Advances in Chemical Engineering - Vol 30 Multiscale Analysis

Advances in Chemical Engineering,
Volume 30: Multiscale Analysis

by Guy B. Marin
• ISBN: 0120085305
• Pub. Date: October 21, 2005
• Publisher: Academic Press
CONTENTS
CONTRIBUTORS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
PREFACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
A Review of Multiscale Analysis: Examples from Systems Biology,

Materials Engineering, and Other Fluid–Surface Interacting Systems
DIONISIOS G. VLACHOS
I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
II. Deterministic, Continuum Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
A. Hierarchy of Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
B. Solving Deterministic, Continuum Differential Equation Models: Techniques and
Status. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
III. Overview of Discrete, Particle Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
A. Hierarchy of Stochastic Models for Well-mixed, Chemically Reacting Systems . . 9
B. Solving Master Equations Stochastically: Monte Carlo Methods . . . . . . . . . . . 10
IV. Classification of Multiscale Simulation Approaches . . . . . . . . . . . . . . . . . . . . . 12
V. Hybrid Multiscale Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
A. Onion-type Hybrid Multiscale Simulations and Algorithms . . . . . . . . . . . . . . 15
B. Application of Onion-type Hybrid Multiscale Simulation to Growth
of Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
C. Applications of Onion-type Hybrid Multiscale Simulation to Other Areas . . . . . 23
D. Multigrid-type Hybrid Multiscale Simulations . . . . . . . . . . . . . . . . . . . . . . 24
E. An Example of Multigrid-type Hybrid Multiscale Simulation for Growth under
Large Length Scale Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
F. Challenges in Hybrid Multiscale Simulations . . . . . . . . . . . . . . . . . . . . . . . 28
VI. Coarse Graining of Stochastic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A. Temporal Upscaling of KMC Simulation in Well-mixed Systems. . . . . . . . . . . 32
B. Spatial Upscaling of Distributed (Lattice) KMC Simulation. . . . . . . . . . . . . . 36
C. Spatiotemporal Acceleration of Distributed (Lattice) KMC Simulation . . . . . . . 38
VII. Multiscale, Stochastic Modeling of Biological Networks. . . . . . . . . . . . . . . . . . . 40
A. Spatially Well-mixed Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
B. Spatially Distributed Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
VIII. Systems Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A. Sensitivity and Identifiability Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
B. Parameter Estimation from Experimental Data and Finer Scale Models . . . . . . 51
C. Model Reduction and Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
D. Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
v
vi CONTENTS
IX. Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Quantifying Physics and Chemistry at Multiple Length-scales using

Magnetic Resonance Techniques
LYNN F. GLADDEN, MICHAEL D. MANTLE AND ANDREW J. SEDERMAN
I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
II. Principles of MR Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
A. Spatially Unresolved And Spatially Resolved
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
B. Nuclear Spin Relaxation Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
C. Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
D. Temperature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
E. The k-space Raster. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
F. Fast Data Acquisition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
III. Recent Developments in MR as a Tool in Chemical
Engineering Research. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
A. ‘‘Ultra-fast’’ Imaging of Velocity Fields . . . . . . . . . . . . . . . . . . . . . . . . . . 93
B. Multiple Images From a Single Excitation . . . . . . . . . . . . . . . . . . . . . . . . 96
C. Imaging Rotating Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
D. ‘‘Ultra-fast’’ Diffusion Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
E. Gas-phase MR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
IV. Reaction Engineering: From Catalyst to Reactor . . . . . . . . . . . . . . . . . . . . . . . 103
A. MR Spectroscopy of Catalysts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
B. Micro-imaging and Molecular Diffusion Studies of Formed Catalyst Pellets . . . . 105
C. Single-Phase Flow in Fixed-Bed Reactors . . . . . . . . . . . . . . . . . . . . . . . . . 110
D. Measuring Chemical Composition and Mass Transfer in Fixed-Bed Reactors:
In Situ Studies of Reactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
E. Two-Phase Flow in Fixed-Bed Reactors. . . . . . . . . . . . . . . . . . . . . . . . . . 119
F. Hydrodynamic Transitions in Fixed-Bed Reactors. . . . . . . . . . . . . . . . . . . . 123
V. Future Prospects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Modeling of transport and transformation processes in porous and

