e)

Figure 6.4. Illustrations of five different collision mechanisms: (a) orthokinetic; (b)
accelerative-correlated; (c) accelerative-independent; (d) differential sedimentation; and
(e) perikinetic.

collide. A summary of these different collision mechanisms and of the regime where
each mechanism dominates is given in Table 6.1. Of particular interest in the current
chapter are the orthokinetic, accelerative-correlated, and accelerative-independent
mechanisms dealing with the effect of turbulence on particle collision rate.

Table 6.1. Summary of collision mechanisms and regimes where they are valid or dominant

Regime where valid or

Mechanism Description dominant

Perikinetic Particles collide under Brownian Small particle Peclet

motion due to impulse from collision number (Pe p < O(1))
with fluid molecules
Orthokinetic Particles follow fluid streamlines at Small Stokes number
centroid and collide due to finite size (StK 1)
Accelerative-correlated Particles mostly follow fluid Intermediate Stokes
streamlines but drift at curves due to number
finite inertia, leading to collisions
Accelerative-independent Particles collide after traveling several Large Stokes number
eddy length scales, with velocity that (StK 1)
has low correlation with local fluid
velocity
Differential sedimentation Particles of different sizes and masses Small Froude number
fall under gravity at different (Fr ≤ O(1))
terminal velocities, such that faster
particles run into slower particles
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

190 Particle Dispersion in Turbulent Flows

6.3.2. Orthokinetic Collisions (Small Stokes Numbers)

Early work on orthokinetic particle collision in laminar shear flow with shear rate
γ̇ was performed by Smoluchowski (1917), who found that the rate of collisions
per unit volume between a group of particles with radius r1 and a second group of
particles with radius r2 is given by

ṅc12 = 43 γ̇ n1 n2 (r1 + r2 )3 = α12 n1 n2 , (6.3.1)

where ni is the number of particles with radius ri per unit volume and α12 is called
the collision kernel. The rate of collisions of a set of particles with radius ri with itself
is obtained from the Smoluchowski theory as

ṅcii = 23 γ̇ n2i (2ri )3 , (6.3.2)

which differs from (6.3.1) by a factor of 12, which is necessary to avoid double-
counting collisions among like particles. Hu and Mei (1998) argue that (6.3.2) still
contains self-collisions, and that the proper expression should be

ṅcii = 23 γ̇ ni (ni − 1/V )(2ri )3 , (6.3.3)

where V is the volume of the particulate flow domain. For large volumes and large
numbers of particles, niV 1 and (6.3.2) reduces to (6.3.1).
Saffman and Turner (1956) derived a well-known expression for orthokinetic
particle collision rate in turbulent flows. The derivation made a number of signifi-
cant simplifications, including: (1) the particles are neutrally buoyant spheres with
diameter much smaller than the turbulence Kolmogorov length scale η = (ν 3 /ε)1/4 ,
(2) the particle response time is small compared with the Kolmogorov time scale, (3)
the turbulence is isotropic and has a large Reynolds number, (4) there are no hydro-
dynamic interactions between the particles, (5) the concentration field is uniform,
(6) Stokes drag is the only force acting on the particles, and (7) the gradient of the
fluctuating fluid velocity is normally distributed. The second of these assumptions
implies that the Stokes number StK based on the Kolmogorov length and velocity
scaling, defined by

StK ≡ muK /3π μdη, (6.3.4)

must be small compared with unity. In this equation, the Kolmogorov velocity scale
is defined by uK = (νε)1/4 and ε is the turbulence dissipation rate per unit mass.
The argument put forth by Saffman and Turner examines transport over a
spherical surface S with radius σ ≡ r1 + r2 , centered at the centroid of particle 1, as
shown in Figure 6.5. If wr denotes the radial component of the velocity of particle 2
relative to particle 1 and x is a point on the surface S, then collision occurs for all x
values for which wr (x) < 0. The collision kernel is given by
+ ,
α12 = − wr da , (6.3.5)
S(wr <0)

where the brackets · denote the expected value, or mean. Assuming that the
particles travel at the same velocity as the surrounding fluid, so that wr can be
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.3. Particle Collision Models 191

S σ
Figure 6.5. Schematic showing definition of the sphere sur-
r1
face S of radius σ = r1 + r2 surrounding particle 1, along with x1
streamlines of a local straining flow relative to the particle.

interpreted as the fluid velocity relative to particle 1, the continuity equation for an
incompressible flow requires that
 
wr da + wr da = 0. (6.3.6)
S(wr <0) S(wr >0)

From this requirement, it follows that

 
1
− wr da = |w | da, (6.3.7)
S(wr <0) 2 S r

where the integration on the right-hand side is over the entire sphere. Substituting
(6.3.7) into (6.3.5) gives

1
α12 = |wr | da = 2π (r1 + r2 )2 |wr |. (6.3.8)
2 S
From the assumption that the turbulence is isotropic, so that wx2  = wy2  = wz2 , it
follows that
 
|wr | = sin2 θ cos2 ϕ wx2 + sin2 θ sin2 ϕ wy2 + cos2 θ wz2  = |wx |. (6.3.9)

Saffman and Turner assume that the particle is much smaller than the Kol-
mogorov scale and use a Taylor series to write wx = σ (∂u/∂x), where u is the x-
component of the turbulence fluctuation velocity. The usual scaling relationship
for inertial-range isotropic turbulence gives the mean-square velocity gradient as
(∂u/∂x)2  = ε/15ν (Tennekes and Lumley, 1972, 66–67). We recall that if X is a
normally distributed random variable with zero mean, then |X | has a half-normal
distribution with mean value given by |X | = (2/π )1/2 X 2 1/2 . Applying this result
to the velocity gradient, the expressions given earlier can be combined to write
|wr | = σ (2ε/15πν)1/2 , so that (6.3.8) becomes

α12 = (8π /15)1/2 (r1 + r2 )3 (ε/ν)1/2 . (6.3.10)

This expression is similar to an earlier expression by Camp and Stein (1943)

based on the same assumptions, in which the collision kernel was obtained as

α12 = 43 (r1 + r2 )3 (Etot /ν)1/2 . (6.3.11)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

192 Particle Dispersion in Turbulent Flows

Using the generalized version of this expression proposed by Pedocchi and

Piedra-Cueva (2005) (see also Clark, 1985), we define Etot as the total energy dis-
sipation rate per unit mass, which can be associated with the mean-flow energy
dissipation rate Em = 2ν D̄ : D̄ and the turbulence energy dissipation rate ε by

Etot = Em + ε. (6.3.12)

Here, D̄ = (1/2)[∇ ū + (∇ ū)T ] is the rate of deformation tensor for the mean velocity
field ū and D̄ : D̄ denotes the scalar product of D̄ with itself. The traditional overbar
is used here for the mean velocity field and related quantities in the turbulent flow
for brevity. For turbulence with zero mean flow, (6.3.11) is similar to the Saffman-
Turner expression, with a difference in the coefficient value of about 3%. For a
laminar shear flow of the form ū = γ̇ y ex , (6.3.11) reduces to the Smoluchowski
expression (6.3.1).

6.3.3. Accelerative-Independent Collisions (Large Stokes Numbers)

In the opposite extreme of a flow with high values of StK , an expression was pro-
posed by Abrahamson (1975) in which particles move with random, independent
motions as they are flung about by distant eddies of the turbulent flow. Because the
particle Stokes number is large, the particles travel long distances with a velocity
that is relatively uncorrelated to the local fluid velocity. Because of the intended
application to large Stokes number particles with independent motion, the rela-
tive radial velocity in the Abrahamson expression is estimated not using a Taylor
expansion from the local fluid velocity, as done by Saffman and Turner, but
instead using the mean-square velocity for each particle. Again using the expres-
sion |X | = (2/π )1/2 X 2 1/2 for a normally distributed random variable X, we can
write |wr | = (2/π )1/2 [v12  + v22 ] 1/2 , where v1 and v2 are velocity components of
particles 1 and 2, respectively, in one direction (say, the x-direction). The turbu-
lence is assumed to be isotropic, so the mean-square velocity magnitude for particle
1 is given by v1m
2
 = 3 v12 , and similarly for the second particle. Substituting this
approximation for the mean absolute value of the radial velocity into (6.3.8) gives
the Abrahamson collision kernel as

α12 = (8π /3)1/2 (r1 + r2 )2 v1m 2 + v2m
2 . (6.3.13)

6.3.4. Accelerative-Correlated Collisions (Intermediate Stokes Numbers)

Many researchers have proposed expressions for particle collision rate at intermedi-
ate Stokes numbers, which is dominated by the accelerative-correlated mechanism
associated with gradual particle drift across the fluid streamlines. Early work on this
problem for small particles is reported in the second part of the paper by Saffman
and Turner (1956), followed by work from Williams and Crane (1983), Yuu (1984),
Kruis and Kusters (1997), and others. The essential problem can be expressed by
writing the collision kernel in the form

α12 = (8π /3)1/2 (r1 + r2 )2 wshear
2 + waccel
2 , (6.3.14)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.3. Particle Collision Models 193

1.1

Particle kinetic energy

1200
1.0 Abrahamson

Collision frequency
1000
0.9
800
0.8

0.7 600
DNS Simulation
0.6 400

0.5 200 Saffman-Turner

0 2 4 6 8 0 2 4 6 8 10
StK StK
(a) (b)
Figure 6.6. Plots showing DNS results for variation of (a) particle kinetic energy and (b)
collision frequency as a function of Stokes number based on Kolmogorov scaling. The latter
plot also compares DNS results to predictions of the Saffman-Turner theory (6.3.10) and
of the Abrahamson kinetic theory (6.3.13). [Reprinted with permission from Sundaram and
Collins (1997).]

where wshear is the relative velocity between the particles due to the fluid
shear flow and waccel is the relative
 velocity due to particle dispersion. Setting
waccel = 0 and wshear = (r1 + r2 ) ε/5ν reduces to the Saffman-Turner result (6.3.10)
for orthokinetic
 collisions at small Stokes numbers, and setting wshear = 0 and
waccel = v1m 2  + v 2  reduces to the Abrahamson result (6.3.13) for accelerative-
2m
independent collisions at high Stokes numbers. A wide range of expressions have
been derived giving estimates of wshear and waccel at intermediate values of the
Stokes number, many of which are summarized by Meyer and Deglon (2011). These
expressions, many of which are fairly complex, differ from each other primarily by
the choice of particle-fluid interaction forces that are included in the model and the
assumptions made about the particle size relative to turbulence length scales.
A key challenge in collision rate modeling at intermediate values of the Stokes
number arises from the need to account for the occurrence of particle clustering in
the region between turbulent eddies, as discussed in the first section of this chapter.
This tendency of the particles to cluster is referred to as “preferential concentration”
in the literature, as statistical models of particle motion in the turbulent flow must
account for the fact that particles have a higher probability of being located within
these interstitial regions than within the eddy structures. In order to provide data for
turbulent particle collision rate and to examine the validity of various approxima-
tions, direct numerical simulations have been performed by Sundaram and Collins
(1997), Wang et al. (1998), and Zhou et al. (1998) for homogeneous turbulence
with a single particle size and by Zhou et al. (2001) for a mixture of particles of
two different sizes. Simulation results from Sundaram and Collins (1997) are shown
in Figure 6.6a for particle kinetic energy and collision rate as a function of Stokes
number StK . The particle kinetic energy decreases with increase in Stokes number,
which is consistent with the filtering effect of the particle inertia as discussed in
Section 6.1. The collision frequency obtained from direct numerical simulation is
observed in Figure 6.6b to be higher than the value predicted by the Saffman-Turner
expression (6.3.10) for orthokinetic collision and lower than the value predicted by
the Abrahamson kinetic theory approximation (6.3.13) for accelerative-independent
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

194 Particle Dispersion in Turbulent Flows

r
Preferential concentration index, D
0.12

0.10

0.08 Figure 6.7. Plot showing variation of the

preferential concentration index Dr as
0.06 a function of Stokes number based on
Kolmogorov scaling. [Reprinted with per-
0.04 mission from Sundaram and Collins
(1997).]
0.02

0
0 2 4 6 8
St K

collisions. However, at low Stokes numbers the DNS results decrease substantially,
moving toward the Saffman-Turner expression predictions, and at high Stokes num-
bers the DNS results appear to approach the Abrahamson expression predictions,
although in both cases the Stokes number becomes neither high enough nor low
enough for the results to converge to these limits.
Sundaram and Collins (1997) developed a measure of the extent of particle clus-
tering by examining the statistics of the probability distribution P(m) of the number
of particles, denoted by m, located within a grid cell at a given time. The particle
number distribution is obtained by examining all grid cells within the flow domain
at a given time. For a randomly distributed system of noninteracting particles, the
variable m exhibits a Poisson distribution Pr (m) = Mm e−M /m!, where M ≡ m is
the expected value of the random variable m, given by the ratio of the total number
of particles to the total number of grid cells. A measure Dr of particle preferen-
tial concentration is defined based on the difference between the actual distribu-
tion of m, denoted by P(m), and the Poisson distribution for randomly distributed
particles as

Dr ≡ [P(m) − Pr (m)]2 , m = 0, 1, . . . . . (6.3.15)
m

A plot showing DNS results for Dr as a function of Stokes number is given

in Figure 6.7. The preferential concentration measure goes to zero for both very
large and very small Stokes numbers, and attains a maximum value at StK = 1.
This observation is consistent with previous results of Wang and Maxey (1993) and
Squires and Eaton (1991).
Jin et al. (2010) reported results of a comparison between a large eddy simulation
(LES) of particle collision rate in homogeneous turbulence and a full direct numerical
simulation of the same problem. Results are also compared to a filtered set of the
direct numerical simulation data in order to examine separately whether differences
are due to spatial resolution or other errors associated with LES. This study concludes
that the LES and filtered DNS approaches provide reasonably accurate predictions of
the particle collision rate when StK > 3, but that these approaches exhibit significant
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.4. Dynamic Models for Particle Dispersion 195

error for small and intermediate values of StK . The error in collision rate prediction
was as much as 60% for StK = 0.5. Of course, the accuracy of LES predictions in
general will depend on the filter size (or, equivalently, the grid size) relative to the
Kolmogorov length scale for the specific problem under examination. The reason
for the difference between LES and DNS predictions at moderate and low values of
Stokes number was attributed by Jin et al. (2010) to the development of preferential
concentration (or clustering) of particles by the small-scale eddies.

6.4. Dynamic Models for Particle Dispersion

In the context of a simulation using the discrete-element method, diffusion, disper-
sion, and collision of particles are directly computed by solving the equations of
motion of each particle given an appropriate fluid flow. The DEM approach may
be used together with a direct numerical simulation of turbulence, in which case
no additional assumptions are necessary, beyond those typically used to obtain the
various forces on the particles in DEM (as discussed in Chapters 3–5). In particular,
with DNS we do not need to introduce any model for the turbulent fluctuation field,
at least for a computation employing the one-way interaction assumption, because
it is directly computed. On the other hand, other turbulence simulation approaches,
such as LES and RANS, do not simulate all of the scales of the turbulent motion,
but instead introduce “closure models” for the effect of these unresolved scales on
quantities such as stress and diffusion rate in the larger scales of the turbulent flow.
Because these unresolved scales may also have a significant influence on the particle
transport and on particle collisions, it is necessary to introduce some model of the
unresolved scales of turbulent motion that can be added to the resolved motion for
use in transporting the particulate phase. This model of the unresolved turbulent
motion is not expected to be accurate in time and space (as this would be the same
as a DNS), but rather it is expected to possess the necessary statistical properties
so as to produce the same statistical measures of the particle dynamics as the actual
turbulence. In this and the next section, we shall simply refer to this model of the
unresolved turbulence scales as a subgrid-scale fluctuation model.
The choice of an appropriate subgrid-scale model is dependent on exactly which
aspect of the particle transport the user desires to simulate. In particular, two very
different categories of problems might be delineated – those in which the particle
dynamics depend on interaction with other particles (type 1) and those in which it
does not (type 2). In the latter (type 2) case, the focus in model development is to
introduce subgrid-scale fluctuations that have the correct kinetic energy and time
scale as seen by individual traveling particles. In the former (type 1) case, it is also
essential that the fluctuations seen by two nearby particles are correlated in such
a manner that both particles see similar fluctuations. This requirement of spatial
correlation of fluctuations of nearby particles is of key importance, for instance, in
prediction of particle collision and adhesion processes. Put another way, the subgrid-
scale fluctuation field for particle transport problems of the first type must produce
the phenomenon of particle preferential concentration (due to clustering of par-
ticles in the interstitial spaces between the eddies), whereas this is not essential
for problems of the second type. As a consequence, whereas the turbulent fluc-
tuations for problems of the second type must have the correct energy and time
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

196 Particle Dispersion in Turbulent Flows

scales, fluctuations for problems of the first type must also have the correct spatial
structure.
The current section deals with problems of the second type, in which particle
interaction is not important, and specifically with a class of models that are called
stochastic Lagrangian models (SLMs). These are very powerful and useful models for
addressing problems such as particle diffusion, and other problems where the kinetic
energy and time scales of individual particles are of central importance. The most
common SLM approach uses a stochastic differential equation of the Langevin form
to approximate the fluctuating fluid velocity field of a turbulent flow. The properties
of the Langevin equation for modeling turbulent flow fluctuations were examined
in detail by Thomson (1987), and the resulting SLM is sometimes referred to as the
Thomson model. In this model, the governing equation for one component of the
Lagrangian fluctuation velocity u+ (t ) is approximated by solution of the stochastic
differential equation
  1/2
+ u+ 2u 2
du = − dt + dξ , (6.4.1)
τL τL

where τL is the Lagrangian integral time scale given in (6.1.1) and the differential
dξ is a Gaussian-distributed random variable with zero mean and variance equal
to the time step dt. The solution of (6.4.1) for u+i (t ) has zero mean and variance
+ 2 
ui (t )  = ui , where u is the root-mean-square of the Eulerian fluctuation velocity.
2

A component u+ (t ) of the Lagrangian velocity has zero mean and variance

equal to 2q/3, where q is the turbulent kinetic energy per unit mass. Two basic
statistical measures of the time variation of the Lagrangian velocity field are the
autocorrelation function ρ(s) and the second-order structure function D(s), defined
by

ρ(s) ≡ u+ (t + s) u+ (t )/u+2 , D(s) ≡ |u+ (t + s) − u+ (t )|2 . (6.4.2)

The structure function is the variance of the velocity change over a time interval of
length s, and the
! ∞ autocorrelation function is related to the Lagrangian integral time
scale as τL = 0 ρ(s) ds. These two measures are related to each other by

D(s) = 2u2 [1 − ρ(s)]. (6.4.3)

The Langevin equation (6.4.1) can be used to derive analytic solutions for these
two measures as (Pope, 2000)

ρ(s) = exp(−s/τL ), D(s) = C0 ετL [1 − exp(−s/τL )] , (6.4.4)

where for convenience we take s ≥ 0. The result for ρ(s) agrees well with results
from direct numerical simulation of homogeneous turbulence by Yeung and Pope
(1989). The solution for D(s) follows from (6.4.3) after using the expression (6.1.1)
for τL . For time increments s τL , (6.4.4) reduces to the asymptotic approximation

D(s) ∼ C0 sε, (6.4.5)

in agreement with Kolmogorov’s similarity hypothesis (Kolmogorov, 1962; Biferale

et al., 2008).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.4. Dynamic Models for Particle Dispersion 197

Pope (2002) notes three shortcomings of the Langevin model (6.4.1): (1) it
contains a single time scale τL ; (2) it has no Reynolds number dependence; and (3) the
slope of ρ(s) is discontinuous at s = 0. These shortcomings are particularly apparent
when one compares the prediction (6.4.4) for the second-order structure function
with data from direct numerical simulation. Whereas the ratio D(s)/εs obtained
from the prediction (6.4.4) is independent of Reynolds number, values obtained
from the DNS data of Yeung et al. (2006) for isotropic turbulence exhibit strong
variation with the microscacle Reynolds number Reλ = u λ/ν, where λ = 15ν u2 /ε
is the Taylor microscale (Pope, 2011). These shortcomings in the Langevin stochastic
Lagrangian model can be addressed by solving a stochastic differential equation for
the fluctuating acceleration field a+ (t ), instead of velocity, and then solving for
velocity using du+ = a+ . This approach was first demonstrated by Sawford (1991),
whose acceleration-based stochastic Lagrangian model can be written in a form
suggested by Pope (1992) as
     12
+ η a+ u+ 2 1 1
da = − 1 + S dt − dt + 2a + dξ , (6.4.6)
τS ηS τS ηS ηS τS

where a2 = u2 /τc ηc is the square of the root-mean-square acceleration fluctuation
and the time scales τS = 4τL /Ĉ0 and ηS = Ĉ0 τη /2a0 are related to the integral time
scale τL and Kolmogorov time scale τη = (ν/ε)1/2 , respectively. The coefficient a0
is obtained by an empirical fit to DNS data as a0 = 0.13Re0.64 λ . The coefficient Ĉ0
is also obtained by a fit to DNS data, giving Ĉ0 = 7. The use of two time scales in
the Sawford model introduces a Reynolds number dependency associated with the
ratio of these time scales, addressing the first two of the shortcomings listed by Pope
(2002).
Another statistical measure of accuracy of the subgrid-scale fluctuation model is
its ability to predict the correct accleration probability density function in turbulent
flows, as shown based on experimental data in Figure 6.2. The Sawford model yields
a probability density function with Gaussian form, which is inconsistent with the
fat-tail form of the experimentally observed function. Modified acceleration-based
stochastic Lagrangian models, developed by Reynolds (2003a) and by Lamorgese
et al. (2007), are designed to yield the correct form of the acceleration probability
density function. Reynolds proposed a stochastic differential equation for both the
acceleration fluctuation a+ and the local turbulence dissipation rate ε+ , which is
defined such that ε+  = ε is the total dissipation rate per unit mass. He assumed
that the logarithm of the normalized dissipation rate χε ≡ ln(ε+ /ε) can be modeled
by an Uhlenbeck-Ornstein process of the form
  12
dt 2σχ2
dχε = −(χε − χε ) + dW (6.4.7)
τχ τχ

where dW is a Gaussian random variable with zero mean and variance dt. Here, χε is
Gaussian-distributed random variable that is normalized such that exp(χε ) is unity,
so its mean is given by χε  = −σχ2 /2. The time scale τχ is given by τχ = 2σu2 /C0 ε and
the variance is fit from DNS data in terms of the microscale Reynolds number Reλ
as σχ2 = −0.354 + 0.289 ln(Reλ ). Reynolds solves (6.4.7) together with a stochastic
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

198 Particle Dispersion in Turbulent Flows

100

10-1

Figure 6.8. Probability density function for the

PDF

10-2
x-component of acceleration comparing DNS
data (A) with SLM predictions using the
10-3
A, D Langevin equation (B), the Sawford model
(C), and the Reynolds model (D). [Reprinted
with permission from Sala (2012).]
10-4
B, C

10-5
0 5 10 15
ax/<ax2>1/2

differential equation for the acceleration fluctuation of the form

 
1 1 1 dσ u+
da+ = − a+ dt −
a|ε
+ − dt
τS ηS σa|ε dt τS ηS
(  
1 1 1
+ 2σu2 + dξ , (6.4.8)
τS ηS τS ηS

where the time scales τS and ηS have the same meaning as defined earlier with
reference to Sawford’s model. The variance of velocity and acceleration are given
√
by σu2 = 2q/3 and σa2 = a∗0 ε ε/ν, respectively. The constant a∗0 = 3.3 is obtained by
fitting to the DNS data of Yeung and Pope (1989). The conditional acceleration
variance σa/ε is given by

σa|ε = a∗0 (ε/ < ε >)3/2 a2η , (6.4.9)

where aη = (< ε >3 /ν)1/4 is the Kolmogorov acceleration scale. The differential
dξ is also a Gaussian random variable with zero mean and variance dt, but it is
independent of the variable dW . A plot is shown in Figure 6.8 comparing direct
numerical simulation results for the acceleration probability density function with
predictions from the Langevin equation (6.4.1), the Sawford model (6.4.6), and the
Reynolds model (6.4.7), (6.4.8), and (6.4.9). The Reynolds model predictions are
found to be in excellent agreement with the DNS data.
The use of stochastic Lagrangian models for solving for diffusion of heavy par-
ticles in a turbulent flow was examined by Sawford and Guest (1991), among many
later investigators, by coupling the stochastic differential equation (6.4.1) for the
fluid velocity fluctuations to the particle momentum equation. These authors par-
ticularly examined the ability of the SLM approach for dealing with the so-called
inertia and crossing-trajectories effects discussed by Csanady (1963). The inertia
effect is associated with the observation that particle inertia has the effect of filtering
out fluctuations with high eddy Stokes number, as discussed in Section 6.1. As a con-
sequence, the energy spectrum of the particle will differ from that of the surrounding
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.5. Dynamic Models for Particle Clustering 199

-5/3
k 100
10-2

10-1
10-3
E(k)

PDF
DNS
-2
10
10
-4
SLM

-5
SVS 10
-3 SLM
10

10-6 10
-4
DNS

10 20 30 40 50 60 0 10 20

k φs /<φs2>1/2
(a) (b)
Figure 6.9. Comparison of (a) the energy spectrum and (b) the probability density function of the
dispersion measure ϕs /ϕs2 1/2 between direct numerical simulation (DNS) for homogeneous turbulence
and the Reynolds stochastic Lagrangian method (SLM). [Figure (b) reprinted with permission from
Sala (2012).]

turbulent fluid. The crossing-trajectories effect is associated with the fact that particle
drift across fluid streamlines, as discussed in Section 6.2, causes the time scale that
a particle remains in a given turbulent eddy to be smaller than the time scale of a
typical fluid element within the eddy. Whereas the fluid element time scale is asso-
ciated with decay of the eddy or exchange of fluid between eddies, the particle time
scale has the additional factor that particles drift relative to the surrounding fluid.
This particle drift might occur either as a result of the particle’s own momentum or
from imposition of an external force on the particle, such as gravity.

6.5. Dynamic Models for Particle Clustering

The discussion of stochastic Lagrangian models in the previous section focused on
the statistics of fluid element motion in time under this model, using measures such
as the temporal second-order structure function defined in (6.4.2). Different levels
of stochastic models were introduced, each of which is found to accurately predict
progressively more detailed temporal stochastic measures in comparison to direct
numerical simulation studies. On the other hand, there was no discussion of the
spatial structure of the turbulence generated by these stochastic Lagrangian models.
In general, SLM predictions do not possess the correct spatial structure because the
random forcing terms used for the different particles are not spatially correlated. As
a consequence, the velocity field generated by stochastic Lagrangian models, when
interpolated back to a grid, is found to consist of eddies with a size close to the grid
scale.
For instance, a comparison between the energy spectrum of a direct numeri-
cal simulation for forced homogeneous turbulence and a solution of the Reynolds
stochastic Lagrangian model is shown in Figure 6.9a. In this calculation, the stochastic
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

200 Particle Dispersion in Turbulent Flows

Lagrangian model (6.4.7)–(6.4.9) was solved on a set of Lagrangian fluid elements

and then interpolated onto a set of fixed grid nodes using the conservative scat-
tered data interpolation method of Monaghan (1985). The DNS results exhibit
the expected E(k) ∝ k−5/3 wavenumber scaling in the inertial range, and more
rapid energy decay in the dissipation range. The results of the Reynolds stochas-
tic Lagrangian model exhibit entirely the opposite trend, with increasing energy
as the grid cell size becomes smaller. Comparison of velocity fields shows that the
velocity generated by the Reynolds model consists of random fluctuations with char-
acteristic size approximately equal to the grid cell size. Other stochastic Lagrangian
models have also been examined and found to produce similar results.
The lack of correct spatial structure in stochastic Lagrangian models has obvious
significance for processes in which particles interact with each other in the turbulent
flow, as processes such as particle collision are dependent on correct modeling of
particle relative velocity. However, spatial structure also has a significant impact on
the particle dispersion itself. For instance, it is argued in Section 6.2 that particle
drift across fluid streamlines can be measured by the ϕs parameter defined in (6.2.6),
which is associated with the cross product of the acceleration and velocity fields.
While the Reynolds stochastic Lagrangian method has been shown to agree well
with the probability density functions for both the fluid velocity and acceleration
obtained from direct numerical simulations (as shown, for instance, in Figure 6.8),
it does not follow that this method also agrees well with the statistics of the a × u
field. A plot of the probability density function of the dispersion measure ϕs /ϕs2 1/2 is
given in Figure 6.9b, comparing results from a direct numerical simulation with those
of the Reynolds stochastic Lagrangian method. From the definition of ϕs in (6.2.6),
this ratio is independent of Stokes number and entirely depends upon turbulent
fluid velocity field. The Reynolds model exhibits a superstatistical distribution with
a fat tail for this dispersion measure similar to that for the acceleration field. By
contrast, the DNS result in Figure 6.9b decays much more rapidly, and is closer in
nature to a Gaussian distribution. This observation seems to suggest that the part
of the acceleration field that gives rise to the fat-tailed acceleration distribution
does not have an impact on particle dispersion, as characterized by drift of particles
across fluid streamlines. The major reason for the difference in the distribution of
ϕs in Figure 6.9b is the fact that in the direct numerical simulations, regions of high
acceleration magnitude and regions of high values of the ϕs measure both tend
to be located in regions immediately surrounding coherent turbulent vortices, as
illustrated in Figure 6.3, whereas in a stochastic Lagrangian model these quantities
are randomly distributed across the field with no coherent structure.
The difficulty with use of the stochastic Lagrangian method for modeling tur-
bulent dispersion and clustering of interacting particles was noted by Ayyalaso-
mayajula et al. (2008), who emphasized the importance of spatial structure in the
turbulence model. Ayyalasomayajula et al. made a first step toward development
of a structure-based approach by proposing a very simple two-dimensional vortex
model for simulating particle dispersion. The model positioned the vortex structures
on a two-dimensional Cartesian array, where each vortex structure was a line vortex
of infinite length. The strength of the vortex structures was varied in a stochastic
manner in accordance to the Lagrangian integral time scale τL , and the spacing
between the vortices was set to be equal to the integral length scale L . Even though
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

6.5. Dynamic Models for Particle Clustering 201

this two-dimensional model is quite simple, Ayyalasomayajula et al. (2008) found

that it exhibits several desirable features, such as prediction of the fat-tailed type of
probability density function for acceleration observed experimentally by La Porta
et al. (2001) and others, as shown in Figure 6.2.
This concept of developing a stochastic model based on vortex structures that
generate the turbulent fluctuations was developed further in recent work by Sala
(2012) and Sala and Marshall (2013) for homogeneous turbulence. These authors
proposed a stochastic vortex structure (SVS) model, in which the turbulent fluctu-
ating vorticity field is approximated by a set of three-dimensional vortex structures
that have certain features set based on integral-scale length and time scales and
other features selected as random variables with a prescribed probability distribu-
tion function. The total number of vortex structures in the computational domain, N,
is set so that the average distance between the vortex centers is proportional to the
Lagrangian integral length scale L . The vortex structures have finite length, which
is also proportional to L , and a core radius that is proportional to the Kolmogorov
length scale η = (v 3 /ε)1/4 . Each vortex structure has a finite life span that is set equal
to the integral time scale τL . The maximum vortex strength max is proportional
to the ratio 2L /τL . The constants of proportionality for these scaling parameters
are specified so as to yield the prescribed kinetic energy; however, sensitivity tests
indicate that the various flow measures are not sensitive to different combinations
of these constants giving the same kinetic energy.
The flow field is initiated for this model by placing N vortex structures in the com-
putational domain, where the vortex centroid position and orientation are selected
randomly using a uniform probability distribution. For homogeneous turbulence the
position of each vortex structure is fixed in time, because there is no mean flow. The
initial vortex structures are each assigned an initial age τ0n , chosen as a random
variable for which the ratio τ0n /τL has a uniform distribution between 0 and 1. The
age τn (t ) of the nth vortex structure increases with time as

τn = τ0n + t − t0n , (6.5.1)

where t0n is the time at which the vortex structure is introduced. When τn ≥ τL ,
the nth vortex is removed from the flow field and a new vortex is introduced with
t0n = t and τ0n = 0. This new vortex is again located with centroid position and vortex
orientation selected as random variables with uniform probability distribution. The
vortex strength n (t ) is specified as
⎧
⎨5τn /τL
⎪ for 0 ≤ τn /τL < 0.2
n (t ) = max An 1 for 0.2 ≤ τn /τL ≤ 0.8 (6.5.2)
⎪
⎩
1 − 5(τn /τL − 0.8) for 0.8 < τn /τL ≤ 1

so that the strength ramps up gradually near the beginning of the vortex life, remains
constant throughout the middle part of the vortex life, and ramps back down near the
end of the vortex life. The coefficient An is a normally distributed random variable
with zero mean and unit variance, which is set once for each vortex at the time that
the vortex is initialized.
Computational results comparing scaling measures for the SVS velocity field
with DNS are given in Figure 6.10. In Figure 6.10a, the energy spectrum predicted
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

202 Particle Dispersion in Turbulent Flows

-5/3
k
10-2

10-3
E(k)

DNS
-4
10 SLM

-5
SVS
10

10-6

10 20 30 40 50 60

k
(a) (b)
Figure 6.10. (a) Comparison of the energy spectrum between direct numerical simulation
(DNS) and the stochastic vortex structure (SVS) model. (b) The probability density function
of the dispersion measure ϕs /ϕs2 1/2 between direct numerical simulation (DNS) for homo-
geneous turbulence and the stochastic vortex structure method. [Figure (b) reprinted with
permission from Sala (2012).]

by SVS is observed to be very close to DNS for wavenumbers within the inertial
range, for which the standard E ∼ k−5/3 scaling is observed. For high wavenumbers
in the dissipation range, the SVS energy spectrum falls off more rapidly than DNS,
which is expected, as the SVS model is based on integral scaling. The SVS model
agrees well with DNS results for probability density function of the acceleration
fluctuations, as shown by Sala (2012). The method also generates predictions for
the dispersion measure ϕs /ϕs2 1/2 that agree reasonably well with DNS data, as
shown in Figure 6.10b. Investigation of the flow field indicates that the regions with
largest magnitude of this dispersion measure are located in a ring surrounding the
large-scale vortex structures.
Figure 6.11 plots the number of particle collisions as a function of time for a DEM
computation of particle transport in homogeneous turbulence with a Stokes number

106 SLM

DNS
105
Figure 6.11. Comparison of the prediction for
number of particle collisions as a function of
104 SVS dimensionless time from direct numerical sim-
ulation (DNS) for homogeneous turbulence
Nc 10 3
in a (2π )3 domain, the stochastic Lagrangian
model (SLM) of Reynolds (2003), and the
102
stochastic vortex structure (SVS) model, for
a computation with Stokes number StL = 0.5
based on integral-scale variables. [Reprinted
101
from Sala and Marshall (2013).]

100
0 2 4 6 8 10
t / τL
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

References 203

StL = 0.5 based on the integral time and length scales. The plot compares results
obtained using direct numerical simulation, the stochastic vortex structure model
of Sala (2012) and Sala and Marshall (2013), and the stochastic Lagrangian model
of Reynolds (2003a). The turbulent flow was allowed to attain a quasi-equilibrium
state in a preliminary computation with no particles, and then the particles were
added and the flow computation was restarted with the same initial conditions and
values of computational and flow parameters for the three computations. The data
in Figure 6.11 indicate that the Reynolds SLM approach predicts nearly an order of
magnitude more particle collisions by the end of the computation than DNS, whereas
the SVS model predictions are very close to the DNS results. This plot is presented
in a semi-log form in order to more clearly denote the differences between the
different models; however, the computed results for number of collisions increase
nearly linearly in time. Although the models used by Ayyalasomayajula et al. (2008)
and Sala and Marshall (2013) were applied to homogeneous turbulence and have not
yet been tested over a large range of particle transport and flow conditions, the basic
approach employed in the stochastic vortex structure model appears to have promise
for future development in treating dispersion problems with interacting particles.

REFERENCES

Abrahamson J. Collision rates of small particles in a vigorously turbulent fluid. Chemical

Engineering Science 30, 1371–1379 (1975).
Ayyalasomayajula S, Gylfason A, Collins LR, Bodenschatz E, Warhaft Z. Lagrangian
measurements of inertial particle accelerations in grid generated wind tunnel turbulence.
Physical Review Letters 97, 144507 (2006).
Ayyalasomayajula S, Warhaft Z, Collins LR. Modeling inertial particle acceleration statistics
in isotropic turbulence. Physics of Fluids 20, 095104 (2008).
Bec J, Celani A, Cencini M, Musacchio S. Clustering and collisions of heavy particles in
random smooth flows. Physics of Fluids 17, 073301 (2005).
Bec J, Biferale L, Cencini M, Lanotte A, Musacchio S, Toschi F. Heavy particle concen-
tration in turbulence at dissipative and inertial scales. Physical Review Letters 98, 084502
(2007).
Beck C. Superstatistics: Theoretical concepts and physical applications. In Anomalous
Transport: Foundations and Applications, R. Klages, G. Radons, I. M. Sokolov, editors,
Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim, Germany, 2008.
Biferale L, Boffetta G, Celani A, Devenish BJ, Lanotte A, Toschi F. Multifractal statistics of
Lagrangian velocity and acceleration in turbulence. Physical Review Letters 93(6), 064502
(2004).
Biferale L, Bodenschatz E, Cencini M, Lanotte AS, Ouellette NT, Toschi F, Xu H. Lagrangian
structure functions in turbulence: A quantitative comparison between experiment and direct
numerical simulation. Physics of Fluids 20, 065103 (2008).
Camp TR, Stein P. Velocity gradients and internal work in fluid motion. Boston Society of
Civil Engineers 30, 219–237 (1943).
Clark MM. Critique of Camp and Stein’s RMS velocity gradient. Journal of Environmental
Engineering 111(3), 741–754 (1985).
Csanady GT. Turbulent diffusion of heavy particles in the atmosphere. Journal of the Atmo-
spheric Sciences 20, 201–208 (1963).
Falkovich G, Pumir A. Intermittent distribution of heavy particles in a turbulent flow. Physics
of Fluids 16(7), L47–L50 (2004).
Ferry J, Balachandar S. A fast Eulerian method for disperse two-phase flow. International
Journal of Multiphase Flow 27, 1199–1226 (2001).
Garcı́a M. Développement et validation du formalisme Euler-Lagrange dans un solveur par-
allèle et non-structuré pour la simulation aux grandes échelles – TH/CFD/09/1. PhD thesis,
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

204 Particle Dispersion in Turbulent Flows

Université de Toulouse – Ecole doctorale : Mécanique, Energétique, Génie civil, Procédés

(2009).
Germano M, Piomelli U, Moin P, Cabot WH. A dynamic subgrid-scale eddy viscosity model.
Physics of Fluids A 3(7), 1760–1765 (1991).
Grits B, Pinsky M, Khain A. Investigation of small-scale droplet concentration inhomo-
geneities in a turbulent flow. Meteorol Atmos Phys 92, 191–204 (2006).
Hu KC, Mei R. Particle collision rate in fluid flows. Physics of Fluids 10(4), 1028–1030
(1998).
Jin G, He G-W, Wang L-P. Large-eddy simulation of turbulent collision of heavy particles in
isotropic turbulence. Physics of Fluids 22, 055106 (2010).
Kolmogorov AN. A refinement of previous hypotheses concerning the local structure of
turbulence in a viscous incompressible fluid at high Reynolds number. Journal of Fluid
Mechanics 13, 82–85 (1962).
Kruis FE, Kusters KA. The collision rate of particles in turbulent flow. Chemical Engineering
Communications 158, 210–230 (1997).
La Porta A, Voth GA, Crawford AM, Alexander J, Bodenschatz E. Fluid particle acceleration
in fully developed turbulence. Nature 409, 1017–1019 (2001).
Lamorgese AG, Pope SB, Yeung PK, Sawford BL. A conditionally cubic-Gaussian stochastic
Lagrangian model for acceleration in isotropic turbulence. Journal of Fluid Mechanics 582,
423–448 (2007).
Launder BE, Sharma BI. Application of the energy dissipation model of turbulence to the
calculation of flow near a spinning disc. Letters in Heat and Mass Transfer 1(2), 131–138
(1974).
Lee S, Lee C. Intermittency of acceleration in isotropic turbulence. Physical Review E 71(5
Pt 2), 056310 (2005).
Meyer CJ, Deglon DA. Particle collision modeling: A review. Minerals Engineering 24, 719–
730 (2011).
Monaghan JJ. Extrapolating B splines for interpolation. Journal of Computational Physics
60(2), 253–262 (1985).
Mordant N, Crawford AM, Bodenschatz E. Three-dimensional structure of the Lagrangian
acceleration in turbulent flows. Physical Review Letters 93(21), 214501 (2004).
Pedocchi F, Piedra-Cueva I. Camp and Stein’s velocity gradient formalization. Journal of
Environmental Engineering 131(10), 1369–1376 (2005).
Pope SB. A stochastic-Lagrangian method for acceleration in turbulent flows. Physics of
Fluids 14(7), 2360–2375 (1992).
Pope SB. Turbulent Flows. Cambridge University Press, Cambridge (2000).
Pope SB. A stochastic Lagrangian model for acceleration in turbulent flows. Physics of Fluids
14(7), 2360–2375 (2002).
Pope SB. Simple models of turbulent flows. Physics of Fluids 23, 011301 (2011).
Reynolds AM. On the application of nonextensive statistics to Lagrangian turbulence. Physics
of Fluids 15(1), L1–L4 (2003a).
Reynolds AM. Superstatistical mechanics of tracer-particle motions in turbulence. Physical
Review Letters 91(8), 084503 (2003b).
Reynolds AM, Mordant N, Crawford AM, Bodenschatz E. On the distribution of Lagrangian
accelerations in turbulent flows. New Journal of Physics 7, 58 (2005).
Saffman PG, Turner JS. On the collision of drops in turbulent clouds. Journal of Fluid
Mechanics 1, 16–30 (1956).
Sala K. Analysis of stochastic methods for predicting particle dispersion in turbulent flows.
MS Thesis, University of Vermont, Burlington, Vermont. (2012).
Sala K, Marshall JS. Stochastic vortex structure method for modeling particle clustering and
collisions in homogeneous turbulence. Physics of Fluids 25(10), 103301 (2013).
Sawford B. Reynolds number effects in Lagrangian stochastic models of turbulent dispersion.
Physics of Fluids 3(6), 1577–1586 (1991).
Sawford BL, Guest FM. Lagrangian statistical simulation of the turbulent motion of heavy
particles. Boundary-Layer Meteorology 54, 147–166 (1991).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-06 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 10:55

References 205

Sawford BL, Yeung PK, Borgas MS, Vedula P, La Porta A, Crawford AM, Bodenschatz
E. Conditional and unconditional acceleration statistics in turbulence. Physics of Fluids
15(11), 3478–3489 (2003).
Smoluchowski M von. Versuch einer mathematischen Theorie der Koagulationkinetik
kollider lösungen. Z. Phys. Chem. 92, 129–168 (1917).
Squires KD, Eaton JK. Preferential concentration of particles by turbulence. Physics of Fluids
A 3, 1169–1178 (1991).
Sundaram S, Collins LR. Collision statistics in an isotropic particle-laden turbulent suspension.
I. Direct numerical simulations. Journal of Fluid Mechanics 335, 75–109 (1997).
Tennekes H, Lumley JL. A First Course in Turbulence. MIT Press, Cambridge, MA (1972).
Thomson DJ. Criteria for the selection of stochastic models of particle trajectories in turbulent
flows. Journal of Fluid Mechanics 180, 529–556 (1987).
Voth GA, Satyanarayan K, Bodenschatz E. Lagrangian acceleration measurements at large
Reynolds numbers. Physics of Fluids 10(9), 2268–2280 (1998).
Wang L-P, Maxey MR. Settling velocity and concentration distribution of heavy particles in
homogeneous isotropic turbulence. Journal of Fluid Mechanics 256, 27–68 (1993).
Wang L-P, Wexler AS, Zhou Y. On the collision rate of small particles in isotropic turbulence.
I Zero-inertia case. Physics of Fluids 10(1), 266–276 (1998).
Williams JJE, Crane RI. Particle collision rate in turbulent flow. International Journal of
Multiphase Flow 9(4), 421–435 (1983).
Yeung PK, Pope SB. Lagrangian statistics from direct numerical simulations of isotropic
turbulence. Journal of Fluid Mechanics 207, 531–586 (1989).
Yeung PK, Pope SB, Sawford BL. Reynolds number dependence of Lagrangian statistics in
large numerical simulations of isotropic turbulence. Journal of Turbulence 7, N58 (2006).
Yeung PK, Pope SB, Kurth EA, Lamorgese AG. Lagrangian conditional statistics, acceler-
ation and local relative motion in numerically simulated isotropic turbulence. Journal of
Fluid Mechanics 582, 399–422 (2007).
Yuu S. Collision rate of small particles in a homogeneous and isotropic turbulence. AICHE
Journal 30(5), 802–807 (1984).
Zhou Y, Wexler AS, Wang L-P. On the collision rate of small particles in isotropic turbulence.
II. Finite inertia case. Physics of Fluids 10(5), 1206–1216 (1998).
Zhou Y, Wexler AS, Wang L-P. Modelling turbulent collision of bidisperse inertial particles.
Journal of Fluid Mechanics 433, 77–104 (2001).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7 Ellipsoidal Particles

Although the assumption that particles are spherical is frequently a useful approxi-
mation, there are many applications in which the deviation of the particulate matter
from a spherical shape plays a major role in the system dynamics. For instance, red
blood cells have the shape of biconcave disks with a flattened center. This shape
allows blood to easily flow within blood vessels with concentrations of 40–45% by
volume without jamming. Fuel particles are often of a nonspherical shape, partic-
ularly for biowaste combustion processes (Džiugys and Peters, 2001). Soil particles
are often highly irregular in shape, particularly for soils formed of minerals with
anisotropic crystal structure or for soils in dry regions where there is not much water
erosion (Knuth et al., 2012). Key work on modeling of aerosol systems formed of
nonspherical particles is reported by Gallily and Cohen (1979) and Fan and Ahmadi
(2000). Such models can be used for aerosol dynamics problems such as inhalation
of small asbestos fibers suspended in air, or modeling of snow fall or dispersion of
mineral or soot aerosols. Liquid crystal phase transitions can be modeled as bifurca-
tions in a mixture of rod- and platelike particles (Camp and Allen, 1996). Solis and
Martin (2010) have recently reported that a solution of metallic platelets subjected
to a biaxial oscillating magnetic field can exhibit an amazing variety of transitions
as a function of the forcing frequency, including formation of vortex arrays that
produce a flow field reminiscent of natural convection flows, only without the need
for a temperature difference.
In cases such as those discussed here, the nonspherical particles can be rea-
sonably approximated as ellipsoids. Indeed, in a large number of cases two of the
ellipsoid axes will be approximately equal and the particles can be further approx-
imated as spheroids. Hence, cylindrical particles of grass biofuel or short carbon
nanotubes in suspension might be treated as high aspect ratio prolate spheroids (Yin
et al., 2003), whereas red blood cells or platelets might be treated as oblate spheroids
(Chesnutt and Marshall, 2009). Although inexact, this idealization allows us to intro-
duce important features of nonspherical particles into the flow computation, such
as the higher packing limits, shape effects on particle rolling, particle alignment
both within the flow and within agglomerates (Chesnutt and Marshall, 2010), and
effects on force chain transmission of stress through particle collisions (Campbell,
2011).