multiphase bodies
JURAJ KOSEK, FRANTIŠEK ŠTĚPÁNEK AND MILOŠ MAREK
I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
II. Methodology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
A. Representation of Multiphase Media. . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
B. Structure Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
C. Morphological Characterization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
CONTENTS vii
D. Digital Reconstruction of Multiphase Media . . . . . . . . . . . . . . . . . . . . . . . 145

E. Calculation of Effective Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
F. Effective-scale Transport Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
III. Transformations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
A. Skeletonization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
B. Phase Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
C. Chemically Reactive Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
IV. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
A. Multi-scale Reconstruction of a Catalyst Pellet. . . . . . . . . . . . . . . . . . . . . . 175
B. Reconstruction of Closed-cell Polymer Foam Structure. . . . . . . . . . . . . . . . . 179
C. Polymer Particle Morphogenesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
D. Granulation and Dissolution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
E. Simulation of CO Oxidation on Reconstructed
Catalytic Washcoat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
V. Outlooks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
A. Biological Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
B. Materials Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
VI. Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
Spatially Averaged Multi-Scale Models for Chemical Reactors

SAIKAT CHAKRABORTY AND VEMURI BALAKOTAIAH
I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
A. A Brief History of Chemical Reactor Models. . . . . . . . . . . . . . . . . . . . . . . 208
B. Multi-scale Nature of Homogeneous and Catalytic Reactors . . . . . . . . . . . . . 211
C. Different Approaches to Multi-scale Averaging or Dimension Reduction. . . . . . 214
II. Spatial Averaging of Convection–diffusion–reaction Models using the
L–S Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
III. Spatially Averaged Models for Describing Dispersion Effects in Tubes and
Packed Beds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
A. A Hyperbolic-Averaged Model for Describing Dispersion Effects in
Tubes/Capillaries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
B. Multi-mode Hyperbolic Averaged Models for Describing Dispersion Effects
in Chromatographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
C. A Hyperbolic Model for Describing Dispersion Effects in Monoliths with
Diffusion into the Solid Phase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
IV. Spatially Averaged Multi-mode (Multi-scale) Models for Homogeneous Reactors . . . 239
A. Isothermal Tubular Reactors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
B. Loop and Recycle Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
C. Tank Reactors (CSTRs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
D. Non-isothermal Reactor Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
E. Multiple Reactions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
F. Examples Illustrating Use of Multi-mode Homogeneous Reactor Models. . . . . . 260
V. Spatially Averaged Multi-Mode (Multi-Scale) Models for Catalytic Reactors . . . . . . 273
A. Wall-catalyzed Reactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
B. Coupled Homogeneous and Wall-catalyzed Reactions . . . . . . . . . . . . . . . . . 277
viii CONTENTS
C. Non-isothermal Reactor Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

D. Examples Illustrating Use of Multi-mode Catalytic Reactor Models. . . . . . . . . 279
VI. Accuracy, Convergence and Region of Validity of Multi-mode/Multi-scale
Averaged Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
A. Accuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
B. Convergence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
C. Regularization of the Local Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
VII. Summary, Conclusions, and Recommendations for Future Work . . . . . . . . . . . . . 293
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
INDEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
CONTENTS OF VOLUMES IN THIS SERIAL . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
CONTRIBUTORS
Numbers in parentheses indicate the pages on which the authors’ contribution begin.
VEMURI BALAKOTAIAH, Department of Chemical Engineering, University of

Houston, Houston, TX 77204-4792, USA (205)
SAIKAT CHAKRABORTY, Department of Chemical Engineering, University of
Houston, Houston, TX 77204-4792, USA (205)
LYNN F. GLADDEN, Department of Chemical Engineering, University of
Cambridge, Pembroke Street, Cambridge CB2 3RA, UK (63)
JURAJ KOSEK, Department of Chemical Engineering, Prague Institute of Chemical
Technology, Czech Republic (137)
MICHAEL D. MANTLE, Department of Chemical Engineering, University of
MILOS MAREK, Department of Chemical Engineering, Prague Institute of
Chemical Technology, Czech Republic (137)
ANDREW J. SEDERMAN, Department of Chemical Engineering, University of
FRANTIŠEK ŠTĚPÁNEK, Department of Chemical Engineering, Imperial College
London, UK (137)
DIONISIOS G. VLACHOS, Department of Chemical Engineering and Center for
Catalytic Science and Technology, University of Delaware, Newark,
Delaware 19716, USA (1)
ix
PREFACE
The present volume being the first one I have been editing, I would like to
take the opportunity to comment briefly on the needs and criteria a series such
as ‘‘Advances in Chemical Engineering’’ should meet today.
In the first preface of this series, almost 50 years ago, the founding editors
raised the issue of ‘‘the flood of information’’ created by ‘‘the practioners of the
chemical engineering art’’. Communication both within and among scientific
communities defines the borders of such a community and constitutes a major
activity of any scientist, next to research as such of course. Complementary to the
very important oral presentations and discussions at seminars or conferences, the
scientific press has from the very beginning of print as medium been very in-
strumental in this: verba volent, scripta manent. The emergence of Information
and Communication Technology (ICT) in general and Internet in particular has
led to a tremendous increase of the amount of information that is available and
the frequency at which it is exchanged. I am convinced that this does not decrease
the added value of the so-called archival publications, on the contrary. This
holds even more so for a series offering a stage to scholars who, upon invitation,
are capable and willing to spend time to report in a broader context on their
personal contributions to a field. Any paper in ‘‘Advances in Chemical Engi-
neering’’ should allow to assess the state-of-the-art in a particular domain and to
develop a feeling of its further evolution without claiming to be exhaustive.
Going beyond the limits imposed by the ‘‘regular’’ scientific journals while not
imposing those typical of a text book is part of the success recipe I have in mind.
The subjects covered are not limited to the classical chemical engineering
disciplines. Contributions connecting chemical engineering to related scientific
fields, either providing a fundamental basis or introducing new concepts and
tools, are encouraged.
Of course applications of chemical engineering receive special attention. A
balance between well-developed areas such as process industry, transformation
of materials, energy and environmental issues and areas where applications of
chemical engineering are more recent or emerging is aimed at.
The theme of the present volume ‘‘Multiscale Analysis’’ has been introduced
about a decade ago and is now reaching a stage where a first balance can be
made and further research directions should be decided. Which are the dom-
inant and most successful concepts or methodologies? How do these relate to
our ‘‘classics’’? How and where should they be applied next?
The selection of the contributions was among others guided by the concern
not to make the gap between the different scales too large. The reader will not
be confronted with quantum mechanics at one side of the spectrum nor with
chemical plants or even the environment on the other side. Bridging the gap
xi
xii PREFACE
between the phenomena occurring on the scale of a catalytic site and those on the
scale of a reactor or, even smaller, that of a polymer is sufficiently challenging
and allows, if not to answer, at least to address the above questions. Maintaining
a strong connection with reality, i.e. experimental data was another selection
criterion. Experimental validation remains the corner stone of any theoretical
development and very powerful experimental techniques are emerging.
First, a broad overview is provided by Dion Vlachos of the University of
Delaware. An important example of experimental techniques is discussed in
depth by Lynn Gladden and coworkers from the University of Cambridge.
Coming from the medical world, Magnetic Resonance techniques can now
provide even quantitative answers to problems our community is faced with.
The modeling issue is discussed further in the paper coming from the Prague
Institute of Chemical Technology and Imperial College, London. Finally, the
limitations of the classic reactor engineering models are outlined in a paper from
the University of Houston by contrasting the intuitive averaging over length
and timescales they are based upon with the rigorous Liapunov–Schmidt meth-
od. The authors have made an effort to provide examples when appropriate.
References to ‘‘a jar containing soup and meat balls’’ or to ‘‘the wall of a
champagne glass’’ provocatively illustrate the broadness of the applications of
chemical engineering.
This makes me return to the first preface of this series and even to the very
first sentences of it. The danger of fragmentation of our field, some of us are so
afraid of, was presented as an opportunity: ‘‘The chemical engineer ministers to
an industry of far-flung interests. Its products range from soap to plutonium,
from gasoline to paper, from antibiotics to cement. It flourishes on change: new
products, processes, methods, and applications; new needs are created and
foreseen. Versatile men with breath of interest in science and commerce have
been demanded and the challenge of the field has found for it such men.’’ I leave
it up to the reader to appreciate the flavor of the ‘‘old’’ American, the list of
applications, the used gender. Most striking and still very much a topic of the
day, however, is the frontier spirit expressed by these lines. A spirit which can be
summarized by the device of a 16th century scholar, Pieter de Zuttere, who lived
and preached in the Low Countries and in particular in the city of Ghent:
‘‘Cesse le vieux, s’il appert mieux’’ in old French, or in his and his contem-
porary Lowys Elsevier’s native tongue:
‘‘Als beter can blycken, dat oude sal wijcken’’.
GUY B. MARIN
GHENT, BELGIUM
March 2005
A REVIEW OF MULTISCALE ANALYSIS: EXAMPLES FROM
SYSTEMS BIOLOGY, MATERIALS ENGINEERING, AND
OTHER FLUID–SURFACE INTERACTING SYSTEMS
Dionisios G. Vlachos
Department of Chemical Engineering and Center for Catalytic Science and Technology
(CCST)
University of Delaware, Newark, DE 19716, USA
I. Introduction 2
II. Deterministic, Continuum Models 6
A. Hierarchy of Models 7
B. Solving Deterministic, Continuum Differential Equa-
tion Models: Techniques and Status 8
III. Overview of Discrete, Particle Models 8
A. Hierarchy of Stochastic Models for Well-mixed,
Chemically Reacting Systems 9
B. Solving Master Equations Stochastically: Monte
Carlo Methods 10
IV. Classification of Multiscale Simulation Approaches 12
V. Hybrid Multiscale Simulation 14
A. Onion-type Hybrid Multiscale Simulations and Al-
gorithms 15
B. Application of Onion-type Hybrid Multiscale Simu-
lation to Growth of Materials 17
C. Applications of Onion-type Hybrid Multiscale Sim-
ulation to Other Areas 23
D. Multigrid-type Hybrid Multiscale Simulations 24
E. An Example of Multigrid-type Hybrid Multiscale
Simulation for Growth under Large Length Scale
Gradients 26
F. Challenges in Hybrid Multiscale Simulations 28
VI. Coarse Graining of Stochastic Models 32
A. Temporal Upscaling of KMC Simulation in Well-
mixed Systems 32
B. Spatial Upscaling of Distributed (Lattice) KMC
Simulation 36
C. Spatiotemporal Acceleration of Distributed (Lattice)
KMC Simulation 38
VII. Multiscale, Stochastic Modeling of Biological Networks 40
A. Spatially Well-mixed Systems 40
B. Spatially Distributed Systems 44
VIII. Systems Tasks 45
A. Sensitivity and Identifiability Analyses 46
B. Parameter Estimation from Experimental Data and
Finer Scale Models 51
C. Model Reduction and Control 53
1
Advances in Chemical Engineering, vol. 30 Copyright r 2005 by Elsevier Inc.
ISSN 0065 2377 All rights reserved
DOI 10.1016/S0065-2377(05)30001-9
2 DIONISIOS G. VLACHOS
D. Bifurcation 54
IX. Outlook 55
Acknowledgments 55
References 56
Abstract
Multiscale simulation is an emerging scientific field that spans many