206
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.1 Particle Dynamics 207

^^ Particle
^y y ^x

Figure 7.1. Coordinate frames for transport of

^^x
an ellipsoidal particle: global, particle, and co-
y Co-moving
moving frames. ^^z ^z

z x
Global

7.1. Particle Dynamics

In describing the motion of ellipsoidal particles, it is useful to employ different refer-
ence frames for handling different aspects of the dynamics of the system. Specifically,
we introduce a global inertial frame x = [x y z], a particle frame x̂ = [x̂ ŷ ẑ],
and a co-moving frame x̂ˆ = [ x̂ˆ ŷˆ ẑˆ ], all of which are assumed for convenience
to be Cartesian frames. The origin of both the co-moving and particle frames is
the particle centroid. The axes of the co-moving frame are parallel to the corre-
sponding axes of the inertial frame, and hence do not move in time. The axes of
the particle frame are fixed to the particle, and therefore rotate relative to the iner-
tial frame as the particle rotates. These three reference frames are illustrated in
Figure 7.1.
The equation of the surface of an ellipsoidal particle in the particle frame is

x̂2 ŷ2 ẑ2

+ + − 1 = 0, (7.1.1)
a2E b2E c2E

where aE , bE , and cE are semiaxis lengths in the x̂, ŷ, and ẑ directions, respectively.
Since the co-moving and particle frames are both Cartesian and have the same origin,
they can be mapped into each other using a coordinate rotation of the form
ˆ
x̂ = Ax̂, (7.1.2)

where A is an orthonormal matrix (such that the transpose of A is equal to its

inverse). The rotation of the particle frames in three dimensions can be decomposed
into a sequence of three separate rotations, each about a single angle. These angles,
denoted by ϕ, θ, and ψ, are called the Euler angles, and their definition is illustrated
in Figure 7.2. The intersection of the co-moving x̂ˆ − ŷˆ plane and the rotating x̂ − ŷ
plane is called the line of nodes. The particle frame can then be rotated back onto the
co-moving frame by a series of three steps. The first step is a rotation of the particle
frame (x̂, ŷ, ẑ) by an angle ψ about the ẑ axis, so that at the end of the rotation the x̂
axis lies on the line of nodes. The second step is a rotation of the particle frame by
an angle θ about the new x̂ axis, so that at the end of the rotation the ẑ and ẑˆ axes
are coincident. The third step involves a rotation of the particle frame by an angle ϕ
about the new location of the ẑ axis, at the end of which all the axes of the particle
frame will be coincident with those of the co-moving frame.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

208 Ellipsoidal Particles

^^
z

^z y^
θ Figure 7.2. Illustration showing the co-moving
and particle coordinate systems and the associ-
^^ ated Euler angles ϕ, θ, and ψ. The intersection
^x y of the x̂ˆ − ŷˆ plane and the x̂ − ŷ plane is the dot-
ted line called the line of nodes.
^^ ψ
x φ

Line of nodes

The rotation matrix A can be expressed in terms of the Euler angles as

⎡ ⎤
cos ψ cos ϕ − cos θ sin θ sin ψ −sin ψ cos θ − cos θ sin ϕ cos ψ sin θ sin ϕ
A = ⎣cos ψ sin ϕ + cos θ cos ϕ sin ψ −sin ψ sin ϕ + cos θ cos ϕ cos ψ −sin θ cos ϕ ⎦.
sin θ sin ψ sin θ cos ψ cos θ
(7.1.3)

However, it is often easier to express A in terms of a set of four Euler parameters

(or quaternions) ε1 , ε2 , ε3 , and η, defined by
ϕ−ψ θ ϕ−ψ θ
ε1 = cos sin , ε2 = sin sin ,
2 2 2 2
(7.1.4)
ϕ+ψ θ ϕ+ψ θ
ε3 = sin cos , η = cos cos ,
2 2 2 2
where by trigonometric identity
ε12 + ε22 + ε32 + η2 = 1. (7.1.5)
The rotation matrix A can be written in terms of the Euler parameters as
⎡   ⎤
1 − 2 ε22 + ε32 2(ε1 ε2 + ε3 η) 2(ε1 ε3 − ε2 η)
⎢   ⎥
A = ⎣2(ε2 ε1 − ε3 η) 1 − 2 ε32 + ε12 2(ε2 ε3 + ε1 η) ⎦ . (7.1.6)
 2 
2(ε3 ε1 + ε2 η) 2(ε3 ε2 − ε1 η) 1 − 2 ε1 + ε2 2

The rotation rate  of a particle can be written in terms of its components in the
particle frame as
 = x̂ ex̂ + ŷ eŷ + ẑ eẑ . (7.1.7)
The rate of change of the Euler parameters can be expressed in terms of the com-
ponents of  in (7.1.7) as (Hughes, 1986, 26)
⎡ ⎤ ⎡ ⎤
dε1 /dt ηx̂ − ε3 ŷ + ε2 ẑ
⎢ ⎥ ⎢ ⎥
⎢ dε2 /dt ⎥ 1 ⎢ ε3 x̂ + ηŷ − ε1 ẑ ⎥
⎢ ⎥= ⎢ ⎥
⎢ dε /dt ⎥ 2 ⎢ −ε  + ε  + η ⎥ . (7.1.8)
⎣ 3 ⎦ ⎣ 2 x̂ 1 ŷ ẑ ⎦
dη/dt −ε1 x̂ − ε2 ŷ − ε3 ẑ
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.2 Fluid Forces 209

The equations of motion for a particle consist of the linear momentum equation
for the particle velocity v and the angular momentum equation for the rotation rate
. However, because it is much easier to write the equations for the torques acting
on the particle in the particle coordinate frame, we utilize a component form of the
angular momentum equation expressed in the rotating particle coordinate frame.
The resulting equations of motion are given by
dv
m= FF + FA , (7.1.9)
dt
dx̂  
Ix̂ − ŷ ˆ Iŷ − Iˆ = MF,x̂ + MA,x̂ , (7.1.10a)
dt
dŷ  
Iŷ − ˆ x̂ Iˆ − Ix̂ = MF,ŷ + MA,ŷ , (7.1.10b)
dt
dˆ  
Iˆ − ˆ ŷ Iˆ − Iŷ = MF,ẑ + MA,ẑ , (7.1.10c)
dt
where the mass moments of inertia of an ellipsoidal particle in the particle frame
are Ix̂ = 15 m(b2E + c2E ), Iŷ = 15 m(a2E + c2E ), and Iẑ = 15 m(a2E + b2E ). The moments in
(7.1.10) are written in terms of the components in the particle coordinate frame.

7.2. Fluid Forces

In this section we examine the fluid forces and torques acting on an isolated ellip-
soidal particle traveling relative to an external fluid flow at low particle Reynolds
number, ReP ≡ ρ f |v − u|d/μ < O(1), where d = 2(aE bE cE )1/3 is the effective par-
ticle diameter, ρ f is the density of the fluid, u is the undisturbed fluid velocity at the
particle centroid, and μ is the fluid viscosity. An expression for the hydrodynamic
drag of an ellipsoidal particle was derived by Oberbeck (1876) for a uniform Stokes
flow and by Brenner (1964) for an arbitrary external Stokes flow. Following Happel
and Brenner (1963, 220–222) and retaining only the first-order term in the fluid
velocity (neglecting derivatives of the fluid velocity), we can write
Fd = μ K̂ · (u − v), (7.2.1)
where K̂ is the particle frame translation tensor. The particle-frame translation tensor
for an ellipsoid is a diagonal matrix given by (see Gallily and Cohen, 1979)
 
ex̂ ⊗ ex̂ eŷ ⊗ eŷ eẑ ⊗ eẑ
K̂ = 16π aE bE cE + + , (7.2.2)
χ0 + a2E α0 χ0 + b2E β0 χ0 + c2E γ0
where ex̂ , eŷ , and eẑ are unit vectors in the particle coordinate system and ⊗ denotes
the tensor (or dyadic) product. The coefficients α0 , β0 , γ0 , and χ0 are given by
 ∞  ∞
dλ dλ
χ0 = aE bE cE , α0 = aE bE cE  2  ,
0  0 aE + λ 
 ∞  ∞ (7.2.3)
dλ dλ
β0 = aE bE cE  2  , γ0 = aE bE cE  2  ,
0 bE + λ  0 cE + λ 
where
   1 2
= a2E + λ b2E + λ c2E + λ / . (7.2.4)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

210 Ellipsoidal Particles

An expression for the torque acting on an ellipsoidal particle in the particle

frame at small particle Reynolds numbers was derived by Jeffery (1922) and can be
expressed in terms of the coefficients in (7.2.3) as
16π μ aE bE cE  2    
MF,x̂ =  2  bE − c2E dẑŷ + b2E + c2E (wẑŷ − x̂ ) , (7.2.5a)
3 bE β0 + cE γ0
2

16π μ aE bE cE  2   
MF,ŷ =  2  [ cE − a2E dx̂ẑ + c2E + a2E (wx̂ẑ − ŷ )], (7.2.5b)
3 cE γ0 + aE α0
2

16π μ aE bE cE  2   
MF,ẑ =  2  [ aE − b2E dŷx̂ + a2E + b2E (wŷx̂ − ẑ )], (7.2.5c)
3 aE α0 + bE β0
2

where
   
1 ∂uî ∂u jˆ 1 ∂uî ∂u jˆ
dî jˆ ≡ + , wî jˆ ≡ −
2 ∂x jˆ ∂xî 2 ∂x jˆ ∂xî
are components of the rate of deformation tensor and the vorticity tensor, respec-
tively, in the particle frame. The first term in the brackets in (7.2.5) is associated
with rotation of the ellipsoidal particles by the straining of the external flow, and the
second term is associated with particle rotation caused by the difference between
the particle rotation rate and the rotation rate of the external fluid.
The coefficients in (7.2.3) can be further simplified for the case where the parti-
cles are approximated as being spheroidal in shape, for which two axes of the ellipsoid
have the same length. For definiteness, we assume that bE = cE and that the x̂ coor-
dinate axis is coincident with the particle symmetry axis. The particle aspect ratio
is defined by β ≡ aE /bE , so that prolate and oblate spheroids correspond to cases
with β > 1 and β < 1, respectively. We further define a dimensionless coefficient
χ̄0 ≡ χ0 /b2E . Evaluation of the integrals in (7.2.3) for spheroidal particles gives
α0 = 2(1 − β0 ), β0 = γ0 = (−β 2 + χ̄0 /2)/(1 − β 2 ), (7.2.6)
where the coefficient χ̄0 is given by
 
β β − (β 2 − 1)1/2
χ̄0 = − ln for β > 1 (7.2.7a)
(β 2 − 1)1/2 β + (β 2 − 1)1/2
and
  
2β π β
χ̄0 = − tan−1 for β< 1. (7.2.7b)
(1 − β 2 )1/2 2 (1 − β 2 )1/2

Expressions for a variety of other forces acting on ellipsoidal particles, as well as

corrections to the drag force, have been derived in the literature. For instance, a slip
correction factor for ellipsoidal particles at nonzero Knudsen numbers was derived
by Dahneke (1973). A summary and evaluation of correlations for modification of
the drag force to account for finite Reynolds numbers was given by Chhabra et al.
(1999). An extension of Saffman’s expression for lift force on a particle in a shear
flow was developed by Harper and Chang (1968) for a three-dimensional body of
arbitrary shape. Fan and Ahmadi (2000) gave expressions for the Brownian force on
an ellipsoidal particle, and they also discussed related literature on slip correction
factors for a rotating ellipsoidal particle.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.3 Collision Detection and Contact Point Identification 211

7.3. Collision Detection and Contact Point Identification

One of the most challenging aspects of simulating motion of ellipsoidal particles is
the problem of contact detection. A contact detection algorithm must perform three
separate tasks:

(1) Determine when two particles collide with each other,

(2) Determine the contact points on each particle, along with the associated surface
unit normal vectors at the contact points, and
(3) Determine the amount of overlap between the particles.

For spherical particles these tasks are trivial, but for ellipsoidal particles they
are far from simple. Moreover, it is important that the contact detection algorithm
performs these tasks in a manner that is accurate, stable, and efficient. Accuracy
is vital to ensure that all contacts are identified as soon as they occur and no false
contacts are reported. If the contact detection algorithm misses a collision when it first
occurs and then identifies it at a later time, the amount of overlap can become very
large and the resulting rebound velocities can exceed the impact velocity, leading to
instability in the numerical calculation. Correct identification of the contact point and
unit normal is important to ensure that momentum is conserved during the collision.
Computational efficiency of the contact detection algorithm is essential because
contact detection typically takes up the largest fraction of the overall computational
time for calculations with nonspherical particles. DEM simulations with ellipsoidal
particles can take an order of magnitude, or more, longer than simulations with the
same number of spherical particles, and this difference is entirely controlled by the
efficiency of the contact detection scheme.
Because contact detection is so time consuming, it is worthwhile to first reduce
the number of particle pairs that must be inspected by using simple, inexpensive
tests to see if it is possible for contact to occur. It is convenient to retain a short
list for each particle of other particles that are sufficiently close to it that a collision
may have occurred during the time step. This “collision list” might contain 20 or so
particles that are close to a given particle, and collisions of the given particle are
only considered with particles contained in this list. In a second test, the distance
i j = |xi − x j | between a given particle i and each particle j in its collision list is
determined. The pair interaction is sorted into one of the following categories:

Case A: i j > max(aEi + aE j , bEi + bE j ), (7.3.1)

Case B: i j < min(aEi + aE j , bEi + bE j ), (7.3.2)

Case C: max(aEi + aE j , bEi + bE j ) ≥ i j ≥ min(aEi + aE j , bEi + bE j ). (7.3.3)

In Case A, the collision cannot occur. In Case B, the collision definitely occurs
and it remains to identify the collision point and amount of particle overlap. In
Case C, it is possible that the collision may have occurred, and so we must investigate
further using the full contact detection algorithm.
Numerous algorithms have been proposed in the literature for contact detection
of ellipsoidal particles, as well as for the related two-dimensional problem of contact
of elliptical particles. A summary follows of some of the major approaches used for
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

212 Ellipsoidal Particles

Ellipse 1 Figure 7.3. Illustration of the contact detection algo-

B rithm of Ting (1992) for two-dimensional elliptical par-
ticles. The intersection points of the ellipses are denoted
A and B, and the midpoint of the line segment AB is the
C contact point C.
A
Ellipse 2

this purpose. The discussion is based on the review by Lin and Ng (1995), updated
to include more recent results.

7.3.1. Two-Dimensional Algorithms

Many of the early contact detection algorithms were developed for two-dimensional
problems involving colliding elliptical particles, but do not extend well to ellipsoidal
particles in three-dimensions. For instance, Ting (1992) proposed a very simple
contact detection algorithm illustrated in Figure 7.3, in which the equations for the
surface of two ellipses are solved simultaneously to find the (x,y) positions of the two
intersection points, labeled A and B. If there is no real-valued solution of this system
of equations then the ellipses are not in contact. If a real-valued solution exists, then
the midpoint C of the line segment AB is chosen as the contact point. This method
requires finding the roots of a fourth-degree polynomial for each particle pair at each
time step. It suffers from a high degree of ill-conditioning that can lead to problems
with round-off errors if the amount of overlap of the ellipses is small. The biggest
fault of this method, however, is the fact that it is quite complicated to extend to
three dimensions.
In an alternative approach proposed for arbitrarily shaped two-dimensional
particles by Hogue and Newland (1994), each particle is represented in polar form
where the particle surface is discretized by a set of line segments and the particle
interior is composed of the union of a set of triangles, as shown in Figure 7.4. In order
to check for particle collision, each vertex point of a neighboring particle is examined
to see if it lies within a triangle of the given particle. Determination of whether a
point P is inside a triangle T is performed by drawing lines from the point to each
vertex of the triangle, as illustrated for point P in Figure 7.4. Three subtriangles

Figure 7.4. Triangularization of the particle using the

polar representation proposed in the contact detection
P algorithm of Hogue and Newland (1994).
1
2 3
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.3 Collision Detection and Contact Point Identification 213

Ellipsoid 1
Figure 7.5. Illustration of contact point iden- A
tification using the common normal method.
δΝ
B

Ellipsoid 2

are formed by drawing lines from the point P to two vertices of the triangle T, as
indicated by the dashed lines in Figure 7.4. If the sum of the area of these three
subtriangles is greater than the total area of triangle T, then P is outside of triangle
T. If the sum of the subtriangle areas is equal to the total area of triangle T, then P is
inside T or lying on the surface of T. Although it is possible to extend this scheme to
a tetrahedral discretization of three-dimensional particles, the method has a number
of drawbacks that make it undesirable to do so for particles with smoothly curving
surfaces, such as ellipsoidal particles. Specifically, because the maximum amount
of particle overlap is usually very small, even slight displacement of the particle
surface by the discretization method can have significant consequences for accurate
detection of particle contact. Also, because each vertex of a neighboring particle
must be checked to see if it lies in each triangle of the given particle, the method
can be quite time consuming, even if the set of vertices and triangles that must be
examined is further reduced. This problem becomes even more significant in three
dimensions.

7.3.2. Algorithms Based on a Common Normal Vector

The unit normal vectors to the surfaces of colliding ellipsoidal particles at the contact
point must be parallel, because the particle surfaces have the same tangent plane at
the contact point. Among other things, this requirement is necessary to ensure that
each particle receives an equal and opposite force from the collision, as required by
Newton’s second law. An illustration of the common unit normal algorithm for con-
tact detection is shown in Figure 7.5. The algorithm seeks to find two points, A and B,
located on ellipsoid 2 and ellipsoid 1, respectively, such that the unit normal vectors
to the ellipsoid surface at the two points are oriented along the same line. The equa-
tions of the two ellipsoids are written in terms of the global coordinates (x, y, z) as

f1 (x, y, z) = 0, f2 (x, y, z) = 0. (7.3.4)

The position vectors of points A and B are denoted by x2 and x1 , respectively,

corresponding to the coordinate positions (x2 , y2 , z2 ) and (x1 , y1 , z1 ). The outward
unit normal to a surface f (x, y, z) = 0 is given by n = ∇ f / |∇ f |. Because the outward
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

214 Ellipsoidal Particles

unit normal vectors of the two ellipsoids must point in opposite directions at the
points A and B, we can write
∇ f1 (x1 ) ∇ f2 (x2 )
n1 (x1 ) + n2 (x2 ) = + = 0, (7.3.5)
1 2
   
where 1 ≡ ∇ f1  and 2 ≡ ∇ f2 . Not only must the unit normal vectors at points
A and B be parallel to each other, but they must also be parallel to a line connecting
the two points, as shown in Figure 7.5. The unit tangent vector along this line is
given by t = (x2 − x1 )/, where  ≡ |x2 − x1 | is the distance between the points.
This condition then reduces to
x − x1 ∇ f2 (x2 )
t − n2 = 2 − = 0. (7.3.6)
 2
The vector equations (7.3.5) and (7.3.6) provide a set of six nonlinear equations
for the six coordinates of points A and B. After solving for these coordinates,
contact is detected if point A is inside ellipsoid 1 and point B is inside ellipsoid 2, or
in mathematical terms
f1 (x2 ) ≤ 0 and f2 (x1 ) ≤ 0. (7.3.7)
Once the contact points are found and contact is decided to have occurred, the
normal overlap δN is set equal to the contact point separation distance  .
An evaluation of the common unit normal contact detection algorithm is pre-
sented by Lin and Ng (1995). They found that although this algorithm can be used to
accurately detect contact and identify contact points in both two and three dimen-
sions, it is neither as accurate nor as efficient as algorithms based on geometric
potentials, which are reviewed in the next section.

7.3.3. Algorithms Based on Geometric Level Surfaces

The geometric concept of level surfaces of an ellipsoidal shape defines a family of
similar surfaces emanating outward and inward from the surface of an ellipsoid,
as shown in Figure 7.6a for an ellipsoid with surface E. The contact point of one
ellipsoid (E1 ) in a pair is found by examining its intersection with the level surfaces
of the other ellipsoid (E2 ), and vice versa. The method for obtaining these contact
points goes by many names in the literature, different versions of which can be
found papers by Ting (1992), Lin and Ng (1995), and Perram et al. (1984). The
current discussion follows the presentation of Chesnutt and Marshall (2009), and is
based on the geometric arguments of Alfano and Greer (2003), Chan (2001), and
Schneider and Eberly (2003).
The surface of a general ellipsoid can be written as a quadratic equation of the
form
Pk (x) = xT Ak x + bTk x + ck = 0, (7.3.8)
where x is the position vector and k has the value 1 or 2 for ellipsoids E1 and E2 ,
respectively. This equation can be expressed in a four-dimensional space using a
generalized position vector X = [x y z 1] in the canonical form
XQXT = 0, (7.3.9)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.3 Collision Detection and Contact Point Identification 215

α<0

Α Β
α1,min α2,min

E E2
α>0 E1

(a) (b)
Figure 7.6. (a) Level surfaces of ellipsoid with surface E, where interior and exterior corre-
spond to α < 0 and α > 0, respectively. (b) Contact points correspond to the point on each
ellipsoid surfaces that yields the smallest value of the potential α of the opposing level-surface
ellipsoid family.

where Q is the 4 × 4 symmetric characteristic matrix of the ellipsoid. In terms of the

coefficients qi j of Q, the coefficients in (7.3.8) can be written as
⎡ ⎤
q11 q12 q13
⎣
Ak = q12 q22 q23 ⎦ , (7.3.10a)
q13 q23 q33

bk = 2[q14 q24 q34 ]T , (7.3.10b)

ck = q44 . (7.3.10c)

Walls are treated as degenerate ellipsoids, where (7.3.9) reduces to the equation
of a plane by setting q11 , q22 , q33 , q12 , q13 , and q23 equal to zero. If any vector X exists
such that (7.3.9) is satisfied for two different ellipsoids, with different characteristic
matrices Q1 and Q2 , then the corresponding point x lies on both ellipsoids.
In order to determine if two ellipsoids intersect, equation (7.3.9) for ellipsoid E1
is multiplied by a scalar λ. Subtracting the same equation for another ellipsoid E2
yields
X(λQ1 − Q2 )XT = 0. (7.3.11)

A nontrivial solution of (7.3.11) exists if and only if the matrix (λQ1 − Q2 ) is singu-
lar, from which it follows that λ must be an eigenvalue of the matrix Q−1 1 Q2 . The
corresponding characteristic equation for λ requires finding the roots of the fourth
degree polynomial
 
det λI − Q−1 1 Q2 = 0. (7.3.12)

This polynomial admits four roots. As demonstrated by Alfano and Greer (2003)
and proved by Chan (2001), the two ellipsoids intersect at more than one point
(without completely penetrating through each other) if and only if two of the roots
of (7.3.12) are complex conjugates of each other.
Once contact detection has been verified, it remains to identify the contact
points. For this purpose, the set of level surfaces of ellipsoid E2 is parameterized in
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

216 Ellipsoidal Particles

terms of a parameter α (called the geometric potential) by

P2 (x) = α, (7.3.13)
where the regions α < 0 and α > 0 correspond to the interior and exterior of E2 ,
respectively. In order to find the contact point on ellipsoid E1 , we seek the minimum
value of α for which a vector x exists that is a solution of both (7.3.13) and P1 (x) = 0,
where the latter condition ensures that the point x is on the surface of ellipsoid E1 .
This condition is illustrated in Figure 7.6b.
In order to minimize α in (7.3.13) subject to the constraint P1 (x) = 0, a function
F (x, τ ) is defined as
F (x, τ ) = P2 (x) + τ P1 (x), (7.3.14)
where τ is a Lagrange multiplier. Differentiating F (x, τ ) with respect to τ gives
∂F
= P1 . (7.3.15)
∂τ
Setting this τ -derivative equal to zero recovers the constraint P1 (x) = 0. Setting the
gradient of F to zero gives
∇P2 + τ ∇P1 = 0 (7.3.16)
for some τ , which implies that the gradients of P1 and P2 are parallel. Using (7.3.8)
to obtain the gradients of P1 and P2 , (7.3.16) becomes
2(A2 + τ A1 )x + (b2 + τ b1 ) = 0. (7.3.17)
Solving (7.3.17) for x gives
1 1
x = − (A2 + τ A1 )−1 (b2 + τ b1 ) = y (τ ) , (7.3.18)
2  (τ )
where (τ ) = det(A2 + τ A1 ) and the components of the vector y(τ ) are cubic poly-
nomials in τ . Substituting (7.3.18) into the constraint P1 (x) = 0 yields a sixth-degree
polynomial in τ given by
y(τ )T A1 y(τ ) + (τ ) bT1 y(τ ) + 2 (τ ) c1 = 0. (7.3.19)
After calculating the real roots τn (n ≤ 6) of this polynomial, the corresponding
points x on the ellipsoid E1 are calculated from (7.3.18) and the corresponding
value of the potential α is obtained from (7.3.13). The contact point on ellipsoid E1
corresponds to the root that yields the minimum value of the potential α, denoted
as α2,min . The procedure is then repeated with the ellipsoids switched to obtain the
corresponding contact point on ellipsoid E2 .
It is noted that after the initial collision has occurred, two adhesive particles may
remain in contact even if the ellipsoidal surfaces of the particles do not overlap. This
occurs by the action of the adhesive force deforming the boundary of the particles
as they are pulled apart, which is referred to as particle necking. For this reason, it is
necessary to maintain a list of particle collision pairs at the previous time step and to
repeat the contact point identification process described here for all particle pairs on
this list, even if the ellipsoidal surfaces are no longer in contact. Only if the normal
overlap δN has a negative value that is less than the critical overlap −δC identified in
Section 4.2 are the particles determined to have broken apart and dropped from the
collision list.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.4 Contact Forces 217

7.4. Contact Forces

Collision and adhesion forces and torques acting on two colliding particles consist of
components that are both normal and tangential to the particle surface at the contact
point. A unique unit normal vector n is defined for a contact between particles i and
j as the average of the outward unit normal of particle i at its contact point and the
inward unit normal of a colliding particle j at its contact point. Similar to the case
of spherical particles, the normal force −Fn n for ellipsoidal particles results from a
combination of elastic repulsion, dissipation, and adhesive force. Because the unit
normal does not necessarily pass through the centroid of the ellipsoidal particles,
the normal force produces a torque −Fn ri × n on the particle, where ri is the vector
from the centroid of particle i to the contact point on particle i. The sliding motion
of the particles gives rise to a sliding force Fs tS and corresponding torque Fs ri × tS ,
where tS is a unit vector in the direction of relative motion of the particle surfaces
at the contact point projected onto a plane orthogonal to n. A torque also acts on
the particles due to resistance from twisting motion, given by Mt n. Additionally, as
for spherical particles, rolling motion is resisted by a rolling torque Mr tR × n, where
tR is the direction of rolling velocity, to be defined later in this section. Combining
these various forces and torques gives the net collision and adhesion force FA and
torque MA acting on an ellipsoidal particle i as

FA = −Fn n + Fs tS , MA = −Fn ri × n + Fs ri × tS + Mr tR × n + Mt n (7.4.1)

7.4.1. Geometry of Colliding Particles

Any point P1 on the surface of an ellipsoid E1 admits a curvature within any planar
slice of the surface passing through P1 . Since there are infinitely many such slices,
there are infinitely many curvatures κ of the ellipsoid surface that can be defined at
P1 . The principal curvatures are given by κ1 = 1/R1 and κ1 = 1/R1 , where R1 and R1
are the maximum and minimum values of the radius of curvature of any planar slices
passing through P1 , respectively. If P1 and P2 are the contact points for two colliding
ellipsoids E1 and E2 , and we similarly define principal curvatures κ2 = 1/R2 and
κ2 = 1/R2 at point P2 on ellipsoid E2 , then the principal relative radii of curvature
R and R are defined by

1 1 1 1 1 1

≡  + , 
≡  +  . (7.4.2)
R R1 R2 R R1 R2

The effective radius Re at the contact point is defined as the geometric mean of
the principal relative radii of curvature, or

Re = (R R )1/2 . (7.4.3)

The effective radius Re plays a similar role to the radius of the same name denoted
by R for spherical particles, defined in (3.1.1).
The normal overlap δN is defined as the distance between the contact points
of two colliding particles. The normal overlap is taken to have a positive sign for
intersecting particles and a negative sign for nonintersecting (e.g., necking) particles.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

218 Ellipsoidal Particles

The sum of the displacements normal to the contact surface for two colliding particles
can be written as a function of local coordinates x and y as

uz1 + uz2 = δN − Ax2 − By2 , (7.4.4)

where x = y = 0 at the contact point and the coordinate z measures distance nor-
mal to the tangent plane within the contact region. The coefficients A and B in
(7.4.4) are related to the principal radii of curvature defined in (7.4.2) by (Johnson,
1985, 85)
1 1
A= , B= . (7.4.5)
2R 2R

7.4.2 Hertz Theory for Ellipsoidal Particles

The contact region that forms on collision of nonadhesive ellipsoidal bodies has
an elliptical shape, with semimajor and semiminor axes a and b, respectively. The
pressure distribution within the contact region can be written in terms of the local
coordinates x and y in the direction tangent to the contact surface as

p = p0 [1 − (x/a)2 − (y/b)2 ]1/2 , (7.4.6)

where the total elastic normal force Fne on the particle is given by

Fne = 23 p0 π ab. (7.4.7)

Solution for the elastic displacement field associated with the pressure distribu-
tion (7.4.6) using a normal force prescribed over an elliptical region on the surface of
an elastic half-space yields expressions for the coefficients A and B and the normal
overlap δN as (see Johnson, 1985, for details)
p0 b
A= [K(ec ) − E(ec )], (7.4.8)
E a2 e2c

p0 b
B= [(a2 /b2 )E(ec ) − K(ec )], (7.4.9)
E a2 e2c

p0
δN = b K(ec ), (7.4.10)
E
1/2
where ec ≡ (1 − b2 /a2 ) is the ellipticity of the contact region and K(ec ) and E(ec )
are the complete elliptic integrals of the first and second kinds, respectively, defined
by
 π/2  π/2 
dθ
K(ec ) ≡  , E(ec ) ≡ 1 − e2c sin2 θ dθ . (7.4.11)
0 1 − e2c sin2 θ 0

An equation for the aspect ratio of the elliptical contact region is obtained by
taking the ratio of A and B using (7.4.8) and (7.4.9), and using (7.4.5), to write
B R (a/b)2 E(ec ) − K(ec )
=  = . (7.4.12)
A R K(ec ) − E(ec )
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

7.4 Contact Forces 219

Because ec is a function of ab, the entire right-hand side of (7.4.12) is a function only
of aspect ratio of the contact region. The ratio R /R is known from the location of
the contact point on the two ellipsoids, so (7.4.12) provides a nonlinear equation for
the contact region aspect ratio ab. In addition to aspect ratio, specification of the
geometry of the contact region also requires that we specify a mean contact region
radius, given by
ae ≡ (ab)1/2 . (7.4.13)
To obtain an expression for ae , we take the cube of (7.4.13) to write
a3e = (ab)3/2 = (b/a)3/2 a3 . (7.4.14)
Using (7.4.8) and (7.4.9), we can write
1 p b
(AB)1/2 = = 0 2 2 C(ec ), (7.4.15)
2Re E a ec
where C(ec ) = {[(a/b)2 E(ec ) − K(ec )] [K(ec ) − E(ec )]}1/2 . Solving for p0 from
(7.4.7) and plugging into (7.4.15) gives
1 3Fne C(ec )
= . (7.4.16)
2Re 2π E a3 e2c
Solving for a3 in (7.4.16) and substituting into (7.4.14), and then taking the cube
root of the resulting equation, gives the mean contact region radius as
 
3Fne Re 1/3
ae = F1 (ec ), (7.4.17)
4E
where F1 (ec ) = (4C(ec )/π e2c )1/3 (b/a)1/2 is a function only of the contact region ellip-
ticity ec . The result for spherical particles is recovered as b/a → 0, which corresponds
to F1 (0) = 1 .
The equation for the normal overlap can be written in terms of the normal elastic
force Fne by solving for p0 from (7.4.7) and substituting into (7.4.10) to obtain
3Fne
δN = K(ec ). (7.4.18)
2π Ea
Because a = (ab)1/2 (a/b)1/2 , the result (7.4.17) can be used in (7.4.18) to write
 1/3
9Fne2
δN = F2 (ec ), (7.4.19)
16E 2 Re
where F2 (ec ) = (2/π )(b/a)1/2 K(ec )/F1 (ec ). The spherical particle result is recovered
as ec → 0, for which F2 (0) = 1. Solving for Fne from (7.4.19) gives
3/2
Fne = KH δN , (7.4.20)
where the stiffness coefficient KH is
4ER1/2
[F2 (ec )]−3/2 .
e
KH = (7.4.21)
3
Equation (7.4.20) is the same result for elastic normal force as obtained in Sec-
tion 3.2 for spherical particles, but the stiffness coefficient definition differs by the
factor F3 (ec ) ≡ [F2 (ec )]−3/2 , which accounts for the effect of contact region ellipticity.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

220 Ellipsoidal Particles

ec 1.4

Apect ratio and Ellipticity

0.8

Correction factors
F2 -3/2
1.2
0.6

0.4 1
b/a F1
0.2 F2
0.8

0
1 10 100 1 10 100
R'/R'' R'/R''
(a) (b)
Figure 7.7. Plots showing characteristics of contact between ellipsoidal particles as a function
of the ratio R /R of the maximum and minimum principal relative radii of curvature. (a)
Ellipticity ec (long dashed line) and aspect ratio b/a of the contact region (solid line). Also
shown as a dashed line is the approximation (7.4.22) for the aspect ratio. (b) Correction factors
F1 (ec ) (solid line) and F2 (ec ) (dashed line) for the mean contact region radius and the normal
overlap. Also shown is the correction [F2 (ec )]−3/2 for the stiffness coefficient (long dashed
line).

Figure 7.7 plots various terms and corrections for collision of ellipsoidal particles
as a function of the ratio R /R of the maximum and minimum principal relative
radii of curvature. The ellipticity ec is solved iteratively from (7.4.12) and plotted
as a function of R /R , along with the corresponding contact region aspect ratio
b/a. The ratio a/b for the contact region is observed to be significantly less than the
corresponding ratio R /R based on the maximum and minimum curvatures of the
ellipsoids as a whole. Also plotted is an approximation due to Johnson (1985) for
contact region aspect ratio, given by

b/a = (R /R )−2/3 , (7.4.22)

which is found to be in excellent agreement with the exact solution. The correction
factors F1 (ec ) and F2 (ec ) for the effective contact region radius ae and the normal
overlap δN are plotted in Figure 7.7b as functions of R /R . Also shown is the
correction factor [F2 (ec )]−3/2 for the stiffness coefficient KH shown in (7.4.21). As
each of these correction factors approaches unity, the results approach those for
spherical particles. It is clear from Figure 7.7b that for ellipsoids with modest aspect
ratio, there is a fairly small difference in the normal force between the exact solution
and that obtained for spherical particles with the same effective radii and normal
overlap. Indeed, even for R /R as high as 10, the relative error in normal force would
only have been about 11% had we simply used the original Hertz formula (3.2.8) for
spherical particles. Most of the other contact forces can similarly be approximated
using the expressions for equivalent spherical particles to within reasonable accuracy,
particularly if the particles are not too elongated.
Many investigators using the discrete element model with ellipsoidal particles
simply use the corresponding contact force expressions for spherical particles, unless
the particles are extremely elongated into needle- or disklike shapes. The major
differences in the dynamics of spherical and ellipsoidal particles are that (1) the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

References 221

torque associated with the normal force causes a rotational motion for ellipsoidal
particles when a normal collision occurs, which is not present for spherical particles;
and (2) it is more difficult for ellipsoidal particles to roll for an appreciable distance
along a flat surface because there is a potential energy hill that must be overcome
with each half-rotation of the particle that is not present for spherical particles.
Many practitioners feel that simulations with spherical particles allow too much
rolling motion compared to “real” (generally nonspherical) particles, and for this
reason ellipsoidal particles are often a desirable choice to obtain more realistic levels
of particle rolling.

REFERENCES

Alfano S, Greer ML. Determining if two solid ellipsoids intersect. Journal of Guidance,
Control, and Dynamics 26, 106–110 (2003).
Brenner H. The Stokes resistance of an arbitrary particle. IV. Arbitrary fields of flow. Chemical
Engineering Science 19(10), 703–727 (1964).
Camp PJ, Allen MP. Hard ellipsoid rod-plate mixtures: Onsager theory and computer simu-
lations. Physica A 229, 410–427 (1996).
Campbell CS. Elastic granular flows of ellipsoidal particles. Physics of Fluids 23, 013306 (2011).
Chan K. A simple mathematical approach for determining intersection of quadratic surfaces.
Proceedings of the American Astronautical Society, AAS Paper 01–358, July–Aug. (2001).
Chesnutt JKW, Marshall JS. Blood cell transport and aggregation using discrete ellipsoidal
particles. Computers & Fluids 38, 1782–1794 (2009).
Chesnutt JKW, Marshall JS. Structural analysis of red blood cell aggregates under shear flow.
Annals of Biomedical Engineering 38(3), 714–728 (2010).
Chhabra RP, Agarwal L, Sinha NK. Drag on non-spherical particles: An evaluation of avail-
able methods. Powder Technology 101, 288–295 (1999).
Dahneke BE. Slip correction factors for non-spherical bodies. III. The form of the general
law. Journal of Aerosol Science 4(2), 163–170 (1973).
Džiugys A, Peters B. An approach to simulate the motion of spherical and non-spherical fuel
particles in combustion chambers. Granular Matter 3, 231–265 (2001).
Fan FG, Ahmadi G. Wall deposition of small ellipsoids from turbulent air flows: A Brownian
dynamics simulation. Journal of Aerosol Science 31, 1205–1229 (2000).
Gallily I, Cohen AH. On the orderly nature of the motion of nonspherical aerosol particles.
Journal of Colloid and Interface Science 68, 338–356 (1979).
Happel J, Brenner H. Low Reynolds Number Hydrodynamics. Martinus Nijhoff Publishers,
The Hague (1963).
Harper EY, Chang I-D. Maximum dissipation resulting from lift in a slow viscous shear flow.
Journal of Fluid Mechanics 33(2), 209–225 (1968).
Hogue C, Newland D. Efficient computer simulation of moving granular particles. Powder
Technology 78, 51–66 (1994).
Hughes PC. Spacecraft Attitude Dynamics. John Wiley & Sons, Inc., New York (1986).
Jeffery GB. The motion of ellipsoidal particles immersed in a viscous fluid. Proceedings of the
Royal Society of London A 102, 161–179 (1922).
Johnson, K.L., Contact Mechanics, Cambridge University Press (1985).
Knuth MA, Johnson JB, Hopkins MA, Sullivan RJ, Moore JM. Discrete element modeling of
a Mars Exploration Rover wheel in granular material. Journal of Terramechanics 49, 27–36
(2012).
Lin X, Ng T-T. Contact detection algorithms for three-dimensional ellipsoids in discrete
element modelling. International Journal for Numerical and Analytical Methoids in Geome-
chanics 19, 653–659 (1995).
Oberbeck A. Ueber stationäre Füssigkeitsbewegungen mit Berücksichtigung der inneren
Reibung. Journal für die reine und angewandte Mathematik 81, 62–80 (1876).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-07 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 11:15

222 Ellipsoidal Particles

Perram JW, Wertheim MS, Lebowitz JL, Williams GO. Monte Carlo simulations of hard
spheroids. Chemical Physics Letters 105(3), 277–280 (1984).
Schneider PJ, Eberly DH. Geometric Tools for Computer Graphics. Morgan Kaufmann
Publishers, San Francisco (2003).
Solis KJ, Martin JE. Isothermal magnetic advection: Creating functional fluid flows for heat
and mass transfer. Applied Physics Letters 97, 034101 (2010).
Ting JM. A robust algorithm for ellipse-based discrete element modeling for granular mate-
rials. Computers and Geotechnics 13(3), 175–186 (1992).
Yin C, Rosendahl L, Kaer SK, Sørensen H. Modelling the motion of cylindrical particles in a
nonuniform flow. Chemical Engineering Science 58, 3489–3498 (2003).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8 Particle Interactions with Electric

and Magnetic Fields

Particle interactions with electric and magnetic fields are important for a wide range
of industrial applications, including mixing and separation processes in microfluidic
flows and biological assay processes, electrospray coating processes, particle sepa-
ration devices such as electrostatic precipitators, nanoparticle dispersion and manu-
facturing processes, and electrostatic classification process for aerosol-generation or
particle-sizing applications. In these processes, as well as in many others, particles
are significantly influenced by forces arising from particles traveling in electric or
magnetic fields, as well as from charged or magnetic particles interacting with each
other. Interactions of charged particles immersed in an electrolytic solution result
in short-range interaction force due to ionic shielding, as discussed in Chapter 4.
However, in an aerosol there are no ions available, and the resulting electric and
magnetic forces between particles decay slowly in space.
This chapter discusses the physics of electric and magnetic forces on parti-
cles, as well as computational methods for simulation of the electric field, that can
be used in conjunction with DEM computations. A description of the forces and
torques imposed on a particle in both DC and AC electric fields is presented in
the first section, including the phenomenon of dielectrophoresis. The second section
discusses different methods by which particles in electric fields become charged,
including contact electrification and de-electrification. The third section discusses
forces and torques on particles in a magnetic field, including the phenomenon of
magnetophoresis. Review of particle physics concludes in the sixth section, which
examines the agglomerate chain structure of particles in an electric field.
Methods for computation of the electric field are the focus of the fourth and
fifth sections, including the boundary element method (BEM) and the fast multipole
method (FMM). The FMM provides a method for rapid computation of the electric
field induced by charged particles. The BEM can be used to solve for the electrostatic
field induced by macroscopic bodies immersed in the flow, such as electrodes, or the
modification of an external electric field by dielectric surfaces. The bodies considered
with the BEM are generally much larger than the particles. The total electric field
to which the particles are exposed is the sum of an external electric field (typically
prescribed), the electric field induced by other particles (obtained by FMM), and
the electric field induced by macroscopic bodies in the flow (obtained by BEM), as
illustrated in Figure 8.1. Summing all of these different fields yields the total electric
223
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

224 Particle Interactions with Electric and Magnetic Fields

E0

FMM: for predicting

E field induced by
particles

BEM: for predicting

surface charge density
DEM: for predicting of domain boundaries
particle motion

Figure 8.1. Illustration of the roles of DEM, FMM, and BEM for prediction of the dynamics
and adhesion of particles near a body in the presence of an electric field.

field vector, which is used in DEM to compute the forces and torques on the particles
and to move the particles in time. In the fourth and fifth sections of this chapter, we
present general computational approaches for FMM and BEM applicable to systems
consisting of a large number of particles immersed in a fluid flow in the presence of
conducting or dielectric “macroscopic” bodies of arbitrary shape.

8.1. Electric Field Forces and Torques

8.1.1. Coulomb Force and Dielectrophoresis

A particle with charge q placed in an electric field E is subject to the Coulomb
force

FCL = q E. (8.1.1)

An electric dipole is formed of positive and negative charges ±q separated by a

displacement d in the limit as |d| → 0 and q → ∞, as shown in Figure 8.2. If the
negative charge is located at a position x and the positive charge is located at a
position x + d, the resulting force on the dipole is given by

Fdipole = lim [−qE(x) + qE(x + d)]. (8.1.2)

d→0

qE(x+d)
E(x+d) +
+
+q
d
Centroid
-
-
–q E(x) –qE(x)

(a) (b)
Figure 8.2. Illustration of a particle with negative charge −q and a particle with positive
charge +q in an electric field, separated by a displacement d. (a) Electric field on the particles;
(b) Coulomb forces acting on the particles.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.1. Electric Field Forces and Torques 225

Expanding E(x + d) in a Taylor series about x gives

E(x + d) = E(x) + d · ∇E + · · · . (8.1.3)
Substituting (8.1.3) into (8.1.2) gives
Fdipole = lim qd · ∇E. (8.1.4)
d→0

For a dipole, we require that |d| → 0 and the point charge q → ∞ in such a manner
that the product qd approaches a constant p, called the dipole moment. The resulting
force on the dipole is obtained from (8.1.4) as
Fdipole = p · ∇E. (8.1.5)
There is also a moment exerted on the dipole about the centroid point, located at
the midpoint of the line segment connecting the two charge points. The moment
Mdipole is obtained by summing the moment exerted by the Coulomb force on the
two charge points making up the dipole about the centroid, giving
 
Mdipole = 12 d × qE(x + d) + − 12 d × (−q)E(x). (8.1.6)
The lever arm from the centroid is d/2 for the positively charged particle and −d/2
for the negatively charged particle. Taking the limit qd → p while d → 0, (8.1.6)
becomes
Mdipole = p × E. (8.1.7)
The dipole moment of a conductive particle is caused by migration of charge on
the particle surface, whereas for a dielectric particle, the dipole moment is generated
by charge migration within the particle volume. Particles can possess two different
kinds of dipole moments. The first type is a permanent dipole, for which the dipole
exists within the material independent of other fields. An example of a permanent
dipole is a polarized molecule, such as a water molecule. A second example is a
material called an electret, which can be formed from certain dielectric materials that
are first heated above their melting temperature and then cooled while exposed to a
strong electric field.
The second type of dipole moment is an induced dipole, for which the polariza-
tion arises as a result of exposure of the particle to an electric field, but vanishes when
the electric field is removed. Polarization occurs within dielectric materials due to
displacement of positive charges in the “upstream” direction relative to the stream-
lines of the electric field vector, and displacement of negative charges in the opposite
direction (Figure 8.3). The value of the induced dipole moment for a dielectric spher-
ical particle of radius r p immersed in a constant (DC) electric field E = E0 ex can
be determined by recalling the solution of the Laplace equation ∇ 2  = 0 governing
the electrostatic potential , where E = −∇, both inside and outside of a sphere,
given in spherical polar coordinates by

−E0 r cos θ + (C1 /r2 ) cos θ for r > r p
= . (8.1.8)
−C2 r cos θ for r ≤ r p
The constants of integration C1 and C2 can be determined from the boundary con-
ditions on the sphere surface. If the sphere has permittivity ε p and the surrounding
fluid medium has permittivity ε f , then continuity of the tangential component of the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

226 Particle Interactions with Electric and Magnetic Fields

rp
− +
− θ ++
− Figure 8.3. Dielectric sphere with permittivity ε p
− + x immersed in a uniform electric field E in a medium
E − εp with permittivity ε f .
− +
− +
εf

electric field vector implies that  must be continuous across the sphere surface, or

1 (r p , θ ) = 2 (r p , θ ), (8.1.9)
where the subscript 1 denotes the fluid side of the interface and the subscript 2
denotes the particle side. The second boundary condition requires continuity of the
normal component of the electric displacement vector D = ε E, so that
∂1 ∂
εf (r p , θ ) = ε p 2 (r p , θ ). (8.1.10)
∂r ∂r
Substituting (8.1.8) into (8.1.9) and (8.1.10) yields two equations for the coeffi-
cients C1 and C2 as
εp − ε f 3ε f
C1 = r3p E0 , C2 = E0 . (8.1.11)
ε p + 2ε f ε p + 2ε f
The first term in the outer (r ≥ r p ) solution (8.1.8) for 1 represents the uniform
electrostatic field and the second term has the same form as the potential function
for an electric dipole with dipole moment p = 4π ε f C1 ex . Substituting (8.1.11) for
C1 into this equation, we find that in the outside region r ≥ r p the electric field has
the form of a uniform electric field vector plus a point dipole with induced dipole
moment
p = 4π ε f KCM r3p E, (8.1.12)
where the coefficient KCM , known as the Clausius-Mossotti function, is given by
εp − ε f
KCM = . (8.1.13)
ε p + 2ε f
The value of KCM is limited by (8.1.13) to the range −0.5 ≤ KCM ≤ 1. The value of
the vacuum permittivity ε0 is 8.85 × 10–12 C2 /N m2 . The relative permittivity ε is
commonly used to reflect the strength of the electrostatic field produced in different
materials by a fixed potential relative to that produced in a vacuum under identical
conditions. For most materials, relative permittivity falls in the range 1 < ε < 10;
for example, ε is 1.00059 for air, 2.1 for Teflon, 2.4–2.7 for polystyrene, and 3.9–
4.3 for quartz. Some materials exhibit much higher values of relative permittivity;
for example, ε is equal to 80 for pure water at room temperature and 86–173 for
titania.
The process by which an induced dipole moment forms in a dielectric parti-
cle exposed to an electric field is called dielectrophoresis, and the force resulting
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.1. Electric Field Forces and Torques 227

from substitution of (8.1.12) into (8.1.5) is called the dielectrophoretic (DEP) force.
Because the induced dipole moment in (8.1.12) is proportional to E, the resulting
DEP force can be written as

FDEP = 2π ε f KCM r3p ∇E 2 , (8.1.14)

where E is the magnitude of the electric field vector E. The result (8.1.14) implies
that particles with KCM < 0 are attracted to regions with high electric field strength
and particles with KCM > 0 are repelled from regions with high electric field strength.
A similar derivation can be performed for a conducting particle immersed in
a fluid with a constant (DC) electrostatic field. In this case the boundary condition
(8.1.10) is replaced by the condition that the normal component of the current density
vector J = σ E is continuous at the sphere surface, or

∂1 ∂
σf (r p , θ ) = σ p 2 (r p , θ ), (8.1.15)
∂r ∂r
where σ p and σ f are the electrical conductivities of the particle and fluid, respec-
tively. Solving for C1 from (8.1.9) and (8.1.15) and substituting into (8.1.8) gives the
Clausius-Mossotti function for a conducting sphere as
σp − σ f
KCM = . (8.1.16)
σ p + 2σ f

The total electric force on a particle is given by FE = FCL + FDEP . For spherical
particles with no permanent dipole and with induced dipole moment aligned with
the electric field, as in (8.1.12), there is no electric torque on the particle (ME = 0).
Particles respond to any electric field to which they are exposed. This electric field
could originate from an electrode immersed in the fluid, or it could come from the
electric field emitted by other nearby particles. The electric field emitted by nearby
particles could in turn be associated with particle charge, or it could itself be induced
by response to an electric field generated elsewhere. A further complication is that
the presence of nonuniformity in the incident electric field causes the induced field to
be more complex than a simple dipole, but instead to consist of the sum of a dipole, a
quadrupole, and higher order multipoles. As a consequence, the problem of a cloud
of particles immersed in a fluid in the presence of an electrode is highly nonlinear,
with the electric field induced by polarization of each particle influencing that induced
by every other particle. Nevertheless, this problem can often be simplified with
reasonable accuracy by treating each particle as a monopole (for a charged particle)
plus a dipole, ignoring higher-order multipoles, and lagging the interaction terms
between the particles in time during the computation. An exception occurs for
cases where the particles are actually touching each other, which is discussed in
Section 8.5.

8.1.2. Dielectrophoresis in an AC Electric Field

In an AC electric field, the presence of ohmic energy loss due to oscillation of
the ambient electric field introduces some important changes in the nature of the
induced dipole moment. Let us reconsider the problem of a uniform electric field
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

228 Particle Interactions with Electric and Magnetic Fields

past a sphere considered in the previous section, but this time with an oscillating
electric field vector given by
E(t ) = Re[E0 exp( jωt )] ex , (8.1.17)
√
where j ≡ −1. In an oscillating electric field, the boundary condition (8.1.10) is
replaced by
∂qs
Jr1 − Jr2 + = 0 at r = r p , (8.1.18)
∂t
where Jr is the radial component of the current density vector J = σ E and qs is the
free electric surface charge, defined by
qs = ε f Er1 − ε p Er2 at r = r p . (8.1.19)
Assuming that all variables oscillate harmonically in time in proportion to exp( jωt )
and substituting (8.1.19) into (8.1.18), the normal flux boundary condition at r = r p
can be written as
ε∗f Er1 (r p , θ ) = ε∗p Er2 (r p , θ ), (8.1.20)
where the complex dielectric constants ε∗f and ε∗p are defined by
ε∗f = ε f − jσ f /ω, ε∗p = ε p − jσ p /ω. (8.1.21)
Equation (8.1.20) has the same form as (8.1.10) with the replacement of the permit-
tivities by the complex dielectric constants. Substituting the inner and outer solutions
(8.1.8) into the boundary conditions (8.1.9) and (8.1.20) gives a solution for the coef-
ficient C1 in (8.1.8) as
ε∗p − ε∗f
C1 = r3p E0 . (8.1.22)
ε∗p + 2ε∗f

The second term in the outer flow solution of (8.1.8) has the form of an induced
dipole with a complex-valued effective dipole moment
p∗eff (t ) = 4π ε f KCM
∗
r3p E(t ), (8.1.23)
∗
where KCM is the complex Clausius-Mossotti function

∗
ε∗p − ε∗f
KCM = . (8.1.24)
ε∗p + 2ε∗f
A real-valued effective dipole moment can be defined by
peff (t ) = Re[p∗eff (t ) exp( jωt )] = 4π ε f Re[KCM
∗
] r3p E. (8.1.25)
Equations (8.1.17) and (8.1.25) give expressions for the instantaneous electric field
E(t ) and effective dipole moment peff (t ), both of which oscillate with the driving
frequency ω of the AC electric field. The instantaneous DEP force and torque are
given by
FDEP (t ) = peff (t ) · ∇E, (8.1.26a)

MDEP (t ) = peff (t ) × E. (8.1.26b)

 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.1. Electric Field Forces and Torques 229

Substituting (8.1.25), the time-averaged DEP force can be written as

∗
FDEP  = 2π ε f r3p Re[KCM (ω)] ∇Erms
2
, (8.1.27)

where Erms = E · E1/2 is the root-mean-square electric field strength.

The torque vanishes for the case of a unidirectional electric field, as given
by (8.1.17). However, many applications involve electric fields that are circularly
polarized, in which the magnitude of E is constant in time but the vector direction
rotates in a circular manner. For a right-polarized electric field, (8.1.17) is replaced
by

E(t ) = Re[E0 (ex − jey ) exp( jωt )] . (8.1.28)

If E∗ (t ) is the complex electric field vector, given by the term in brackets in (8.1.28),
and p∗eff (t ) is the complex dipole moment defined by (8.1.23), then the DEP torque
on the particle is given by

MDEP (t ) = Re[p∗eff (t )] × Re[E∗ (t )], (8.1.29)

where for the right-polarized electric field, (8.1.23) gives

 
Re[p∗eff (t )] = 4π ε f r3p Re[KCM
∗
]Re[E∗ ] − Im[KCM
∗
]Im[E∗ ] . (8.1.30)

Substituting (8.1.30) into (8.1.29) gives

∗
 
MDEP = −4π ε f r3p Im[KCM ] Im[E∗ ] × Re[E∗ ] . (8.1.31)

From (8.1.28), Im[E∗ ] × Re[E∗ ] = E02 ez , so the DEP torque on the particle for this
right-polarized field is obtained from (8.1.31) as
∗
MDEP = −4π ε f r3p Im[KCM ] E02 ez , (8.1.32)

where ez is a unit vector along the axis of polarization. Although the electric field
varies in time, the time dependence cancels out from the cross-product of the imag-
inary and real components of E∗ (t ) due to the identity sin2 (ωt ) + cos2 (ωt ) = 1.

8.1.3. Application to Particle Separation and Focusing

The complex Clausius-Mossotti function given in (8.1.24) can be rearranged as
  
∗
σp − σ f jωτ0 + 1
KCM = , (8.1.33)
σ p + 2σ f jωτMW + 1

where
ε p + 2ε f εp − ε f
τMW = , τ0 = . (8.1.34)
σ p + 2σ f σp − σ f

The parameter τMW , called the Maxwell-Wagner relaxation time, characterizes the
decay of free charge at the sphere surface in response to the AC electric field. Taking
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

230 Particle Interactions with Electric and Magnetic Fields

(a) (b)
Figure 8.4. Plots showing the real and imaginary parts of the Clausius-Mossotti functions as
a function of the dimensionless parameter ωτMW for cases with (a) Eε = 1/2 and Eσ = −1/3
and (b) Eε = −1/3 and Eσ = 1/2.

∗
the real and imaginary parts of KCM , we find
   
∗
σp − σ f 1 εp − ε f ω2 τMW
2
Re[KCM (ω)] = + , (8.1.35a)
σ p + 2σ f 1 + ω2 τMW2 ε p + 2ε f 1 + ω2 τMW
2

 
∗
εp − ε f σp − σ f ωτMW
Im[KCM (ω)] = − . (8.1.35b)
ε p + 2ε f σ p + 2σ f 1 + ω2 τMW
2

As ωτMW → 0, the real part of the Clausius-Mossotti function approaches the value
Eσ ≡ (σ p − σ f )/(σ p + 2σ f ) given in (8.1.16) for a conducting particle in a DC field
and the imaginary part vanishes. As ωτMW → ∞, the real part of the Clausius-
Mossotti function approaches the value Eε ≡ (ε p − ε f )/(ε p + 2ε f ) given in (8.1.13)
for a dielectric particle in a DC field and the imaginary part again vanishes. Examples
∗
showing variation of both the real and imaginary parts of KCM (ω) as a function of
frequency are given in Figure 8.4. For the case shown in Figure 8.4a where Eε > 0 and
∗
Eσ < 0, the value of real part of KCM (ω) is negative for ωτMW 1 and positive for
∗
ωτMW 1 and the imaginary part of KCM (ω) is everywhere positive. The opposite
occurs when Eε < 0 and Eσ > 0, as shown in Figure 8.4b. Most of the variation
∗
in the real and imaginary parts of KCM (ω) occurs in the interval 0.1 < ωτMW < 10;
outside of this interval the imaginary part is close to zero and the real part is nearly
constant.
Cases such as those shown in Figure 8.4, where ε p − ε f and σ p − σ f are of
opposite signs, are of particular interest in applications, as in such cases the sign
∗
of Re[KCM (ω)] changes as the frequency is varied from a low to a high value.
By adjusting the permittivity and conductivity of the fluid, the DEP force can be
used to induce a differential drift on particles with different values of ε p and σ p ,
including in some cases a drift velocity in different directions. This observation is the
basis of the dielectrophoretic method used by several researchers for separation of
small particles of different sizes or different materials from each other, or even for
separation of biological cells of different types (Fiedler et al., 1998; Chen and Du,
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.2. Mechanisms of Particle Charging 231

1.25

Focusing fraction
1.0

Electrode
0.75
Channel
0.50
U
0.25
Electrode
0
0 5 10 15 20

Applied voltage (V)

(a) (b)
Figure 8.5. (a) Dielectrophoretic particle focusing system used to transport particles in a
microchannel into a thin stream by Yu et al. (2005). (b) Results with latex particles for the
focusing fraction, defined as the final width of the particle stream divided by the tube width, as
a function of applied voltage at frequency 10 kHz (filled circles) and 30 kHz (open squares).

2007; Zhang et al., 2006). In a related application, dielectrophoresis has been used for
particle focusing, in which particles that are initially dispersed in a microchannel flow
are maneuvered into a concentrated stream, typically at the center of the channel
(Morgan et al., 2003; Yu et al., 2005). Focusing is typically achieved by wrapping
electrodes around the outside of the microchannel and oscillating the potential of the
electrodes at a set frequency. For example, the particle focusing system proposed
by Yu et al. (2005), shown in Figure 8.5a, consists of an elliptical microchannel
surrounded at regular intervals by ring-like elliptical electrodes. The electrodes are
attached to conductors running along the length of the channel in an alternating
manner so as to produce a two-phase AC electric field. The DEP force for cases
∗
with Re[KCM (ω)] < 0, or so-called negative DEP, acts to move particles away from
the regions near the electrodes with high electric field strength toward the region in
the center, at which the electric field strength is at a minimum. The negative DEP
phenomenon thus produces a thin stream of particles at the channel center, the
thickness of which is limited by diffusion related to particle collision. Figure 8.5b
plots the ratio of the width of the particle stream after focusing to the tube width,
which we call the focusing fraction, as a function of peak-to-peak applied voltage
based on data from Yu et al. (2005). Cases with different driving frequencies appear
to give similar results, although the time required to achieve focusing depends on
frequency. These results illustrate the increase in DEP-induced particle drift as the
electric field strength increases.

8.2. Mechanisms of Particle Charging

Determination of particle charge is essential for accurately predicting the electric
force on particles. The principal mechanisms by which a solid particle acquires
charge in an aerosol include field charging, diffusion charging, contact electrifica-
tion (tribocharging), and thermionic emission (known as the Edison effect). Of
particular interest for many manufacturing processes are liquid droplet aerosols,
for which spray charging and electrolytic charging during the processes of atomiza-
tion/nebulization are of most importance, along with electrokinetic streaming due
to the relative motion between solid and fluid phases. Among these mechanisms,
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

232 Particle Interactions with Electric and Magnetic Fields

q=0 q=qmax

(a) (b)
Figure 8.6. Schematic diagrams of the electric field lines near (a) an uncharged particle and
(b) a particle at saturation charge.

the field and diffusion charging mechanisms are feasible and easily operational for
production of highly charged aerosols, either of solid or liquid phases. These two
mechanisms both depend on the collisions of particles with gaseous ions, which
are generated by unipolar or bipolar chargers. The difference lies in that the field
charging is due to the particle-ion collisions influenced by an applied external field,
whereas diffusion charging is due to collisions driven by random Brownian motion
of ions. Also of importance for problems with frequent particle collisions is the phe-
nomenon of contact electrification, which results in a transfer of electrons from one
material to another upon contact. The current section discusses and compares these
different methods for particle charging in aerosol flows.

8.2.1. Field Charging

Field charging occurs when gas ions are driven by an external electric field to collide
with and adhere to the particle surface. This method of charging is dominant for
particles larger than about 1 µm. For a spherical particle with radius rp and with a
charge q placed in a uniform external electric field E = E0 ex , the resultant electric
field around the particle is a summation of the external field E, the field E p due to
the effect of induced polarization of the particle, and the field Eq due to the particle
charge, so the total electric field vector is given by ET = E + E p + Eq . The field at
a point x due to a dipole of strength p at x0 , with r = |x − x0 |, is given by (Jackson,
1962)
3n(p · n) − p
Ep = . (8.2.1)
4π ε f r3

Substituting the expression (8.1.12) for the induced dipole moment strength into
(8.1.36) gives
r3p q
ET = E + KCM [3(E · n)n − E] + n, (8.2.2)
r3 4π ε f r2
where n is a unit vector defined by n = (x − x0 )/|x − x0 |. Figure 8.6 illustrates the
electric field lines around a particle both with no charge and with the saturation
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.2. Mechanisms of Particle Charging 233

charge. Because the ions in the air migrate along the electric field lines in the
direction of ET , the ions only reach the particle surface (and hence charge the
particle) if the electric field lines terminate on the particle with negative value of
the radial component ET · n on the particle surface. Using (8.2.2) together with
E0 = |E|, the radial component Er of the electric field at the particle surface (r = r p )
can be calculated as
q
Er |r=r = E0 cos θ (2KCM + 1) + , (8.2.3)
p 4π ε f r2p

where θ is the angle between E and the radial direction n. If Er ≥ 0 for all θ, no
electric field line can reach the particle surface, implying that no ions can impact
with the particle. Using this condition, the particle saturation charge by the field
charging method can be estimated by setting cos θ = −1 in (8.2.3) to obtain

3ε p
qmax = 4π ε f r2p E0 (2KCM + 1) = 4π ε f r2p E0 . (8.2.4)
ε p + 2ε f

The time variation of the particle charge via the field-charging method is commonly
approximated by the equation
t t
q f (t ) = qmax = qmax εf , (8.2.5)
t +τ t +4
eN0 Z

where the characteristic field-charging time τ is dependent on the gas phase permit-
tivity ε f , the ion concentration N0 (with a magnitude of 1013 /m3 or greater), and the
electric mobility of the ion Z (with a value of 10−4 − 10−3 m2 V· s). Equation (8.2.5)
is known as the Pauthenier and Moreau-Hanot equation, which was developed to
model charging of dust particles in the continuum regime.

8.2.2. Diffusion Charging

Diffusion charging occurs when ions subject to a random Brown motion collide
with and adhere to a particle surface. The diffusion-charging mechanism is domi-
nant for ultra-fine particles, where the continuum approximation breaks down and
the free-molecular regime is possible (see Chapter 9). White (1951) derived a sim-
plified diffusion charging model for particles in the free-molecular regime. Under
equilibrium diffusion charging, a particle of diameter r p possessing n elementary
charges (i.e., qd = ne) has an electrostatic surface potential  = ne/4π ε f r p . The
number concentration of ions in the gas, Ni , satisfies the Boltzmann distribution
Ni = N0 exp(−e/kB T ), where kB is the Boltzmann constant and T is absolute
 tem-
perature. The mean thermal velocity of ions of mass mi is given by v̄i = 8kB T/π mi .
From molecular collision theory, the rate of ion collisions on the particle can be esti-
mated as dn/dt = π r2p Ni v̄i , provided the ion size is much smaller than the particle
size. Substituting for Ni and v̄i gives
   
dn ne2 8kB T 1/2
= π r p N0 exp −
2
. (8.2.6)
dt 4π ε f r p kB T π mi
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

234 Particle Interactions with Electric and Magnetic Fields

Particle diameter,
Figure 8.7. Number of elementary charges on dust particles in an electrostatic precipitator
with external electric field E = 1.5 × 105 V/m, ion concentration Ni = 1013 /m3 , t = 1s, and ion
mean velocity v̄i = 240 m/s. The experimental data from Jin (2013) was measured using an
electric low pressure impactor (ELPI) device, while the calculation is based on the Pauthenier
and Moreau-Hanot equation for field charging and the White equation for diffusion charging.

Integrating this equation in time, with initial condition n(0) = 0, gives the particle
charge due to diffusion charging as
 
4π ε f r p kB T v̄i r p e2 N0t
qd (t ) = ln 1 + . (8.2.7)
e 4ε f kB T

This equation was first derived by White (1951). In contrast to the relationship
q f ∝ d2 for field charging, (8.2.7) yields a relationship qd ∝ d for diffusion charging.
The total charge of a particle in an electric field is the sum of both field and diffusion
charging.
The number of elementary charges on a particle, q/e, can be measured as a func-
tion of particle size in an electrostatic precipitator by using an electrical low-pressure
impactor (ELPI) device (Glover and Chan, 2004). Jin (2013) reports experimental
data for particle charging with electric field E = 1.5 × 105 V/m, ion concentration
Ni = 1013 /m3 , t = 1s, ε f = 8.854 × 10−12 F · m−1 , and mean ion velocity v̄i = 240 m/s.
As shown in Figure 8.7, the sum of White’s formula (8.2.7) for diffusion charging and
the Pauthenier and Moreau-Hanot formula (8.2.5) for field charging can accurately
predict the experimental data of Jin (2013) for the combined charge of particles over
a wide range of sizes. For instance, for 1 µm diameter particles, Figure 8.7 indicates
that the particle charge will be approximately 110 e, which corresponds to a surface
charge density of 35 charges/µm2 , or 56 µC/m2 . Because the particle charge for
field charging varies in proportion with the surface area, both of which increase
as d2 , it follows that the surface charge density for particles under field charging
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.2. Mechanisms of Particle Charging 235

+ + + + + +
+ + + +
+ + + +
+ + + +
+ + + + + +

(a)
Figure 8.8. Mechanisms for (a) con-
tact electrification and (b) contact de-
electrification between two spheres of +
+ + +
+
+ + +
+ +
both different and identical materials + + + +
+ + + +
(adapted from Soh et al., 2012). + + + + + +

+ + + + + +
+ + + +
+ + + +
+ + + +
+ + + + + +
Discharge Discharge
+ + +
+ + + +
+
+ +

+ + +
+ + +
+
+ + +

Charges transfer to gas atmosphere

(b)

is independent of particle diameter. Thus, this value of 35 charges/µm2 can help to

approximately determine the magnitude of the field charge for particles with various
sizes. Brocilo et al. (2003) recently made a thorough review of the various kinds
of field and diffusion charging models across the broad range of Knudsen numbers
for free-molecular, transition, and continuum regimes. The results of this study con-
firm the validity of combining the White and Pauthenier-Moreau-Hanot equations,
though other more detailed models are also available.

8.2.3. Contact Electrification

When two particles, made of different materials, are brought into contact and then
separated, an electric charge is usually transferred from one particle to the other
across the contact region, as illustrated in Figure 8.8a. This phenomenon is called
contact electrification or contact charging. It is termed as impact charging when the
contact is fast and in a normal direction, as in a particle collision, and it is termed
as tribocharging, or frictional electrification, when the surfaces of two materials are
rubbed relative to each other, as when a particle rolls or slides along a surface. Figure
8.9 provides a chart, known as the triboelectric series, which illustrates the tendency
of different kinds of materials to acquire a positive or negative charge during contact
electrification. As can be seen from this figure, polar materials such as glass, mica,
fur, and silica easily become positively charged, whereas nonpolar materials such
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

236 Particle Interactions with Electric and Magnetic Fields

Poly(vinylidene chloride) (Saran)

Poly(methyl methacrylate) (PMMA)

Poly(vinyl chloride) (PVC)

Poly(vinyl alcohol) (PVA)

Polyacrylonitrile (PAN)

Polypropylene (PP)
Polyethlene (PE)
Polystyrene (PS)
Nature rubber
Epoxy resion

Teflon (PTFE)
Aluminum
Nylon 6.6

Copper
Amber
Cotton
Paper
Glass

Wood
Silica

Silver
Wool

Steel
Mica

Gold
Silk
Fur

More positive More negative

Figure 8.9. The triboelectric series indicating the tendency of materials for acquiring a positive
or negative charge on contact, based on the series proposed by McCarty and Whitesides (2008).

as Teflon (PTFE), PVC (polyvinyl chloride), PS (polystyrene), and other polymers

tend to become negatively charged upon contact (McCarty and Whitesides, 2008).
Although contact electrification has been studied since antiquity, the underlying
physical mechanisms that control charge separation during particle collisions are
still topics of active discussion. In part, this is because multiple mechanisms may
contribute to the movement of charge in different systems, or even in the same
system. Contact electrification is typically classified into three categories based on the
properties of the contacting materials, that is, metal-metal contact, metal-insulator
contact, and insulator-insulator contact (Matsusaka and Masuda, 2003). For metal-
metal contact, the resulting charge of two colliding particles depends on the electron
transfer from one metal to another. The charge change after separation, qc , is
approximated by the following simple equation
W2 − W1
,qc = Cp (8.2.8)
e
where W1,2 are the work functions of metals 1 and 2, respectively, and Cp is the
capacitance between the two particle surfaces at the equilibrium contact region sep-
aration distance δ and contact area S, given by Cp = ε f S/δ. For instance, a 4 mm
chromium sphere colliding with Ag or Rh particles produces a charge density of
about 0.06 charges/µm2 , or 0.01 µC/m2 , which is so small as to be almost unnotice-
able. For metal-insulator contacts, a similar relationship between qc and the work
function difference between metal and insulator, (WM − WI ), can be assumed, and
interpreted to suggest that the charge carrier is also an electron (Soh et al., 2012).
However, the mechanism appears not to be universal for all types of insulators.
Tribocharging of insulator-insulator contacts is of considerable importance in
applications such as the flow of granular matter, the processing of pharmaceutical
powders, the movements of small particles in dust storms and dust clouds, and so on.
Though the separation of charges by the transfer of electrons plays an important role
in metal-metal contacts, the contact electrification of insulators does not correlate
with bulk electronic properties. Horn et al. (1993) note a correlation between contact
electrification of insulators and the pH of insulating materials. Whitesides and co-
workers confirmed that the charge carriers in insulator-insulator contacts are the
mobile counterions (McCarty and Whitesides, 2008; Thomas et al. 2008; Apodaca
et al., 2010). They introduced a “rolling sphere tool,” in which a rotating bar magnet
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.3. Magnetic Field Forces 237

causes a ferromagnetic steel sphere to roll along a circular path on an insulator

surface. The charging density that was achieved by the sphere in this experiment
was about 200 charges/µm2 , or 32 µC/m2 , which is about five times larger than the
maximum charge obtained by field and diffusion charging in an electrostatic field
(Thomas et al. 2008).

8.3.4. Contact De-electrification

The amount of contact charging achievable in a system of colliding particles even-
tually reaches an asymptotic limit, which implies the possibility of a discharging (or
de-electrification) process that balances the charging process (Soh et al., 2012). This
discharging process is illustrated in Figure 8.8b, where the contact of two previ-
ously charged insulator particles with the same polarity results in the discharging of
both particles. Contact de-electrification occurs when the electric field strength in
the gap within the contact region between two colliding charged particles exceeds
the breakdown voltage of the gas (called the Paschen limit). Above the breakdown
limit, the gas molecules ionize and generate ions (either positive or negative) and
electrons. Depending on the electrical polarity of the particle surfaces, either the
positive ions or the electrons are attracted to the particle surface while the other
ones are repelled from the surface. Attraction of either positive ions or electrons
to the particle surface causes the particles to discharge. Soh et al. (2012) report
that both Nylon and Teflon particles lost approximately 30% of their initial charge
after contact, whereas without particle contact the charge of individual Nylon or
Teflon particles only decreases by about 12% over a 10 hour period. Continuation of
research to clarify physical mechanisms and validate prediction models for contact
electrification and de-electrification is important in order to better account for the
significant role of these phenomena in many particle flow systems.

8.3. Magnetic Field Forces

A particle with charge q traveling with velocity v in a space with electric field E and
magnetic flux density B experiences the Lorentz force

FLOR = q(E + v × B), (8.3.1)

The first term within parentheses is simply the Coulomb force given in (8.1.1), and
the second force arises from the motion of a charged particle within a magnetic field.
This force underlies the operation of electric motors and generators. The magnetic
flux density is related to the volume magnetization vector M and the magnetic field
intensity vector H by

B = μ0 (H + M), (8.3.2)

where μ0 is the permeability of free space and has the value μ0 = 4π · 10−7 H/m. In a
linear magnetic material M and H are related by M = χM H, where χM = (μM /μ0 ) −
1 is the magnetic susceptibility and μM is the magnetic permeability. Substituting
this relationship into (8.3.2) gives

B = μM H. (8.3.3)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

238 Particle Interactions with Electric and Magnetic Fields

A noncharged particle in a magnetic field can also exhibit a nonzero force due to
the phenomenon of magnetophoresis, or MAP. We consider a spherical particle of
linear magnetic material with radius r p immersed in a constant external magnetic
field vector H0 = H0 ex . The vector H has zero curl both inside and outside the sphere,
so a potential ψ can be defined such that H = −∇ψ. Because H has zero divergence,
it follows that ψ satisfies the Laplace equation ∇ 2 ψ = 0 everywhere throughout the
space. The Laplace equation has solutions inside and outside the sphere, given in
spherical polar coordinates by

−H0 r cos θ + (A1 /r2 ) cos θ for r > r p
ψ= . (8.3.4)
−A2 r cos θ for r ≤ r p
If the sphere has magnetic permeability μ p and the surrounding medium has mag-
netic permeability μ f , then continuity of the tangential component of the magnetic
field intensity vector H implies that ψ must be continuous across the sphere surface,
or

ψ1 (r p , θ ) = ψ2 (r p , θ ). (8.3.5)

The second boundary condition requires continuity of the normal component of the
magnetic flux density vector B = μM H, so that
∂ψ1 ∂ψ
μf (r p , θ ) = μ p 2 (r p , θ ). (8.3.6)
∂r ∂r
These two boundary conditions yield solutions for the two coefficients A1 and A2 as
μp − μ f 3μ f
A1 = r3p H0 , A2 = H0 . (8.3.7)
μ p + 2μ f μ p + 2μ f
In the region outside of the sphere, the magnetic field vector has the form of a
uniform magnetic field H0 plus a magnetic dipole with induced dipole moment

m = 4π μ f KM r3p H0 . (8.3.8)

The coefficient KM is the magnetic Clausius-Mossotti function, given by

μp − μ f
KM = . (8.3.9)
μ p + 2μ f
The derivation has a similar form to that given for the induced dipole moment of a
sphere in a DC electric field, with the magnetic permeabilities replacing the electric
permittivities.
The force on a magnetic dipole is given by

Fdipole = μ f m · ∇H. (8.3.10)

Substituting the effective dipole moment given in (8.3.8) yields the magnetophoretic
(MAP) force as

FMAP = 4π μ f KM r3p H0 · ∇H0 = 2π μ f KM r3p ∇H02 . (8.3.11)

As we see from the discussion presented here, under the assumption of a linear
magnetic material, such that M = χM H, the influence of constant magnetic field
on particles closely parallels that of a DC electric field. However, many materials of
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.4. Boundary Element Method 239

n SB
VE x
Figure 8.10. A body with bounding surface SB immersed in x'
an external electric field E0 . The potential field and permit- VI
tivity are denoted, respectively, by I and εI interior to the
body and by E and εE exterior to the body. εI
εE

interest either behave in a nonlinear manner, where the volume magnetization vector
M is not proportional to H, or else exhibit eddy currents as a response to fluctuating
magnetic fields, and this analogy breaks down. In general, the problem of particles in
electric and magnetic fields exhibits a rich physics, particularly when consideration is
given to nonspherical or nonhomogeneous particles, but such topics are outside the
scope of the current book. More extensive consideration of the electromechanics of
particles of different types exposed to electric and magnetic fields can be found in
Jones (1995).

8.4. Boundary Element Method

Section 8.1 discusses how particles, even if they carry no charge, are sensitive to
the electric field vector and its gradient. In conducting a DEM simulation with
charged particles, it is therefore necessary to compute the electric field to which
the particles are exposed. This electric field is influenced by two features of the
flow field – the domain boundaries and the presence of other particles. Domain
boundaries may include electrode surfaces that emit an electric field into the flow,
or dielectric surfaces that nevertheless alter the electric field within the domain. An
electric field may also be introduced by the presence of charged particles, each of
which generates an electric field that decays slowly with distance away from the
particle. Even if the other particles are uncharged, they will exhibit an induced
dipole in the presence of an electric field that will influence forces and torques
on nearby particles. The electric field induced by charged particles is particularly
challenging for DEM simulation because the decay rate of the electrostatic potential
is sufficiently slow that the force induced by a single particle will have significant
influence throughout the flow field, thereby considerably slowing the computation
of the electric field in the presence of many particles. The next section discusses a
computational method for accelerating the computation in the presence of charged
particles. The current section is concerned with use of the boundary element method
(BEM) for simulation of the influence of domain boundaries on the electric field.
After a general introduction to BEM, we discuss two particular problems that occur
when using BEM together with DEM for a bounded particulate flow in an electric
field, as well as methods by which these problems might be resolved.

8.4.1. General Boundary Element Method

We consider a flow within a three-dimensional space VI with boundary surface SB
having outward unit normal n, as shown in Figure 8.10. The region exterior to SB is
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

240 Particle Interactions with Electric and Magnetic Fields

denoted by VE . When the electrostatic assumption applies, the electric field vector E
can be written in terms of a potential  as E = −∇, where  satisfies ∇ 2  = 0. The
potential field (x) at any point x within the interior domain VI can be generated by
placing a sheet charge of strength q(x) along the boundary SB , which can be written
in terms of the local surface charge density qs (x) on SB as q(x) = −qs (x)/εE , where
εE is the permittivity of the external medium. Using the method of Green’s functions
(Marshall, 2001, chapter 12) yields

(x) − 0 (x) = q(x ) G(x − x ) da , (8.4.1)
SB

where 0 (x) is the potential generated by a prescribed external field E0 = −∇0 and
G = G(x − x ) is the Green’s function of the Poisson equation in three-dimensional
space, given by
1
G(x − x ) = − . (8.4.2)
4π |x − x |
The source sheet strength is related to the jump in the gradient of the potential
across SB , and can be written as
q(x) ≡ ∂ (E − I )/∂n, (8.4.3)
where I and E are values of the electrostatic potential just inside and just outside
SB , respectively, as shown in Figure 8.10. Taking the normal derivative of (8.4.1) and
evaluating this equation at a point x on the surface SB yields an integral equation for
q(x) as

1 ∂G 
(∂E /∂n + ∂I /∂n) − ∇0 · n = q(x ) da , (8.4.4)
2 SB ∂n
where ∂/∂n on the boundary SB is set equal to the average of its value on either
side of the boundary. However, this problem specification is not complete, because
as shown in (8.4.3), q(x) is related to the normal gradients of I and E .
The next step in the formulation of the problem depends on whether the material
in region VI is a dielectric or a conductor. For convenience, in the following discussion
we assume that the material in the exterior region VE is a dielectric. As noted
in Section 8.1 the boundary conditions for conducting and dielectric bodies are
different. In particular, the surface of a conducting body is everywhere equipotential,
such that for any point x on the bounding surface SB of a conducting body the
potential satisfies the Dirichlet boundary condition  = B , where B is a constant
value. Consequently, when evaluated at a point x ∈ SB on the surface of a conducting
body, (8.4.1) becomes

B − 0 (x) = q(x )G(x − x ) da . (8.4.5)
SB

Employing the three-dimensional Green’s function and using q(x) = −qs (x)/εE ,
(8.4.5) becomes

qs (x ) 
B − 0 (x) = da , (8.4.6)
SB 4π εE r

where r = |x − x |. Equation (8.4.6) is a Fredholm integral equation of the first

kind for qs (x). For problems such as determination of the potential field around an
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.4. Boundary Element Method 241

electrode, it can be assumed that the value of surface potential B is known as part
of the problem specification. However, in some situations the value of B might not
be known a priori, such as for the case of an isolated conductor. In such cases, we
might instead know the total charge qB on the body surface SB . This total charge can
be used to pose an additional constraint as

qB = qs (x ) ds , (8.4.7)
SB

which is sufficient to make the solution of (8.4.6) determinate.

By contrast, the potential function on the interface between two dielectric bod-
ies, in the absence of a permanent surface charge, satisfies a Neumann boundary
condition of the form

εE ∂E /∂n − εI ∂I /∂n = 0, (8.4.8)

which is a generalization to arbitrary bodies of the boundary condition given in

(8.1.10). Substituting this boundary condition into (8.4.3), the surface sheet strength
can be written as

q = [(εI − εE )/εI ]∂E /∂n. (8.4.9)

Substituting (8.4.9) into (8.4.4) gives a Fredholm integral equation of the second
kind for q(x) as

1 εE + εI ∂G 
− q(x) − ∇0 · n = q(x ) ds . (8.4.10)
2 εE − εI SB ∂n

The integral equations (8.4.6) for a conducting body and (8.4.10) for a dielectric body
must be solved numerically to obtain the surface sheet strength q(x), or equivalently
the surface charge density qs (x). Equation (8.4.6) for conducting bodies and equation
(8.4.10) for dielectric bodies can be discretized using flat panels on the body surface
on which the source sheet strength is constant, using a procedure analogous to that
discussed by Hess and Smith (1967) for aerodynamic applications. This discretization
results in a matrix equation of the form Aq = b, where

Ai j = G (xi − x j ) da, bi = B − 0 (xi ) for conductors, (8.4.11a)
SB, j
  
∂G 1 εE + εI
Ai j = (x − x j ) + δ da, bi = −∇0 · n for dielectrics.
SB, j ∂n i 2 εE − εI i j
(8.4.11b)

In this equation, δi j is the Kronecker delta, which is equal to 0 if i = j and 1 if

i = j. After q is obtained by solution of the matrix equation, the potential field at
any point x ∈ VI is obtained by substitution into (8.4.1) and evaluating the resulting
integral using the panel discretization. The matrix A is dependent only on the relative
location of the body panels. The electric field generated by charged particles or by
the induced field of particle dipoles is accounted for in the earlier equations by the
prescribed field 0 , which enters into the vector b. Consequently, when BEM is
applied to nondeforming bodies using the LU-decomposition method, the matrix A
needs to be generated and decomposed only once at the beginning of the calculation.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

242 Particle Interactions with Electric and Magnetic Fields

8.4.2. Pseudoimage Method for Particles near an Electrode Surface

When the BEM approach for computation of the electric field is used together
with a computational method such as DEM for the particles, a number of problems
with the combined computational approaches can occur that are related to the
relative sizes of the particles and the panels used to discretize the surface SB . These
problems in combining BEM and DEM were examined by Liu et al. (2010), who
also recommended a number of effective approaches for resolving these problems
without significantly increasing the computational time or memory requirements.
The first problem arises from the fact that a particle located near a surface admits
an image field over the surface, where this image field has a length scale on the order
of the distance between the particle and the surface. As the particle approaches
the surface, the length scale associated with this image field becomes progressively
smaller, decreasing to a length on the order of the particle diameter for a particle
touching the surface. If the particle diameter is much smaller than the size of the
panels on the bounding surface SB , as would typically be the case, the inability of
the panel discretization to resolve the small length scale of the image field results in
a considerable loss in accuracy of the electric field computation.
The resolution to this problem proposed by Liu et al. (2010) is based on the
observation that if image is the length scale associated with the image field, such that
image = O(d) as the particle approaches the surface, and κ is the mean curvature at
some point on the surface SB , then κ image = O(κd). Assuming that the particle size
is sufficiently small that κd 1, we find that as the particle approaches the surface
SB , its image field becomes increasingly similar to the image of a particle over a flat
surface. Put another way, if Eimage is the electric field associated with the exact image
field and Eflat is the image field associated with the same particle located at the same
distance from a flat surface, then the characteristic length scale diff of the difference
field Eimage − Eflat must satisfy diff / image = O(κd)−1 1 as the particle approaches
the surface SB .
The resolution for this problem proposed by Liu et al. (2010) is to add a pseu-
doimage field Eflat for a flat surface to the electric field E for all particles that lie
within some distance crit = O(κ −1 ) of SB . For instance, for a flat conducting surface
the image of a point charge is given simply by a charge of the opposite sign located
at the reflection of the original charge (Jackson, 1962, 27). The problem of a point
charge image over the boundary between two dielectrics is discussed by Sometani
(2000). The difference between the exact image field Eimage over the curved bound-
ary SB and the approximate image field Eflat for a flat surface is computed using the
boundary element method, which is achieved simply by including the pseudoimage
field in the prescribed potential 0 . Because the difference field has characteristic
length scale diff = O(κ −1 ), it can be well resolved by the surface panel discretization
even as the particle moves very close to the body surface.
The utility of this pseudoimage method can be demonstrated by comparing the
classic analytical solution for a point charge Q outside of a grounded conducting
sphere with radius r p (Figure 8.11a) to the numerically computed result obtained
using BEM. If the point charge is located at a point (x, y) = (−x0 , 0) and the sphere
is centered at the origin, the exact image of Q has strength QI,exact = r p Q/x0 and is
located at (xI,exact , yI,exact ) = (−r2p /x0 , 0). The distance between the point charge and
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.4. Boundary Element Method 243

10
1.2
r/rp =1.2
1 8

q, x0 q', x0' r 0.8

6
θ E 0.6 r/rp =1.05 E
O
4
0.4
2
0.2

0 0
2.6 2.8 3 3.2
θ
(a) (b)
Figure 8.11. Results for a point charge q outside of a grounded conducting sphere of radius r p : (a) two
circular arcs used for validation of the pseudoimage method in the x-y plane; (b) comparison of the
electric field magnitude on the two arcs computed both with the pseudoimage method (open circles) and
with the standard BEM (filled triangles). The exact solution is indicated by a dashed line. [Reprinted
with permission from Liu et al., 2010.]

the sphere surface is b = x0 − r p , where it is assumed that b r p . The electrostatic

field is computed in three different ways: (a) analytically using the exact image field,
(b) numerically using the standard boundary element method with no prescribed
images, and (c) numerically using BEM with the pseudoimage approach. For the two
numerical solutions, the sphere surface is approximated by 400 quadrilateral panels,
with B = 0. For standard BEM calculations, 0 (x) is set equal to the potential field
of the point charge Q. For the pseudoimage method, 0 (x) is set equal to the sum
of the potential field of the point charge Q plus an approximate image with strength
QI = −Q located at xI = −x0 + 2b.
The magnitude of the electric field is compared in Figure 8.11b for the three solu-
tions described earlier as functions of angle θ along circular arcs with radii r/r p = 1.05
and 1.2, where the point charge is located at angle θ = π . For the arc at r/r p = 1.2,
both the standard BEM solution and the solution using the pseudoimage BEM are
close to the exact solution. However, for the closer arc at r/r p = 1.05, the standard
BEM solution exhibits significant deviation from the exact result, giving a value of
E at θ = π of approximately half the exact solution. By contrast, the pseudoimage
method maintains excellent accuracy for all values of θ , even as we approach the
point charge position at θ = π. Even though the approximate source image location
and strength do not correspond to the exact image location or strength, the difference
between the electric fields induced by the exact and approximate images has a length
scale that is on the order of the sphere radius, which is large compared to the panel
size, and hence this difference field can be well approximated by the BEM approach.

8.4.3. Problems with DEP Force Near Panel Edges

An additional problem occurs as particles move close to panels placed on the surface
of a curved body due to variation of the electric field along the panel surface,
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

244 Particle Interactions with Electric and Magnetic Fields

(a) (b)
Figure 8.12. Variation of the electric field strength at one particle radius away from the
cylinder surface as a function of circumferential direction, computed (a) without and (b) with
panel subdivision. Peak points are closest to panel edges while trough points are closest to
panel centers. [Reprinted with permission from Liu et al. (2010).]

particularly near the panel edges. This variation of the electric field due to panel
discretization is illustrated in Figure 8.12a, which shows the magnitude E of the
electric field vector E for the problem of a charged cylinder of unit radius with a
constant surface potential in an unbounded domain. There is no imposed electric
field far away from the cylinder, so the resulting value of E should be independent of
angle around the cylinder. The cylinder surface is discretized in the circumferential
direction using 20 panels with uniform size, with panel edges located at angles
θn = nπ /10, where n = 0, 1, . . . , 19. The electric field is calculated on a circle with
radius r = 1.05, and the resulting value of E is plotted against θ in Figure 8.12a.
Instead of being constant, the value of E is observed to change periodically with
distance over the surface, with maximum values coinciding with positions of the
panel edges and minimum values coinciding with the panel centers. Particles placed
in this electric field would tend to move toward the panel edges for cases with
positive DEP and toward the panel centers for cases with negative DEP. Of course,
particle drift along the panels is mitigated by the particle inertia and, for adhesive
particles, by the particle adhesive forces that bind particles to the panel surfaces. In
the case of positive DEP, the particle is also attracted to the dipole and, for a charged
particle, monopole image over the panel surface, which may be sufficiently strong to
overcome the tendency to drift along the panel. The percentage variation in electric
field strength along the panel decreases with decrease in the product panel κ, where
panel is the arc length of the panels and κ is the surface curvature. This electric field
fluctuation decays with distance on the order of panel from the body surface, so this
error is primarily important for particles colliding with the body.
The fluctuations in electric field strength are caused by the fact that the panels do
not exactly lie on the surface of the body, but instead only the panel vertices are on
the body surface. This edge effect can be reduced by increasing the number of panels,
which has the effect of reducing the change in angle between panels. Of course, since
the computational cost of the BEM matrix inversion with N panels varies with O(N 3 ),
assuming direct matrix inversion, increase in the number of panels rapidly increases
computational time and memory requirements. Rather than increase panel numbers,
Liu et al. (2010) proposed a panel subdivision method that resolves the problem with
negligible additional computational cost. In this method, the original panels along
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.5. Fast Multipole Method for Long-Range Forces 245

the body surface are used to compute the electric field on particles that are not close
to the body. For near-surface particles, the panels within the region of the surface
closest to each near-surface particle (called the “near-surface region”) are identified.
These “parent” panels are each subdivided into “child” panels, each of which has
the same strength as the corresponding parent panel. However, instead of lying
on the plane of the parent panels, the vertices of the child panels are moved to lie
on the body surface SB . The electric field on each near-surface particle is calculated
using the child panels for panels within the near-surface region and the original
parent panels for panels outside of this region. This subdivision process significantly
reduces the panel edge error by bringing the location of the panel points closer to
the body surface, without the expense of recomputing the panel strengths for the
subpanels.
Figure 8.12 shows the variation in E with θ for cases both with and without
subdivision of the near-surface panels for the constant-potential cylinder problem,
where five child panels are used in the circumferential direction for each parent panel.
Comparison of Figure 8.12a and 8.12b shows that the panel subdivision procedure
reduced the peak-to-peak variation of the electric field magnitude by a factor of
nearly sixty compared to the case with no subdivision.