disciplines, including physics, chemistry, mathematics, statistics, chemical
engineering, mechanical engineering, and materials science. This review
paper first defines this new scientific field and outlines its objectives. An
overview of deterministic, continuum models and discrete, particle models
is then given. Among discrete, particle models, emphasis is placed on
Monte Carlo stochastic simulation methods in well-mixed and spatially
distributed systems. Next, a classification of multiscale methods is carried
out based on separation of length and time scales and the computational
and mathematical approach taken. Broadly speaking, hybrid simulation
and coarse graining or mesoscopic modeling are identified as two general
and complementary approaches of multiscale modeling. The former is
further classified into onion- and multigrid-type simulation depending on
length scales and the presence or not of gradients. Several approaches,
such as the net event, the probability weighted, the Poisson and binomial
t-leap, and the hybrid, are discussed for acceleration of stochastic sim-
ulation. In order to demonstrate the unifying principles of multiscale
simulation, examples from different areas are discussed, including systems
biology, materials growth and other reacting systems, fluids, and statistical
mechanics. While the classification is general and examples from other
scales and tools are touched upon, in this review emphasis is placed on
stochastic models, their coarse graining, and their integration with con-
tinuum deterministic models, i.e., on the coupling of mesoscopic and
macroscopic scales. The concept of hierarchical multiscale modeling is
discussed in some length. Finally, the importance of systems-level tools
such as sensitivity analysis, parameter estimation, optimization, control,
model reduction, and bifurcation in multiscale analysis is underscored.
I. Introduction
A decadal report recently issued by the National Research Council (NRC),