8.5. Fast Multipole Method for Long-Range Forces

As noted in the previous section, the presence of charged particles within the com-
putational domain induces an electric field that decays slowly with distance away
from each particle. The electric field induced by a charged particle subject to an
external electric field can be approximated using the effective moment method by
the sum of a monopole and dipole field, given respectively by Emonopole = qr/4π ε r3
and Edipole = −∇(p · r/4π ε r3 ). Here q and p denote the particle charge and dipole
moment, respectively, r is the vector from the particle centroid to a space point
x, and r = |r| is the distance from the particle centroid to the space point x. The
magnitude of the electric field vector induced by a dipole decays as O(r−3 ), which is
sufficiently fast that dipoles beyond a relatively small distance from the space point x
can be ignored. The electric field magnitude for an electrostatic source (monopole)
decays with distance as O(r−2 ), which is slow enough that a large group of charged
particles will generate a significant electric field even at fairly large distances. For a
system with N particles, the cost of computing the pair-wise electrostatic interactions
to determine the electric field strength at all particle locations varies in proportion
to O(N 2 ), which for large N can easily become the limiting factor in determining
the computational size. This section introduces the fast multipole method (FMM),
which is a popular and effective approach for accelerating the computation of the
electric field induced by charged particles. This method is also used in other fields
governed by a convolution integral with a kernel that decreases slowly in space, such
as computation of gravitational interactions of massive bodies or the computation
of fluid flows using vortex methods.
The objective of the fast multipole method is to obtain an approximation for
the electric field induced by a group of particles (the source) at a “target” point
x located sufficiently far from the source charges. Let us assume that the various
charged particles within the flow field are sorted into some box structure, so that
at a given level of the box structure all particles are contained in exactly one box.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

246 Particle Interactions with Electric and Magnetic Fields

Let us consider a box numbered within this box structure that contains N charged
particles and has centroid located at x̂ . If q(xi ) is the charge of a particle located at
position xi within box and the interaction kernel is denoted by K(r ) (= r /4π ε r3 ),
where r ≡ x − x̂ and r = |r |, the electric field E (x) generated by the N source
particles at a target point x can be expressed in terms of a multipole expansion as
+∞ 
 +∞ 
+∞
(−1)m+n+k ∂ m+n+k
E (x) = I ,mnk K(x − x̂ ). (8.5.1)
m!n!k! ∂xm ∂yn ∂zk
m=0 n=0 k=0

The box moments I ,mnk of a box are defined by

N

I ,mnk = Qi (xi − x )m (yi − y )n (zi − z )k . (8.5.2)
i=1

The box moments are independent of the target point location, and can thus be
evaluated once per time step for each box. On the other hand, the interaction kernel
K(x − x̂ ) depends only on the difference between the box centroid location x̂ and
the target point location x, and it is independent of the number of particles in the
box.
The speed-up from the multipole acceleration approach comes about from the
fact that there are many fewer boxes than there are charged particles. For instance,
suppose there are a total of N particles contained in M boxes at a given level of the box
structure, where on average N̄ = N/M. For each time step we desire to find the value
of E on each particle induced by all of the other particles. Using the multipole accel-
eration procedure, the first step is to determine the box number in which each particle
is contained. There are several ways to do this sorting procedure, depending on the
box structure, but with an efficient algorithm and a standard box structure, it usually
can be accomplished in O(N) computations. Second, we compute the box moments
I ,mnk for each box, which also requires O(N) computations. Third, we compute and
store the derivatives of the interaction kernel K(x − x̂ ) that appear in (8.5.1) for
each box. Although the range of the indices m, n, and k in (8.5.1) is unbounded,
in a computation these indices would be limited such that m + n + k ≤ H,
where a small value of H is typically assumed so that only a few terms of the mul-
tipole expansion are used. Therefore omitting the factor of H, computation of the
derivatives requires O(M) computations for each target point. A tree-code boxing
structure is proposed by Barnes and Hut (1986) in which the value of M increases
in proportion to log N, and the resulting fast multipole method therefore requires a
computation count of O(N log N). Of course, the FMM has a much higher overhead
than does the direct calculation O(N 2 ) approach, and so we typically find that the
method is only worthwhile compared with direct calculation if N is sufficiently large.
In the FMM presented in this section the electric field is computed at each
particle. As shown by Greengard and Rokhlin (1987), additional speed-up can be
gained for large numbers of particles by also using a local expansion procedure, in
which the electric field is computed on a grid covering the flow and then interpolated
onto the particles via a local Taylor expansion. This local expansion procedure
reduces the operational count for the FMM to O(N), but unless N is very large the
speed-up gained with the local expansion approach may not be worth the additional
overhead and coding complexity.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.5. Fast Multipole Method for Long-Range Forces 247

Figure 8.13. Illustration of the Barnes-

Hut box family with n p = 1.

The fast multipole method can be optimized by using a theoretical error bound
developed by Salmon and Warren (1994) for an H-order multipole expansion, in
which the absolute error eH in the expansion is found to satisfy the bound

1 BH+1 BH+2
eH (x) ≤ (H + 2) H+1 − (H + 1) H+2 , (8.5.3)
(d − bmax )2 d d

where d ≡ |x − xl | is the distance between the target point and the centroid of box
, bmax is the maximum distance between the box centroid and any other point in
the box, and BH is defined by
Nl

BH ≡ qi |xi − xl |H . (8.5.4)
i=1

By specifying the computational precision eH (x) and the multipole order H, (8.5.3)
can be used to determine the lower limit of d for which to apply an H-order multipole
expansion and the maximum size of the box that can be used to achieve the stated
precision.
Prior to implementation of the fast multipole method, it is necessary to form
a family of boxes, such that all particles are contained within at most one box at
each level of the box family. Many different types of box families are used by
different investigators, but two primary examples are discussed here. An illustration
of the Barnes-Hut (1986) algorithm is shown in two dimensions in Figure 8.13. This
algorithm starts by placing a single box over the flow field, which constitutes the
first level of the box tree-structure. At the next level, the box is divided into four
equal-area parts in two dimensions, or eight equal-volume parts in three dimensions.
We call the original box the “parent” box and the different smaller boxes after
subdividing the “child” boxes. The number of particles in each child box is now
counted, and any child box that contains more than a prescribed number n p particles
is again divided into four parts. The procedure is repeated until all boxes at the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

248 Particle Interactions with Electric and Magnetic Fields

Figure 8.14. Illustration of the Clarke-

Tutty box family, with the same points
as in Figure 8.13 and n p = 3.

smallest level contain n p particles or less. Barnes and Hut set n p = 1, but it is often
convenient to use a higher value, say, n p = 20–100, so as to reduce the number of
boxes. Structured box families such as this example have the advantage that they are
easy to construct. The disadvantage of this box family is that the number of particles
per box can vary widely, including generation of some boxes with no particles. This
can particularly become a problem for highly clustered particles, as is often observed
in the presence of particle adhesive force.
An alternative example is the Clarke-Tutty (1994) box family, shown in Fig-
ure 8.14, which uses a tree structure that is designed to maintain approximately the
same number of particles in each box at each level of the structure. The Clarke-Tutty
box family can be generated by a recursive approach in which the particles within
the flow field are initially contained within a single box, which constitutes the first
level of the box structure. This box is then divided along the largest dimension into
two “child” boxes, in such a manner that each child box contains approximately an
equal number of particles. The procedure is repeated until the number of particles
in each box is approximately equal to a prescribed value n p .
Although the Clarke-Tutty box family requires a slightly more complex algo-
rithm to form, the resulting box family is well suited for parallel processing appli-
cations and for use with optimized fast multipole methods due to the fact that it
possesses the minimal number of boxes for a give value of n p and the property that
each box at each level has approximately the same number of particles. For exam-
ple, the set of points shown in Figure 8.13 requires 49 boxes using a Barnes-Hut
box structure with n p = 1, of which 19 are empty. The same set of points in Fig-
ure 8.14 with a Clarke-Tutty box structure and n p = 3 requires nine boxes, where at
the smallest level each box contains either two or three particles. Had we conducted
the Barnes-Hut division in Figure 8.13 with n p = 3, we would have required 16 boxes.
Although these differences might not seem large for a small number of particles, the
differences in number of boxes can become quite significant with a large number of
particles, particularly in three dimensions with highly clustered particles.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.6. Electrostatic Agglomeration Processes 249

Figure 8.15. Schematic diagram of a particle attached to a pla-

nar surface with an upward electric field.

8.6. Electrostatic Agglomeration Processes

Electric field forces between oppositely charged particles will move the particles
toward one another. Even for uncharged particles placed in an electric field, DEP
force between a particle and surrounding particles will make the particles adhere
together. A similar phenomenon occurs with DEP force between a particle and
its image over a conducting or dielectric surface. This electrostatic adhesion force
acts in concert with the van der Waals adhesion force discussed in Chapter 4; how-
ever, the electrostatic force has a very long range compared to the van der Waals
force.

8.6.1. Relative Importance of Electrostatic and van der Waals

Adhesion Forces
The question of the relative importance of electrostatic and van der Waals forces on
particle adhesion on a substrate has long been an issue of debate in the literature.
Although it might sound academic, the problem is of a great deal of importance
for processes that require control of particle adhesion. For instance, it is necessary
to remove charged particles from the surface of semiconductor wafers during the
manufacturing process so as to avoid unwanted electrostatic discharge. Electrostatic
removal processes would be effective for cleaning particles from the wafer if elec-
trostatic forces dominate particle adhesion, but these same measures would make
adhesion by van der Waals forces much worse. Similar control problems arise in
electrophotographic printing processes, in which it is necessary to electrostatically
deposit and remove particles from the paper during different parts of the process
(Rimai et al., 2006).
Different studies examining the problem of electrostatic particle attachment
to a plane surface have been performed using different techniques to measure the
adhesion force, including atomic force microscopy (Zhou et al., 2003). Mathematical
models for electrostatic adhesion processes (Rimai and Quesnel, 2002; Feng and
Hays, 2003; Tang et al., 2006) typically include three types of electrostatic forces
involved in particle adhesion: (1) the Coulomb force, (2) the attraction between a
charged particle and its image charge (of the opposite sign) over the surface, and
(3) DEP attraction between the induced dipole on the particle and its image over
the surface. In the presence of an upward electric field, as shown in Figure 8.15,
expressions for these three terms can be written as (Feng and Hays, 2003)

α1 q2 1
FE,adh = − + α2 qE − α3 π ε d2 E 2 , (8.6.1)
4π ε d 2 4
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

250 Particle Interactions with Electric and Magnetic Fields

400
Detacthment force (nN)
Figure 8.16. Plot of detachment force for a parti-
cle from a plane surface as a function of particle
diameter, showing total detachment force (filled
circles, solid line), van der Waals force (open
200
circles, long dashed line), and electrostatic force
(filled triangles, short dashed line). For d = 2 µm,
the electrostatic force was so small that the total
and van der Waals forces are both shown as a sin-
gle open circle. [Reprinted with permission from
0
0 4 8 12
Rimai and Quesnel (2002).]
Particle diameter (μm)

where the coefficients α1 , α2 , and α3 are functions of the dielectric coefficients of the
two media. A negative term in (8.6.1) is attractive and a positive term is repulsive.
The critical van der Waals force for detachment of a particle from a surface is given
by JKR theory by (4.2.37), or

FC = 32 π γ d. (8.6.2)

If the particle surface charge density qs is assumed to be constant, then the particle
charge q can be written as a function of diameter d as

q = π qs d2 . (8.6.3)

Substituting (8.6.3) into (8.6.1), all three terms of the electrostatic force are found
to vary in proportion to d2 . By contrast, the critical van der Waals force in (8.6.2)
varies linearly in d.
As a consequence of the linear dependence of van der Waals force and the
quadratic dependence of electrostatic force on d, it is expected that the van der
Waals force would dominate for sufficiently small particles and the electrostatic
force would dominate for sufficiently large particles. A study of the dependence
of the detachment force of charged particles from a surface in an electric field on
the particle diameter is reported by Rimai and Quesnel (2002). Data are shown
in Figure 8.16 for the total detachment force, the van der Waals force, and the
electrostatic force as a function of particle diameter in the range 2 µm ≤ d ≤ 13 µm.
The total detachment force in this range is found to vary approximately linearly with
d and to be primarily due to the van der Waals force. Other studies, such as Zhou et
al. (2003), note that the relative importance of van der Waals and electrostatic force
for adhesion depends critically on factors such as surface roughness, where rough
surfaces can have significantly lower van der Waals forces than smooth surfaces
without significant change in electrostatic force.

8.6.2. Particle Chain Formation

Uncharged particles immersed in an electric field exhibit an induced dipole moment
that arises from charge polarization, as discussed in Section 8.1. The dipole moments
from one particle can interact strongly with the induced dipole moments from sur-
rounding particles to influence the structure of agglomerates formed of the particles.
The net effect of this interaction is to cause the particles to align into a chain-like
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.6. Electrostatic Agglomeration Processes 251

Figure 8.17. Schematic diagram of chain-like

agglomerate formed of uncharged particles in an
electric field E, with particles lined up such that
the induced dipoles of neighboring particles are
aligned.

structure, in which the positively charged side of one induced dipole is attracted to
the negatively charged side of the induced dipole on a neighboring particle, and vice
versa, as illustrated in Figure 8.17. The electric field strength is high in the regions
between the particles where the negative and positive sides of the dipoles meet, and
it is low along the sides of the particles.
If the dipole moment on a given particle (labeled particle A in Figure 8.18a)
is given by p = pey , the electrostatic potential field generated by the dipole can be
written in polar coordinates as
p cos θ
= . (8.6.4)
4π ε f r2
The electrostatic force F acting on a second particle (labeled particle B in Figure
8.18a) located at position (r, θ ) has components
3p2 6p2
Fr = − (3 cos2 θ − 1), Fθ = − sin θ cos θ . (8.6.5)
4π ε0 r 4 4π ε0 r 4

B
Fr<0
r Fθ>0 N Fθ>0
θ
y Fr>0 Fr>0
Fθ<0 Fθ<0
S
x P A
Fr<0

(a) (b)
Figure 8.18. Schematic illustrating the sign of the interaction force between two polarized
particles with the same dipole moment p = pey : (a) coordinate system; (b) regions with
positive and negative signs of the radial force Fr .
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

252 Particle Interactions with Electric and Magnetic Fields

Figure 8.19. Numerical computation of the force vector

that would be imposed on a particle, with force vectors
drawn at the particle centroid location, near the tip of
a particle chain. [Reprinted with permission from Liu
et al. (2010).]

 radial component Fr is negative within a conical region with boundary cos θ ∼ ±

The
1/3 located both above and below the particle, labeled N and S in Figure 8.18b.
In this figure, θ = 0 at the N polar axis. Outside of these two cones, Fr is positive. A
second particle B lying in either of the regions N or S will be attracted toward the
original particle A, whereas a particle B lying outside of these conical regions will
be repelled from particle A. The tangential force component Fθ is oriented so as to
always push particle B toward the equatorial plane of particle A. If A and B collide
as particle B is drawn inward at a poles of A, the van der Waals adhesion force
between the two particles can be sufficient to overcome the tangential electrostatic
force. It is noted that this simple discussion does not account for the fact that the
induced dipoles of the two particles will also be influenced by the mutual interaction
of the particles, and hence will not point exactly in the y-direction. Nevertheless, this
simple thought experiment illustrates how one particle might be drawn toward the
poles of a second particle and be repelled from the equatorial region of the particle.
The electrostatic force F(x, y) that would be exerted on a free-moving uncharged
particle traveling near the tip of a particle chain is plotted in Figure 8.19, from Liu
et al. (2010). The particle chain in this example is forced from five particles aligned
with the direction of an external electric field, and all particles are required to have
the same dipole moment p = pey . The plot shows the force vector at the centroid
position of the free-moving particle. The particle radius r p is set equal to 0.1, and
consequently no force vectors are shown with centroid positions within a distance
2r p of the centroid of any of fixed particles in the chain. The region with highest
force magnitude is located near the tip of the particle chain, and it is oriented so as
to draw the free-moving particles onto the end of the chain, hence increasing the
chain length. There are a few small regions along the sides of the chain (indicated
by dashed curves) where a small force points inward toward the particle chain, and
particles moving into one of these regions might be the site of a dendritic branch
from the main chain. Outside of these dashed regions, the force vector is oriented
so as to repel particles from the chain.
The force calculation shown in Figure 8.19 is for illustrative purposes only. In
an actual particle chain, the mutual interaction of the particle dipoles significantly
alters the electric field of the chain. For instance, Jones (1995, 146) estimates that the
effective dipole moment of each particle in a chain of length N increases by a factor
of 2.4 for N = 2, 4.0 for N = 3, and 5.95 for N = 4 compared with what it would
have been for a particle exposed only to the ambient electric field. Of course, this
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

8.6. Electrostatic Agglomeration Processes 253

Figure 8.20. Example of particle dendrite formation for fly ash

from a power plant collecting on a 20 µm diameter polyprope-
lene fiber at three different times. [Reprinted with permission
from Huang et al. (2006).]

1 2 3

mutual interaction effect is included in most DEM simulations, because as discussed

previously in this chapter the induced field of each particle is computed based on the
sum the electric field generated by the electrodes and that generated by the other
particles, where the latter includes both monopole sources from charged particles
and induced dipole sources.
Because the chain-like particle agglomerates are observed to form with either
single-chain or branched structures, they are sometimes referred to as pearl chains
or dendrites, respectively. An example of particle chain formation is shown in Fig-
ure 8.20 for the problem of collection of polarized fly ash particles in air on a
cylindrical fiber oriented orthogonal to the ambient velocity and electric field vectors
(Huang et al., 2006). The particles are observed to form dendritic structures, which
primarily extend outward from the cylinder but also exhibit multiple branches as
they extend away from the fiber. Examples of particle chain formation have also
been observed in DEM simulations. For instance, particle capture by a cylinder in
the presence of an external electric field has been examined using DEM by Park and
Park (2005) and Liu et al. (2010). Simulation results of electrostatic particle capture
by a cylinder from Liu et al. (2010) are shown in Figure 8.21. These simulations
employ a combination of the pseudoimage BEM approach for the electric field,

Figure 8.21. Results of DEM simulations of Liu et al. (2010) for chain formation of uncharged
particles in an air flow past a cylinder subject to a uniform external electric field. [Reprinted
with permission from Liu et al. (2010).]
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

254 Particle Interactions with Electric and Magnetic Fields

including particle images and particle mutual electrostatic interactions. The presence
of particle chain formation is quite noticeable in the figure, with preferential direction
oriented away from the cylinder.
Finally, we would be remiss if we did not note that the simple assumption that
the electrostatic interaction between the particles can be represented using only the
induced dipole fields breaks down as the particles move very close to each other.
Although this assumption produced qualitatively correct results for the particle
electrostatic agglomeration, as we have seen, many investigators have pointed out
that the induced field for touching particles is not accurately represented by a simple
monopole-dipole representation, but instead higher-order multipoles must also be
used. Details of the higher-order calculation can be found in references such as
Washizu and Jones (1996a,b), Techaumnat and Takuma (2003), and Nakajima and
Matsuyama (2002). Collision of electrically charged particles in a surrounding gas
or liquid medium can also exhibit complications due to the electrical breakdown
of the fluid if the electric field strength between the particles exceeds the Paschen
limit, resulting in formation of an arc between the particles (Jones and Saha, 1990).
However, such phenomena are outside the scope of the present text.

REFERENCES

Apodaca MM, Wesson PJ, Bishop KJM, Ratner MA, Grzybowski BA. Contact electrification
between identical materials. Angewandte Chemie-International Edition 49, 946−949 (2010).
Barnes JE, Hut P. A hierarchical O(NlogN) force calculation algorithm. Nature 324(4), 446–
449 (1986).
Brocilo D, Eng B, Eng M. Electrode geometry effects on the collection efficiency of submi-
cron and ultrafine dust particles in wire-plate electrostatic precipitators. PhD dissertation,
McMaster University, p. 41–61 (2003).
Chen D, Du H. A dielectrophoretic barrier-based microsystem for separation of microparti-
cles. Microfluidics and Nanofluidics 3, 603–610 (2007).
Clarke NR, Tutty OR. Construction and validation of a discrete vortex method for the
two-dimensional incompressible Navier-Stokes equations. Computers Fluids 23(6), 751–
783 (1994).
Feng JQ, Hays DA. Relative importance of electrostatic forces on powder particles. Powder
Technology 135–136, 65–75 (2003).
Fiedler S, Shirley SG, Schnelle T, Fuhr G. Dielectrophoretic sorting of particles and cells in a
microsystem. Analytical Chemistry 70, 1909–1915 (1998).
Glover W, Chan HK. Electrostatic charge characterization of pharmaceutical aerosols using
electrical low-pressure impaction (ELPI). Journal of Aerosol Science 35, 755–764 (2004).
Greengard L, Rokhlin V. A fast algorithm for particle simulations. Journal of Computational
Physics 73, 325–348 (1987).
Hess JL, Smith AMO. Calculation of potential flow about arbitrary bodies. Progress in
Aerospace Sciences 8, 1–138 (1967).
Horn RG, Smith DT, Grabbe A. Contact electrification induced by monolayer modification
of a surface and relation to acid–base interactions. Nature 366, 442–443 (1993).
Huang B, Yao Q, Li SQ, Zhao HL, Song Q, You CF. Experimental investigation on the
particle capture by a single fiber using microscopic image technique. Powder Technology
163, 125–133 (2006).
Jackson JD. Classical Electrodynamics. John Wiley & Sons, New York (1962).
Jin X. Research on the capture technology of fine particles in electrostatic precipitator. Master
Thesis, Tsinghua University, Beijing (2013).
Jones TB. Electromechanics of Particles. Cambridge University Press, Cambridge UK (1995).
Jones TB, Saha B. Nonlinear interactions of particles in chains. Journal of Applied Physics
68(2), 404–410 (1990).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-08 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 13:46

References 255

Liu G, Marshall JS, Li S-Q, Yao Q. Discrete-element method for particle capture by a body
in an electrostatic field. International Journal for Numerical Methods in Engineering 84(13),
1589–1612 (2010).
Marshall JS. Inviscid Incompressible Flow, John Wiley & Sons, New York (2001).
Matsusaka S, Masuda H. Electrostatics of particles. Advanced Powder Technology 14(2),
143–166 (2003).
McCarty LS, Whitesides GM. Electrostatic charging due to separation of ions at interfaces:
Contact electrification of ionic electrets. Angewandte Chemie-International Edition 47,
2188–2207 (2008).
Morgan H, Holmes D, Green NG. 3D focusing of nanoparticles in microfluidic channels. IEE
Proceedings – Nanobiotechnology 150(2), 76–81 (2003).
Nakajima Y, Matsuyama T. Electrostatic field and force calculation for a chain of identical
dielectric spheres aligned parallel to uniformly applied electric field. Journal of Electrostatics
55, 203–221 (2002).
Park HS, Park YO. Simulation of particle deposition on filter fiber in an external electric field.
Korean Journal of Chemical Engineering 22(2), 303–314 (2005).
Pauthenier MM, Moreau-Hanot M. Generateurs de haute tension a courants gazeux. Journal
De Physique 5, 193–196 (1937).
Rimai DS, Quesnel DJ. The adhesion of spherical particles: Contributions of van der Waals
and electrostatic interactions. The Journal of Adhesion 78, 413–429 (2002).
Rimai DS, Weiss DS, de Jesus MC, Quesnel DJ. Electrophotography as a means of microfab-
rication: The role of electrodynamic and electrostatic forces. C. R. Chimie 9, 3–12 (2006).
Salmon JK, Warren MS. Skeletons from the treecode closet. Journal of Computational Physics
111, 136–155 (1994).
Soh SL, Kwok SW, Liu HL, Whitesides GM. Contact de-electrification of electrostatically
charged polymers. Journal of the America Chemical Society 134, 20151−20159 (2012).
Sometani T. Image method for a dielectric plate and a point charge. European Journal of
Physics 21(6), 549–554 (2000).
Tang T, Hui C-Y, Jagota A. Adhesive contact driven by electrostatic forces. Journal of Applied
Physics 99, 054906 (2006).
Techaumnat B, Takuma T. Calculation of the electric field for lined-up spherical dielectric
particles. IEEE Transactions on Dielectrics and Electrical Insulation 10(4), 623–633 (2003).
Thomas SW, Vella SL, Kaufman GK, Whitesides GM. Patterns of electrostatic charge and dis-
charge in contact electrification Angewandte Chemie-International Edition 47, 6654−6656
(2008).
Velev OD, Bhatt KH. On-chip micromanipulation and assembly of colloidal particles by
electric fields. Soft Matter 2, 738–750 (2006).
Washizu M, Jones T. Dielectrophoretic interaction of two spherical particles calculated by
elquivalent multipole-moment method. IEEE Transactions on Industry Applications 32(2),
233–242 (1996a).
Washizu M, Jones T. Generalized multipolar dielectrophoretic force and electrorotational
torque calculation. Journal of Electrostatics 38, 199–211 (1996b).
White HJ. Particle charging in electrostatic precipitation. AIEE Transactions 70, 1186–1191
(1951).
Yu C, Vykoukal J, Vykoukal DM, Schwartz JA, Shi L, Gascoyne PRC. A three-dimensional
dielectrophoretic particle focusing channel for microcytometry applications. Journal of
Microelectromechanical Systems 14(3), 480–487 (2005).
Zhang L, Tatar F, Turmezei P, Bastemeijer J, Mollinger JR, Piciu O, Bossche A. Contin-
uous electrodeless dielectrophoretic separation in a circular channel. Journal of Physics:
Conference Series 34, 527–532 (2006).
Zhou H, Götzinger M, Peukert W. The influence of particle charge and roughness on particle-
substrate adhesion. Powder Technology 135–136, 82–91 (2003).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9 Nanoscale Particle Dynamics

Nanotechnology is expected to have major impacts in medicine, catalysis, electronics,

and materials science. The basic science and engineering of nanoparticles increas-
ingly play a central role in many engineering fields, including the environment (air
pollution, climate change), energy utilization (nanoenergetic or nanocatalytic mate-
rials, fossil fuel combustion, and fly ash formation), and biotechnology (medical
diagnostics and drug delivery). Despite the fact that selective use of nanoparticles
dates back to ancient times, the term nanoparticle appears to have first been used at
the international symposium on small particles and inorganic clusters in the 1980s
(Goesmann and Feldmann, 2010). In certain fields, nanoparticles are also called
ultrafine particles, particularly in the older literature in energy and environmental
engineering (Madler and Friedlander, 2007). Typically, a nanoparticle has a diam-
eter ranging in size from about 1 to 100 nm, which is several orders of magnitude
smaller than the micrometer-size particles that have been the focus of the major-
ity of the book. This small size gives nanoparticles a number of unique properties
and it significantly modifies some of the computational approaches that must be
used to simulate nanoparticle systems. Modeling of nanoparticle interaction with
a surrounding fluid is particularly influenced in cases where the particle size is of
the same order of magnitude as, or smaller than, the mean-free-path of the fluid, at
which point noncontinuum effects become important.
A topic of significant interest in the nanoparticle literature deals with meth-
ods for synthesis of nanoparticles with specific desired properties. There is a large
literature on methods to produce particles with desired coatings, electric and mag-
netic properties, hollow and solid cores, various crystalline structures, and so on. A
second area of great interest has to do with the self-assembly of nanoparticles into
larger assemblages or agglomerates, as well as ways to control these self-assembly
processes. Related to this area is the problem of how nanoparticles can be dispersed
over a broad region during a manufacturing process with a desired agglomerate size
and structure. Meanwhile, the rapid development of nanotechnology requires us
to develop more strict control devices for high-efficiency capture of nanoparticles,
which is important considering their serious environmental and health effects.
All of these issues benefit from the ability to accurately model and simulate
nanoparticle motion and agglomeration under different flow fields. This chapter
describes various types of interparticle interactions (van der Waals, electrostatic,
256
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.1. Continuum and Free-Molecular Regimes 257

etc.) and transport phenomena (drag, diffusion, thermophoresis, etc.) at the nano-
scale, which are necessary for accurate modeling of nanoparticle flows at a particle
scale. The focus of the chapter is on nanoparticle aerosols. This choice is made in
part because the mean-free-path of many gases is on the same scale as the nanopar-
ticle diameter, whereas that of a liquid is of the scale of the molecular radius, which
is of the order of 0.1 nm. Consequently, nanoparticles in a liquid can generally be
modeled using standard continuum theory, whereas nanoparticles in a gas require
special treatment to account for noncontinuum effects. Another reason to focus on
nanoparticle aerosols is that many of the most interesting nanoparticle transport
problems occur in aerosols. In particular, many commodity nanoparticle products
are made by flame aerosol processes including carbon blacks, fumed silica, pig-
mentary titania, and optical fibers. Compared with wet chemistry approaches (e.g.,
sol–gel), the aerosol technology has so far proved advantageous for manufacture
of commercial quantities of nanoparticles due to the high-throughput production,
the short processing time, and the simplicity of the one-step process (Madler and
Friedlander, 2007). Nanoparticles with complex composition and most mixed oxide
ceramics can be manufactured more cheaply and with less environmental harm by
aerosol technology than by wet-chemistry processes (Feng et al., 2006).
The first section of this chapter discusses modification to the fluid force on
nanoparticles due to noncontinuum effects, including drag, diffusion, and ther-
mophoretic effects. The second section discusses nanoparticle interaction problems,
including evaluation of adhesive contact models and electrostatic interaction for
nanoparticle systems. The third section discusses Brownian coagulation of nanopar-
ticles, including particle sintering following adhesive contact.

9.1. Continuum and Free-Molecular Regimes

Equation (5.1.15) defines the Knudsen number as the ratio of the fluid mean-free-
path to the particle diameter, or Kn ≡ λ/d. The Knudsen number serves as a measure
for the accuracy of the assumptions of continuum theory. The value of λ can be
obtained from (5.1.11), and has a value of about 68 nm for air at room temperature.
Nanoparticles therefore possess a value of Knudsen number that is of order unity
for large nanoparticles, or significantly larger than unity for small nanoparticles. The
case with Kn 1 is called the free-molecular regime, for which case the flow about
the particle differs significantly from that obtained by continuum theory. The case
with Kn  1 is called the transition regime, and for this case the continuum theory still
approximately holds but it is subject to a number of modifications having to do with
breakdown of the no-slip boundary condition and the presence of Brownian motion.
An estimate of scaling parameters for particles with diameter ranging from
10 nm to 1 µm is given in Table 9.1. This estimate is obtained using fluid length and
velocity scales typical of a suspended particle in the atmosphere. For small particle
sizes, the fluid scaling might be set equal to that of the Kolmogorov (dissipation) scale
eddies of the atmospheric turbulent motion, for which the length and time scales
are given by η = (ν 3 /ε)1/4 and τK = (ν/ε)1/2 , where ν is the fluid kinematic viscosity
and ε is the turbulence energy dissipation rate per unit mass. In atmospheric turbu-
lence, the energy dissipation varies in the range ε = 10−3 − 10−1 m2 /s3 , depending
on height and atmospheric conditions. Assuming an average value of ε = 0.01 m2 /s3
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

258 Nanoscale Particle Dynamics

Table 9.1. Values of the particle Peclet number, slip correction factor, and particle time scale
for nanoparticles with specific gravity S = 2.5 in air at standard temperature and pressure. The
mean free path of air is λ = 68 nm, and the fluid convective time scale based on the typical
Kolmogorov scale of atmospheric turbulence is τ f = 38 ms. The Tabor parameter is
computed for quartz particles with adhesion length scale δA = 0.165 nm

Particle Particle Knudsen Slip Tabor

diameter, time scale, number, correction parameter,
d (nm) τ p (ns) Pe−1/2
p Kn factor, CC λT

10 0.77 24,520 6.8 0.044 0.12

20 3.1 6130 3.4 0.089 0.15
50 19 981 1.4 0.21 0.21
100 77 245 0.68 0.37 0.27
200 309 61 0.34 0.54 0.33
500 1929 9.8 0.14 0.75 0.45
1000 7716 2.5 0.068 0.86 0.57

gives L = 0.76 mm and U = 20 mm/s. In order to estimate the Tabor parameter in

Table 9.1, the relationship γ = A/24π δ 2 between surface energy density γ and adhe-
sion length scale δ is used, where A is the Hamaker constant for the material. The
value of the Tabor parameter is computed for quartz particles with δ = 0.165 nm,
elastic modulus E p = 72 GPa, and Hamaker coefficient A = 6.5 × 10−20 J.
The particle time scale listed in Table 9.1 is given by τ p = m/3π dμ. The listed
values can be compared to the fluid time scale for this problem, τ f = L/U = 0.04 s,
resulting in Stokes numbers StK ranging from 2 × 10−8 for 10 nm particles to
2 × 10−4 for 1 µm particles. The Knudsen number ranges from 0.07 to about 7
over this range of particle diameters, with the resulting Cunningham slip correction
factors CC ranging from 0.86 for 1 µm particles to 0.04 for 10 nm particles. It was
observed in Section 5.5 that the ratio of Brownian force to drag force is proportional
to Pe−1/2
p , where Pe p is the particle Peclet number. Values of Pe−1/2
p listed for this
problem in Table 9.1 range from order unity for 1 µm particles to much greater than
unity for 10 nm particles. These scaling results indicate that both surface slip and
Brownian motion have a strong influence on the transport of nanoparticles, with the
result that the fluid forces acting on nanoparticles are much different from those for
micrometer-size particles, for which these effects are generally of small importance.
The collision and adhesion forces between particles are characterized by the Tabor
parameter λT , as discussed in Chapter 4. Estimated values for the Tabor parameter
listed in Table 9.1 vary from about 0.1 to 0.6 as particle diameter varies from 10
to 1000 nm. These values lie between the regions covered by the DMT and JKR
models, but they are still well above the range of Tabor parameter values for which
the more extreme assumptions of the Bradley model would be appropriate. More
discussion of the effect of small size on nanoparticle collision and adhesion forces is
given in the remaining sections of the chapter.

9.1.1. Drag Force

One of the early consequences of the breakdown of the continuum assumption
at finite values of the Knudsen number is the development of slip on the inter-
face between the particle and the surrounding gas. The effects of surface slip were
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.1. Continuum and Free-Molecular Regimes 259

discussed in Section 5.1.3, where we showed that the effect of surface slip on the
drag force can be accounted for using slip correction factor CC (also known as the
Cunningham correction factor). Using the empirical Millikan expression for CC ,
given in (5.1.22), the drag force becomes
−3πdμ(v − u)
Fd = −3πdμ(v − u)CC = . (9.1.1)
1 + Kn[A1 + A2 exp(−A3 /Kn)]
The selection of the parameters A1 , A2 , and A3 has been reviewed by Li and Wang
(2003a,b). Equation (9.1.1) simplifies to the original Stokes equation for the con-
tinuum region when Kn → 0, for which the drag force is proportional to particle
diameter d. In the free-molecule regime (Kn 1), this equation asymptotically
approaches
−3π d2 μ(v − u)
Fd = . (9.1.2)
λ(A1 + A2 )
In this large Knudsen number limit, the drag force is proportional to d2 , rather
than d.
Using kinetic theory, Epstein (1924) argued that the drag force on an ultrafine
particle in the free-molecule regime can be expressed as a linear combination of terms
resulting from diffuse and specular scattering effects. Introducing an accommodation
coefficient α to represent the fraction of gas molecules that are reflected diffusively
from the particle, the drag force on a particle in the free-molecular regime can be
written as
(
2& πα ' 2 2π kB T
Fd = − 1 + d ρf (v − u), (9.1.3)
3 8 Mg

where Mg is the mass of a gas molecule. The coefficient α can be evaluated experi-
mentally, but it is usually near 0.9 for momentum transfer (Friedlander, 2000).
In Section 5.1.3, it is noted that kinetic theory also provides a relationship
(5.1.12) between the molecular diameter σM and the fluid viscosity μ. Substitut-
ing this relationship
√ into (9.1.2), with the mean-free-path given by (5.1.11), or
λ = kB T/ 2 π σ 2 P, and P = NkB T = ρ f kB T/Mg, where N is the number density
of gas molecules and ρ f is the fluid mass density, the Stokes-Cunningham drag force
for a nanoparticle in the free-molecular regime can be written as
(
15π 2π kB T
Fd = − d ρf
2
(v − u). (9.1.4)
16(A1 + A2 ) Mg

Comparison of (9.1.3) and (9.1.4) yields an approximation for the sum A1 + A2 as

A1 + A2 = 45π / [4(8 + π α)] ≈ 3.274. This approximation is compared to values of
the sum A1 + A2 from various references in Table 9.2, and it is found to be within
about 1.5% of the empirically obtained values.
Using kinetic theory and considering the influence of van der Waals interactions
(via the Lennard-Jones potential) on momentum transfer during collision of a gas
molecule with a particle in the specular and diffuse scattering limits, a rigorous
generalized expression for drag force of spherical particles in the entire range of
the Knudsen number was derived by Li and Wang (2003a,b). Their theoretical
expression agrees well with the empirical correlation (9.1.1) using the parameter
values proposed by Allen and Raabe (1982), given in Table 9.2. We recommend
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

260 Nanoscale Particle Dynamics

Table 9.2. Comparison of theoretical approximation for A1 + A2 obtained from

Equations (9.1.3) and (9.1.4) with empirical values obtained in the literature

Source A1 A2 A3 A1 + A2

Theory 3.274
Davies (1945) 2.514 0.80 0.55 3.314
Fuchs (1964) 2.49 0.84 0.435 3.33
Allen and Raabe (1982) 2.310 0.942 0.298 3.252

that (9.1.1) with the Allen-Raabe coefficients be used for numerical computations,
both because of its ease of implementation and because of its close agreement with
theoretical and experimental results. Figure 9.1 plots the slip correction factor CC as
a function of the particle diameter in air in a range from 1 nm to 100 µm, for which
the Knudsen number reduces from 70 to 7×10–4 , respectively. In particular, for
nanoparticles with diameter less than 50 nm, the Epstein theory (9.1.3) predicts the
drag with less than a 20% difference compared to the empirical Stokes-Cunningham
formula (9.1.1) in the free-molecular regime. Using (9.1.3) and (5.1.12), an expression
for CC for nanoparticles in the free-molecular regime is obtained as
√ 2
32 2 σM ρf & πα '
CC = 1+ d. (9.1.5)
45Mg 8

9.1.2. Brownian Force

In Section 5.4, the particle diffusion coefficient under Brownian motion is given
by the Stokes-Einstein expression Db = kB T/3π μ d. A generalized expression for

CC

d (nm)
Figure 9.1. The slip correction factor CC as a function of particle diameter for spherical par-
ticles in both the free molecular regime and the continuum regime (with T = 293.13 Kρ f =
1.204 kg/m3 , M = 4.81 × 10–26 kg, σM = 0.37 nm).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.1. Continuum and Free-Molecular Regimes 261

Brownian diffusion coefficient that accounts for particle surface slip can be written
as
kB T
Db = . (9.1.6)
3π μ dCC
Using the expression (9.1.5) for the slip correction factor in the free-molecular
regime, (9.1.6) becomes

1 kB T Mg
Db = 2  √ d−2 . (9.1.7)
1 + πα 2π ρf
3 8

Assuming that the absolute pressure P and temperature T are related to ρ f by the
ideal gas law, the Brownian diffusion coefficient in (9.1.7) satisfies Db ∝ P−1 T 3/2 d−2 ,
which is similar to the diffusion coefficient of a molecule. For a larger particle in the
continuum regime, we would instead obtain Db ∝ T 1/2 d−1 . This relationship suggests
that a nanoparticle in the free-molecular regime behaves much like an individual
molecule, in the sense that the kinetic energy of the particle is dominated by the
thermal energy. Knowing the particle diffusion coefficient Db , the mean-square
displacement (MSD) for Brownian motion of the nanoparticles can be predicted
using (5.4.12).

9.1.3. Mean-Free-Path of Nanoparticles

When a fluid molecule hits a particle, the particle velocity and direction persists to
a large extent due to the much larger inertia of the particle. This is in contrast to
the case where a molecule collides which another molecule of a similar type, which
often results in reversal of molecule travel direction with a single collision. As a
consequence, it requires collisions with a large number of fluid molecules to change
the direction of a nanoparticle. The collision of molecules with a nanoparticle occurs
quite frequently, for example, about 1012 collisions per second for a 10 nm particle,
so the change in direction of a nanoparticle can occur over a relatively short length
scale. The average distance a particle travels before it changes direction is defined
as the particle mean-free-path (Fuchs, 1964). The root-mean-square velocity of a
nanoparticle of mass m due to thermal motion in one direction is given by

kB T
v̄ p,rms,r = . (9.1.8)
m
The response time of a particle in fluid is given by τ p = m/(3π μdCC ), with Cc given
by (9.1.5) for the free-molecular regime. The particle mean-free-path λ p is equal to
the product of the root-mean-square velocity and the response time, giving

mkB T
λ p = v̄ p,rms,r · τ p = . (9.1.9)
3π μ dCC
The particle mean-free-path plays an important role in particle deposition processes.
When λ p is larger than d, particle diffusion may exceed inertial impact to become
the dominant mechanism for deposition as well as interparticle collision (Madler
et al., 2006). Figure 9.2 illustrates the ratio of the particle mean-free-path to the
particle diameter over a range of particle sizes from 1 nm to 1 µm with an assumed
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

262 Nanoscale Particle Dynamics

T = 20°C
ρp = 1000 kg/m3
Particle mean free path
10
Particle diameter

Figure 9.2. Ratio of particle mean free path to particle

1 diameter ranging from 1 nm to 1 µm. Particles smaller
than about 20 nm travel an average distance greater
than their own diameter before changing directions
0.1 due to random collisions with gas.

1 10 100 1000
Particle diameter d (nm)

particle density of 1000 kg/m3 . The λ p /d ratio decreases with increasing particle size.
Nanoparticles with diameter less than about 20 nm travel an average distance that
is more than their own diameter under Brownian motion before changing direction.

9.1.4. Thermophoretic Force

Thermophoresis describes particle drift in the presence of a temperature gradient
resulting from the gradient in root-mean-square velocity of the surrounding gas
molecules. The molecular root-mean-square velocity increases with temperature in
proportion to T 1/2 , so that molecular collisions on the “hot” side of the particle import
more momentum than do collisions with the slower-moving molecules on the “cold”
side of the particle. As illustrated in Figure 9.3, the molecular collisions on the hot
side push the particle toward the cold side and those on the cold side push the particle
toward the hot side, with the result that the net thermophoretic force on the particle
is oriented opposite to the direction of the temperature gradient. This effect was first
reported by Tyndall (1870), who noticed a particle-free region around a heated metal
wire that was visible as a dark space in the bright background of scattered light from
the particles. Thermophoresis is of major importance in a variety of nanoengineering

Figure 9.3. Schematic diagram of a particle (large

center sphere) immersed in a gas. The small dots
represent the gas molecules that are randomly fluc-
tuating. The gradient in the mean molecular veloci-
ties produces a force on the particle that pushes the
particle against the temperature gradient toward the
colder region.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.1. Continuum and Free-Molecular Regimes 263

applications. For instance, in the flame synthesis process for nanoparticle films,
thermophoresis is used as a deposition method for porous thin-film manufacture
(Tolmachoff et al., 2009; Zhang et al., 2012c). A modified chemical vapor deposition
process has been used to manufacture optical fiber by deposition of particles on the
tube wall (Park et al., 1999). The semiconductor industry uses thermophoresis to
avoid “light point defects” on wafers and lithographic masks by particle deposition
(Stratmann et al., 1988; Ye et al., 1991). Thermophoresis can also be used as a high-
efficiency capture mechanism to remove nanoparticles from high-temperature gas
streams, where the collecting medium is cooler than the surrounding gas.
Similar to the particle drag force, the thermophoretic force is strongly dependent
on the Knudsen number. In the free-molecule regime (Kn 1 1), Waldmann (1959)
proposed an expression for thermophoretic force as
(
8 κtr d2 2 2π Mg
Fth = − ∇T = − κ d2 ∇T, (9.1.10)
15 c̄ 15 kB T tr

where the mean speed c̄ of gas molecules is given by c̄ = 8kB T/π Mg and the gas
thermal conductivity κtr pertains to the translational part of the thermal energy. κtr
is linearly related to the fluid viscosity μ as κtr = 15kB μ/(4Mg ). The thermophoretic
force is often counteracted by the fluid drag on the particle, and under equilibrium
conditions the combination of these two forces leads to a terminal particle drift
velocity that can be derived by equating (9.1.10) and (9.1.3) to obtain

κtr ∇T 3μ ∇T
vth = −  πα
 =−  πα
 . (9.1.11)
5 1+ 8
kB N T 4ρ f 1 + 8
T

From the equation derived using the Waldmann theory, the thermophoretic drift
velocity for particles in the free-molecular regime is observed to be independent of
the particle size. However, the Waldmann analysis employs an assumption that the
particles collide as rigid bodies, which is of questionable validity for nanoparticles.
Li and Wang (2004) examined the influence of nonrigid body collision by accounting
for van der Waals or other long-range forces between the particles and surrounding
gas molecules, and these authors find that long-range forces play a significant role
for particles with diameter less than about 10 nm. This influence was expressed in
terms of reduced collision integrals in the kinetic theory approach. The resulting
expression for the thermophoretic force obtained with this modified theory is
(
 
2 2π mr  (1,1) (1,2) 
Fth = − κtr d2 ∇T 5avg −6avg , (9.1.12)
15 kB T

where mr ≡ m p Mg/(m p + Mg ) is the reduced mass of the gas molecule and a particle
(1,1) (1,2)
and avg and avg are the average reduced collision integrals for specular and
diffuse scattering, respectively (Li and Wang, 2004; Li, 2005). The corresponding
thermophoretic velocity is given by
 (1,2)

6avg κtr ∇T
vth = − 1− (1,1) . (9.1.13)
5avg NkB T
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

264 Nanoscale Particle Dynamics

Figure 9.4. Ratio of thermophoretic drift velocity with nonrigid collisions (i.e., van der Waals
interaction) to that with rigid body collisions for Cu2 O and soot particles. (Based on calcula-
tions of Li and Wang, 2004.)