entitled Beyond the Molecular Frontier: Challenges for Chemistry and Chemical
A REVIEW OF MULTISCALE ANALYSIS 3
Engineering (NRC, 2003a), advances 13 ‘‘Grand Challenges’’ for the field.

‘‘Advancing Chemical Theory and Modeling’’ is viewed as one of the critical,
enabling technologies. Quoting from the report: ‘‘Chemistry covers an enor-
mous span of time and space from atoms and molecules to industrial-scale
processing. Advances in computing and modeling could help us connect phe-
nomena at the electronic and molecular scale to the commercial processing.’’ In
the information and communications NRC report and in recent roadmaps,
multiscale analysis is repeatedly identified as the emerging computational and
mathematical science that could enable design and control of complex engi-
neering systems (Thompson, 1999; NRC, 2003b).
The foundations of transport phenomena, reaction engineering, thermody-
namics, and nonlinear analysis, along with significant advances in numerical
analysis of differential equations at the continuum level and the increase in
computational power, have shaped for the most part the first engineering process
modeling paradigm of chemical sciences of the 20th century (the BSL paradigm
of continuum conservation equations and continuum constitutive relations
(Bird et al., 1960)). An outcome of this long-time effort has been the widespread
use of computational fluid dynamics (CFD) simulation that nowadays routinely
assists the design of many industrial processes.
The rapid growth in computational speed over the past decades has enabled a
molecular-based approach to product and process engineering. Molecular sim-
ulations such as molecular dynamics (MD) and Monte Carlo (MC) algorithms
have emerged as preeminent computational tools for science and engineering
research. Additional discrete particle simulations, such as Brownian dynamics
(BD), lattice Boltzmann (LB), direct simulation Monte Carlo (DSMC), and
dissipative particle dynamics (DPD), have attempted to bridge information
from the molecular to the mesoscopic scale, but often in a phenomenological
manner, as the rules of coarse graining are not fully established. At the other
end of the modeling spectrum, quantum mechanical (QM) calculations, such as
ab initio and density functional theory (DFT), in conjunction with transition
state theory (TST), have extended the realm of simulation to smaller scales by
providing electronic structure information such as potential energy surfaces
(PESs) and activation energies that are used in molecular simulations. The
advances in molecular and quantum mechanics theory and simulation have
established the second modeling paradigm (the molecular and quantum modeling
paradigm).
Multiscale simulation is emerging and will unquestionably become the third
modeling paradigm. The idea of multiscale modeling is straightforward: one
computes information at a smaller (finer) scale and passes it to a model at a
larger (coarser) scale (see Fig. 1) by leaving out degrees of freedom as one moves
from finer to coarser scales. Within this context, the most common goal of
multiscale modeling is to predict the macroscopic behavior of an engineering
process from first principles (upscaling or bottom-up approach). This approach
has its roots in the work of Newton, Hooke, Bernoulli, Einstein, Bodenstein,
Length
Scale (m)
10-2 Mesoscopic CFD