(1,1) (1,2)
In the limit of rigid body collision, both avg and avg are equal to unity and
mr approaches the molecular mass M, so that (9.1.12) reduces to the Waldmann
equation (9.1.10) and (9.1.13) reduces to (9.1.11).
Figure 9.4 shows the ratio of vth with nonrigid body collisions to that with rigid
body collisions for Cu2 O and soot particles, obtained using (9.1.11) and (9.1.13).
Soot is assumed to be composed of pure carbon and its density is chosen to be
1,800 kg/m3 . The coefficient α in the Waldmann formula is chosen to be 0.9 for both
Cu2 O and soot, but the corresponding coefficient in Li and Wang’s formula is varied.
For ultrafine particles, with diameter around 1 nm, the value of the thermophoretic
velocity obtained from the Waldmann theory is significantly larger than that obtained
from the more refined Li-Wang analysis. For particles with diameter larger than 30
nm, the predictions of the two theories are within 10% of each other.
Compared with the free-molecular regime, the analysis of thermophoresis for
intermediate Knudsen numbers (in the transition regime) is relatively complicated.
Most theories for thermophoretic force with arbitrary Kn are derived by solution
of the Boltzmann equation, or some approximation of the Boltzmann equation.
Similar to (9.1.10), a general form for thermophoretic force at intermediate Knudsen
numbers can be written as
√
π κ d2
Fth = − h(Kn, κ21 ) tr ∇T, (9.1.14)
8 c̄
where h(Kn, κ21 ), sometimes called the dimensionless thermophoretic force, is a
function of the Knudsen
 number Kn and the gas/particle thermal conductivity
ratio κ21 , and c̄ = 8kB T/π Mg. Combining (9.1.14) with the expression (9.1.1) for
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.1. Continuum and Free-Molecular Regimes 265

Figure 9.5. Schematic illustration of nanoparticle deposition adjacent to the substrate during
thin film production.

drag force, the equilibrium thermophoretic velocity on a particle with broad Kn

range can be predicted. Expressions for h(Kn, κ21 ) were proposed by Sone and
Aoki (1983), Yamamoto and Ishihara (1988), and Loyalka (1992) on the basis of
the linearized Boltzmann equation. Talbot and co-workers (Talbot et al., 1980;
Talbot, 1981) present an empirical formula for h(Kn, κ21 ) based on the assumption
of near-continuum conditions. Comparisons between different expressions for ther-
mophoretic force are given by Talbot (1981), Davis and Schweiger (2002), Li (2005),
and Madler and Friedlander (2007).

9.1.5. Competition between Diffusion and Thermophoresis

during Deposition
We consider particle deposition in an axisymmetric stagnation-point flow with strain
rate s = U0 /2h, where U0 is the characteristic velocity along the symmetry axis at a
distance h from the impingement surface. The axisymmetric Heimenz flow admits
√
a boundary layer with characteristic thickness δBL = 2.8 ν/s, where ν is the fluid
kinematic viscosity (White, 2006). As shown in Figure 9.5, the boundary layer can be
divided into three layers based on comparison of the particle thermophoretic velocity
and the axial gas velocity. The convection-controlled layer is the outermost layer of
the boundary layer, in which the axial gas velocity dominates the particle transport.
The thermophoresis-controlled layer is the layer closest to the impingement surface,
in which the gas velocity is small compared with the particle thermophoretic velocity.
This inner layer has a thickness of between 2 and 10% of the total boundary layer
thickness, depending on the temperature gradient. Between these two regions is the
transition layer, in which the gas velocity is of the same order of magnitude as the
thermophoretic velocity.
Assuming that the particle concentration is a function only of the axial coordinate
x, the equation for particle transport via convection, thermophoresis, and diffusion
in the boundary layer is given by
 
dφ d dφ
(u + vth ) m = Db m , (9.1.15)
dx dx dx
where φm is particle mass concentration and vth is the particle thermophoretic veloc-
ity. Particle deposition on the wall is controlled by the thermophoretic-controlled
layer within which u vth , where vth is assumed to be approximately constant within
this inner layer. Defining a Peclet number based on the thermophoretic velocity as
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

266 Nanoscale Particle Dynamics

Peth = vth δBL /Db , (9.1.15) can be integrated in x to obtain the particle deposition
flux ṁp in the thermophoretic-controlled layer as

ePeth
ṁp = φm1 vth , (9.1.16)
ePeth − 1
where φm1 denotes the mass concentration at the top of the thermophoretic-
controlled layer. When Peth 1, the fraction in (9.1.16) quickly approaches unity,
giving ṁp = φm1 vth , so that the particle mass flux is controlled entirely by ther-
mophoretic transport. For Peth < O(1), the deposition flux is influenced both by the
thermophoretic force and by Brownian diffusion. The fluid convection is important
for determining the coefficient φm1 in the equation.

9.2. Nanoparticle Interactions

When two nanoparticles move close to each other, they experience a number of dif-
ferent types of interaction forces. For particles on the larger end of the nanoparticle
size spectrum, the length scale of adhesive forces remains small compared with the
particle diameter and it is reasonable to use the adhesive contact models discussed
in Chapter 4, which assume that adhesive forces have no influence until the particles
collide. However, for particles on the smaller end of the nanoparticle size spectrum,
the length scale of adhesive forces is a significant fraction of the particle diameter
and action of adhesive forces prior to collision becomes important.
This section first discusses the collision and adhesion forces between nanoparti-
cles in the large end of the size spectrum, followed by a discussion of the influence
of long-range effects of adhesive force for very small nanoparticles. Here, we refer
to “long range” as having a length scale on the order of the particle diameter. The
final two parts of this section discuss electrostatic Coulomb and dipolar interac-
tions between nanoparticles, which play a major role in the synthesis, design, and
manipulation of nanostructured materials.

9.2.1. Collision of Large Nanoparticles

The accuracy of the continuum approximation for nanoscale collision problems with
particles in the large end of the nanoparticle size spectrum, given approximately
by d > 20 nm, has been examined by a number of investigators using molecular
dynamics (MD) to simulate the atomic interactions in nanoparticle collision prob-
lems (Wenning and Müser, 2001; Luan and Robbins, 2005). Luan and Robbins (2005)
examined both normal and tangential (frictional) forces for contact between a par-
ticle (with either a spherical or cylindrical shape) and a flat substrate. Results are
shown in Figure 9.6 for a cylinder with radius of 100 atomic diameters (approximately
30 nm) colliding on a flat substrate. The MD simulation results are compared with
continuum theory predictions obtained using the Maugis-Dugdale adhesive contact
model. Use of this model is consistent with the range of Tabor parameter values
for nanoparticles in this size range listed in Table 9.1. Figure 9.6 compares the MD
simulations with continuum theory for normal displacement, contact region radius,
and tangential friction force as a function of normal force. The normal displacement
results exhibit excellent agreement between the MD simulations and continuum
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.2. Nanoparticle Interactions 267

(a) 4
Displacement
3

δN / σM
2
Bent commensurate
Bent incommensurate
1
Amorphous
Stepped
0
(b)
Figure 9.6. Comparison of the predictions Contact Radius
of the Maugis-Dugdale continuum theory
(solid lines) and molecular dynamics simula- 15
tions (symbols) for different atomic arrange-

a / σM
ments with a 30 nm diameter cylinder col- 10
liding with a flat substrate, for (a) normal
displacement; (b) contact region radius; and
(c) frictional force as a function of the 5
dimensionless normal force. The dashed and
dotted lines in (c) are best fits to the data.
0
[Reprinted with permission from Luan and (c)
Robbins (2005).] Friction

0.004
Fs / (lRE)

0.002

0
0 0.005 0.01 0.015
Fn / (lRE)

theory. The contact region radius obtained from MD solutions is about 10–20%
higher than the continuum theory predictions, but the trend is the same between
the two. It is noted that the Maugis-Dugdale model is approximate, so this modest
disagreement might well be due to inaccuracy in solution of the continuum problem.
Whereas the normal force solutions agree well between MD simulation and
continuum theory, the frictional sliding force obtained from MD is highly sensitive
to the atomic arrangement within the nanoparticle, as shown in Figure 9.6c. Although
the MD simulations for materials with crystalline atomic structure agree reasonably
well with the continuum predictions, the frictional force is much smaller for materials
with random amorphous atomic structure. The study by Wenning and Müser (2001)
used MD simulations to examine the relationship between frictional force Fs and
normal load Fn under static friction conditions for contacts with radius of curvature
ranging between 1 and 15 nm. Whereas the Amonton theory used in continuum
mechanics descriptions assumes that friction force is proportional to normal load,
Wenning and Müser only observe linear behavior for crystalline (commensurate)
atomic arrangements. For amorphous atomic arrangements, the friction force varies
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

268 Nanoscale Particle Dynamics

1
10

0 Slope = 0.97 Figure 9.7. Plot showing normalized sliding

Fs (nN)

10
force Fs versus normal load Fn (in nano-
Newtons) for nanoparticles with commensu-
-1 rate structure (circles, solid line) and amor-
10
Slope = 0.63 phous structure (squares, dashed line), along
with best-fit lines, based on the MD simula-
-2
10 tion results of Wenning and Müsser (2001).
The upper (solid) line has slope 0.97 and the
-3 lower (dashed) line has slope 0.63.
10 -1 2
10 100 101 10
Fn (nN)

more closely to the 2/3rds power of the normal load (Figure 9.7), and yields results
for friction force that are considerably smaller than for crystalline materials. This
observation is supported by experimental results with friction force microscopes
(Schwartz et al., 1997).
Normal force models similar to the DMT model were used by Eggersdorfer et al.
(2010) and Higashitani et al. (2001) for simulation of the restructuring and break-up
of soft agglomerates of nanoparticles under a shear flow. The normal force in this
model is given by
⎧
⎪ A
⎪
⎪ kN δN + ηN (vR · n) − 2 R δN > 0
⎪
⎪ 6δ
⎨
Fn = A (9.2.1)
⎪ − R −δC ≤ δN ≤ 0
⎪
⎪ 6(δ + δ ) 2
⎪
⎪
N
⎩
0 δ < −δC

where the stiffness kN can be written as kN = 4E a/3 using the Hertz contact result
(3.2.11), the normal dissipation coefficient ηN is determined by the TTI model given
in (3.3.12), and δ is the minimum separation distance within the contact region, with
value ranging between 0.15 and 0.40 nm. For convenience in simulation, Higashitani
et al. (2001) proposed a linear empirical relationship of kN = π ri Ei , where E =
Ei /2(1 − νi2 ). The first line in Equation (9.2.1), valid for δN > 0, is just the DMT
model (4.2.15), modified by addition of the normal damping term. The simulation
of Eggerdorfer et al. (2010) indicates that the variation of elastic moduli has no
influence on the agglomerate breakage time or general behavior. This observation is
in agreement with other studies of DEM of adhesive particles. For instance, for the
problem of particle capture by a circular cylinder, Li and Marshall (2007) observed
that the particle dynamics is insensitive to the value of the elasticity parameter
 (≡ E/ρ pU02 ), defined as the ratio of elastic force to particle inertia. The second
line in (9.2.1), valid for −δC ≤ δN ≤ 0, is an approximate expression of the van der
Waals force obtained by applying Hamaker theory for two spheres separated by a
small distance. The critical pull-off distance δC is arbitrarily set as 10 nm in these
papers, where van der Waals force drops to less than 1% of the maximum attractive
force.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.2. Nanoparticle Interactions 269

Tangential forces play a significant role in the formation of particle agglomerates.

Pantina and Furst (2005) experimentally demonstrated the existence of tangential
forces between bonded nanoparticles and the capability of these bonds to supporting
bending moments, and then presented a possible explanation for the tangential forces
in terms of the JKR theory for adhesive surfaces. They related the single bond rigidity
to the work of adhesion derived from the JKR theory and to the elastic modulus
of the particles. Becker and Briesen (2008) developed tangential force models for
DEM simulations including both rolling and twisting frictions, in which several model
parameters were determined by the experiments of Pantina and Furst (2005). The
more detailed effects of adhesion on the sliding, twisting, and rolling frictions are
summarized in Sections 4.2.3 and 4.2.4.

9.2.2. Collision of Small Nanoparticles

As particles get increasingly smaller within the nanoparticle size range, the models
used for adhesive contact forces and torques begin to break down. As discussed
earlier, this occurs first with the models for tangential resistance, as indicated by
observations that the atomic arrangement within the nanoparticle has a significant
influence on the sliding resistance. At very small particle sizes, the models for normal
force begin to break down in cases where the particle diameter is about an order
of magnitude larger than the length scale of the van der Waals force, characterized
by the equilibrium gap thickness δ. For such cases, the primary difference with the
theory discussed in Chapter 4 is the fact that the van der Waals force begins to have
a significant influence on the particle dynamics prior to collision of the particles.
For nanoparticles with diameters smaller than about 10 nm, the Tabor parameter
is sufficiently small (Table 9.1) that we can adopt the assumption employed in the
DMT model and consider the van der Waals force and the repulsive elastic force
separately, so that the particle deformation is assumed to have negligible effect
on the van der Waals force. The interaction energy ws−s (h) between two spherical
particles with radii r1 and r2 and separation distance h is given by

A12 2r1 r2 2r1 r2
ws−s = − +
6 (r1 + r2 + h) − (r1 + r2 )
2 2 (r1 + r2 + h)2 − (r1 − r2 )2
 
(r1 + r2 + h)2 − (r1 + r2 )2
+ ln , (9.2.2)
(r1 + r2 + h)2 − (r1 − r2 )2

where the Hamaker coefficient A12 between media 1 and 2 can be determined by
(4.1.7). For two equal-size particles (r1 = r2 = d/2), this equation reduces to
  
A d2 d2 d2
ws−s = − + + ln 1 − . (9.2.3)
6 2(d + h)2 − 2d2 2(d + h)2 (d + h)2
The van der Waals force between the spheres is given by
dws−s A 1
FvdW = − =− , (9.2.4)
dh 6d x (x + 1)3 (x + 2)2
2

where x = h/d is a dimensionless separation distance. Equation (9.2.4) indicates

that the van der Waals force at a certain separation distance (relative to particle
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

270 Nanoscale Particle Dynamics

100
10-1 Coulomb Dipolar Force
-2
10 vdW Force (MD) Figure 9.8. Interaction forces between
-3 vdW Force (Hamaker Approach)
10 two co-oriented 3 nm TiO2 nanopar-
10-4 ticles at different separation distances.
Forces (nN)

Data are for MD simulations of the

10-5
dipolar force (+ symbols) and the
10-6 van der Waals force (O symbols), and
10-7 for prediction of the van der Waals
force from Equation (9.2.3) (solid line).
10-8
[Reprinted with permission from Yan
10-9 et al. (2010).]
10-10
3 8 13 18 23 28
Interparticle distance (nm)

diameter) scales inversely with particle diameter. Figure 9.8 shows the interaction
forces between two approaching 3 nm charge-neutral anatase TiO2 nanoparticles at
273 K as a function of the interparticle center-to-center distance (d + h), calculated
by a molecular dynamics simulation with a Matsui-Akaogi potential that includes
electrostatic, van der Waals, and repulsive interactions (Matsui and Akaogi, 1991).
The van der Waals force predicted by (9.2.4) agrees well with that obtained by the
MD simulation. At sufficiently large interparticle separation distance, the adhesive
force is dominated by the dipole-dipole interaction of the nanoparticles, which is
discussed in the next section.
The significance of the long-range effect of van der Waals force can be evaluated
by comparing the van der Waals interaction given in (9.2.3) with the particle kinetic
energy. As noted at the beginning of this chapter, nanoparticle dynamics prior to
particle collision is dominated by a balance between the drag force and the Brownian
force. For very large nanoparticles with diameter of roughly 1,000 nm, the particle
inertial force is of a similar order of magnitude to the drag and Brownian forces, but
for much smaller nanoparticles the particle inertia is significantly smaller than drag
or Brownian force. Taking the ratio of |ws−s | from (9.2.3) with the energy associated
with Brownian motion, given by (3/2)kB T , gives the ratio of the adhesive energy to
the particle kinetic energy as
  
WA A 1 1 1
= + + ln 1 − . (9.2.5)
WB 9kB T 2(1 + x)2 − 2 2(1 + x)2 (1 + x)2
This ratio is plotted as a function of dimensionless separation distance x = h/d in
Figure 9.9, assuming room temperature (T = 300K) and values of the Hamaker
constant for particles interacting in air of A = 4 × 10−20 J and 4 × 10−19 J, where
the latter value applies primarily to metallic particles. The energy ratio is equal to
unity for xcrit = 0.1 with the smaller value of Hamaker coefficient, with Brownian
motion dominating adhesion for values of x > xcrit . For the larger value of Hamaker
coefficient, the critical separation distance increases to xcrit = 0.34.
Fanelli et al. (2006) developed a model based on DEM to simulate the agglom-
erate behavior of nanoparticles with size below 4 nm. They used a Born repulsive
force model for the elastic force instead of the elastic force expression obtained
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.2. Nanoparticle Interactions 271

2
10

1
10
Energy ratio

0
10

-1
10

10-2

-3
10
10-2 10-1 100
x
Figure 9.9. Ratio of the adhesive energy to the Brownian thermal energy of a nanoparticle
as a function of the dimensionless particle separation distance x = h/d. Plots are given both
for Hamaker constants of 4 × 10−20 J (solid line) and of 4 × 10−19 J (dashed-dotted line). The
point where the two energies are equal is indicated by a dashed line.

from Hertz theory. Results of MD simulations for the normal contact force of two
identical nanoparticles are reported by Sun et al. (2013) using particle diameters of
1.95, 4.01, 6.13, or 8.24 nm. Results of these simulations comparing the MD results
with predictions of the JKR and DMT models are shown in Figure 9.10. The DMT
predictions agree reasonably well with the MD predictions, but the JKR predictions
indicate that the normal force is too low. This difference between DMT and JKR is
consistent with what would be expected from consideration of the Tabor parameter
values for particles of this small size. This work focuses on the initial stage of particle
contact. At later times additional bonding mechanisms can result in sintering of the
particles, as discussed in Section 9.3.

9.2.3. Long-Range Interparticle Electrostatic Forces

A theoretical calculation of nanoparticle charging from Sahu et al. (2012) based on
the attachment of ions to particles suggests that particles smaller than 10 nm are
charged primarily by diffusion charging, whereas the experimental charge measure-
ments shown in Figure 9.11 confirm that particles smaller than 7 nm carry no charge.
Particles smaller than 25 nm are primarily observed to carry no more than one ele-
mentary unit of charge. In general, it is very difficult to charge nanoparticles by field
charging mechanisms when the saturation charge is less than one elementary charge.
Even without a net charge or the presence of an external electric field, the
nanoparticles themselves can possess permanent dipoles whose interaction leads to
important forces, particularly during particle collisions, as seen in Figure 9.8. Some
dipoles arise from the asymmetry of wurzite crystalline structure of the nanoparti-
cles, such as in CdSe nanocrystals. Blanton et al. (1997) reported values of dipole
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

272 Nanoscale Particle Dynamics

6 10
(a) (b)
5 8 MD
MD JKR
4 JKR
6 DMT
d=4.01 nm
DMT
d=1.95 nm
3

Fn (nN)
Fn (nN)

4
2

1 2

0 0

-1
0.0 0.1 0.2 0.3 0.0 0.1 0.2 0.3
δδNN (nm)
(nm) δN (nm)
δ N (nm)
(a) (b)
70 80
60 (c) (d)
MD MD
50 60 JKR
JKR
d=6.13DMT
nm DMT
40 d=8.24 nm
Fn (nN)

40
Fn (nN)
30
20
20
10
0 0
-10
0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 1.0
δN (nm)
δ N (nm) δδNN (nm)
(nm)
(c) (d)
Figure 9.10 Comparison of the normal forces Fn between silica nanospheres of different
diameter of (a) 1.95; (b) 4.01; (c) 6.13; to (d) 8.24 nm obtained from the MD simulations of
Sun et al. (2013) (symbols) with those predicted by JKR model (solid line) and DMT model
(dashed line).

moments of 3.4 and 4.6 nm CdSe nanocrystals as 25 D (Debye) and 47 D, respectively.

For some nanocrystals with centrosymmetric lattice, for example, ZnSe and TiO2 ,
the asymmetrical distribution of ions at the surface could also create permanent
dipoles, verified by both experiments (Shim and Guyot-Sionnest, 1999) and more
recently by molecular dynamics (MD) simulation results (Yan et al., 2010). Figure 9.8

Figure 9.11. Fraction of particles charged after

exposure to a unipolar ion source, with ion con-
Fraction charged

centration n and charging time t. Open symbols are

used for the data reported by Pui et al. (1988) for
NaCl with nt = 3.0 × 106 , 7.2 × 106 , and 10.3 × 106
s/cm3 indicated by squares, circles, and triangles,
respectively. Filled symbols are used for the data
reported by Adachi et al. (1985) for ZnCl2 with
nt = 0.58 × 106 and 1.2 × 106 s/cm3 indicated by
squares and circles, respectively.
d (nm)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.2. Nanoparticle Interactions 273

Figure 9.12. Self-assembly of TiO2 nanoparticles under dipole-dipole interaction. [Reprinted

with permission from Penn and Banfield, 1999.]

shows the interaction forces between two approaching charge-neutral anatase TiO2
nanoparticles with diameter 3 nm at 273 K as a function of the interparticle center-
to-center distance. The dipole-dipole interaction energy between two nanoparticles
can be expressed by

1 p1 · p2 − 3(n · p1 )(n · p2 )
w12
d−d
= , (9.2.6)
4π ε0 (r1 + r2 + h)3

where p1 and p2 are the permanent dipole moments of the two nanoparticles, r1 and
r2 are the radii, and n is the unit vector between the particle centroids. Equation
(9.2.6) is a close approximation to the MD simulation predictions by Yan et al.
(2010) and Zhang et al. (2011). As seen in Figure 9.8, the electrostatic force resulting
from the dipole-dipole interaction is several orders of magnitude larger than the van
der Waals force at large separations (>10 nm). The additional long-range attraction
provided by dipole-dipole interaction enlarges the effective particle collision radius,
which serves to shorten the collision time scale.
Unlike the isotropic nature of the van der Waals force, the dipole-dipole inter-
action acts to align particles as they collide, which is important for certain nanopar-
ticle self-assembly processes. Figure 9.12 shows an electron microscope photo of
a self-assembled structure of TiO2 nanoparticles from Penn and Banfield (1999).
The particles in this structure are aligned end-to-end due to dipole-dipole inter-
action. Nanoparticles are aligned as they approach each other, resulting in prefer-
ential orientation at the time of attachment. Similar observations of nanoparticle
alignment during self-assembly are reported for CdTe nanoparticles by Tang et al.
(2002). The relative strength of the interaction potential energy compared with the
energy of Brownian fluctuations is an important factor in determining whether this
self-assembled structure will be achieved. A detailed discussion of the relationship
between nanoparticle self-assembly, Brownian motion, and interparticle forces is
given in the review by Bishop et al. (2009).
As the particle scale reduces, thermal fluctuations play an increasingly important
role in the particle dynamics, especially at high temperatures. Figure 9.13 shows
the effect of temperature on the dipole moment of TiO2 nanoparticles, obtained by
molecular dynamics simulation. As the temperature increases from 273 K to 1673 K,
the magnitude of the time-averaged dipole moment dramatically decreases, from
60.1 D to 2.1 D. As the temperature further increases, the dipole moment remains
below 4.0 D with relatively small deviation. The decrease in magnitude of the time-
averaged dipole moment caused by an increase in temperature can be attributed to
the increasing fluctuation of the instantaneous dipole direction. Inset A of Figure 9.13
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

274 Nanoscale Particle Dynamics

70

Time-averaged Dipole Moment (D)

A B:973K
60 A:273K
B
A A
50

40 38 D 38 D

30 C
C:1673K

20 B
B
10 38 D
C
0 C
0 400 800 1200 1600 2000 2400 2800
Temperature (K)
Figure 9.13. Time-averaged dipole moments of a TiO2 nanoparticle at different temperatures.
Insets A, B, and C show the distribution of dipole vectors projected onto the x–y plane along
the time period at different temperatures. [Reprinted with permission from Yan et al. (2010).]

shows the distribution of dipole vectors projected onto the x–y plane at 273 K. The
distribution of the fluctuating dipole directions is relatively narrow and is constrained
to angles close to π4. As the temperature increases to 973 K (Inset B), the direction
of the dipole vector becomes more widely spread out, but the distribution is still
obviously asymmetrical, resulting in a dipole moment value of 21.8 D. When the
temperature further increases to 1,673 K (Inset C), the dipole vectors have widely
different orientations spreading over the entire angular range, with the result that the
time-averaged dipole moment is very small. This phenomenon is caused by the fact
that Brownian motion becomes progressively more important as the temperature
increases, eventually outweighing the energy associated with dipole-dipole interac-
tion, leading to random nanoparticle aggregates rather than an ordered assembly.

9.3. Time Scales of Nanoparticle Collision-Coalescence Mechanism

A great deal has been learned about the formation of solid nanoparticles using
inorganic metal oxides in flame aerosol processes, as well as other kinds of high-
temperature vapor-phase methods, based on the collision-coalescence mechanism.
As illustrated in Figure 9.14, this mechanism is a limiting case in which the chemical
reactions that lead to particle formation are very rapid, so that all particles stick to
each other when they collide (i.e., zero activation energy for nucleation). Therefore,
the competition between the time scales for particle–particle collisions (coagulation)
and coalescence (sintering) eventually determine the morphology of nanomaterials
(Xing et al., 1996; Hawa and Zachariah, 2004). If the characteristic sintering time
is significantly smaller than the characteristic collision time, colliding particles will
merge into a single sphere before another collision event occurs, resulting in the
formation of nearly monodispersed spherical particles. On the other hand, if the
collision time is much smaller than the characteristic sintering time, particle chain
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.3. Time Scales of Nanoparticle Collision-Coalescence Mechanism 275

Spherules

τsin
l>
Reaction Nucleation τco

τco
l<
τsi
n
Vapor Monomers/ Primary
precursor cluster particles
Agglomerates

Figure 9.14. Key time scales related to the collision-coalescence mechanism of vapor-phase
nanoparticle formation.

agglomerates may form. In the case where the sintering time is large compared with
the collision time scale, the time-accurate DEM approach based on an appropriate
adhesive contact model is well suited to describe the restructuring and break-up of
particle agglomerates. If sintering time is small compared with the collision time
scale, a population balance method might be the best way to model the particle
transport and coagulation, in which colliding particles are immediately merged into
a single quasi-particle. This section focuses on determination of the time scales for
different phases of the collision-coalescence mechanism, as well as clarifying the
underlying physics related to the two key processes of collision and sintering for
nanoparticles.

9.3.1. Time Scale of Particle Collisions

Coagulation is a process where small particles collide with each other to form larger
coalesced or aggregated particles. There are various mechanisms that can bring
particles to collide with each other, including Brownian motion, gravitational set-
tling, laminar shear, turbulent flow, and field-induced coagulation. Which mechanism
dominates the process depends on physicochemical properties of the particles and
the suspending fluid as well as the flow conditions. Nanoparticles have small inertia
and tend to have low volume concentrations, so Brownian motion is usually the
dominant collision mechanism for most nanoparticle applications.
During the Brownian coagulation process, the particles move about in a random
manner due to the thermal collisions of the particles with gas molecules. As noted
in our discussion of particle collision by turbulent flows in Section 6.3, the rate of
collisions per unit volume ṅc12 between a group of particles with radius r1 and a
second group of particles with radius r2 can be written in terms of the number of
particles per unit volume of each group, n1 and n2 , as

ṅc12 = α12 n1 n2 . (9.3.1)

The collision kernel α12 depends on the particle radii and relative velocities. When
the collision process is dominated by Brownian motion, the expression for collision
kernel depends on Knudsen number. In the free-molecular regime (Kn 1), the
probability for a gas molecule to collide with the particle is low, and the particles
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

276 Nanoscale Particle Dynamics

therefore move along straight trajectories relatively long distances between colli-
sions. The particles move with a purely ballistic motion between collisions, and the
resulting collision kernel is similar to that for a gas molecule, which is given by the
kinetic theory of gases (Glassman and Yetter, 2008) as
 1/6  1/2  
3 6kB T 1 1 1/2 & 1/3 '2
α12 = + V1 + V21/3 (9.3.2)
4π ρp V1 V2

where ρ p is the particle density and V1 and V2 are the volumes of individual particles
in the two groups. This result can be obtained from the Abrahamson collision kernel
(6.3.13) by replacing the turbulent flow mean-square velocities v1m 2
 and v2m
2
 with
the Brownian motion mean-square velocities 3kB T/m1 and 3kB T/m2 , respectively.
In the continuum regime (Kn 1), the gas molecules collide frequently with
the particle and the particle moves in a random, meandering path. In this regime,
Brownian collision can be described as a Brownian diffusion process, for which the
collision kernel is given by

α12 = 4π Db (r1 + r2 ), (9.3.3)

where Db is determined by the Einstein-Stokes expression (9.1.6). Using the general

form of the Brownian diffusion coefficient for unequal-size particles and substituting
into (9.3.3), the Brownian collision kernel in the continuum regime can be expressed
in terms of the particle volumes as
 
2kB T 1 1 & '
α12 = 1/3
+ 1/3 V11/3 + V21/3 . (9.3.4)
3μ CC V V
1 2

The characteristic collision time for a mono-disperse particle system can be

approximately determined using (9.3.1) as
2
τcol = . (9.3.5)
α11 n1
This collision time scale provides an estimate of the time interval between particle
collisions. It is noted that the collision time scale is different from the contact time
scale τC ∼= 2.868(m2 /E 2 Rv0 )1/5 given in (3.2.22), where the latter represents the
characteristic time that particles remain in contact during a collision.
The effect of particle long-range interactions on collision frequency is a long-
standing problem in aerosol dynamics. In general, long-range interaction refers to
forces with a range of action that is of the order of the particle radius or longer. For
nanoparticles, the long-range interactions mainly include van der Waals interaction,
Coulomb electrostatic forces, and dipole-dipole interaction. Fuchs (1964) proposed
that the effect of long-range interaction be accounted for with use of an enhancement
factor in the equation for collision rate and collision kernel, giving

ṅC = Wenh ṅC0 , α = Wenh α0 , (9.3.6)

where ṅC and α are the collision rate and collision kernel including the effects
of long-range interaction and ṅC0 and α0 are the same quantities for Brownian
coagulation without long-range interaction. Experimental results shown in Fig-
ure 9.15 indicate the rapid growth in value of the enhancement factor as the particle
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.3. Time Scales of Nanoparticle Collision-Coalescence Mechanism 277

10

Ag
Figure 9.15. Experimental enhancement

Enhancement factor
ZnCl2
factors as a function of diameters for Ag, NaCl
ZnCl2 , and NaCl obtained by Okuyama
et al. (1984). The dashed line illustrates the
case with no particle interaction, for which
the enhancement factor is equal to unity.
The experiment confirms the effect of par- Noninteracting model
NoninteractingModel
ticle interaction on collision rate when the
particle size decreases to the nanoscale size 1
range.
10 100
Diameter (nm)

size decreases within the nanoparticle size range. A wide variety of expressions have
been proposed for the enhancement factor Wenh , both for the continuum regime
and for the free-molecule regime with both charged and neutral particles. A review
of various expressions for enhancement factor in these two regimes for problems in
which the particle interaction is governed by a potential field is given by Ouyang et al.
(2012).
Zhang et al. (2011) applied molecular dynamics (MD) to study the role of dipole-
dipole interaction on enhancing the Brownian collision of neutral TiO2 nanopar-
ticles in the free molecular regime. To quantitatively evaluate the enhancement
factor, four characteristic dipole-moment directions/orientations were examined, as
illustrated in Figure 9.16. The parallel-to-path, co-orientated dipoles exhibited the
greatest enhancement factor values, and the parallel-to-path, counter-orientated
dipoles gave the least enhancement factors. Comparatively, the perpendicular-to-
path dipoles, either co-orientated or counter-orientated, straddle the two limiting
cases. The dipole-dipole interaction is found to have a significant influence on the
enhancement factor values, particularly at lower temperatures. Figure 9.17 illustrates
the enhancement factors including both van der Waals and dipole-dipole interaction
forces for temperatures ranging from 273 K to 1,273 K. As the temperature increases,

Initial
velocity
L0 v0 v0
Figure 9.16. Four initial dipole directions/
orientations investigated in the MD sim-
ulation: (a) parallel-to-path, co-oriented
Ld
(a) (b)
dipoles; (b) parallel-to-path, counter-
oriented dipoles; (c) perpendicular-to-path,
co-oriented dipoles; (d) perpendicular-to-
path, counter-oriented dipoles. v0 v0

(c) (d)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

278 Nanoscale Particle Dynamics

1.5

Force (pN)
11 A
1.0

Enhancement factor 9
A 0.5

0.0
7 2.5 5.0 7.5 10.0

2
5

Force (pN)
1 B
B
3 0

-1
1
200 400 600 800 1000 1200 1400 -2
2.5 5.0 7.5 10.0
Temperature (K) Distance (nm)
Figure 9.17. Effect of system temperature on enhancement factors of coagulation. Open cir-
cles indicate the case with only van der Waals force, and the filled square symbols indicate
the case with both dipole-dipole force and van der Waals force. Subplots A and B show the
forces between anatase TiO2 nanoparticles at 273 K and 1,273 K, respectively. [Reprinted
with permission from Zhang et al. (2011).]

the enhancement factor drops rapidly due to the reduction of the particle dipole
moments as a result of thermal fluctuations.

9.3.2. Time Scale of Nanoparticle Sintering

Coalescence of particles via sintering is a common occurrence in nanoparticle pro-
cessing, and it is primarily responsible for determining the final morphology and
size of the particles. As discussed in Section 4.6, sintering is a thermal process that
involves mass transport at the atomic scale. The coalescence of nanoparticles shares
the same physical process as observed for microscale particles, but nanoparticle
sintering also has a number of distinguishing features due to the small particle size.

9.3.2.1. Mass Transport in Sintering

The nature of mass transport during sintering processes depends on the temperature
relative to the melting point of the material. If the particle temperature is above the
melting point, or if the material is amorphous, the mass transport is governed by fluid
advection. If the temperature is below the melting point, atomic diffusion governs
the sintering mass transport. Because most nanoparticle coagulation processes of
interest occur with temperatures below the melting point, the latter process is of
primary interest.
Several different diffusion mechanisms have been reported to describe the
transport mechanism of the atoms (German, 1996). These mechanisms include sur-
face diffusion (SD), volume diffusion (VD), grain-boundary diffusion (GBD), and
evaporation-condensation processes (E-C). An illustration of these various mecha-
nisms is given in Figure 9.18. Surface diffusion is caused by the presence of a concen-
tration gradient on surfaces of different curvature. Volume diffusion refers to the
transport of atoms from the interior of the particle to the surface, as well as transport
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.3. Time Scales of Nanoparticle Collision-Coalescence Mechanism 279

E-C
rb
SD
VD
ra
Figure 9.18. Diffusion mechanisms of mass trans- x
GBD
port in sintering. SD: surface diffusion. VD: vol-
ume diffusion. GBD: grain boundary diffusion.
E-C: evaporation-condensation model.

of vacancies from the surface to the interior. Grain-boundary diffusion occurs at the
grain boundary separating the two particles. The evaporation-condensation process
involves phase change of the atoms under a given vapor pressure, where the fluid
pressure is influenced by local surface curvature due to the effect of surface tension.
All these mechanisms tend to decrease the surface area and surface energy of the
system.
For nanoparticle sintering processes, surface diffusion is usually considered to be
the dominant mechanism due to the small size (and high curvature) of the particles.
Volume diffusion can be neglected as a consequence of the large surface-volume
ratio of the particles. Grain boundary diffusion is complex for nanoscale sintering,
especially when atomic reorientation or alignment occurs (Zachariah and Carrier,
1999; Ding et al., 2009). Lehtinen and Zachariah (2001, 2002) pointed out that
the heat release resulting from the decrease in surface area and surface energy
during sintering acts to increase the particle temperature, which in turn enhances
the coalescence rate of the nanoparticles.

9.3.2.2. Characteristic Sintering Time

Coalescence between two nanoparticles in an aerosol system generally begins with
Brownian collision of the particles and ends when a new (larger) particle is formed
from two colliding particles. The time scale for this whole process, called the coa-
lescence time, ranges from nanoseconds to milliseconds according to the size of the
particles.
Koch and Friedlander (1990) developed a model for the effect of particle coa-
lescence on the change in particle surface area as

dSa 1
=− (S − Sa, final ), (9.4.1)
dt τsin a

where Sa is the total surface area of the colliding particles before sintering, Sa, final
is the particle surface area after sintering, and τ sin denotes the sintering time scale
(or the coalescence time). If the particles are initially perfect spheres, then Sa, final
will be 79% of the initial area Sa . As mentioned in the previous section, viscous
flow governs the sintering mass transport when the particle temperature is above the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

280 Nanoscale Particle Dynamics

material melting point. In this case, the sintering time is affected by surface tension
and material viscosity, so that (Frenkel, 1945)
μ pd
τsin = , (9.4.2)
σp
where μ p represents viscosity of the melted particle material, σ p is its surface tension
of the particle melt-air system, and d is the particle diameter before sintering. The
driving force of coalescence is surface tension, and the melt viscosity acts to resist
coalescence. When the temperature is below the melting point, solid-state diffusion
governs the mass transport and the sintering time scale becomes (Friedlander and
Wu, 1994)
3kB TVp
τsin = , (9.4.3)
64πDσ p a
where kB is the Boltzmann constant, T is particle absolute temperature, D is the
solid-state diffusion coefficient, σ p is the surface tension of the particle, and Vp and
a denote the particle volume and the molecular volume, respectively.
The material properties for the nanomaterial melt, such as diffusion coefficient
and surface tension, are usually quite different from those of the bulk material and
not known a priori. Various empirical formulas have been reported to estimate
the coalescence time. For example, Xiong and Pratsinis (1993) in studies of TiO2
nanoparticles estimated sintering time as
 
3700
τsin = 8.3 × 1024 T d4 exp , d in m; (9.4.4)
T
where d is particle diameter before sintering. Ehrman et al. (1998), also working
with TiO2 nanoparticles, proposed
 
34372
τsin = 1.87 × 109 T d3 exp − , d in m. (9.4.5)
T
Kobata et al. (2004) proposed a similar formula
 
258 kJ/mol
τsin = 1.19 × 1018 T r4p exp , r p in m. (9.4.6)
RT
These formulas all have the basic form of some factor times a term that varies
exponentially in the inverse temperature; nevertheless, the final estimate for the coa-
lescence time can differ widely between the formulas. Some researchers attribute the
difference to the overestimation of coalescence time when the particles are relatively
small (Windeler et al., 1997; Wu et al., 1993). These results imply that sintering at
nanoscales requires further investigation and reconsideration of model assumptions.
Figure 9.19 shows typical Brownian collision time and sintering time (calculated
using Equation (9.4.5)) values as functions of particle diameter for different temper-
atures. The characteristic times exhibit strong dependence on particle size. When the
sintering time is larger than the collision time for a given temperature and particle
size, coalescence between the colliding nanoparticles cannot be completed in the
time between subsequent collisions of the particle. In this case, agglomerates are
formed with soft interparticle bonding and the size of the primary particle remains
unchanged. On the other hand, if the collision time is larger than the sintering time
scale, new particles form with approximately spherical shape.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

9.3. Time Scales of Nanoparticle Collision-Coalescence Mechanism 281

1
1400 K
0.1

0.01

Time (s)
1E-3
Figure 9.19. Typical Brownian collision time and 1600 K
sintering time scales for different temperatures. 1E-4

1E-5
collision time
1E-6 sintering time

5 10 15 20 25 30
Diameter (nm)

9.3.3.3. Lattice Structure in Coalescence

The lattice structure of nanoparticles is no longer uniform during coalescence, during
which it is influenced by the extremely high surface curvature that occurs near the
sintering region. Molecular dynamics simulation is frequently used to study lattice
structure of nanoparticles during coalescence. For example, Zhu (1996) used MD to
examine grain boundary rotation during nanoparticle sintering. Other researchers
have calculated particle coalescence time using MD and made comparisons with
different models (Zachariah and Carrier, 1999; Busesser et al., 2011).
The lattice structure of TiO2 nanoparticles during coalescence was recently
investigated by Zhang et al. (2012b), who found the structure to be sensitive to
particle size and local temperature. For a 3 nm TiO2 (anatase) nanoparticle at a
temperature below the melting point, the typical lattice structure is a crystal core
surrounded by an amorphous shell. When the particle size is below a critical value at
a specific temperature, a structural transition occurs from the core-shell structure to
a completely amorphous structure. Figure 9.20 shows the dependence of the lattice
structure regimes on particle diameter and temperature for TiO2 particles.

Figure 9.20. Lattice structure map of TiO2 (anatase) as a function of temperature and particle
diameter. L: liquid. A(S): solid in amorphous. Core-shell (S): solid in core-shell structure.
(Based on data from Zhang et al., 2012b.)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

282 Nanoscale Particle Dynamics

1.2
2000
1573K
Temperature 1800

Temperature (K)
0.8
1600
a – af
a0 – af Recrystallization 1400
0.4
Surface area 1200

0.0 Surface 1000

diffusion
800
0 4000 8000 12000 16000
Time (ps)
Figure 9.21. Coalescence process between two nanoparticles with initial temperature at
1,573 K.

Sintering between two crystal nanoparticles at a temperature well below the

melting point is initially dominated by surface diffusion with amorphous shell fusion
and neck formation, and subsequently dominated by grain-boundary reorientation
within the later stages of sintering. As the temperature approaches the melting point,
recrystallization of the particle structure occurs. Figure 9.21 reveals the predicted
temperature increase and the surface area reduction during MD simulations of
coalescence of two TiO2 nanoparticles (anatase with core-shell structure) at an initial
temperature of 1573 K. First, the amorphous shell fuses by surface diffusion, then a
new crystal core appears in the center of the newly coalesced particle accompanied
a sharp temperature increase due to the recrystallization process. The surface area
reduction is far below the prediction of the Koch-Friedlander model (9.4.1) at the
start of coalescence due to the fact that the core-shell structure of the nanoparticle
violates the assumption of homogeneity made in the model.

REFERENCES

Adachi M, Kousaka Y, Okuyama K. Unipolar and bipolar diffusion charging of ultrafine

aerosol particles. Journal of Aerosol Science 16(2), 109–123 (1985).
Allen MD, Raabe OG. Re-evaluation of Millikan’s oil drop data for the motion of small
particles in air. Journal of Aerosol Science 13(6), 537–547 (1982).
Becker V, Briesen H. Tangential-force model for interactions between bonded colloidal
particles. Physical Review E 78, 061404 (2008).
Bishop KJM, Wilmer CE, Soh SL, Grzybowski BA. Nanoscale forces and their uses in self-
assembly. Small 5(14), 1600–1630 (2009).
Blanton SA, Leheny RL, Hines MA, Guyot-Sionnest P. Dielectric dispersion measurements
of CdSe nanocrystal colloids: Observation of a permanent dipole moment. Physical Review
Letters 79, 865–868 (1997).
Buesser B, Gröhn A J, Pratsinis S E. Sintering rate and mechanism of TiO2 nanoparticles by
molecular dynamics. Journal of Physical Chemistry C 115, 11030–11035 (2011).
Davies CN, Definitive equations for the fluid resistance of spheres. Proceedings of the Physical
Society 57(4), 259–270 (1945).
Davis EJ, Schweiger G. The Airborne Microparticle: Its Physics, Chemistry, Optics and Trans-
port Phenomena. Springer-Verlag, Heidelberg, pp. 755–807 (2002).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

References 283

Ding L, Davidchack RL, Pan J. A molecular dynamics study of sintering between nanoparti-
cles. Computational Materials Science 45, 247–256 (2009).
Eggersdorfer ML, Kadau D, Herrmann HJ, Pratsinis SE. Fragmentation and restructuring
of soft-agglomerates under shear. Journal of Colloid and Interface Science 342, 261–268
(2010).
Ehrman SH, Friedlander SK, Zachariah MR. Characteristics of SiO2/TiO2 nanocomposite
particles formed in a premixed flat flame. Journal of Aerosol Science 29, 687–706 (1998).
Epstein PS. On the resistance experienced by spheres in their motion through gases. Physical
Review 23, 710–733 (1924).
Fanelli M, Feke DL, Manas-Zloczower I. Prediction of the dispersion of particle clusters in
the nano-scale. I. Steady shearing responses. Chemical Engineering Science 61, 473–488
(2006).
Feng XD, Sayle DC, Wang ZL, Paras MS, Santora B, Sutorik AC, Sayle TXT, Yang Y,
Ding Y, Wang XD, Her YS. Converting ceria polyhedral nanoparticles into single-crystal
nanospheres. Science 312, 1504–1508 (2006).
Frenkel J. Viscous flow of crystalline bodies under the action of surface tension. Journal of
Physics 9, 385–391 (1945).
Friedlander SK. Smoke, Dust and Haze: Fundamentals of Aerosol Dynamics. 2nd ed., Oxford
University Press, New York (2000).
Friedlander SK, Wu MK. Linear rate law for the decay of the excess surface area of a coalescing
solid particle. Physical Review B 49, 3622–3624 (1994).
Fuchs NA, The Mechanics of Aerosols. Macmillan, New York (1964).
German RM. Sintering Theory and Practice. John Wiley & Sons: Wiley-Interscience, New
York (1996).
Glassman I, Yetter RA. Combustion. 4th ed., Academic Press, Burlington, Vermont (2008).
Goesmann H, Feldmann C. Nanoparticulate functional materials. Angew. Chem. Int. Ed. 49,
1362–1395 (2010).
Hawa T, Zachariah MR. Molecular dynamics study of particle–particle collisions between
hydrogen-passivated silicon nanoparticles. Physical Review B 69, 035417 (2004).
Higashitani K, Iimura K, Sanda H. Simulation of deformation and breakup of large aggregates
in flows of viscous fluids. Chemical Engineering Science 56, 2927–2938 (2001).
Kobata A, Kusakabe K, Morooka S. Growth and transformation of TiO2 crystallites in aerosol
reactor. AIChE Journal 37, 347–359 (2004).
Koch W, Friedlander SK. The effect of particle coalescence on the surface area of a coagulating
aerosol. Journal of Colloid and Interface Science 140, 419–427 (1990).
Lehtinen KE, Zachariah MR. Effect of coalescence energy release on the temporal shape
evolution of nanoparticles. Physical Review B 63, 205402 (2001).
Lehtinen KE, Zachariah MR. Energy accumulation in nanoparticle collision and coalescence
processes. Journal of Aerosol Science 33, 357–368 (2002).
Li SQ, Marshall JS. Discrete element simulation of micro-particle deposition on a cylindrical
fiber in an array. Journal of Aerosol Science 38, 1031–1046 (2007).
Li ZG. Gas kinetic theory and molecular dynamics simulation of nanomaterial transfer in
dilute gases. Ph.D. thesis, University of Delaware, pp. 71–89 (2005).
Li ZG, Wang H. Drag force, diffusion coefficient, and electric mobility of small particles.
I. Theory applicable to the free-molecule regime. Physical Review E 68(6), 061206 (2003a).
Li ZG, Wang H. Drag force, diffusion coefficient, and electric mobility of small particles.
II. Application. Physical Review E 68(6), 061207 (2003b).
Li ZG, Wang H. Thermophoretic force and velocity of nanoparticles in the free molecule
regime. Physical Review E 70, 021205 (2004).
Loyalka SK. Thermophoretic force on a single particle. I. Numerical solution of the linearized
Boltzmann equation. Journal of Aerosol Science 23, 291–300 (1992).
Luan B, Robbins MO. The breakdown of continuum models for mechanical contacts. Nature
Letters 435, 929–932 (2005).
Madler L, Friedlander SK. Transport of nanoparticles in gases: Overview and recent advances.
Aerosol and Air Quality Research 7(3), 304–342 (2007).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

284 Nanoscale Particle Dynamics

Madler L, Lall AA, Friedlander SK. One-step aerosol synthesis of nanoparticle agglomerate
films: Simulation of film porosity and thickness. Nanotechnology 17(19), 4783–4795 (2006).
Matsui M, Akaogi M. Molecular dynamics simulation of the structural and physical properties
of the four polymorphs of TiO2 . Molecular Simulation 6, 239–244 (1991).
Okuyama K, Kousaka Y, Hayashi K. Change in size distribution of ultrafine aerosol particles
undergoing Brownian coagulation. Journal of Colloid and Interface Science 101, 98–109
(1984).
Ouyang H, Gopalakrishnan R, Hogan GJ. Nanoparticle collisions in the gas phase in the
presensce of singular contact potentials. Journal of Chemical Physics 137, 064316 (2012).
Pantina JP, Furst EM. Elasticity and critical bending moment of model colloidal aggregates.
Physical Review Letters 94, 138301 (2005).
Park SH, Kim WJ, Kim SS. Thermophoretic transport and deposition of particles in verti-
cal tube flow with variable wall temperature and thermal radiation. KSME International
Journal 13(3), 253–263 (1999).
Penn RL, Banfield JF. Morphology development and crystal growth in nanocrystalline aggre-
gates under hydrothermal conditions: Insights from titania. Geochimica et Cosmochimica
Acta 63(10), 1549–1557 (1999).
Pui DYH, Fruin S, McMurry PH. Unipolar diffusion charging of ultrafine aerosols. Aerosol
Science and Technology 8(2), 173–187 (1988).
Sahu M, Park JH, Biswas P. In situ charge characterization of TiO2 and Cu–TiO2 nanoparticles
in a flame aerosol reactor. Journal Nanoparticle Research 14, 687 (2012).
Schwarz UD, Zwörner O, Köster P, Wiesendanger R. Quantitative analysis of the frictional
properties of solid materials at low loads. II. Mica and germanium sulfide. Physical Review
B 56(11), 6997–7000 (1997).
Shim M, Guyot-Sionnest P. Permanent dipole moment and charges in colloidal semiconductor
quantum dots. Journal of Chemical Physics 111(15), 6955–6964 (1999).
Sone Y, Aoki K. A similarity solution of the linearized Boltzmann equation with application
to thermophoresis of a spherical particle. J. Mecanique Theorique Appliquee 2(1), 3–12
(1983).
Stratmann F, Fissan H, Papperger A, Friedlander S. Suppression of particle deposition to
surfaces by the thermophoretic force. Aerosol Science and Technology 9(2), 115–121 (1988).
Sun WF, Zeng QH, Yu AB, Kendall K. Calculation of normal contact forces between silica
nanospheres. Langmuir 29, 7825−7837 (2013).
Talbot L. Thermophoresis: A review. In Progress Astro Aero, Vol. 74, Rarefied Gas Dynamics,
SS Fisher, editor, New York, AIAA (1981).
Talbot L, Cheng R. K, Schefer RW, Willis DR. Thermophoresis of particles in a heated
boundary layer. Journal of Fluid Mechanics 101, 737–758 (1980).
Tang Z, Kotov N A, Giersig M. Spontaneous organization of single CdTe nanoparticles into
luminescent nanowires. Science 297, 237–240 (2002).
Tolmachoff ED, Abid AD, Phares DJ, Campbell CS, Wang H. Synthesis of nano-phase TiO2
crystalline films over premixed stagnation flames. Proc Combust Inst 32, 1839–1845 (2009).
Tyndall J. On dust and disease. Proc. Royal Inst. 6, 1–14 (1870).
Waldmann L. Über die Kraft eines inhomogenen Gases auf kleine suspendierete Kugeln.
Z. Naturforsch A 14a, 589–599 (1959).
Wenning L, Müser MH. Friction laws for elastic nanoscale contacts. Europhysics Letters 54(5),
693–699 (2001).
White FM. Viscous Fluid Flow, 3rd ed., McGraw-Hill, New York, pp. 149–151 (2006).
Windeler RS, Lehtinen KE, Friedlander SK. Production of nanometer-sized metal oxide
particles by gas phase reaction in a free jet. II. Particle size and neck formation: Comparison
with theory. Aerosol Science and Technology 27, 191–205 (1997).
Wu MK, Windeler RS, Steiner CKR, Börs CK, Friedlander SK. Controlled synthesis of
nanosized particles by aerosol processes. Aerosol Science and Technology 19, 527–548
(1993).
Xing YC, Koylu UO, Rose RDE. Synthesis and restructuring of inorganic nano-particles in
counterflow diffusion flames. Combustion & Flame 107, 85–102 (1996).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-09 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 7:44

References 285

Xiong Y, Pratsinis SE. Formation of agglomerate particles by coagulation and sintering. Part I.
A two-dimensional solution of the population balance equation. Journal of Aerosol Science
24, 283–300 (1993).
Yamamoto K, Ishihara Y. Thermophoresis of a spherical particle in a rarefied gas of a transi-
tion regime. Physics of Fluids 31, 3618–3624 (1988).
Yan W, Li S, Zhang Y, Yao Q, Tse SD. Effects of dipole moment and temperature on
the interaction dynamics of titania nanoparticles during agglomeration. Journal Physical
Chemistry C 114, 10755–10760 (2010).
Ye Y, Pui DYH, Liu BYH, Opiolka S, Blumhorst S, Fissan H. Thermophoretic effect of
particle deposition on a free standing semiconductor wafer in a clean room. Journal of
Aerosol Science 22(1), 63–72(1991).
Zachariah MR, Carrier MJ. Molecular dynamics computation of gas-phase nanoparticle sin-
tering: A comparison with phenomenological models. Journal of Aerosol Science 30, 1139–
1152 (1999).
Zhang YY, Li SQ, Yan W, Yao Q, Tse SD. Role of dipole–dipole interaction on enhancing
Brownian coagulation of charge-neutral nanoparticles in the free molecular regime. Journal
of Chemical Physics 134, 084501 (2011).
Zhang YY, Li SQ, Deng SL, Yao Q, Tse SD. Direct synthesis of nanostructured TiO2 films
with controlled morphologies by stagnation swirl flames. Journal of Aerosol Science 44,
71–82 (2012a).
Zhang YY, Li SQ, Yan W, Tse SD. Effect of size-dependent grain structures on the dynamics
of nanoparticle coalescence. Journal of Applied Physics 111, 124321 (2012b).
Zhang YY, Li SQ, Yan W, Yao Q. Nanoparticle transport and deposition in boundary layer
of stagnation-point premixed flames. Powder Technology 227, 24–34 (2012c).
Zhu H. Sintering processes of two nanoparticles: A study by molecular dynamics simulations.
Philosophical Magazine Letters 73, 27–33 (1996).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10 Computer Implementation and Data Analysis

The physical models used to describe the forces and torques on particles in various
types of DEM simulations are discussed in the previous chapters of this book. This
chapter focuses on numerical aspects related to implementation of these models,
as well as on different ways to characterize and interpret the results of the DEM
simulations.

10.1. Particle Time Stepping

Because the momentum and moment of momentum equations governing parti-
cle motion are ordinary differential equations in time, the first task that must be
addressed in computer implementation of DEM is the discretization of these equa-
tions in time. In order to select an appropriate time step for calculation of the particle
motion, it is necessary to first understand the various time scales involved in the prob-
lem. The largest time scale is usually the fluid convective time scale τF = O(L/U ),
where L and U denote characteristic macroscopic length and velocity scales of the
flow field, respectively. This is itself somewhat of a simplification, as fluid flows often
have very different time and length scales depending on position in the flow (e.g.,
whether or not the particle is in a boundary layer region) and on wavenumber of the
fluctuation in a turbulent flow. For present purposes, however, let us assume that
a distinct fluid velocity and length scale can be defined. The next smallest scale is
typically the particle convective time scale τCP = O(d/U ), where d is the particle diam-
eter. It is important in a DEM simulation that the time step used to move the parti-
cles be significantly less than TCP in order not to miss particle collisions during the
time step. In the presence of particle collisions, we must also set the time step so as
to resolve elastic response of the particle during the collision. The particle contact
time based on the Hertz theory is given in (3.2.22) as τC ∼ = 2.868(m2 /E 2 Rv0 )1/5 . For
an order-of-magnitude analysis, we take m = O(ρ p d ), E = O(E p ), R = O(d), and
3

v0 = O(U ), so that particle contact time scale τC = d(ρ p2 /E 2pU )1/5 , where E p is the
particle elastic modulus. Adhesion forces acting between colliding particles have
approximately the same time scale as the elastic repulsion force.

286
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.1. Particle Time Stepping 287

10.1.1. Numerical Stability

An additional time scale, called the particle aerodynamic time scale τAP = m/μ d, is
related to resolution of the particle response to change in the fluid velocity field.
One way in which this time scale arises is evident by consideration of the numerical
stability of the particle moment equation when computed using an explicit scheme.
To see this, we examine the simple model in which the particle’s inertia is balanced
by its drag force, the governing equation for which is given by (1.2.2) as
dv
m = −3π μd(v − u). (10.1.1)
dt
Nondimensionalizing velocity by U and time by L/U and indicating dimensionless
variables with a prime, (10.1.1) becomes
dv 1
= − (v − u ), (10.1.2)
dt  St
where the Stokes number St is defined by (1.2.5). The differential equation (10.1.2)
is discretized using the simple forward Euler scheme, which is typical of an explicit
difference scheme, giving
 
n+1 t  n t   n
v = 1− v − u , (10.1.3)
St St
where v n = v (tn ) and t  is the dimensionless time step. Now suppose that the
initial velocity value v̂0 is perturbed by an amount δv , so that the initial condition
in the computation becomes

v0 = v̂0 + δv . (10.1.4)

Substituting (10.1.4) into (10.1.3) gives the solution at any time step n as
 
t  n 
vn = v̂n + 1 − δv , (10.1.5)
St

where v̂n is the solution at time tn with the exact initial condition v0 = v̂0 . The
magnitude of the error term in (10.1.5) changes in proportion to an amplification
term |ξ |n , where ξ = 1 − t  /St, which will increase as time (n) increases for |ξ | > 1
and decrease with time for |ξ | < 1. Consequently, the condition for stability of the
forward Euler scheme for this equation is that |ξ | ≤ 1, from which we see that
computations are stable provided that t  ≤ 2St. In terms of dimensional variables,
the stability criterion becomes
2m
t ≤ . (10.1.6)
3π μ d
Although this discussion pertains only to the forward Euler scheme, other explicit
schemes exhibit a similar stability condition that limits t  to be less than some
multiple of the Stokes number. Of course, it is possible to solve (10.1.2) implicitly
using an integrating factor to write the equation as
d  1

[v exp(t  /St)] = u exp(t  /St). (10.1.7)
dt St
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

288 Computer Implementation and Data Analysis

Solving for particle velocity then gives

 t n+1
1
vn+1 = vn exp(−t  /St) + exp(−t n+1 /St) u (τ ) exp(τ /St) dτ. (10.1.8)
St t n

The integral in (10.1.8) must in general be evaluated numerically. For instance, if

we assume that the fluid velocity u is approximately constant during the time step,
(10.1.8) becomes
vn+1 = vn exp(−t  /St) + u [1 − exp(−t  /St)] .
n
(10.1.9)
The implicit scheme (10.1.9), as well as the more general form (10.1.8), is stable for
all values of the time step t  . However, if t  is larger than O(St), the exponential
term in (10.1.9) quickly approaches zero, giving the result that the particle velocity
will always be equal to the local fluid velocity. In this case, the particle response to
changes in the fluid velocity is nearly instantaneous and the inertial response of the
particle is not resolved.

10.1.2. Multiscale Time-Stepping Approaches

The principal challenge in implementing DEM for large numbers of particles is
the fact that the various time scales associated with particle transport and colli-
sion described earlier have very different orders of magnitude. For instance, dust
transport in air is a fairly typical aerosol that is of importance for many different
applications, including fouling of electronics and construction equipment. Assum-
ing an aerosol with 10 µm diameter dust particles (ρ p = 2650 kg/m3 ) in air flow
(μ = 1.8 × 10−5 Pa·s) with fluid length and velocity scales L = 1 cm and U = 1 m/s,
respectively, the Stokes number is obtained as St = 0.08. The fluid time scale is
τF = L/U = 0.01 s, and the particle convection and aerodynamic time scales are
τCP = d/U = 10−5 s and τAP = 2LSt/U = 0.0016 s. Assuming an elastic modulus for
the dust particle (quartz) of E p ∼= 70 GPa, the contact time scale is τC ∼ = 10−8 s. The
ratio of the fluid time scale to the contact time scale in this example is τF /τC = 106 ,
and the ratio of the fluid time scale to the particle time scale is τF /τCP = 103 . This
large difference in the observed time scales indicates a very significant degree of
numerical stiffness in computational solution of the particle transport.
A similar situation exists in a typical colloidal solution, although here the rela-
tionship between the two particle time scales τCP and τAP can often be reversed. Flow
of red blood cells in a large artery is perhaps a typical example. For this estimate,
let us treat red blood cells as rigid particles with effective diameter of approximately
3 µm. The fluid length and velocity scales in an artery correspond to the artery
diameter and mean velocity, given approximately by L ∼ = 1 cm and U ∼ = 10 cm/s,
respectively. The fluid time scale and the particle convection time scale for this
problem are τF ∼ = 0.1 s and τCP ∼ = 3 × 10−4 s. The Stokes number is computed
using the properties of blood plasma, with viscosity μ ∼ = 0.003 Pa · s and density
∼ −6
ρ f = 1000 kg/m , as St = 1.7 × 10 , giving the particle aerodynamic time scale as
3

τAP ∼= 3.3 × 10−7 s. The effective elastic modulus for red blood cells is determined
experimentally by Dulinska et al. (2006) as E p ∼ = 2.6 × 104 Pa, giving the particle
contact time scale as τC ∼ −6
= 1.3 × 10 s. Because red blood cells are relatively soft,
the contact time scale τC is about two orders of magnitude larger for this application
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.2. Flow in Complex Domains 289

than it is for the dust aerosol considered earlier. Similarly, the high viscosity of blood
plasma compared with air makes the aerodynamic time scale τAP about four orders
of magnitude smaller than it is for the aerosol. As a result, the particle aerodynamic
time scale is the smallest time scale in this problem, with a ratio of the fluid to particle
aerodynamic time scales of τF /τAP ∼ = 3 × 105 .
Because of the high numerical stiffness involved in particulate flow systems, it
is necessary to implement various methods to try to accelerate the calculation while
resolving motion at the different time scales listed earlier. In the absence of particle
collisions and for systems with small Stokes number, the relative velocity is known
to obey the scaling estimate |v − u| = O(St U ), as noted in Chapter 1. In this case,
the particle acceleration in the particle momentum equation can be approximated
to leading order in the Stokes number by the fluid acceleration, or dv/dt ∼ = Du/Dt.
This approximation was introduced in Section 6.2 with a more detailed derivation, in
the context of measuring particle dispersion in turbulent flows. For systems with only
particle inertia and particle drag force, the dimensionless particle transport equation
(10.1.2) can be approximated by
v = u − St a + O(St)2 , (10.1.10)
where a is the dimensionless fluid acceleration a ≡ Du /Dt  and DDt is the material
time derivative with respect to fluid elements. The approximation (10.1.10) entirely
removes the stiffness imposed by the particle aerodynamic stability condition (10.1.6)
on the time step, thus allowing use of much larger time steps than would be the case
with the full particle momentum equation. Extension of this method for cases with
other fluid forces, such as lift and added mass force, is discussed by Ferry and
Balachandar (2001) and Ferry et al. (2003). However, this approximation has not yet
been successfully applied for problems where particles collide or adhere with each
other.
A multiple-time step DEM algorithm was proposed by Marshall (2009) using
three time steps – a fluid time step, a particle time step, and a contact time step. For the
particle time step, either the particle collision time τCP or the particle aerodynamic
time τAP is used, whichever is smaller. The algorithm performs time-consuming com-
putational tasks at the longer fluid time step, such as identifying a list of neighboring
particles, while minimizing the number of computations that need be performed at
the shorter collision and particle time steps.

10.2. Flow in Complex Domains

The various models for fluid force and torque on the particle discussed in Chapter
5, as well as the models for electrostatic force discussed in Chapter 8, require that
the fluid velocity and vorticity field and the electric field vector be known at each
particle position at every time step. Because DEM simulations usually require a very
small time increment, as discussed in the previous section, it is necessary to utilize a
fast method for determining these fields if we are to enable simulations with large
numbers of particles. In typical applications, the fluid flow is solved using a grid-
based computational fluid dynamics approach, so that the fluid velocity and vorticity
vectors will be known at the nodes of the grid used for the fluid flow computation.
Many different types of grids are used in computational fluid dynamics for different
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

290 Computer Implementation and Data Analysis

problems and solution approaches, including tetrahedral unstructured grids, multi-

block structured grids, and even Cartesian grids. The last type is used, for instance,
in formulations where the motion of bodies or interfaces within the flow field are
represented using level-set, volume-of-fluid, or related approaches in which the
interface is represented by a jump in fluid density and viscosity in a continuous fluid
domain, with any interfacial force distributed to neighboring grid cells (Sethian and
Smereka, 2003).
In electrostatics, the electric field is often obtained using a boundary element
method, as discussed in Section 8.3. However, if there are M panels used to discretize
the domain boundary using a BEM approach, computation of the electric field at a
point in space will require O(M3 ) computations at the initial time step and O(M2 )
computations at subsequent time steps. For large values of M and large number of
particles, as is typical in BEM with domains of complex shape, it will be too time
consuming to repeat this computation at each particle location at each time step.
An alternative is to evaluate the electric field vector on a grid covering the flow
field and then interpolate onto particle locations from this grid. Usually, the electric
field values on the grid are computed using a larger time interval than that used for
particle motion.
Prior to interpolation of a field from a grid onto a particle centroid location,
it is necessary to identify which grid cell contains the particle. Although grid cell
identification is a simple matter with a Cartesian grid using integer division, this
task can be quite time consuming for other grid types. We therefore recommend
that the grid used for the fluid flow computation be mapped onto a Cartesian grid
prior to the particle transport computation. If the fluid grid is fixed in time, this
mapping can be done once and stored for later use. With fluid velocity and vorticity
fields known on a Cartesian grid, it is a simple matter to determine the grid cell
containing a given particle and to interpolate these fields onto the particle location.
However, because the flow boundaries no longer correspond with the grid bound-
aries on the Cartesian grid, it is necessary to utilize an additional function to indicate
the distance of any point on the grid to the domain boundary. For this purpose we
recommend use of a level-set distance function, which is a function whose absolute
value is equal to the distance to the closest point on the flow boundary and whose
sign changes as one travels across the boundary. In the two sections that follow,
an efficient search algorithm is discussed that can be used to map the flow com-
putation grid onto a Cartesian grid and a method is described for computation of
the level-set distance function used to identify domain boundaries on the Cartesian
grid.

10.2.1. Particle Search Algorithm

A particle search algorithm proposed by Allievi and Bermejo (1997) can be used
to determine which grid cell contains a given space point with arbitrary grid types.
For definiteness, this algorithm is illustrated in the current section for quadrilateral
elements in a two-dimensional space with coordinates (x, y). Let us suppose that the
computational domain D̄ can be decomposed into a set of M nonoverlapping regions
D j , such that the union of all of the D j is equal to D̄. As shown in Figure 10.1, a
mapping F j is defined that maps the unit square D̂ in a two-dimensional space with
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.2. Flow in Complex Domains 291

q 3
4 4 (0,1)
3
Fj ^
Ωj Ω p
y (-1,0)
1 2 (1,0)

x 1 (0,-1) 2

Figure 10.1. Illustration of mapping of a rectangular element into a quadrilateral element,

used in the search algorithm of Allievi and Bermejo (1997).

coordinates (p, q) into the element D j in the x–y plane. We seek to know whether a
given point (X, Y ) is contained within an element D j .
For a quadrilateral element, the x- and y-components of the mapping F j : D̂ →
D j are given by


4 
4
F1 j (p, q) = x = xv,i Gi (p, q), F2 j (p, q) = y = yv,i Gi (p, q), (10.2.1)
i=1 i=1

where (xv,i , yv,i ), i = 1, . . . , 4 denote the positions of the four vertices of D j . The
basis functions Gi are given by

G1 (p, q) = 14 (1 − p)(1 − q), (10.2.2a)

G2 (p, q) = 14 (1 + p)(1 − q), (10.2.2b)

G3 (p, q) = 14 (1 + p)(1 + q), (10.2.2c)

G4 (p, q) = 14 (1 − p)(1 + q). (10.2.2d)

The algorithm consists of two parts, referred to as “search” and “locate.” In the
search part, we seek to determine which grid cell contains the point (X, Y ). The
algorithm is initiated by selecting an initial grid cell D j to examine. This could be
done randomly, but if a simple method presents itself by which an initial cell can
be selected that is close to the point (X, Y), the search algorithm will converge to
the correct cell more rapidly. Once an initial cell is selected, we seek to determine
whether or not the point (X, Y) is in this cell and, if not, which neighboring cell
should next be examined to optimally find this point. For this purpose, we utilize the
iterative expression
* * *
pk+1 pk 1 b2 + b3 pk −a2 − a3 pk X − xk
= + k , (10.2.3)
qk+1 qk  −b1 − b3 qk a1 + a3 qk Y − yk

where xk and yk are the values of x and y corresponding to pk and qk , as determined

from (10.2.1), and

a1 = 14 (xv,2 − xv,1 + xv,3 − xv,4 ), b1 = 14 (yv,2 − yv,1 + yv,3 − yv,4 ),

a2 = 14 (xv,3 − xv,1 + xv,4 − xv,2 ), b2 = 14 (yv,3 − yv,1 + yv,4 − yv,2 ),
a3 = 14 (xv,1 − xv,2 + xv,3 − xv,4 ), b3 = 14 (yv,1 − yv,2 + yv,3 − yv,4 ),
k = (a1 b2 − a2 b1 ) + (a1 b3 − a3 b1 )pk + (a3 b2 − a2 b3 )qk .
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

292 Computer Implementation and Data Analysis

p< 1 1 p 1 p >1
q >1 q >1 q >1

4 5 Box 3 Box 4 Box 5

3
p< 1 1 p 1 p >1
Ωj 6 1 q 1 1 q 1 1 q 1
2
Box 2 ^ Box 6
8 7
1
p< 1 1 p 1 p >1
q< 1 q< 1 q< 1
(a)

Box 1 Box 8 Box 7

(b)
Figure 10.2. Numbering of neighbor grid cells used in the search algorithm of Allievi and
Bermejo (1997): (a) in the original grid in x–y space; (b) in the transformed grid in p-q space.

This expression is simply the Newton-Raphson iteration formula for the mapping
(10.2.1). In the search algorithm, the initial value of p and q is selected as (p0 , q0 ) =
(0, 0), corresponding to the center of cell D̂. The formula (10.2.3) is used to find the
value of (p1 , q1 ). The point (X, Y) is contained within the given grid cell if and only
if (p1 , q1 ) satisfies
− 1 ≤ p1 ≤ 1, −1 ≤ q1 ≤ 1. (10.2.4)
If the condition (10.2.4) is not satisfied, a different grid cell must be selected for
examination. This is done by moving to one of the neighboring cells of the original
grid cell, based on the value of (p1 , q1 ) as indicated in Figure 10.2b. For instance,
if p1 < −1 and q1 < −1, then we move to the neighbor indicated by the number 1
in Figure 10.2a, and so forth. The procedure is repeated until a cell is found that
satisfies (10.2.4).
Once the correct cell is found containing the point (X, Y), it is necessary to
determine the value of p and q corresponding to this point, denoted as (P, Q). This
can be accomplished by iterating on (pk , qk ) using (10.2.3) until some appropriate
convergence threshold is met. Once the transformed point (P, Q) is found, interpo-
lation of any field f (x, y) from the grid onto the point (X, Y ) can be performed using


4
f (X, Y ) = fv,i Gi (P, Q), (10.2.5)
i=1

where Gi are the basis functions given in (10.2.2) and fv,i = f (xv,i , yv,i ) are the values
of f (x, y) on the four vertices of the grid cell.
The search procedure presented here can be used at each time step to locate each
particle within the grid used to compute the fluid flow, in order that the fluid velocity
and vorticity fields can be interpolated onto the particle locations. However, even
with an optimized search algorithm it is usually far too time consuming to perform
such a search at every time step, particularly because DEM computations generally
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.2. Flow in Complex Domains 293

require a large number of time steps. Alternatively, this procedure can be used to
generate a mapping between the fluid computation grid and a Cartesian grid that
covers the entire flow field. Assuming that both grids are fixed in time, such an
approach allows the fluid velocity to be rapidly mapped onto the Cartesian grid and
then interpolated from there onto the particle locations. Although this procedure
involves two interpolation steps, it requires that the time-consuming search step only
be performed once at the beginning of the computation.

10.2.2. Level Set Distance Function

Interpolation of the fluid velocity and vorticity fields onto a Cartesian grid results
in internalization of the flow domain boundaries, with the result that the domain
boundaries no longer coincide with the grid boundaries. In order to include colli-
sions between particles and the domain boundaries in the computations, a level-set
distance function ϕ(x, t ) can be used, which is defined such that ϕ(x, t ) = 0 on the
domain boundary. The absolute value of ϕ(x, t ) is equal to the distance to the nearest
domain boundary and the sign of ϕ(x, t ) is positive inside the flow domain and nega-
tive outside the flow domain. Based on this definition, it is evident that the level-set
function satisfies the Eikonal equation

|∇ϕ| = 1. (10.2.6)

The value of the level-set function is determined on the Cartesian grid points. For
cases where the flow computation grid is fixed, the level-set function can be computed
once at the start of the computation and then interpolated onto the centroid location
of the particles at each time step. Collision of a particle with a domain boundary
(e.g., a wall) occurs when the value ϕi of the level-set function at the centroid of
particle i is less than the particle radius ri . The normal overlap between the particle
and the wall is given by δN = ri − ϕi , and the particle-wall contact point is given by
xC = xi − n̂ϕi , where n̂ = ∇ϕ/|∇ϕ| is the wall unit normal (pointing into the flow
domain).
The level-set distance function can be constructed by solving (10.2.6) using the
fast marching method of Sethian (1996a). In this method, the level-set function
is first initialized at Cartesian grid points that surround the domain boundaries
(i.e., the ϕ = 0 surface). The value of the level-set function is then obtained for other
grid points by solution of (10.2.6) in the discretized form
"  2  −y 2  −z 2 #1/2
+y
max D−x +x +z
i jk ϕ, −Di jk ϕ, 0 +max Di jk ϕ, −Di jk ϕ, 0 +max Di jk ϕ, −Di jk ϕ, 0 =1
(10.2.7)

where D−x +x
i jk and Di jk denote the backward and forward differences in x, and so forth
for the y and z directions. This equation results in a quadratic equation that can be
solved to obtain the level-set function at each Cartesian grid point using given values
of the level-set function on neighboring grid points. The fast marching method solves
(10.2.7) on the remaining Cartesian grid points using only upwind values of ϕ, which
is achieved by propagating information from smaller values of |ϕ| (i.e., from grid
points lying closer to the ϕ = 0 curve) to larger values of |ϕ|. A detailed explanation
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

294 Computer Implementation and Data Analysis

Volume Pq

Figure 10.3. Illustration showing volume Pj (shaded) opposite to node

j. [Reprinted with permission from Marshall and Sala (2013).]
Particle

Node q

of the fast marching algorithm for solving (10.2.7) is given in the monograph by
Sethian (1996b, Chapter 8).

10.3. Measures of Local Concentration

It is often desirable to compute the concentration field for particulate flows. The
concentration field is used as a way to illustrate variation of particle number density
or segregation of particles of different volume within the flow field, and it is an
important step in computing the body force imposed on the fluid flow due to the
particle interfacial force. Computation of the concentration field can often be chal-
lenging, particularly when the particle diameter d is a significant fraction of the grid
increment size x. A comparative study of five different methods for computation
of concentration field is given by Marshall and Sala (2013), using two methods based
on use of a Cartesian grid and three methods based on use of a radial basis function
representation.
The most common approach for computing the concentration field distributes
the volume of each particle to the nodes of the grid cell containing the particle
centroid. The change in the concentration φq, j at the jth node of a grid cell q due
to the nth particle is given by
Vp,n Pj
φq, j = (10.3.1)
G2cell
where Pj , j = 1, . . . , 8, is the volume of the part of grid cell q opposite to node j,
as illustrated in Figure 10.3. In this equation, Vp,n is the volume of particle n and
Gcell is the volume of cell q. The box-counting method given by (10.3.1) satisfies the
discretely conservative property, which requires that the numerical approximation
of the integral of the concentration field over the flow volume is equal to the sum of
the volume of particles in the flow, or


Q 
N
M0 ≡ φq Gcell = Vp,n (10.3.2)
q=1 n=1

where φq is the average concentration value in grid cell q. One difficulty with this
method is that it produces an excessive amount of noise for cases where d/x is not
sufficiently small.
The concentration blob method for computation of the concentration field uses
a radial basis function approximation (Babic, 1997; Zhu and Yu, 2002), in which the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.3. Measures of Local Concentration 295

4
y
2

U 0
0 5 10 15 20
x
Figure 10.4 Plot showing particles of large and small sizes in a backward-facing step flow,
which originates in the small channel on the left and exits the computational domain from the
large channel on the right-hand side of the figure.

concentration at a position x is written in as a sum of concentration “blobs,” in which

the centroid x̂n of blob n coincides with that of particle n. The concentration field
associated with each blob is decomposed as the product of an amplitude coefficient
An and a weighting function f, such that the total concentration field is given by


N
φ(x, t ) = An f (x − x̂n , Rn ) (10.3.3)
n=1

where Rn is a characteristic distance associated with blob n, called the blob radius.
The function f is normalized so that its integral over all space equals unity. A common
weighting function is the Gaussian function
2    
f (x − x̂n , Rn ) = exp − x − x̂ 2 /R2 . (10.3.4)
n n
3π R3n
If the blob amplitude An is set equal to the particle volume Vp,n , the exact integral
of (10.3.3) over the flow field gives
 
N
φ(x, t )dv = Vp,n (10.3.5)
V f low n=1

The concentration blob method yields a smoothly varying concentration function for
arbitrary grid increment size. However, the method does not satisfy the discretely
conservative property (10.3.2).
A third method for concentration field calculation, proposed by Marshall and
Sala (2013), is given by a discretely conservative form of the concentration blob
method, in which instead of setting the blob amplitude equal to the particle volume,
it is specified as
Vp,n
An = , (10.3.6)
)
Q
Gcell f (x j − x̂n , Rn )
j=1

where x j is the location of the centroid of grid cell j and x̂n is the centroid of particle
n. The discretely conservative property (10.3.2) is identically satisfied when the blob
amplitude is set using (10.3.6).
An example illustrating different concentration computation methods is shown
for the backward-facing step flow field shown in Figure 10.4. In this flow, particles are
carried into the computational domain within the small channel on the left-hand side
of the figure. Two particle sizes are used, with a large particle size with dimensionless
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

296 Computer Implementation and Data Analysis

(a)
4
y
2

0
0 5 10 15 20
x
(b)
4
y
2

0
0 5 10 15 20
x
(c)
4
y
2

0
0 5 10 15 20
x
Figure 10.5. Concentration contours obtained using the box-counting method with different
numbers of grid cells: (a) 25 × 5 grid; (b) 50 × 10 grid; (c) 100 × 20 grid. Regions with φ < 0.05
are shaded gray and regions with φ > 0.4 are shaded black.

diameter d1 = 0.16 (nondimensionalized by the inlet channel width L) and small

particles with diameter d2 = 0.08. These two particle sizes are selected to mimic
flow of red blood cells and platelets, respectively, in both size and number ratio in
a two-dimensional inlet channel with fluid Reynolds number ReF = 125. In keeping
with the usual particle migration in blood flows, the larger particles (RBCs) are
primarily located in the central region at the inlet and the small particles (platelets)
are primarily located near the sides (Aarts, 1988). The flow field separates at the
step and reattaches to the top boundary x = 16. However, the pressure gradient
downstream of the step causes a second flow separation along the bottom surface
at about x = 15. Most large particles are swept downstream in the region of highest
flow velocity. However, some large particles and a relatively high number of small
particles become caught in the recirculation region near the top stagnation point and
are carried backward into the recirculation region.
Concentration computations for this flow were conducted with all three methods
described earlier; however, the results with the radial basis function method (10.3.3)
and the discretely conservative form of this method given in (10.3.6) were visually
almost identical, so we only show the former results. Figure 10.5 shows results for the
concentration fields computed using the box-counting method with three different
grid sizes. For the coarsest grid size, the contour lines are poorly resolved and exhibit
jagged corners. The middle-size grid begins to exhibit significant noise, and the finest
grid exhibits considerable noise in the concentration field. This noise arises because
the grid cell size for this grid is close to the large particle diameter, so the number
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.4. Measures of Particle Agglomerates 297

(a)
4
y
2

0
0 5 10 15 20
x
(b)
4
y
2

0
0 5 10 15 20
x
Figure 10.6. Concentration contours obtained using the concentration blob method on a
100 × 20 grid, with blob radius: (a) R = 0.4; (b) R = 0.2. Regions with φ < 0.05 are shaded
gray and regions with φ > 0.4 are shaded black.

of particles in a grid cell is small. Figure 10.6 shows concentration results obtained
using the concentration blob approach on the finest grid (100 × 20), with blob radius
R = 0.4 in Figure 10.6a and R = 0.2 in Figure 10.6b. Figure 10.6a exhibits a smooth
concentration field that appears to be consistent with the particle locations in Fig-
ure 10.4. The results with R = 0.2, on the other hand, exhibit significant noise because
the blob radius has been reduced to significantly less than the distance between
neighboring particles. In general, both the box-counting and concentration blob
methods are capable of generating accurate concentration fields provided that blob
radius and grid increment used in the computations are carefully selected, although
in general the results of the concentration blob computations are smoother than
those for the box counting method.

10.4. Measures of Particle Agglomerates

The structure of an agglomerate is related to the number and mass of particles
making up the agglomerate, the volume of the region occupied by the agglomerate,
the relative locations of contact points of touching particles, and (for nonspherical
particles) the relative orientations of these particles. A variety of different mea-
sures are utilized for analysis of these different features of agglomerate structure.
The following discussion presents some of these measures for systems composed of
spheroidal particles, which are used because although spheroids still offer a rela-
tively simple description of the particle geometry, they have the additional feature
of a preferential orientation that is not present in spherical particles.

10.4.1. Particle Count and Orientation Measures

A number of simple agglomerate measures can be developed by tracking the agglom-
erate size distribution, which gives the percentage of agglomerating particles con-
tained in agglomerates consisting of some number n particles. Dividing the total
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

298 Computer Implementation and Data Analysis

number of particles contained in all agglomerates by the number of agglomerates

gives the average number of particles per agglomerate.
The orientation of the particle symmetry axis can be used to measure tendency
for particle alignment along certain preferential directions, such as the maximum
direction of stretch in a shear flow. Defining a unit vector m along the particle
symmetry axis, particle orientation measures ζx , ζy , and ζz in each of the three
coordinate directions can be introduced as
1    1    1 
N N N
ζx = mx , ζy = my , ζz = |mz |, (10.4.1)
N N N
i=1 i=1 i=1

where N is the total number of particles. A value of ζx equal to unity indicates that
the particle symmetry axis is aligned along the x-direction (streamwise), whereas
a value of ζx of zero indicates that the particle symmetry axis is orthogonal to the
streamwise direction.

10.4.2. Agglomerate Orientation Measures

Chesnutt and Marshall (2010) introduced a number of orientational measures for
agglomerates formed of spheroidal particles. A symmetry-axis-angle orientation mea-
sure OI is defined based on the angle θi j between the symmetry axes of two touching
spheroidal particles, labeled i and j. This measure provides information about the
relative orientation of particles within an agglomerate. This orientation measure is
defined by

1    
N N
OI = ai j cos θi j , (10.4.2)
2NT
i=1 j=1

where NT is the number of touching particle pairs and the indicator coefficient ai j
equals unity if particles i and j are touching each other and zero otherwise. When
OI = 1 the symmetry axes of all particles are parallel, and when OI = 0 the symmetry
axes of all touching particle pairs are perpendicular.
Information about the location of the particle contact points is given by the
contact-point orientation measure OII . This measure is defined in terms of the angle
ϑi j between the symmetry axis of particle i and the line joining the center of particle
i and its contact point with another particle j. This measure is defined by

1    
N N
OII = ai j cos ϑi j . (10.4.3)
2NT
i=1 j=1

When OII,i j ≡ |cos ϑi j | = 1, the contact point of particle i with particle j is located
at one end of the symmetry axis (a “pole”) of particle i. For OII,i j = 0, the contact
point of particle i with particle j is located on the “equator” of particle i.

10.4.3. Equivalent Agglomerate Ellipse

The orientation measures introduced in the previous section deal with the relative
orientation of different particles contained in an agglomerate. In the current section,
we describe a method to instead examine the orientation of the agglomerate itself.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.4. Measures of Particle Agglomerates 299

Equivalent
agglomerate
ellipse

(a) (b)
Figure 10.7. (a) Illustration showing the extrema points of a set of (two-dimensional) elliptical
particles, along with the equivalent agglomerate ellipse shown as a heavy solid line, where the
ellipse center is marked by an x. (b) Equivalent aggregate ellipses (bold) fit to aggregates in
a close-up of a 2D channel flow.

Specifically, the size, shape, and orientation of an agglomerate is characterized by

fitting an ellipse to the projection of the agglomerate along a plane formed of two
coordinates, such as the x–y or x-z planes. The first step prior to fitting this ellipse is
to identify a set of extrema points corresponding with the intersection of the particle
principal axes with the particle surface. For spheroidal particles projected into the
y-z plane, the extrema points along the particle equator are chosen to be the farthest
points in the y- and z-directions, respectively, as shown in Figure 10.7a. The extrema
points for all particles in the agglomerate are projected onto the projection plane.
An ellipse can be fit to the set of extrema particles on the projection plane
using the direct least squares method described by Fitzgibbon et al. (1999), which
was subsequently modified by Halir and Flusser (1998) to improve the numerical
stability. Using this method, an ellipse can be expressed as a quadratic form as

F (x, y) = ax2 + bxy + cy2 + dx + ey + f = 0, (10.4.4)

where the requirement that the general conic be of elliptical shape is satisfied subject
to the inequality constraint

b2 − 4ac < 0. (10.4.5)

The least-square fit of the elliptic curve to the given set of extrema points results
in a constrained minimization problem, the numerical solution of which is simplified
if we scale the ellipse coefficients so that the right-hand side of (10.4.5) has a specific
value. For instance, if this value is set equal to −1, then the constraint becomes

4ac − b2 < 1. (10.4.6)

The coefficients in (10.4.4) can be obtained using a direct least-squares method with
the constraint (10.4.6) imposed by the Lagrange multiplier method, as described by
Halir and Flusser (1998). An example showing the equivalent agglomerate ellipses
(in bold) computed using this method for elliptical particles in a two-dimensional
shear flow is shown in Figure 10.7b.
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

300 Computer Implementation and Data Analysis

(a) (b) (c)

Figure 10.8. Examples of three particle agglomerates of complex shape, with agglomerate
compactness increasing from left to right. [Reprinted with permission from Eggersdorfer
et al. (2011).]

A third orientation measure is defined by Chesnutt and Marshall (2010) in terms

of the orientation of the particles contained within an agglomerate relative to the
orientation of the agglomerate itself. Let Jk denote the set of Mk particles contained
in an agglomerate k, and let αik be the angle between the symmetry axis of particle
i and the major axis of the equivalent ellipse of agglomerate k. A particles-within-
agglomerate orientation measure OIII,k is defined by
1 
OIII,k = |sin αik |. (10.4.7)
|Jk |
i∈Jk

When |sin αik | = 1 the symmetry axis of particle i is orthogonal to the major axis
of the equivalent agglomerate ellipse, and when |sin αik | = 0 the symmetry axis of
particle i is parallel to the major axis of the equivalent ellipse. The average of OIII,k
over all agglomerates is
NA
1 
OIII = OIII,k , (10.4.8)
NA
k=1

where NA is the number of agglomerates.

10.4.4. Agglomerate Fractal Dimension

One of the distinguishing features of particle agglomerates is that they exhibit com-
plex geometry as the number of particles in the agglomerate increases. Experimental
results from Eggersdorfer et al. (2011) showing complex particle agglomerates with
varying degrees of compactness are shown in Figure 10.8. This geometry can be
characterized by measures such as the number of particles in an agglomerate, the
area shaded by the agglomerate when projected onto a plane, or some other measure
of the size of an agglomerate. A common size measure is the radius of gyration RG ,
which is defined by

N
R2G = |xi − x̄|2 , (10.4.9)
i=1

where N is the number of particles in the agglomerate, xi is the centroid position for
the ith particle, and x̄ is the centroid of all of the particles in the agglomerate. The
radius of gyration provides a measure of the effective radius of the agglomerate.
As the number of particles within an agglomerate grows large, certain relation-
ships are observed to form between the number of particles in the agglomerate and
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

10.4. Measures of Particle Agglomerates 301

7 7

Acetylene
Propylene
6 6
Ethylene
Propane

5 5
ln (N )

ln (N )
4 4

3 3

2 2
0 1 2 3 2 3 4 5 6

ln (RG /d) ln (Aa /Ap)

(a) (b)
Figure 10.9. Plots showing power-law scaling for agglomerates formed of soot particles from
different combustible gases obtained by plotting the logarithm of the number of particles
versus (a) the logarithm of the radius of gyration RG normalized by the particle diameter
d and (b) the logarithm of the agglomerate projected area Aa normalized by the projected
particle area A p = π d2 /4. [Reprinted with permission from Köylü et al. (1995).]

the various measures of agglomerate size. These relationships usually take the form
of a power law with noninteger exponent, which is characteristic of geometric objects
called fractals (e.g., Mandlebrot, 1982). An important power-law relationship occurs
between the number of particles in the agglomerate and the radius of gyration RG ,
which has the form
N = kG (RG /r p )d f . (10.4.10)
In this expression, it is assumed that the N particles making up the agglomer-
ate have a uniform radius r p . The coefficient kG is called the fractal pre-factor and
d f is the fractal dimension of the agglomerate. An example showing experimental
data for N and RG for agglomerates of soot particles obtained from flames gen-
erated with different combustible gases is shown in Figure 10.9a, from Köylü et al.
(1995). Although there is some natural variability in the agglomerates, the data agree
reasonably well with the power-law expression (10.4.10) over two orders of magni-
tude of variation in agglomerate size. Similar results are reported by a number of
other investigators, including agglomerates generated using numerical simulations
by Brasil et al. (1999). Based on a review of many studies reported in the literature,
Brasil et al. (1999) report typical values of the coefficient kG to be in the range 1.2
to 3.5. Reported values of the fractal dimension d f vary between about 1.5 and 2.3
depending on the agglomerate formation process, although the most common value
is about 1.8.
It is often very difficult to experimentally determine RG for particle agglomerates
formed in the laboratory, whereas it is a simple matter to estimate agglomerate size
using the projected area of the agglomerate onto a planar surface. For this reason,
many experimental studies instead plot the number of particles in an agglomerate
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

302 Computer Implementation and Data Analysis

as a function of the agglomerate projected area, which is typically normalized by

the particle projected area A p = π d2 /4. A second power-law expression is typically
observed between these two variables, given by

N = ka (Aa /A p )α , (10.4.11)

where ka and α are unknown coefficients. A typical plot of this type is shown in
Figure 10.9b, for which case ka = 1.05 and α = 1.09. A procedure is described by
Brasil et al. (1999) for recovering the coefficients in the expression (10.4.10) for
the three-dimensional agglomerate from quantities measured from the agglomerate
projection, such as the coefficients in (10.4.11).
Alternatively, an expression of the form (10.4.10) can be plotted using the two-
dimensional radius of gyration RG,2D obtained directly from the projected image
of the agglomerates. Köylü et al. (1995) examined simulated particle agglomerates
with two orders of magnitude variation in N and found that the three-dimensional
gyration radius RG,3D is proportional to that obtained from the two-dimensional
projection, such that

RG,3D /RG,2D = 1.24 ± 0.01. (10.4.12)

As a consequence, the fractal dimension d f in (10.4.10) will be approximately the

same whether the expression is written in terms of RG,3D or RG,2D .

10.4.5. Particle Packing Measures

Some of the traditional measures used for characterization of packed beds are also
useful for particle agglomerates. One such measure is called the coordination num-
ber, which is defined as the average number of spheres in contact with any given
sphere. The coordination number is a simple topological measure that is commonly
used in computational studies to assess the degree of particle contact within a flow
(Yang et al., 2000). Experimental studies typically do not have sufficient informa-
tion on the positions and contact of all spheres, so this measure is not as frequently
reported in the experimental literature.
Another measure of the particle spacing is obtained by computing the average
number of neighboring particles, N(r), whose centroids are located within a radial
distance r from a given particle centroid. The pair distribution function g(r) is defined
as
1 dN
g(r) = , (10.4.13)
4π ρ0 r2 dr
where ρ0 is the average particle number per unit volume, which is related to the
particle concentration φ by ρ0 = 6φ/π. The function g(r), which is also known as
the radial distribution function, can be interpreted as the probability distribution of
finding the center of a particle at a distance r from a reference particle center. The
radial distribution function is widely used for characterizing packing chain structures
(Aste et al., 2005; Zou et al., 2009). A plot of the radial distribution function as a
function of distance r is shown in Figure 10.10, based on computational data from
Aste et al. (2005) for a large packed bed of particles. A large peak in the radial
distribution function curve is observed at r ∼ = d, which corresponds to neighboring
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

References 303

12 Secondary peak

First peak
10 r = 3d

Figure 10.10. Radial distribution func- 8

g (r)
Tertiary peak

Second peak
tion of dense packing of particle chains,

Third peak
showing structures associated with sec- 6
ondary and tertiary peaks (based on r = 2d
computational data of Aste et al., 2005, 4
for packed particle beds).
2

0
1 2 3 4 5
r/d

particles that are in contact with the given particle (first-neighbors). A series of
additional
√ peaks are observed for higher values of r, with the second peak occurring
at r = 3d and the third peak at r = 2d. As shown by Zou et al. (2009), the second
peak is associated with particles that form a planar rhomboid structure between
the given particle and two first-neighbor particles that are in contact with the given
particle as well as with each other (see inset of Figure 10.10). The second peak is
caused by a fourth particle that is touching the two first-neighbor particles, thus
completing the rhombus shape. The third peak corresponds to structures in which
three particles are located along a line, which include the given particle, a first
neighbor that is in contact with the given particle, and a second neighbor that is in
contact with the first neighbor. These peaks occur even though the particle packing
shows no signs of a crystalline structure. For very large r, the value of g(r) approaches
unity.

REFERENCES

Aarts PAMM, van den Broek SAT, Prins GW, Kuiken GDC, Sixma JJ, Heethaar RM. Blood
platelets are concentrated near the wall and red blood cells, in the center in flowing blood.
Arterioscler 8, 819–824 (1988).
Allievi A, Bermejo R. A generalized particle search-locate algorithm for arbitrary grids.
Journal of Computational Physics 132, 157–166 (1997).
Aste T, Saadatfar M, Senden TJ. Geometrical structure of disordered sphere packings. Phys-
ical Review E 71, 061302 (2005)
Babic M. Average balance equations for granular materials. International Journal of Engi-
neering Science 35, 523–548 (1997)
Brasil AM, Farias TL, Carvalho MG. A recipe for image characterization of fractal-like
aggregates. Journal of Aerosol Science 30(10), 1379–1389 (1999).
Chesnutt JKW, Marshall JS. Structural analysis of red blood cell aggregates under shear flow.
Annals of Biomedical Engineering 38(3), 714–728 (2010).
Dulinska I, Targosz M, Strojny W, Lekka M, Czuba P, Balwierz W, Szymonski M. Stiffness of
normal and pathological erythrocytes studied by means of atomic force microscopy. Journal
of Biochemical and Biophysical Methods 66, 1–10 (2006).
Eggersdorfer ML, Kadau D, Herrmann HJ, Pratsinis SE. Multiparticle sintering dynamics:
From fractal-like aggregates to compact structures. Langmuir 27, 6358–6367 (2011).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-10 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 13:59

304 Computer Implementation and Data Analysis

Ferry J, Balachandar S. A fast Eulerian method for disperse two-phase flow. International
Journal of Multiphase Flow 27, 1099–1226 (2001).
Ferry J, Rani SL, Balachandar S. A locally implicit improvement of the equilibrium Eulerian
method. International Journal of Multiphase Flow 29, 869–891 (2003).
Fitzgibbon A, Pilu M, Fisher RB. Direct least square fitting of ellipses. IEEE Transactions on
Pattern Analysis and Machine Intelligence 21, 476–480 (1999).
Halir R, J Flusser. Numerically stable direct least squares fitting of ellipses. The Sixth Interna-
tional Conference in Central Europe on Computer Graphics and Visualization ’98 1, 125–132
(1998).
Köylü ÜÖ, Faeth GM, Farias TL, Carvalho MG. Fractal and projected structure properties
of soot aggregates. Combustion and Flame 100, 621–633 (1995).
Mandlebrot BB. The Fractal Geometry of Nature. W.H. Freeman and Company, New York
(1982).
Marshall JS. Discrete-element modeling of particulate aerosol flows. Journal of Computational
Physics 228, 1541–1561 (2009).
Marshall JS, Sala K. Comparison of methods for computing the concentration field of a
particulate flow. International Journal of Multiphase Flow 56, 4–14 (2013).
Mousel J, Marshall JS. Aggregate growth and breakup in particulate suspension flow through
a micro-nozzle. Microfluidics and Nanofluidics 8(2), 171–186 (2010).
Sethian JA. A fast marching level set method for monotonically advancing fronts. Proceedings
National Academy of Sciences 93, 1591–1595 (1996a).
Sethian JA. Level Set Methods and Fast Marching Methods. Cambridge University Press,
Cambridge UK (1996b).
Sethian JA, Smereka P. Level set methods for fluid interfaces. Annual Review of Fluid Mechan-
ics 35, 341–372 (2003).
Yang RY, Zou RP, Yu AB. Computer simulation of the packing of fine particles. Physical
Review E 62, 3900–3907 (2000).
Zhu HP, Yu AB. Averaging method of granular materials. Physical Review E 66, 021302
(2002).
Zou LN, Cheng X, Rivers ML, Jaeger HM, Nagel SR. The packing of granular polymer chains.
Science 326, 408–410 (2009).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11 Applications

Adhesive particle flows are involved in a wide variety of applications, including

industrial, energy, environmental, biological, and geological processes of many types.
Rather than trying to treat a large number of such processes here, we restrict atten-
tion to four specific types of particle flows that are of particular interest. This interest
arises in part because the particles in these systems exhibit interesting behavior that
has important consequences in many different applied systems. Although this chap-
ter is by no means a comprehensive coverage of particle flow phenomena, it will
perhaps serve as an introduction to some of the interesting and unusual behaviors
that particulate systems can exhibit.

11.1. Particle Migration in Tube and Channel Flows

The problem of a suspension of particles carried within a tube or channel by a
fluid flow arises in a large number of problems. For example, this flow is found in
pneumatic intake systems used to feed fuel particles (e.g., coal or biomass pellets)
into a combustion unit, in the flow of blood through the cardiovascular system (where
the red blood cells and other types of blood cells are treated as particles), in flow of
digestive fluids through the intestines, in tar-sand excavation and oil extraction, and
in many other process in which a particulate slurry is transported along a channel or
pipeline. For the sake of simplicity in the current discussion let us refer to flow in a
tube, but similar phenomena occur also with channel flows. It is also convenient to
simplify the problem by neglecting gravity, or else considering a vertically oriented
flow in which gravity points in the direction of flow. We simplify the problem further
to consider laminar flow in a straight tube, so that the fluid streamlines are all aligned
in the direction of the tube axis and the particles are simply carried downstream by
the fluid flow. However, as the particles are transported along the tube axis by the
fluid flow, a very interesting thing happens – they begin to drift laterally in the tube.
This phenomenon is known as particle migration. As a result of particle migration,
the concentration profile of the particles within the tube becomes nonuniform, where
the particle concentration will become higher either at the tube center or at some
other finite radial location and it will be come lower in other regions, typically near
the tube walls, even if the concentration is uniform at the tube entrance.

305
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

306 Applications

Particle migration is most often observed in low Reynolds number laminar flows,
as fluid turbulence tends to rapidly mix the particles so as to eliminate concentration
nonuniformities. In flows where significant particle migration occurs, it can have an
important influence on the flow mechanics. One important example of this occurs in
blood flow, for which the migration of red blood cells (RBCs) toward the low-shear
region at the center of the tube (i.e., the blood vessel) leads to the phenomena of
marginalization and plasma skimming, discussed in Section 1.3.4. The first of these
phenomena causes the different constituents of blood (e.g., the red and white blood
cells and the platelets) to have highly nonuniform concentration distributions within
the blood vessel. These nonuniformities influence the overall blood flow resistance
as well as the reaction rates between the white blood cells and platelets with the
endothelial cells along the vessel walls. The plasma skimming phenomenon occurs
when an upstream channel with strong particle migration enters into a bifurcation
with nonuniform flow rates. The outlet channel with low flow rate tends to primarily
entrain fluid with low RBC concentration from near the blood vessel walls, whereas
the outlet channel with high flow rate will tend to entrain more of the RBC-rich fluid
from within the center of the channel. As a result, the average RBC concentration
within the outlet channels can be very different from each other.
It is known that there can be no lateral particle transport in flows with vanishing
particle Reynolds number and no particle collisions (Bretherton, 1962). The effects
of particle inertia and of collisions on particle migration are examined in the following
sections, followed by a discussion of the enhancement of particle migration rate that
can occur due to waviness of the tube wall.

11.1.1. Inertial Particle Migration in Straight Tubes

We consider a dilute particulate suspension undergoing a fully developed laminar
flow in a tube, in which the flow is oriented vertically upward so that gravity is
oriented opposite to the direction of flow. If the particles are heavier than the fluid,
the downward drift of the particles will induce a negative relative slip velocity of
the particles, vs < 0, and the particles will lag the fluid flow. Both the Saffman
and Magnus lift forces discussed in Section 5.2 are oriented such that particles
will be drawn inward toward the tube axis when vs < 0. If the particles are lighter
than the fluid, the opposite occurs and the Saffman lift force drives the particles
toward the tube wall. A well-known experimental study performed by Segré and
Silberburg (1962) for neutrally buoyant particles in tube flow shows that for very
dilute suspensions and finite particle Reynolds numbers, particles collect at a radial
position of approximately 60% of the tube radius. This phenomenon is called the
tubular pinch effect, and is characterized by the particles collecting at some radial
position midway between the wall and the tube axis. An example illustrating this
effect is shown in Figure 11.1, based on the experimental data of Matas et al. (2004).
The lateral force on a neutrally buoyant particle due to weakly nonlinear inertia
for two-dimensional laminar shear flow was examined by Ho and Leal (1974) using
a perturbation analysis. This paper shows that both shear rate variation and the
lateral boundaries at the channel walls lead to an inward radial force for particles
close to the wall. This inward force acting on particles close to the wall combined
with the outward lateral force on particles near the channel center is consistent with
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.1. Particle Migration in Tube and Channel Flows 307

1 0.2

0
0.1

-1
0
-1 0 1 0 0.2 0.4 0.6 0.8 1.0
(a) (b)
Figure 11.1. Experimental results for (a) particle distribution and (b) histogram of radial
locations for particles of diameter d flowing with mean velocity U in a tube of diameter D with
tube Reynolds number ReF = U D/ν = 67 and D/d = 9. [Reprinted with permission from
Matas et al., (2004).]

the experimental observations of Segré and Silberburg (1962). Extensions to this

asymptotic theory have been developed for Poiseuille flow in a tube (Schonberg
and Hinch, 1989), for particles that are not neutrally buoyant (Hogg, 1994), and for
higher Reynolds numbers (Asmolov, 1999). An experimental study by Matas et al.
(2004) showed that the radial position at which the particles collect moves outward
toward the wall as the Reynolds number is increased. These authors also observe
appearance of a second inner ring on which particles collect, located closer to the
center than the primary ring.

11.1.2. Collision-Induced Particle Migration

The effect of finite particle concentration on lateral migration in a tube was examined
by Han et al. (1999) and Hampton et al. (1997). Experimental results from Han
et al. (1999) are plotted in Figure 11.2 for a nominal particle Reynolds number
of 0.2 with four different values of the average concentration. For small average
concentration values, the particles collect preferentially at a finite radial position. As
the average concentration increases, the particles increasingly drift to the low-shear
region near the tube center. For moderate average concentration values there are
local concentration maxima at both a finite radial location and at the tube center. At
large values of the average concentration, the only concentration maximum occurs
at the tube center, which exhibits a local cusp in concentration value.
The tendency for particles to preferentially drift toward the center of the tube as
the average concentration increases is caused by the phenomenon of shear-induced
migration. This phenomenon was initially identified by Leighton and Acrivos (1987),
who also developed a scaling theory to estimate the particle diffusive flux. The dif-
fusion associated with this phenomenon comes from irreversible particle interac-
tions (e.g., collisions), the rates of which vary as a function of the shear rate, the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

308 Applications

0.3

0.2

φ D

0.1 A
C

B
0
0 0.2 0.4 0.6 0.8 1.0

r / rp
Figure 11.2. Plot of concentration profiles in tube flow with nominal particle Reynolds number
of 0.2, for bulk concentrations of φ0 = 0.06 (Line A, solid), 0.1 (Line B, dashed), 0.2 (Line C,
solid), and 0.28 (Line D, dashed). Based on experimental data of Han et al. (1999).

particle concentration, and the effective viscosity of the multiphase flow. Specifi-
cally, Leighton and Acrivos argue that a particle placed in a flow field with a nonzero
gradient in the particle concentration φ or shear rate γ̇ will experience a higher rate
of collisions with other particles on the side of the particle with higher concentra-
tion or higher shear rate, where the excess rate of interactions on this side of the
particle is proportional to −r p ∇(γ̇ φ). Assuming that the displacement of the given
particle after each collision is proportional to the particle radius r p , then the diffusive
flux associated with the concentration or shear rate gradient should be proportional
to −r2p φ∇(γ̇ φ).
The governing equation for particle concentration φ in the continuum particulate
flow theory has the form
Dφ
+ ∇ · N = 0, (11.1.1)
Dt
where N is the particle diffusive flux. Phillips et al. (1992) developed a constitutive
equation for N that has three terms, where
N = Nb + Nc + Nμ . (11.1.2)
In this equation, Nb is the diffusive flux from Brownian motion, given by
Nb = −Db ∇φ, (11.1.3)
where Db is the Brownian diffusion coefficient defined in (5.4.11). The second term in
(11.1.2) is the shear-induced diffusion related to concentration or shear rate gradient.
Using the Leighton-Acrivos scaling, we can write
Nc = −Kc r2p (φ 2 ∇ γ̇ + φ γ̇ ∇φ), (11.1.4)
where Kc is a constant. The third term in (11.1.2) is a second type of shear-induced
diffusion that is related to the fact that a gradient in particle concentration induces
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.1. Particle Migration in Tube and Channel Flows 309

Figure 11.3. Result of a DEM simulation of particle transport in a tube with a wavy wall
and periodic boundary conditions. This figure shows only a slice of the particles within a
section passing through the channel centerline. The particles initially fill the tube with uniform
concentration, but over time the particles drift toward the central part of the tube. [Reprinted
with permission from Hewitt and Marshall (2010).]

an associated gradient in the effective viscosity of the two-phase mixture. Assuming

that the shear-induced collision frequency is proportional to ϕ γ̇ and that the net
drift associated with viscosity gradient is proportional to the change in effective
viscosity over the particle radius divided by the effective viscosity, or r p (∇μeff )/μeff ,
the diffusive flux associated with viscosity variation can be written as
r2p dμeff
Nμ = −Kμ φ 2 γ̇ ∇φ, (11.1.5)
μeff dφ
where Kμ is a constant. A typical empirical formula relating effective viscosity with
concentration is that proposed by Krieger (1972):

μeff = μ(1 − φ/φm )−1.82 , (11.1.6)

where φm is the maximum packing concentration, equal to approximately 0.68 for

spheres.
Comparison of the Phillips et al. (1992) theory with experimental data is reported
by Hampton et al. (1997) for pressure-driven flow in a circular tube and by Koh et al.
(1994) for flow in a rectangular channel. The predictions for concentration profile
are in reasonable agreement with the data, except in the immediate vicinity of the
channel center. The Phillips et al. theory predicts a sharp peak in concentration at
the tube or channel center, predicting that the centerline concentration will always
be equal to the maximum packing concentration φm no matter what the value of the
average concentration within the tube. This prediction at the centerline disagrees
with the experiments, particularly for lower values of the average concentration. An
extension to this theory that does not require that concentration approach φm at the
tube centerline was proposed by Nott and Brady (1994), who also present a validation
of their theory using Stokesian dynamics simulations. This theory was extended for
Brownian suspensions by Frank et al. (2003) and to account for streamline curvature
effects by Kim et al. (2008).

11.1.3. Particle Migration in the Presence of Wavy Tube Walls

It was reported by Hewitt and Marshall (2010) that the inward migration of parti-
cles in a pressure-driven tube flow is dramatically accelerated if the tube has wavy
walls. An example is shown in Figure 11.3, showing results of a DEM simulation
with periodic end conditions, where the particle concentration was initially uni-
form throughout the tube. The interaction of the suspended particles with the flow
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

310 Applications

oscillations induced by the wavy walls causes a rapid inward particle drift, leading
the particles to collect within a region away from the tube walls.
To see why this rapid migration occurs, we recall the approximation (10.1.10),
which for small particle Stokes numbers gives the particle drift velocity vd = v − u
in dimensionless form as
vd = −St a , (11.1.7)

where velocity is nondimensionalized by the mean fluid velocity U, acceleration

is nondimensionalized by U 2 /D, and D is the tube mean diameter. The furrow-
averaged radial location of particle n is defined as
 z
1 n
r̄n (zn ) ≡ r (z)dz, (11.1.8)
λF zn −λF n
where (rn , zn ) denotes the instantaneous radial and axial coordinates of particle n and
λF is the wavelength of the furrow. Assuming small wave amplitude, η ≡ A/D 1,
the fluid acceleration at the particle centroid can be expanded in a Taylor series
about the furrow-averaged position r̄n (t ) to write

 
      ∂a 
a (rn ) = a (r̄n ) + (rn − r̄n )   +··· , (11.1.9)
∂r r=r̄n
where the omitted terms are higher order in the small parameter η. Two functions,
vd1 and vd2 , called the first and second particle drift velocities, are defined by

vd1 ≡ −St a (r̄ ), vd2 ≡ −St [(r − r̄ )∂a /∂r ], (11.1.10)

where the total drift velocity is the sum of vd1 and vd2 . The first drift velocity is the
average of the fluid acceleration along the tube at the furrow-averaged radial position
of the particle. The second drift velocity is related to the correlation between the
radial oscillation of the particle about the furrow-averaged location and the variation
of the radial acceleration gradient.
In a steady flow, the definition of a streamline can be integrated to write the
radial perturbation from the furrow-averaged position as

r − r̄ = (u /w  )dz ,
 
(11.1.11)

where u and w are the radial and axial components of the fluid velocity, respectively.
Using the lubrication approximation for small flow Reynolds numbers and η 1,
the fluid velocity components can be approximated as
2r αη 2
u = − [1 − (r /h )2 ] sin(αz ), w = [1 − (r /h )2 ], (11.1.12)
h3 h2
where α ≡ kD, k = 2π /λF is the wave number of the wall corrugation, and h(z) is
the tube diameter at axial position z. Substituting (11.1.12) into (11.1.9), (11.1.10),
and (11.1.11) gives the radial component of the two parts of the drift velocity to
leading order in η as

vd1 = −2St (αη)2 r̄ (1 − r̄2 )(3 − 7r̄2 ), (11.1.13a)

vd2 = 2 St (αη)2 r̄ (1 − r̄2 )(1 − 5r̄2 ). (11.1.13b)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.2. Particle Filtration 311

1.5

0.5
v'd
0

–0.5

–1

–1.5
0 0.2 0.4 0.6 0.8 1
r'

Figure 11.4. Plot showing first particle drift velocity vd1 (dashed line), second particle drift

velocity vd2 (dashed-dotted line), and net particle drift velocity vd = vd1
 
+ vd2 (solid line) in
the radial direction for low Stokes number particle transport in a tube given by the lubrication
theory. All velocities are normalized by St (αη)2 . [Reprinted with permission from Hewitt and
Marshall (2010).]

The particle drift velocities, normalized by St (Ak)2 , are plotted in Figure 11.4 as
functions of radius. The magnitude of the first drift velocity is significantly greater
than that of the second drift velocity within the central part of the tube. The drift
velocities each change sign within the outer part of the tube, and in this region both
drift velocities have similar magnitude. The net drift velocity is observed to have a
negative sign throughout the tube, implying that all particles will drift toward the
tube center at a radial velocity proportional to St (Ak)2 . Extension of this lubrication
theory result to finite flow Reynolds numbers is examined computationally using
DEM by Hewitt and Marshall (2010), who find that with increase of the Reynolds
number the particles drift to a region with nonzero radius within the tube, rather
than drifting to the tube axis.

11.2. Particle Filtration

Collision of particles on surfaces and their subsequent adhesion is an important
area of research for applications ranging from air or water filtration to fouling of
microfluidic circuits (Perry and Kandlikar, 2008). Problems of this type involve
particles suspended in a fluid flow past a blunt body, such as a fiber or the fin at the
end of a microchannel. The fluid drag force attempts to move the particles around
the body, but drift due to particle inertia allows the particles to move closer to the
surface of the bluff body at conditions of finite Stokes number than they would if
the particles followed the fluid flow streamlines. Collisions of the particles with the
surface occur due to a combination of the particle drift within the region of high
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

312 Applications

DEM domain
4

y 0

–2

–4
–10 –5 0 5 10
x
Figure 11.5. Fluid flow past a 10 µm diameter fiber at a fiber Reynolds number ReF = 0.178.
The inner region used for DEM simulations is indicated by a dashed line. [Reprinted with
permission by Li and Marshall (2007).]

streamline curvature in the stagnation-point flow and the finite particle size effect
within the high shear regions along the particle sides, as discussed in Section 1.3.1.

11.2.1. Fiber Filtration

Fiber filtration, with either a single fiber or a fiber array, has been studied by a
number of investigators as being representative of particle interaction with bluff
bodies immersed in the flow field. The initial stages of particle collection on the
fiber can be adequately handled using the one-way coupling approximation, because
the flow field is not very much affected by the particles. As the particle deposits on
the fiber build up over time, it is necessary to utilize some method to account for
modification of the fluid flow near the fiber by the particle deposits. A typical flow
field used to study particle capture on a fiber is plotted in Figure 11.5. Because the
time interval and number of particles that can be practically simulated with DEM is
limited, it is common in filtration computations to use DEM only in a small region
around the fiber, indicated by dashed lines in Figure 11.5. It is also common to
accelerate the process of particle collection by significantly increasing the particle
concentration at the inlet to the computational domain compared to what is observed
in practice.
A sketch of the fluid streamlines in flow past a cylindrical fiber at low Reynolds
number is shown in Figure 11.6, along with the motion of a select particle. The
fiber surface is divided into three regions in this sketch. At the front of the fiber
(Region A) is a region in which particles that disperse rapidly from the fluid flow,
that is, particles with intermediate or large Stokes numbers, will collide with the

B
Figure 11.6. Sketch showing fluid streamlines in the flow past a
A cylindrical fiber as well as motion of a select particle. The fiber
C
surface is divided into three regions: (A) rebound region, (B)
collision region, and (C) no-impaction region.
B
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.2. Particle Filtration 313

t = 10 t = 300
2 2
1 1

0 0
Y

Y
–1 –1

–2 –2
–4 –2 0 2 4 –4 –2 0 2 4
X X
t = 100 t = 500
2 2
1 1

0 0
Y

Y
–1 –1

–2 –2
–4 –2 0 2 4 –4 –2 0 2 4
X X
t = 200
2
1

0
Y

–1

–2
–4 –2 0 2 4
X
Figure 11.7. Time series showing results of a DEM simulation of particles collecting on
the front surface of a cylindrical fiber. The agglomerate observed in the dashed ellipse at
t = 300 has been stripped off and carried downstream by t = 500. [Reprinted with permission
from Li and Marshall (2007).]

fiber. These particles will strike the fiber surface at a large angle and with relatively
high velocity, so they will be most likely to bounce from the fiber surface. Smaller
particles with smaller Stokes numbers will tend to more closely follow the fluid flow.
Even if these particles are aimed directly toward the fiber far upstream, they will
tend to be swept aside by the fluid flow near the fiber front surface and collide
with the fiber closer to the shoulders of the cylinder, as indicated by Region B in
Figure 11.6. Because these particles are smaller than those colliding in Region A,
they have lower inertia and are less inclined to bounce following collision with the
fiber surface. For noncharged fibers it is rare to find particles adhering to the rear
of the fiber, indicated by Region C in Figure 11.6, as the particles instead tend to be
swept downstream with the fluid flow and do not often enter this region.
A time series showing results of a DEM simulation by Li and Marshall (2007)
of particle collision and adhesion on an array of cylindrical fibers is shown in Fig-
ure 11.7. For the case examined, the adhesive surface energy and the normal dissi-
pation coefficient of the particles were sufficiently high that even particles colliding
near the cylinder leading edge did not bounce significantly from the surface. The
particles are observed to continue collecting on the front face of the fiber, leading
to development of a particulate structure that projects a length L(t ) upstream of the
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

314 Applications

Maximum length of dendrites (μm)

25 Simulation (φ=1000) Experimental
No. 1
Case 3 No. 2
20

15
Experimental

10

5 Simulated

0
0 200 400 600 800 1000 1200 1400
Dimensionless time, t
Figure 11.8. The variation of the length L of the dendritic structure upstream of the fiber
with dimensionless time, comparing DEM simulations of Li and Marshall (2007) (line) with
experimental results of Huang et al. (2006) (symbols). [Reprinted with permission from Li
and Marshall (2007).]

cylinder leading edge. Variation of the length of this particulate structure with time as
computed by the DEM simulations is compared with experimental results obtained
by Huang et al. (2006) in Figure 11.8, showing reasonable agreement between the
computational results and experimental data. The soft-sphere DEM method used
by Li and Marshall allows for particle agglomerates to break off of the fiber if they
become too long to sustain the fluid drag force exerted on them by the flow field.
This break-off can be seen in Figure 11.7, for instance, by noting the agglomerate
present within the dashed elliptical region at dimensionless time t = 300 that is no
longer present at time t = 500, but instead has broken off of the fiber due to the
high shear stress in this region and has been carried downstream out of the field of
view. The agglomerate break-off process allows the particles adhering to the fiber to
approach a state of statistical equilibrium after long time. This attainment of a quasi-
equilibrium state associated with intermittent break-off of particle agglomerates is a
key observation of the experimental study of Huang et al. (2006), but such a state is
not allowed by many previous models of particle filtration processes, which assume
that once a particle collides with the surface it is determined to have adhered to the
surface and the adhesive bond is “frozen” in place (Kanaoka et al., 1983; Ramarao
et al., 1994; Konstandopoulos, 2000).
An important measure of fiber filtration processes is the fiber efficiency η,
defined as the fraction of incident particles that collide with the surface minus
the fraction of incident particles that rebound from the surface. A plot compar-
ing predictions for single-fiber efficiency from the soft-sphere DEM simulation by
Yang et al. (2013) with experimental data from Rembor et al. (1999) is given in Fig-
ure 11.9. The fiber efficiency is plotted as a function of Stokes number over an interval
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.2. Particle Filtration 315

0.4

Single fiber efficiency dp=2.83um, df=30um, DEM

dp=1.77um, df=20um, DEM
0.3 dp=2.83um, df=30um, Exp
dp=1.77um, df=20um, Exp

0.2

0.1

0.0
0.1 1 10
St
Figure 11.9. Comparison of soft-sphere DEM simulation predictions for single-fiber capture
efficiency (lines with open symbols) with experimental data of Rembor et al. (1999) (filled
symbols), for different particle sizes as a function of Stokes number. Square symbols are for
particle diameter 2.83 µm and fiber diameter 30 µm, and circular symbols are for particle
diameter 1.77 µm and fiber diameter 20 µm. [Reprinted with permission from Yang et al.
(2013).]

0.35 < St < 7. The DEM results are again found to agree closely with the corre-
sponding experimental data. In addition to the Stokes number, the fiber efficiency
depends on the adhesive surface energy of the particles. In order to model the ten-
dency of particles to rebound from the fiber surface upon collision, it is convenient to
nondimensionalize the particle surface energy by its inertia, which gives the adhesion
parameter Ad, defined by

2γ
Ad = , (11.2.1)
ρ pU 2 d

where d is the particle diameter. A plot of DEM predictions for the single fiber
efficiency as a function of Stokes number and adhesion parameter is given in Fig-
ure 11.10, from Yang et al. (2013). The fiber efficiency is observed to be large within
a region with large Stokes number (between 1 and 10) and large adhesion parameter
(between 1 and 100). A Stokes number greater than unity is necessary in order for
the incident particles to have a sufficiently large drift velocity relative to the fluid to
collide with the fiber surface. A sufficiently large adhesion parameter is necessary so
that these colliding particles will stick to the surface.
An important component in determining the fiber efficiency is the probability
that a colliding particle will stick to the fiber surface, called the sticking probability h.
Several empirical fits for sticking probability have been proposed, including those of
Hiller and Löffler (1980) and Ptak and Jaroszcyk (1990), in which h is expressed as
a function of Stokes number, flow Reynolds number, and particle-to-fiber size ratio.
While these correlations fit the data from which they were developed, they miss
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

316 Applications

1.0 1.0

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
10
10
ber
Ad 1
hes 1 um
ion
par ke sn
am 0.1 0.1 Sto
ete
r
Figure 11.10. Plot showing variation of the single-fiber efficiency η as a function of Stokes
number and adhesion parameter, obtained from soft-sphere DEM simulations. [Reprinted
with permission from Yang et al. (2013).]

the important roles played by the particle adhesive force and the particle inertia.
Based on a set of simulations using a soft-sphere DEM approach, Yang et al. (2013)
found that the computational results for sticking probability when ReF < 16 can be
collapsed onto a power-law expression as a function of the adhesion parameter Ad,
given by

0.0558 Ad5/3 when Ad < 5.65
h= . (11.2.2)
1 when Ad > 5.65
A plot showing the collapse of the computational results is given in Figure 11.11. Data
are plotted both with a fixed particle size and different Stokes numbers, as well as with
a fixed Stokes number and different particle sizes. Higher flow Reynolds numbers
change the location of particle collision, resulting in significantly higher sticking
probabilities than indicated by this correlation. It is noted that these computational
results were obtained using a fixed value of the particle restitution coefficient. As
shown in Figures 3.6 and 3.8, the restitution coefficient is a function of particle impact
velocity, and hence of the Stokes number, although in the viscoelastic regime the
dependence is often quite weak.

11.2.2. Enhancement of Filtration Rate by Particle Mixtures

A number of approaches have been proposed to enhance capture rate of particles by
fibrous filters. In a study of capture of fly ash particles resulting from coal combustion,
Xu et al. (2010) examine the role of simultaneous capture of small and large particles
on the capture efficiency. This study observed formation of fly ash particles in both
a coarse mode (with diameter of approximately 1 µm or larger) and a submicron
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.2. Particle Filtration 317

1 St=1.74
dp=1.0 um
dp=2.0 um St=1.09
dp=2.5 um Re>30 St=1.33
St=1.53
Sticking probability
dp=4.0 um too large
dp=5.0 um St=1.89
St=2.00
0.1 St=2.18
St=2.98
St=3.63
St=4.36
St=5.08
St=5.81
0.01

Ad=5.65

1E-3
0.1 1 10
Adhesion parameter
Figure 11.11. Plot showing collapse of computational results for sticking probability as a
function of adhesion parameter. Solid symbols represent data from 2.6 µm diameter parti-
cles with different Stokes numbers, and open symbols represent different size particles with
St = 1.74. [Based on data from Yang et al. (2013).]

fine mode. Because the fine particles in this combustion process have a relatively
higher surface energy than the coarse particles, due to the alkali-rich composition
of the small particles, they act as a type of adhesive between the coarse particles
and the collecting cylinder (i.e., the fiber). A demonstration of this adhesive effect
was given by Li et al. (2011), who reported results of DEM simulations for capture
of particles by a fiber with a mixture of small and large particle sizes. As shown
in Figure 11.12a, the number of large particles deposited onto the fiber at a given
Number of deposited big particles

500
γS / γB = 0.01
γS / γB 0.1
400 TH
1.0
10.0
300 TL

200

100

0
0 100 200 300 400
Time /T
(a) (b)
Figure 11.12. Plots showing effect of a set of small particles with surface energy density γS on deposition
of a set of big particles with surface energy γB onto a fiber. (a) Number of deposited big particles as
a function of γS /γB obtained from DEM simulations with all other parameters held fixed. (b) Plot
showing big and small particles deposited onto fiber front surface. [Reprinted with permission from Li
et al. (2011).]
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

318 Applications

Figure 11.13. Configuration for the capture of par-

ticles by a cylinder positioned at offset distance d
between two charged infinite plates, with a uniform
fluid flow between the plates. [Reprinted with per-
mission from Liu et al. (2010).]
(a) (b)

time increases significantly as the surface energy ratio γS /γB is increased, where γS
and γB are the surface energy densities of the small and big particles, respectively.
The computations are performed by varying γS with a constant value of γB . At high
values of γS /γB , the presence of the small particles also changes the shape of the
region of deposited particles on the fiber, as shown in Figure 11.12b for γS /γB = 10.
Specifically, comparing Figure 11.12b with the results in Figure 11.7 with only a
single particle size, it is evident that the region of deposited particles at the front of
the cylinder covers a larger region on the cylinder surface in Figure 11.12b than in
the case without the small particles present. A similar phenomenon was observed
experimentally by Xu et al. (2010) and Kupka et al. (2009).

11.2.3. Enhancement of Filtration Rate by Electric Fields

Precharging of particles before they come into contact with the fiber may initially
enhance particle deposition on the fiber surface, but after a certain time the repulsive
interaction between deposited and incident particles will limit further deposition. On
the other hand, approaches such as electrically charging the fiber itself or exposing
the entire filter to an external electric field have been shown to be highly effective
methods for enhancing the efficiency of fibrous filters (Wang, 2001). A well-known
experimental study illustrating the effect of a background DC electric field on particle
deposition on a fiber is reported by Oak et al. (1985) for a single fiber placed in the
electric field generated by two parallel charged plates (Figure 11.13). Shifting of the
cylinder by a distance d from the halfway plane between the two plates leads to a
net charge on the cylinder with surface potential E0 d.
A companion DEM simulation study was performed by Liu et al. (2010) under
the same conditions, with the electric field generated by the cylinder represented
using a boundary element method as described in Section 8.3. The results for cylinder
capture efficiency are compared in Figure 11.14 with the experimental data of Oak
et al. (1985) as a function of a dimensionless parameter Wp = d/a ln(4b/π a), which
represents the ratio of the electric field strength induced by the cylinder to that of
the external uniform electric field. The simulation results and experimental data for
capture efficiency are close to one another up to a critical value of Wp of about 12,
beyond which the cylinder voltage is sufficiently high that the air in the experiments
becomes ionized. This effect is not included in the computations. Photographs taken
by an electronic microscope during the experiments exhibited different deposition
patterns at high and low voltage values, which supports the observations of particle
deposition in the simulations shown in Figure 11.15. For the low-voltage case, most
particles are captured on the side of the cylinder with y > 0, which is attributed to
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.2. Particle Filtration 319

Figure 11.14. Variation of the capture efficiency

η with the dimensionless parameter Wp =
d/a ln(4b/π a), showing the DEM simulation pre-
dictions (solid line) and the experimental data
(symbols) of Oak et al. (1985). The point Wp ∼ = 12
where the two results deviate corresponds to the
point where air starts to become ionized by the
electric field, which was not included in the simu-
lations. [Reprinted with permission from Liu et al.
(2010).]

the fact that the electric field is relatively higher on the upper half of the cylinder.
For the high-voltage case, the electric field is sufficiently strong that particles deposit
on both sides of the cylinder. For both cases, particles agglomerate in the form of
particle chains pointing radially outward into the fluid, due to the presence of strong

Figure 11.15. Snapshots from a DEM simulation of particle deposition on a cylinder exposed
to a low voltage electrostatic field (top) and a high voltage electrostatic field (bottom).
[Reprinted with permission from Liu et al. (2010).]
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

320 Applications

DEP force on the (uncharged) particles. These chains can grow to a length scale of
the order of the cylinder radius.

11.3. Rotating Drum Mixing Processes

Mixing and separation processes play a central role in many applications involving
particulate flow. Mixing processes are necessary both to combine different ingredi-
ents into a homogeneous mixture and to maintain a well-mixed suspension in the
presence of gravity or mean particle drift. A common device used to mix particles in
granular flows is the rotating drum (or rotating kiln), which is used in applications
such as mixing pharmaceutical components, washing and drying operations, mixing
components in food production, cement processing, and particle coating operations.

11.3.1. Flow Regimes

Let us consider a horizontal drum of radius R and length L partially filled with
particles of diameter d, in which the drum axis is orthogonal to the direction of
gravitational acceleration. The drum rotates at a rate D , where for the present
we take D to be a constant. The primary dimensionless parameter characterizing
the flow is the Froude number Fr ≡ D R2 /g. A second important dimensionless
parameter is the filling degree J, which is equal to the percentage of the drum volume
filled with particles at the maximum packing concentration φm , or J = Vdrum φm /NVp
where N is the total number of particles in the drum with particle volume Vp . A
third parameter that influences flow regimes is the particle-wall dynamic friction
coefficient μW , which is used in place of μ f in the sliding resistance equation (3.5.7).
Several different papers have reported different regimes of particle motion
within the rotating drum as a function of these parameters, based both on exper-
iments (Henein et al., 1983; Mellmann, 2001) and on computations (Yang et al.,
2008). Although there are a number of differences between the various regime clas-
sifications, the classifications suggested by Mellmann (2001) serve well to illustrate
the different particle flow transitions as the rotation rate of the system is changed.
Schematic diagrams of these different regimes and a listing of the parameter ranges
for which they are valid are given in Table 11.1 from Mellmann, and plots illustrating
several of these regimes using the DEM simulations of Yang et al. (2008) is given
in Figure 11.16. An analysis giving criterions for transition between these various
regimes is given by Mellmann (2001), which is not repeated here.
Mellmann categorized the particle behavior into three types of motion (or
regimes) – slipping, cascading, and cataracting – each of which has several subtypes.
The slipping regime occurs for small Froude numbers (Fr < 10−4 ), and is character-
ized by motion of the particles as a block, with no relative particle motion (and hence
no mixing). For cases where the particle-wall dynamic friction coefficient μW is less
than a critical value μW,C , a sliding motion occurs where the particles continuously
slide over the rotating drum at a small angle of repose β and with no particle motion.
If μW is greater than μW,C but less than the coefficient of static friction μS , a periodic
surging motion is observed in which the particle mass rotates with the drum up to a
critical angle βmax , beyond which the particles slide back down as a single mass to a
lower angle or repose βmin , and the cycle repeats. Even though a periodic particle
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.3. Rotating Drum Mixing Processes 321

Table 11.1. Listing of different flow regimes for a rotating drum, along with values of key parameters.
[Reprinted with permission from Mellmann (2001)]

Slipping regime Cascading regime Cataracting regime

Basic form
subtype Sliding Surging Slumping Rolling Cascading Cataracting Centrifuging

Schematic

Physical Slipping Mixing Crushing Centrifuging

process
Froude Fr <10−4 10−5 < Fr<10−3 10−4 <Fr<10−2 10−3 <Fr<0.1 0.1<Fr<1 Fr > 1
number,
Fr
Filling J <0.1 J >0.1 J <0.1 J >0.1 J >0.2
degree, J
Wall μW <μW,C μW >μW,C μW > μW,C μW >μW,C
friction
coeff, μW
Application No use Rotary kilns and reactors; rotary dryers Ball mills No use
and coolers; mixing drum

motion is observed in this latter motion relative to a fixed frame, the particles do
not move relative to each other. Because the primary purpose of a rotating drum is
to mix the particles either with each other or with a surrounding fluid, the slipping
regime is not useful for practical applications.
The cascading regime occurs over a broad range of Froude numbers in the
interval 10−5 < Fr < 0.1, for cases with sufficient wall friction. This regime is char-
acterized by either continuous or periodic tumbling of particles along the top of the
bed from the high side to the low side of the particle bed. Mellmann identified three
subtypes of motion within this regime, referred to as slumping, rolling, and cascading.
In the slumping motion, the particle tumbling occurs periodically as a sequence of
“avalanches.” The particle bed is raised with the rotating drum, and then a top layer
of particles avalanche downward from the top of the raised bed to the low portion of

Figure 11.16. Computed particle flow

patterns at different rotation speeds
showing different flow regimes: (a)
slumping, (b) slumping-rolling transi-
tion, (c) rolling, (d) cascading, (e)
cataracting, and (f) centrifuging. Shad-
ing represents particle velocity. The slid-
ing and surging regimes were not exam-
ined in the computations. [Reprinted
with permission from Yang et al.
(2008).]
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

322 Applications

the bed. This avalanche causes the angle of repose of the bed to be lower, allowing
the cycle to repeat. In the rolling motion, the particles roll continuously from the
upper part to the lower part of the particle bed, with the bed surface approximately
flat. The particle flow exhibits two distinct regions – a “fluid-like” cascading region
of rapidly moving particles along the top of the bed and a “solid-like” plug-flow
region of slowly transported particles within the rest of the bed, which are slowly
raised upward by the drum rotation. At higher rotation rates a cascading motion is
observed, which is similar to the rolling motion except that the top of the particle
bed is no longer flat, but instead arches upward within the raised part of the bed. The
rolling and cascading motions generate strong mixing, and as a consequence most
practical applications of rotating drums operate in the cascading regime.
At very high rotation rates, corresponding to Fr > 0.1, the particles exhibit a
cataracting regime, in which the friction from the rotating drum is sufficiently large
to carry the particles nearly up to the top of the drum. Depending on the rotation
rate, they will either fall back downward in a cataracting motion onto the lower part
of the bed, or they will remain on the drum and spin back down the other side in
a centrifuging motion. Relatively little mixing occurs in a very rapid centrifuging
motion, since the particles circulate with the rotating drum in nearly a solid-body
motion.

11.3.2. Mixing and Segregation

Within the cascading regime, the DEM simulations of Yang et al. (2008) show that
the particle velocity magnitude v normalized by the mean wall speed, given by
v ∗ = v/D R, can be collapsed for a different rotation rates onto a single log-normal
distribution of the form

P(v ∗ ) = a exp{−b[ln(v ∗ /v0∗ )]2 }, (11.3.1)

where a = 2.622, b = 1.510, and v0∗ = 0.227. This distribution is compared with vari-
ous DEM computational data in Figure 11.17.
A study of longitudinal and transverse mixing within a rotating drum is reported
by Finnie et al. (2005) based on DEM simulations, building on previous work by Ding
et al. (2001) and Schutyser et al. (2001, 2002). Because there is no mean motion in the
longitudinal direction for a horizontal drum, the diffusion is well approximated by a
Fickian diffusion approximation. Consequently, mixing of nearly identical particles
in the longitudinal direction is governed by a one-dimensional diffusion equation of
the form
 
∂φ ∂ ∂φ
= D , (11.3.2)
∂t ∂x ∂x

for the concentration φ(x, t ) of a certain particle type and D is a diffusion coefficient
with units m2 /s. A dimensionless diffusion coefficient D can be defined as

D = 2π D/D R2 . (11.3.3)

A plot is given in Figure 11.18a showing estimated values of D from DEM sim-
ulations as a function of Froude number for different filling degree values. The
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.3. Rotating Drum Mixing Processes 323

3.0
40 rpm
2.5 80 rpm
150 rpm
200 rpm
2.0

1.5
p(v*)

1.0

0.5

0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6
Normalized particle velocity, v*
Figure 11.17. Plot showing collapse of the probability distribution for the normalized parti-
cle velocity v ∗ onto a log-normal curve within the cascading regime of rotating drum flow.
[Reprinted with permission from Yang et al. (2008).]

simulations indicate that D is relatively insensitive to Froude number and filling

angle, having a value between about 1.4–1.6 over a broad range of Fr and J values.
A comparison of DEM results and predictions of (11.3.3) is given in Figure 11.18b,
which shows excellent agreement for average particle concentration profiles across
the longitudinal direction along the drum as a function of time.
While a diffusion-like equation can also be used to quantify transverse mixing,
the diffusion coefficient is spatially-variable and dependent on the particle velocities
(Heydenrych, 2001). An alternative “entropic” approach was proposed by Schutyser

2.0 1.0
x10–3

1.8 0.8

1.6 0.6

D′ φ
1.4 0.4
n = 10 revs
J = 20% 20 revs
1.2 0.2
J = 30% 30 revs
40 revs
J = 40%
1.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 –0.5 –0.4 –0.3 –0.2 –0.1 0.0 0.1 0.2 0.3 0.4 0.5
Fr x/L
(a) (b)
Figure 11.18. Results of DEM simulation for longitudinal mixing in a rotating drum: (a)
dimensionless particle diffusion coefficient as a function of Froude number for different
filling degree values and (b) concentration profile along longitudinal direction as predicted
by DEM simulations (symbols) after n drum revolutions and as given by Equation (11.3.2)
(solid lines). [Reprinted with permission from Finnie et al. (2005).]
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

324 Applications

1.0
0.6

0.8
0.5

0.6 0.4
M (n)

S 0.3
0.4
0.2
0.2 J = 20%
0.1 J = 30%
J = 40%
0.0 0.0
0 5 10 15 20 25 30 35 40 0.0 0.5 1.0 1.5 2.0 2.5
n Fr
(a) (b)
Figure 11.19. Plots illustrating DEM results for transverse mixing in a rotating drum: (a)
variation of mixing index M(n) with number of drum rotations n comparing DEM data
(symbols) with exponential fit (11.3.6) (line); (b) dependence of computed transverse mixing
rate S on filling degree and Froude number. [Reprinted with permission from Finnie et al.
(2005).]

et al. (2001, 2002). This approach places a Cartesian grid over the domain of interest,
such that for each grid cell (i, j), a quasi-entropy quantity Si j is defined by

Si j = −kVi j [α1,i j ln α1,i j + α2,i j ln α2,i j ]. (11.3.4)

Here, Vi j is the volume occupied by particles in grid cell (i, j), α1,i j is the fraction of
Vi j occupied by particles of type 1, α2,i j (= 1 − α1,i j ) is the fraction of Vi j occupied by
particles of type 2, and k is a constant. Using this entropy measure, a mixing index
M(n) is defined as a function of the number n of drum rotations as
 Si j (n)
M(n) = , (11.3.5)
S∞
i, j

where S∞ = −kV [α1,0 ln α1,0 + α2,0 ln α2,0 ] is the entropy measure for the perfectly
mixed state, V is the total volume occupied by particles, and αi,0 is the fraction of V
occupied by particles of type i. The constant k cancels between the numerator and
denominator in (11.3.5), so the mixing index is independent of the value chosen for
k. The mixing index definition satisfies the limits M = 0 for a completely unmixed
state and M = 1 for a perfectly mixed state.
A plot showing the variation of M(n) with number of drum rotations using DEM
data from Finnie et al. (2005) is given in Figure 11.19a, starting from a completely
unmixed state. In all cases examined, the mixing index varies exponentially with
number of drum rotations as

M(n) ∼
= 1 − exp(−Sn), (11.3.6)

where the coefficient S is a measure of the rate of mixing (defined in terms of number
of drum rotations, rather than time). A plot of S as a function of Froude number for
cases with three different values of the filling degree is given in Figure 11.19b. These
results indicate that the particles within the rotating drum mix with fewer rotations
for smaller values of the Froude number (i.e., lower rotation rate or smaller drum
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.3. Rotating Drum Mixing Processes 325

1/2 full < 1/2 full > 1/2 full

Poincaré Sections P P
P P P

a b c d e
E
E E E E
D–systems

f g h i j
E E E E E
S–systems

k l m n o
Figure 11.20. Results for two-dimensional rotating tumblers of different shapes, comparing
Poincaré sections (top row) with experimental results (middle and bottom rows) for systems
containing particles with different size (S-systems) or density (D-systems). [Copyright (1999)
National Academy of Sciences. Reprinted with permission from Hill et al. (1999).]

radius) and for smaller filling degree. One exception occurs for the highest filling
degree examined, for which S increases with Froude number for the highest value
of Fr examined.
The analysis presented here is for mixing of two sets of similar particles, such as
might occur with two sets of otherwise identical particles that are painted different
colors. When there are differences in physical characteristics of the particles, such
as size or density, analysis of mixing in the rotating drum becomes much more com-
plicated because both mixing and segregation can occur at the same time. A number
of studies of this segregation phenomenon have been reported in the literature,
principally from Ottino, Lueptow, and their students at Northwestern University.
An early study of rotating drum segregation phenomena was performed by Hill
et al. (1999), who examined radial band formation in two-dimensional tumblers of
different shapes, including circles, ellipses, and squares. The systems examined all
consist of particles of two different properties, either different sizes (S-systems) or
different densities (D-systems), and were operated in the rolling regime. Results
comparing Poincaré sections of the granular flow field with experimental results
for these systems are shown in Figure 11.20. The mixtures were formed of binary
D-systems (with 2 mm glass and steel spheres) and of ternary S-systems (0.8 mm,
1.2 mm, and 2.0 mm glass spheres), shown as cases (f) to (j), and of ternary S-systems
(0.8 mm, 1.2 mm, and 2.0 mm glass spheres), shown as cases (k) to (o). In particular,
Figures 11.20f and 11.20k illustrate the classic radially segregated structure in circular
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

326 Applications

γ = 0.0 N/m γ = 0.033 N/m γ = 0.073 N/m

1 rev

5 revs

Figure 11.21 The effect of liquid-bridge adhesion with different surface tension values on
the granular mixing patterns, from DEM simulation results with J = 40% and D = 15 rpm.
[Reprinted with permission from Liu et al. (2013).]

mixers. In these two cases, the dynamics in the flowing surface layer cause the smaller
(or more dense) particles to move down in the layer, which causes them have a low
probability of reaching the end of the layer, leading to a segregated core. A circular
core is time invariant, because it coincides with the streamlines. The segregation
structures obtained in noncircular mixers are much complicated. For instance, this is
quite evident in the case of a half-filled square mixer, shown in Figures 11.20h and m.
A region with a high concentration of small (dense) particles is located away from
the center of rotation and is nearly separated into two regions spanning the corners
of the square and the center of rotation. The corresponding segregation structure
for the ellipse also shows that the smaller (denser) particles are pulled away from
the center of rotation. The final case shown (Case o) never settled down to an equi-
librium flow but instead persisted as a changing pattern of streaks. This study was
extended to three-dimensional systems by Gilchrist and Ottino (2003) and to systems
with combined size and density differences by Jain et al. (2005). Fiedor and Ottino
(2005) examined the effect of unsteady rotation of a circular drum, in which the
rotation rate has a nonzero mean value plus a time-periodic part. An extensive study
of the relationship between mixing and segregation processes in rotating drums and
aspects of dynamical systems theory is given by Meier et al. (2007).

11.3.3. Cohesive Mixing and Segregation

The study of the mixing and segregation of cohesive particles is of considerable
interest for numerous industrial and agricultural processes. By incorporating liquid
bridging force into the Hertz contact force model, Liu et al. (2013) conducted DEM
simulations of the mixing of cohesive particles in rotating drums. Sample results are
shown in Figure 11.21 for a case with J = 40%, D = 15 rpm, and d = 1.3 mm. The
presence of capillary force caused by a small amount of moisture mixed with the
particles is observed to significantly inhibit mixing between two sets of otherwise
identical colored particles. Chaudhuri et al. (2006) had earlier used both DEM
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.3. Rotating Drum Mixing Processes 327

1.0 0.45

0.8 0.40

0.6 0.35
M (n)

S
0.4 0.30
γ = 0.000 N/m
0.2 γ = 0.033 N/m 0.25
γ = 0.073 N/m
0 0.20
0 2 4 6 8 10 0.0 0.2 0.4 0.6 0.8 1.0
n σ (N/m)
(a) (b)
Figure 11.22. Plots illustrating DEM results for cohesion on mixing in a rotating drum: (a)
effect of cohesion on mixing index M(n); (b) dependence of mixing rate S on surface tension
of liquid bridges. [Reprinted with permission from Liu et al. (2013).]

simulations and experiments to demonstrate that small capillary forces enhance

mixing whereas high values of the capillary force inhibit mixing. Liu et al. (2013)
interpreted the DEM results in terms of the mixing index M(n) and mixing rate S
defined in (11.3.4) and (11.3.5), showing the variation of M(n) and S as a function
of the liquid-gas surface tension (Figure 11.22). As the value of surface tension
increases from 0 to 0.073 N/m (the value for an air-water interface), the mixing index
continues to be well fit by the exponential function shown in (11.3.6), but the mixing
rate S decreases from 0.41 to 0.24 as surface tension increases.
For systems of particles of different sizes (S-systems) or different densities (D-
systems), the processes of segregation and mixing occur simultaneously. Specifically,
capillary forces between particles of a similar type (cohesion) tend to enhance seg-
regation of particles, whereas capillary forces between particles of different types
(adhesion) tend to inhibit (or mitigate) segregation. Li and McCarthy (2003) devel-
oped a phase diagram of the particle-scale mixing/segregation in the asymptotic
(long-time) state by determining which conditions favor adhesive versus cohesive
behavior for binary particle mixtures. They restricted their study to rotation rates
that are sufficiently small that the shearing force is negligible compared to the capil-
lary force Fcap and the gravitational force Fg acting on the particles. The ratio of the
magnitude of capillary and gravitational forces yields the granular Bond number,
Bog , which using (4.5.14) can be written as

Fcap 3σ R cos θ
Bog = = Gf, (11.3.7)
Fg gr3p,i ρ p

where R is the effective radius, rp,i is the radius of particles of type i (where i = 1
or 2), and the coefficient G f , defined in (4.5.15), approaches unity as two particles
approach each other. The values of Bog for each potential particle pair in a binary
mixture are denoted by Bog,1−1 , Bog,1−2 , and Bog,2−2 , where particle type 2 is defined
as the larger particles. The phase boundaries are identified by finding where in the
parameter space of the particle size ratio r p,1 /r p,2 , density ratio ρ p,1 /ρ p,2 , and wetting
angle ratio cos θ1 / cos θ2 different regimes of mixing/segregation are observed. When
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

328 Applications

E-phase E-phase
2 2

cos θ1 / cos θ2

cos θ1 / cos θ2
1 1
M-phase M-phase

E-phase
E-phase
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
rp,1 / rp,2 rp,1 / rp,2
(a) (b)
Figure 11.23 Phase diagram of cohesive mixing/segregation in a binary system with density
ratio (a) ρ p,1 /ρ p,2 = 0.56 and (b) the inverse case with ρ p,1 /ρ p,2 = 1.78. Based on data from
Li and McCarthy (2003).

the granular material is adhesion-dominated, such that Bog,1−2 > {Bog,1−1 , Bog,2−2 },
it is expected that a small amount of segregation will occur because capillary
bonds joining particles of different types dominate other forces. When the mate-
rial is cohesion-dominated, such that Bog,1−2 < Bog,1−1 or Bog,1−2 < Bog,2−2 , capil-
lary forces between particles within a given type will dominate, acting to enhance
segregation of particle types compared to the dry granular case.
Figure 11.23 shows regime diagrams indicating regions of mitigated (M-phase)
and enhanced (E-phase) segregation of particles in a rotating drum for a case with
density ratio ρ p,1 /ρ p,2 = 0.56 (Figure 11.23a) and the inverse case with density ratio
ρ p,1 /ρ p,2 = 1.78 (Figure 11.23b), based on the data of Li and McCarthy (2003).
The white region corresponds to the E-phase and the gray region to the M-phase.
The size of the M-phase region grows when the smaller particle is less dense than
the larger particle and shrinks when it is more dense. These regime diagrams indicate
only whether addition of moisture will enhance or mitigate the degree of segregation
relative to the dry state; they do not indicate anything about the absolute amounts
of mixing/segregation.

11.4. Dust Removal Processes

Removal of dust from surfaces is an essential part of numerous cleaning processes.
It is of particular importance in manufacture of electronic components and in explo-
ration of dusty planetary bodies, such as the Moon and Mars. There are many
approaches that can be used for particle removal, several of which require contact
with the surface, such as brushing dust off the surface or removal from capillary
force by a liquid film. In this section, we examine two noncontact methods for par-
ticle removal from a surface, which illustrate several of the interaction phenomena
discussed in this book.

11.4.1. Hydrodynamic Dust Mitigation

A traditional method for particle removal from a surface is to aim a jet flow at the sur-
face, thereby blowing the particles along the surface and entraining the particles into
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.4. Dust Removal Processes 329

suction
outlet jet inlet
confinement surface
Figure 11.24. Plot of a bounded vortex flow, show-
ing the annular jet inlet at the top and the wall-
normal vortex at the center. The shaded sur-
face is an iso-vorticity magnitude surface and the
lines are streamlines originating from the jet inlet.
[Reprinted with permission from Maynard and
Marshall (2011).]

impingement surface

the jet flow. Fluid dynamics issues associated with jet impingement onto a flat plate
surface have been examined by a number of investigators, including two-dimensional
slot jets (Chiriac and Ortega, 2002; Sakakibara et al., 2001), axisymmetric jets
(Chatterjee, 2008), and annular slot jets (Chattopadhyay, 2004). Particle transport
due to impingement of a circular jet on a surface was examined experimentally by
Masuda et al. (1994), Anderson and Longmire (1995), and Zhang et al. (2002) for
constant jet flow rates and by Otani et al. (1995) and Ziskind et al. (2002) for pulsed
jets. Both pulsed and steady impinging jet flows generate ring-like coherent vortices
that surround the jet and increase shear stress magnitude on the impingement plate.
Impinging jet methods are highly effective at removing particles, but they tend to
scatter the particles on the surface.
A combination of a swirling impinging jet flow within an annular slot and a cen-
tral suction flow was used by Maynard and Marshall (2011) and Huang and Marshall
(2011) as an effective approach for hydrodynamic particle removal from a surface.
This combination of a swirling jet and a suction flow forms a wall-normal vortex
structure immediately below the suction intake that is bounded by the axisymmet-
ric slot jet, forming what is called a bounded vortex flow, as illustrated in the flow
field computation shown in Figure 11.24. The bounded vortex flow is generated by
a nozzle with both inlet and outlet ports, which is held at a distance H above the
impingement surface and oriented normal to this surface. The injection jets on the
nozzle are located in a circular pattern at a radial distance R from the nozzle cen-
ter, with inlet flow oriented both downward toward the impingement surface and
tangentially around the impingement point. The outlet suction port is located at the
center of the nozzle, which is connected to a vacuum line. The bounded vortex flow
is characterized by two key dimensionless parameters – the ratio h/R of separation
distance to nozzle injection radius and the ratio η = Qout /Qin of the outlet suction
flow rate to the total of the inlet jet flow rates. Depending on the value of η, the jet
may be entrained into the suction port with little influence on the near-surface flow
or it may spread radially outward scattering particles along the impingement surface.
However, the flow is most effective when the flow rate ratio is set approximately
equal to unity, for which the shear stress streamlines on the impingement surface
swirl inward just underneath the nozzle, carrying the particles in toward the suction
outlet. The wall-normal vortex acts to direct fluid streamlines originating from the
jet inlet downward into the boundary layer along the impingement plate and then
upward along the vortex core into the suction outlet. The wall-normal vortex flow
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

330 Applications

time
2

y 0 (a)

–1

y 0 (b)

–1

y 0 (c)

–1

–2 1 0 1 1 0 1 1 0 1
x x x
Figure 11.25. Time series (from left to right) showing DEM results for particle locations on
the impingement surface for a bounded vortex with flow rate ratio (a) η = Qout /Qin = 0.5; (b)
η = 1; and (c) η = 2. [Reprinted with permission from Maynard and Marshall (2011).]

generates high shear stress levels along the wall (Phillips, 1985; Phillips and Khoo,
1987; Parras and Fernandez-Feria, 2007), which aids in removal of the particles from
the surface.
A series of numerical simulations of particle removal underneath a boundary
vortex flow are reported by Maynard and Marshall (2011), who use a soft-sphere
DEM approach with near-wall corrections to the fluid-induced forces on the particles
similar to those discussed in Section 5.6. The particles are initially spread evenly on
the surface, and they are followed in time by the simulation until an equilibrium
state is reached. Results for particle positions are shown in Figure 11.25 at three
different values of the flow rate ratio η, with three times shown for each value of
η. For jet-dominated flows with η = 0.5, a particle front forms that gradually moves
outward and an inner island of particles is entrained into the suction outlet. For the
suction-dominated flows with η = 2, particles from across the impingement plane
are drawn inward toward the suction outlet, with the results that no cleaned regions
on the impingement surface are formed. In the case of balanced inflow and outflow
(η = 1), regions of inward- and outward-flowing streamlines near the impingement
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.4. Dust Removal Processes 331

surface are separated by the circular separatrix, so that a circle with constant radius
is formed within which particles are removed and outside which the particles remain
on the surface. Extensions of the bounded vortex method for particle removal have
been reported by Vachon and Hitt (2010, 2011) for turbulent vortex flows and by
Furhmann et al. (2013), who showed that addition of acoustic disturbances enhances
efficiency of the process by helping to break the particle bonds with the impingement
surface.

11.4.2. Electric Curtain Mitigation for Charged Particles

When the particles on the surface are charged, many standard methods for particle
removal do not work well. An exception is the electric curtain device discussed in
Section 1.3.2 and shown in Figure 1.10. The electric curtain is formed of a series
of parallel electrodes embedded into a dielectric material, which are coated by a
thin dielectric film. The electrodes have oscillating potential, creating an n-phase
wave pattern in the electric field that propagates across the dielectric surface. Early
work on electric curtains was reported by Masuda et al. (1972) for confinement
of an aerosol cloud in a ring-type electric curtain device. Electric curtain technol-
ogy was later used as a charged toner conveyer device used in xerographic copy
machines (Melcher et al., 1989; Schmidlin, 1991). The electric curtain technology is
of particular interest for designs for handling dust adhesion during extraterrestrial
exploration of dusty planets, as proposed by Calle et al. (2008, 2009) and Atten et al.
(2009), as dust particles in places such as the Moon and Mars carry significant elec-
trical charge. Other proposed applications of electric curtains include liquid droplet
transport (Kawamoto and Hayashi, 2006), particle size classification (Kawamoto and
Hasegawa, 2004), and particle separation and sorting (Weiss and Thibodeaux, 1984;
Masuda et al., 1987; Machowski and Balachandran, 1997; Kawamoto, 2008).
Schmidlin (1995) described three different modes of motion for individual par-
ticles on the electric curtain, referred to as surfing mode, hopping mode, and curtain
mode. These different modes are illustrated in Figure 11.26, along with plots of time
variation of computed particle horizontal velocity and particle height for each mode.
In the surfing mode, particles roll or slide on the dielectric surface at an average veloc-
ity equal to the wave speed. Consequently, a given particle always experiences the
same phase of the electric curtain traveling wave. In the hopping mode, particles are
transported by successive hops, and between each hop particles remain fixed to the
dielectric surface. Particle hop distance can vary from hop to hop, and in some cases
particles perform several small hops, or even a backwards hop, before resuming
large forward hops. Average forward speed of a particle in hopping mode motion
can vary from the wave speed to much less than the wave speed. In the curtain mode,
particles are suspended above the dielectric surface and travel in a spiraling path
with a mean drift velocity that is much smaller than the electric curtain wave speed.
The motion of an individual particle on an electric curtain is more complex and
irregular than represented by these three modes of motion due to the effects of
collisions and electrostatic interactions between particles, as well as the effects of
adhesion between the particles and the dielectric surface. The amount of data that
can be obtained from experimental studies of multiparticle systems is limited because
it is difficult to determine the charge of each particle. Masuda and Kamimura (1975)
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

332 Applications

1.2 1.5 1.6 0.1 0.28

4
1.2 0.27
1.1 0.05
1 2
0.26
u' y' u' 0 0.8 y' u' 0 y'
1
0.5 0.25
-2 0.4
0.9 -0.05
0.24
0 -4 0
0.8 -0.1 0.23
0 0.5T T 1.5T 2T 0 T 2T 3T 4T 0 8T 16T 24T 32T 40T

(a) (b) (c)

Figure 11.26. Schematic diagram of modes of particle motion on an electric curtain, showing (a) particles
rolling or sliding along the curtain in surfing mode; (b) particles intermittently jumping along the curtain
in hopping mode; and (c) particles spiraling above the electric curtain in curtain mode. Plots show
computed time variation of tangential velocity component (left) and height (right) of particles in each
mode. [Reprinted with permission from Liu and Marshall (2010a).]

examined different approximate methods for modeling the electric field induced
by a traveling electric wave above the electric curtain, which are often used in
computational studies. Computations of nonadhesive particle motion on an electric
curtain were reported by Gartstein and Shaw (1999) and Kawamoto et al. (2006)
using the hard-sphere discrete-element method.
Computations using a soft-sphere discrete-element method with adhesive par-
ticles were performed by Liu and Marshall (2010a,b) for both traveling waves and
standing waves on the electric curtain. The computations compared particle trans-
port for traveling waves both with and without particle-particle collisions. Results
are shown in Figure 11.27 for bins representing the fraction of particles that propa-
gate forward, propagate backward, and have essentially no average forward motion.
At low driving frequencies the particles primarily propagate in the forward direc-
tion, whereas at high frequencies particles primarily propagate in the backward
direction or have essentially no forward velocity. Inclusion of particle-particle colli-
sions and electrostatic interactions increases the prevalence of the majority motion
type, so that when particle-particle interactions are included at low frequencies a
higher fraction of particles moves forward and at high frequencies a higher frac-
tion of particles moves backward. The fraction of nontransported particles is always
decreased by inclusion of particle-particle interactions except within a band of inter-
mediate frequencies, where collision of essentially equal numbers of forward- and
backward-moving particles increases the fraction of particles with essential zero
average velocity.

11.5. Final Comments

The discrete element method has been used for a wide variety of problems involv-
ing both adhesive and nonadhesive particles. The applications described in this
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

11.5. Final Comments 333

20 40 60 80
1.2

Fraction of particles 1

0.8

0.6

0.4

0.2

0
00 0.05
20 0.1
40 0.15
60 0.280 0.2
f (Hz)
f'
Figure 11.27. Variation of fractions of forward-transported (square symbols), backward-
transported (triangle symbols), and nontransported (delta symbols) particles versus frequency
for particle transport on an electric curtain, both with particle-particle interaction (solid lines)
and without particle-particle interaction (dashed lines). [Reprinted with permission from Liu
and Marshall (2010a).]

chapter provide just a few examples of problems in which DEM has been employed
to provide insight into mesoscale particle transport. DEM allows investigators to per-
form studies with sufficient numbers of particles to study particle collective behavior,
influenced by collision and adhesion of particles with each other, while still modeling
the particle interactions at a fundamental level. The various models used in DEM for
particle interaction and for forces induced by fluid, electric, and acoustic fields have
developed over the past several decades to greatly extend the range of problems for
which this method can be reliably employed.
At the same time, there are certain aspects of DEM that still require signif-
icant development. Primary among these is the problem of accurately modeling
hydrodynamic interaction between particles without use of excessive computational
resources. Computational methods used for microscale simulation methods, which
compute the flow field around each particle, require too much time to allow for
simulations with large numbers of particles while still accurately resolving the par-
ticle interaction. The Stokesian dynamics method is one approach for handling this
problem, but this method only works for problems with small values of the flow
Reynolds number.
For turbulent flows, stochastic Lagrangian methods can be used to accurately
model time variation of particle fluctuations in the turbulence. These methods also
yield good predictions for large-scale particle dispersion in the turbulence. Fur-
ther work must be performed to improve modeling of the turbulent flows to yield
improved predictions of particle collision and adhesion rate. Stochastic Lagrangian
models are not accurate for this purpose because they lack spatial correlation in the
random forcing term.
Most DEM simulations have been performed for spherical particles. However,
rolling motion of the particles is highly sensitive to variation in particle shape, and
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

334 Applications

such processes might not be accurately modeled when the particles are represented
by spheres. Models have been presented for ellipsoidal particles that can be used to
partially surmount issues with modeling of rolling motions. Unfortunately, methods
for computation of the collision of nonspherical particles are much more time con-
suming than those for spherical particles. Faster methods need to be developed for
handling arbitrarily shaped particles, as these are encountered in many important
applications.
A challenge with the discrete element method occurs when transporting particles
with very small Stokes numbers. In this case the aerodynamic time step limitation
(11.1.6) requires the particle time step to be extremely small. In the absence of
particle collisions, the fast Euler approximation (11.1.10) can be used to overcome
this problem, but this approximation cannot be used for colliding particles. This time-
stepping limitation particularly becomes a problem for transport of nanoparticles in
liquids. Time scale issues also arise for problems for particle collective phenomena
that take a long time to develop, such as thrombosis development in blood flows.
Typical time scales covered by a DEM simulation with adhesive particles are on the
order of several seconds. Bridging between DEM simulations with results of both
larger-scale models (such as the population-balance model and the discrete parcel
model) and smaller-scale models (including different types of microscale simulation
models) remains of central importance for many problems.

REFERENCES

Anderson SL, Longmire EK. Particle motion in the stagnation zone of an impinging air jet.
Journal of Fluid Mechanics 299, 333–366 (1995).
Asmolov ES. The inertial lift on a spherical particle in a plane Poiseuille flow at large channel
Reynolds number. Journal of Fluid Mechanics 381, 63–87 (1999).
Atten P, Pang HL, Reboud JL. Study of dust removal by standing wave electric curtain for
application to solar cells on Mars. IEEE Transactions on Industry Applications 45, 75–86
(2009).
Bretherton FP. The motion of rigid particles in a shear flow at low Reynolds number. Journal
of Fluid Mechanics 14, 284–304 (1962).
Calle CI, Mazumder MK, Immer CD, Buhler CR, Clements JS, Lundeen P, Chen A, Man-
tovani JG. Controlled particle removal from surfaces by electrodynamic methods for ter-
restrial, lunar and Martian environmental conditions. Journal of Physics: Conference Series
142, 012073 (2008).
Calle CI, Buhler CR, McFall JL, Snyder SJ. Particle removal by electrostatic and dielec-
trophoretic forces for dust control during lunar exploration missions. Journal of Electro-
statics 67, 89–92 (2009).
Chatterjee AJ. Multiple vortex formation in steady laminar axisymmetric impinging flow.
Computer and Fluids 37, 1061–1076 (2008).
Chattopadhyay H. Numerical investigations of heat transfer from impinging annular jet.
International Journal of Heat and Mass Transfer 47, 3197–3201 (2004).
Chaudhuri B, Mehrotra A, Muzzio FJ, Tomassone MS. Cohesive effects in powder mixing in
a tumbling blender. Powder Technology 165, 105–114 (2006).
Chiriac VA, Ortega A. A numerical study of the unsteady flow and heat transfer in a tran-
sitional confined slot jet impinging on an isothermal surface. International Journal of Heat
and Mass Transfer 45, 1237–1248 (2002).
Ding YL, Forster RN, Seville JPK, Parker DJ. Scaling relationships for rotating drums.
Chemical Engineering Science 56, 3737–3750 (2001).
Fiedor SJ, Ottino JM. Mixing and segregation of granular matter: Multi-lobe formation in
time-periodic flows. Journal of Fluid Mechanics 533, 223–236 (533).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

References 335

Finnie GJ, Kruyt NP, Ye M, Zeilstra C, Kuipers JAM. Longitudinal and transverse mixing
in rotary kilns: a discrete element method approach. Chemical Engineering Science 60,
4083–4091 (2005).
Frank M, Anderson D, Weeks ER, Morris JF. Particle migration in pressure-driven flow of a
Brownian suspension. Journal of Fluid Mechanics 493, 363–378 (2003).
Furhmann A, Marshall JS, Wu J-R. Effect of acoustic levitation force on aerodynamic particle
removal from a surface. Applied Acoustics 74, 535–543 (2013).
Gartstein YN, Shaw JG. Many-particle effects in travelling electrostatic wave transport, Jour-
nal of Physics D: Applied Physics 32, 2176–2180 (1999).
Gilchrist JF, Ottino JM. Competition between chaos and order: Mixing and segregation in a
spherical tumbler. Physical Review E 68, 061303 (2003).
Hampton RE, Mammoli AA, Graham AL, Tetlow N, Altobelli SA. Migration of particles
undergoing pressure-driven flow in a circular conduit. Journal of Rheology 41(3), 621–640
(1997).
Han M, Kim C, Kim M, Lee S. Particle migration in tube flow of suspensions. Journal of
Rheology 43, 1157–1174 (1999).
Henein H, Brimacombe JK, Watkinson AP. Experimental study of transverse bed motion in
rotary kilns. Metallurgical Transactions B 14B, 191–205 (1983).
Heydenrych MD. Modelling of rotary kilns. Ph.D. Dissertation, Department of Chemical
Engineering, University of Twente, The Netherlands (2001).
Hewitt GF, Marshall JS. Particle focusing in suspension flow through a corrugated tube.
Journal of Fluid Mechanics 660, 258–281 (2010).
Hill KM, Khakhar DV, Gilchrist JF, McCarthy JJ, Ottino JM. Segregation-driven organization
in chaotic granular flows. Proceedings of the National Academy of Sciences 96(21), 11701–
11706 (1999).
Hiller M, Löffler M. Influence of particle impact and adhesion on the collection efficiency of
fiber filters. German Chemical Engineering 3, 327–332 (1980).
Ho BP, Leal LG. Inertial migration of rigid spheres in two-dimensional unidirectional flows.
Journal of Fluid Mechanics 65, 365–400 (1974).
Hogg AJ. The inertial migration of non-neutrally buoyant spherical particles in two-
dimensional shear flows. Journal of Fluid Mechanics 272, 285–318 (1994).
Huang B, Yao Q, Li SQ, Zhao HL, Song Q, You CF. Experimental investigation on the
particle capture by a single fiber using microscopic image technique. Powder Technology
163, 125–133 (2006).
Huang Y, Marshall JS. Experiments on bounded vortex flows and related particle transport.
Journal of Fluids Engineering 133(7), 071204 (2011).
Jain N, Ottino JM, Lueptow RM. Regimes of segregation and mixing in combined size and
density granular systems: an experimental study. Granular Matter 7, 69–81 (2005).
Kanaoka C, Emi H, Tanthapanichakoon W. Convective diffusional deposition and collection
efficiency of aerosol on a dust-loaded fiber. AIChE Journal, 29(6), 895–902 (1983).
Kawamoto H. Some techniques on electrostatic separation of particle size utilizing electro-
static traveling-wave field. Journal of Electrostatics 66, 220–228 (2008).
Kawamoto H, Hasegawa N. Traveling wave transport of particles and particle size classifica-
tion. Journal of Imaging Science and Technology 48, 404–411 (2004).
Kawamoto H, Hayashi S. Fundamental investigation on electrostatic travelling-wave transport
of a liquid drop. Journal of Physics D: Applied Physics 39, 418–423 (2006).
Kawamoto H, Seki K, Kuromiya N. Mechanism of traveling-wave transport of particles.
Journal of Physics D: Applied Physics 39, 1249–1256 (2006).
Kim JM, Lee SG, Kim C. Numerical simulations of particle migration in suspension flows:
Frame-invariant formulation of curvature-induced migration. Journal of Non-Newtonian
Fluid Mechanics 150, 162–176 (2008).
Koh CJ, Hookham P, Leal LG. An experimental investigation of concentrated suspension
flows in a rectangular channel. Journal of Fluid Mechanics 256, 1–32 (1994).
Konstandopoulos AG. Deposit growth dynamics: Particle sticking and scattering phenomena.
Powder Technology 109, 262–277 (2000).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

336 Applications

Krieger IM, Rheology of monodisperse lattices. Advances in Colloid and Interface Science
3(2), 111–136 (1972).
Kupka T, Zajac K, Weber R. Effect of fuel type and deposition surface temperature on the
growth and structure of an ash deposit collected during co-firing of coal with sewage sludge
and sawdust. Energy & Fuels 23, 3429–3436 (2009).
Leighton D, Acrivos A. The shear-induced migration of particles in concentrated suspensions.
Journal of Fluid Mechanics 181, 415–439 (1987).
Li HM, McCarthy JJ. Controlling cohesive particle mixing and segregation. Physical Review
Letters 90(18), 184301 (2003).
Li SQ, Marshall JS. Discrete element simulation of micro-particle deposition on a cylindrical
fiber in an array. Journal of Aerosol Science 38, 1031–1046 (2007).
Li SQ, Marshall JS, Liu G, Yao Q. Adhesive particulate flow: The discrete element method
and its application in energy and environmental engineering. Progress in Energy and Com-
bustion Science 37(6), 633–668 (2011).
Liu G, Marshall JS. Effect of particle adhesion and interactions on motion by traveling waves
on an electric curtain. Journal of Electrostatics 68, 179–189 (2010a).
Liu G, Marshall JS. Particle transport by standing waves on an electric curtain. Journal of
Electrostatics 68, 289–298 (2010b).
Liu G, Marshall JS, Li SQ, Yao Q. Discrete-element method for particle capture by a body in
an electrostatic field. International Journal for Numerical Methods in Engineering 84(13),
1589–1612 (2010).
Liu PY, Yang RY, Yu AB. DEM study of the transverse mixing of wet particles in rotating
drums. Chemical Engineering Science 86, 99–107 (2013).
Machowski W, Balachandran W. Electrodynamic control and separation of charged particles
using traveling wave field technique. Journal of Electrostatics 40–41, 325–330 (1997).
Masuda S, Kamimura T. Approximate methods for calculating a non-uniform traveling field.
Journal of Electrostatics 1, 351–370 (1975).
Masuda S, Fujibayashi K, Ishida K, Inaba H. Confinement and transportation of charged
aerosol clouds via electric curtain. Electrical Engineering in Japan 92, 43–52 (1972).
Masuda S, Washizu M, Iwadare M. Separation of small particles suspended in liquid by
non-uniform traveling field. IEEE Transactions on Industry Applications IA-23, 478–480
(1987).
Masuda H, Gotoh K, Fukada H, Banba Y. The removal of particles from flat surfaces using a
high-speed air jet. Advanced Powder Technology 5, 205–217 (1994).
Matas JP, Morris JF, Guazzelli E. Inertial migration of rigid spherical particles in Poiseuille
flow. Journal of Fluid Mechanics 515, 171–195 (2004).
Maynard AB, Marshall JS. Particle removal from a surface by a bounded vortex flow. Inter-
national Journal of Heat and Fluid Flow 32(5), 901–914 (2011).
Meier SW, Lueptow RM, Ottino JM. A dynamical systems approach to mixing and segregation
of granular materials in tumblers. Advances in Physics 56(5), 757–827 (2007).
Melcher JR, Warren EP, Kotwal RH. Travelling wave delivery of single component developer.
IEEE Transactions on Industry Applications 25, 956–961 (1989).
Mellmann J. The transverse motion of solids in rotating cylinders: Forms of motion and
transition behavior. Powder Technology 118, 251–270 (2001).
Nott PB, Brady JF. Pressure-driven flow of suspensions: Simulation and theory. Journal of
Fluid Mechanics 275, 157–199 (1994).
Oak MJ, Saville DA, Lamb GER. Particle capture on fibers in strong electric fields. I. Experi-
ment studies of the effects of fiber charge, fiber configuration and dendrite structure. Journal
of Colloid and Interface Science 106, 490–501 (1985).
Otani Y, Namiki N, Emi H. Removal of fine particles from smooth flat surfaces by consecutive
pulse air jets. Aerosol Science and Technology 23, 665–673 (1995).
Parras L, Fernandez-Feria R. Interaction of an unconfined vortex with a solid surface. Physics
of Fluids 19, 067104 (2007).
Perry JL, Kandlikar SG. Fouling and its mitigation in silicon microchannels used for IC chip
cooling. Microfluidics and Nanofluidics 5, 357–371 (2008).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11

References 337

Phillips RJ, Armstrong RC, Brown RA, Graham AL, Abbott JR. A constitutive equation
for concentrated suspensions that accounts for shear-induced particle migration. Physics of
Fluids A 4(1), 30–40 (1992).
Phillips WRC. On vortex boundary layers. Proceedings of the Royal Society A 400, 253–261
(1985).
Phillips WRC, Khoo BC. The boundary layer beneath a Rankine-like vortex. Proceedings of
the Royal Society A 411, 177–192 (1987).
Ptak T, Jaroszcyk T. Theoretical-experimental aerosol filtration model for fibrous filters at
intermediate Reynolds numbers. In Proceedings of the 5th World Filtration Congress, Nice
(1990).
Ramarao BV, Tien C, Mohan S. Calculation of single-fiber efficiencies for interception and
impaction with superposed Brownian motion. Journal of Aerosol Science 25(2), 295–313
(1994).
Rembor HJ, Maus R, Umhauer H. Measurements of single fiber efficiencies at critical values
of the Stokes number. Particle and Particle Systems Characterization 16, 54–59 (1999).
Rubinow SI, Keller JB. The transverse force on a spinning sphere moving in a viscous fluid.
Journal of Fluid Mechanics 11, 447–459 (1961).
Saffman PG. The lift on a small sphere in a slow shear flow. Journal of Fluid Mechanics 22,
385–400 (1965).
Sakakibara J, Hishida K, Phillips WRC. On the vortical structure in a plane impinging jet.
Journal of Fluid Mechanics 434, 273–300 (2001).
Schmidlin FW. A new nonlevitated mode of traveling wave toner transport. IEEE Transac-
tions on Industrial Applications 27, 480–487 (1991).
Schmidlin FW. Modes of traveling wave particle transport and their applications. Journal of
Electrostatics 34, 225–244 (1995).
Schonberg JA, Hinch EJ. Inertial migration of a sphere in Poiseuille flow. Journal of Fluid
Mechanics 203, 517–524 (1989).
Schutyser MAI, Padding JT, Weber FJ, Briels WJ, Rinzema A, Boom R. Discrete particle
simulations predicting mixing behaviour of solid substrate in a rotating drum fermenter.
Biotechnology and Bioengineering 75, 666–675 (2001).
Schutyser MAI, Weber FJ, Briels WJ, Boom RM, Rinzema A. Three-dimensional simula-
tion of grain mixing in three different rotating drum designs for solid-state fermentation.
Biotechnology and Bioengineering 79, 284–294 (2002).
Segré G, Silberberg A. Behaviour of macroscopic rigid spheres in Poiseuille flow. Part 2.
Experimental results and interpretation. Journal of Fluid Mechanics 14, 136–157 (1962).
Vachon N, Hitt DL. A bound-vortex surface impingement method for adhered dust particle
removal. Proc. 40th AIAA Fluid Dynamics Conference, AIAA paper 2010–4297 (2010).
Vachon N, Hitt DL. An experimental investigation of a bound-vortex surface impingement
method for dust removal. Proc. 41st AIAA Fluid Dynamics Conference, AIAA paper
2011–3893 (2011).
Wang CS. Electrostatic forces in fibrous filters: A review. Powder Technology 118(1–2), 166–
170 (2001).
Weiss LC, Thibodeaux DP. Separation of seed by-products by an AC electric field. Journal
of the American Oil Chemists’ Society 61(5), 886–890 (1984).
Xu XG, Li SQ, Li GD, Yao Q. Effect of co-firing straw with two coals on the ash deposition
behavior in a down-fired pulverized coal combustor. Energy & Fuels 24, 241–249 (2010).
Yang M, Li SQ, Yao Q. Mechanistic studies of initial deposition of fine adhesive particles on
a fiber using discrete-element methods. Powder Technology (in press, 2013).
Yang RY, Yu AB, McElroy L, Bao J. Numerical simulation of particle dynamics in different
flow regimes in a rotating drum. Powder Technology 188, 170–177 (2008).
Zhang XW, Yao ZH, Hao PF, Xu HQ. Study on particle removal efficiency of an impinging
jet by an image-processing method. Experiments in Fluids 32(3), 376–380 (2002).
Ziskind G, Yarin LP, Peles S, Gutfinger C. Experimental investigation of particle removal
from surfaces by pulsed air jets. Aerosol Science and Technology 36, 652–659 (2002).
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-11 CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 25, 2013 14:11
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-IND CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 9:24

Index

acoustic radiation pressure, 171 capillary number, 118

acoustic wake effect, 173–175 capillary torque, 22, 118–120
added mass, 142 Clausius-Mossotti function
added mass force, 142–143 complex, 228, 229
adhesion conducting materials, 227
categories of, 3 dielectric materials, 226
adhesion parameter, 15, 315 magnetic, 238
adhesive contact mechanics coalescence kernel, 32
regimes, 88 collision
aerosol reactor, 25 two-particle, 44–45, 56–58, 63–64
agglomerate collision detection, 211–216
definition of, 2 collision force
agglomeration measures, 297 ellipsoidal particles, 218–221
coordination number, 302 normal dissipative, 60, 96–98
equivalent ellipse, 300 normal elastic, 54–56
fractal dimension, 300–302 sliding, 73, 98
orientation, 298 collision kernel, 190
radial distribution function, 302 collision torque
radius of gyration, 300 rolling, 78–79, 100
aggregate twisting, 74, 98
definition of, 2 concentration
local field, 294–297
Bagnold number, 14 mass, 12
Bassett force. See history force volume, 11
BD. See Brownian dynamics contact
blood flow, 23–24 point, 51
Boltzmann constant, 37, 135 radius, 51
Boltzmann equation, 102 region, 18, 51
Bond number contact de-electrification, 237
granular, 327 contact electrification, 235–237
boundary element method contact mechanics
electrostatic, 239 definition, 51
bounded vortex flow. See vortex flow: particle contact-value theorem, 105
mitigation by Coulomb force, 224
Brownian dynamics, 4, 37–38 crossing-trajectories effect, 199
Brownian force, 145–147 Cundall-Strack model, 73
on a nanoparticle, 260–261 Cunningham correction factor. See slip correction
Brownian motion, 37, 145, 308 factor
bubble ring, 8
buoyancy force. See pressure gradient force Debye length, 103
Debye–Hückel approximation, 102
capillary force. See liquid bridge force: capillary DEM. See discrete element method

339
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-IND CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 9:24

340 Index

dendrites, 18 flow past sphere

Derjaguin approximation, 88, 106 Oseen solution, 168
Derjaguin-Landau-Verwey-Overbeek theory, flow singularity
103–107 oseenlet, 168
Derjaguin-Muller-Toporov model, 88–91, 268 rotlet, 169
dielectrophoretic stokeslet, 168
particle focusing, 231 stokeslet image, 170
particle separation, 230 stresslet, 169
dielectrophoretic force, 20, 227 FMM. See fast multipole method
time-averaged, 228 friction
dielectrophoretic torque, 229 between nanoparticles, 267
diffuse layer, 102 dynamic, 70
diffusion charging, 233–235 static, 70
discrete element method, 4, 40–41 friction factor, 131
hard-sphere approach, 40 Froude number, 148, 320
soft-sphere approach, 40
discrete parcel method, 30–32 general dynamic equation, 33
dissipative particle dynamics, 4, 38–39 gorge approximation, 112
DLVO theory. See Derjaguin-Landau- granular flow, 5, 22
Verwey-Overbeek theory
DMT model. See Derjaguin-Muller-Toporov Hamaker coefficient, 83, 84
model Hertz theory. See collision force: normal elastic
Dominik-Tielens model, 100 history force, 145
DPD. See dissipative particle dynamics hysteric contact model, 124
DPM. See discrete parcel method
drag force inertia effect, 184–185, 198
crowding correction factor, 156–165 interaction energy, 82
ellipsoidal particle, 209–210
Faxén correction, 132 JKR model. See Johnson-Kendall-Roberts model
inertia correction factor, 133–135 Johnson-Kendall-Roberts model, 91–94
near wall effect, 151–154
on a nanoparticle, 258–260 kinetic theory, 135
slip correction factor, 136, 258 Knudsen number, 9, 136, 257
Stokes, 5, 131, 166
drift velocity. See slip velocity Lagrangian particle methods, 4
Dugdale approximation, 94 Langevin dynamics, 4
Langevin equation, 37, 145, 196
effective mass, 45 Lattice Boltzmann method, 161
effective moduli, 51 LD. See Langevin dynamics
effective radius, 42, 51 Lennard-Jones potential, 36, 84
ellipsoidal particle, 217 level set function, 293–294
Eikonal equation, 293 level surfaces, 214
elasticity parameter, 42 lift force
elastohydrodynamics, 42–44 inviscid, 143
electric curtain, 20, 331–332 Magnus, 141
electric dipole, 224–226 near wall effect, 156
electric double layer, 100–107 Saffman, 138–140
Euler angles, 207 ligand-receptor binding, 3, 107–110
Euler parameters, 208 liquid bridge, 3, 22, 111
rupture distance, 118
fast Eulerian approximation, 186, 289 liquid bridge force, 326
fast multipole method, 245–248 capillary, 22, 111–116
box structure, 247 effect of roughness, 116–117
error bound, 247 viscous, 117–118
field charging, 232–233 Lorentz force, 237
filtration, 15–18, 311–316 lubrication approximation, 42, 117
enhancement by electric field, 318–320
enhancement by particle size differences, Mach number, 136
316–318 magnetic dipole, 238
first-contact energy loss, 96 magnetophoresis, 238
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-IND CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 9:24

Index 341

Magnus effect, 140 plasticity

Magnus force. See lift force: Magnus hysteric model, 68
marginalization. See particle migration: population balance method, 3, 32–33
marginalization of blood cells pressure gradient force, 141
mass loading, 13 pseudo-image method, 242–243
Maugis parameter, 88, 95
Maugis-Dugdale model, 94–96, 266 radial basis function, 294
Maxwell-Wagner relaxation time, 229 restitution coefficient, 44, 46–48, 58, 60, 64
MD. See molecular dynamics Reynolds number, 10–11, 148
M-D model. See Maugis-Dugdale model rheological models, 53
mean-free-path rolling displacement, 75
of a nanoparticle, 261 rolling velocity, 54, 74–77
of gas molecules, 135 rotating drum
modeling cohesive mixing, 326–328
macroscale, 30 flow regimes, 320–322
mesoscale, 30, 34, 41 noncohesive mixing, 322–326
microscale, 30, 41
molecular dynamics, 4, 36–37, 266 separation distance
momentum coupling minimum, 84
one-way, 13 shear-induced migration. See particle migration:
parameter, 13 collision-induced
two-way, 13 sintering, 25, 120–126
of nanoparticles, 278–282
nanoparticles regime map, 121
collision time scale, 275–278 slip velocity, 10, 185
collisions between, 271 Smoluchowski equation, 32
definition, 256 Stern layer, 101
dipole-dipole interaction, 273 stochastic Lagrangian models, 196–199
drag force. See drag force: on a nanoparticle stochastic vortex structure model, 201–203
free-molecular regime, 257, 261 Stokes drag. See drag force: Stokes
sintering of. See sintering: of nanoparticles Stokes number, 5–6
transition regime, 257 eddy, 184
normal overlap, 51 Kolmogorov scale, 190
Stokes-Einstein relation, 147
orthokinetic motion, 173 Stokesian dynamics, 165–170
Oseen dynamics, 168 surface energy density, 85
Oseen equation, 134 suspension, 5
Oseen tensor, 166, 169
Tabor parameter, 87, 258
panel subdivision method, 244 thermophoretic force, 262–265
particle bed time scale
fluidized, 159–161 fluid convective, 286
packed, 156 nanoparticle collisions. See nanoparticles:
particle chains, 250–254 collision time scale
particle cluster, 166 nanoparticle sintering, 279–280
particle clustering, 183 particle aerodynamic, 287
particle dispersion, 183 particle contact, 286
particle drift measure, 188 particle convective, 286
particle migration time stepping
collision-induced, 307–309 multiple step algorithm, 288–289
effect of wavy walls, 309–311 numerical stability, 287–288
inertial, 306–307 tribocharging. See contact electrification
marginalization of blood cells, 24 triboelectric series, 235
particle search algorithm, 290–293 tubular pinch effect. See particle migration:
particulate flow inertial
dense, 12 turbulence
dilute, 1 acceleration probability density function, 186
PBM. See population balance method inertial range scaling, 183
Peclet number, 148, 149, 258 Kolmogorov scale, 183, 257
thermophoretic, 265 Lagrangian integral scale, 183
 Trim: 7in × 10in Top: 0.375in Gutter: 0.875in
CUUS2101-IND CUUS2101/Marshall & Li ISBN: 978 1 107 03207 1 December 26, 2013 9:24

342 Index

turbulent collisions viscoelasticity, 59

accelerative-correlated, 192–195 viscosity
accelerative-independent, 192 relationship to mean-free-path, 136
classifications, 188 viscous torque, 132
orthokinetic, 190–192 vortex flow
particle mitigation by, 330–331
ultrasound trap, 171 particle motion in, 7–8

van der Waals force, 86 Young-Laplace formula, 112

comparison to electrostatic force,
249–250 zeta potential, 102