Upscaling or Theory/CGMC
10-6 bottom-up KMC,
info traffic Master Eqs.,
10-7 DSMC, LB,
DPD, etc.
Downscaling
Extended or top-down
Trajectory MD info traffic
10-9
Quantum
Mechanics
10-10 TST
10-12 10-9 10-3 103

Time Scale (s)
FIG. 1. Schematic representation depicting scales and various simulators. Most multiscale work
has focused on the simplest, one-way information passing, usually from the finest to the coarsest
scale model. On the other hand, most processes exhibit strong coupling between scales or lack
separation of scales.
and others (Phillips, 2002; Raimondeau and Vlachos, 2002a) who left out many
degrees of freedom to propose continuum-based constitutive equations and
simple models for obtaining answers of interest. In recent times, this goal has
been served well, for example, by equilibrium statistical mechanics with QM-
based potentials and associated molecular (MD and MC) models. I envision an
equally important second goal of multiscale analysis, stemming from the
emerging areas of biotechnology, nanotechnology, and device miniaturization.
This goal is the ability to predict and control phenomena and devices with
resolution approaching nanoscopic scale while manipulating macroscopic (en-
gineering) scale variables such as flow rates, pressures, and temperature (down-
scaling or top-down approach). This manipulation may not happen with active
model-based control but instead by properly designing a system, using multi-
scale model-based information, to function desirably at the molecular level. This
issue is further discussed in the section on systems tasks. Reverse engineering is
yet a third potential goal of top-down information flow: given a desirable
property, it is desirable to predict suitable candidate materials (e.g., multicom-
ponent, multifunctional catalysts) and develop rational ways to synthesize them.
This last goal addresses product-driven engineering that is believed by many to
be the future of chemical sciences (Cussler and Wei, 2003). For the most part,
the last two goals have so far remained elusive but are the ones on which
multiscale modeling and simulation would have the most impact in the next
decade.
Advances in analytical methods, such as scanning probe and high-resolution
transmission electron microscopy, now enable experiments with molecular-level
resolution. Furthermore, data from small ensembles of molecules or single en-
tities (e.g., a living cell) become more common. Effectively utilizing these and
related emerging tools and data to develop new products and processes will be
greatly facilitated by a complementary development in multiscale modeling that
can not only model experimentally observed phenomena, but also aid in the
prediction of new, as of yet, unproven products and processes.
Multiscale simulation is growing so rapidly that it emerges as a new mul-
tidisciplinary scientific field. Figure 2 summarizes the number of publications
over the past decade using the term ‘‘multiscale’’ and ‘‘multi-scale’’ in their title
only or in all title, abstract, and keywords. While the term multiscale means
different things in various fields, the explosion is clear. Two new journals,
Multiscale Modeling and Simulation, A SIAM Interdisciplinary Journal, and
the International Journal on Multiscale Computational Engineering (Begell
House Publishers, NY) started in 2003, point to the rapid evolution of this new
field. There have been many activities that speak to the same fact. Examples
include the recent issues 8 and 9 of the 59th volume of Chemical Engineering
1200
Title, abstract, and keywords
1000
Title
800
No. of papers
600
400
200
0
90-91 92-93 94-95 96-97 98-99 00-01 02-03
Pair of years
FIG. 2. Sum of the number of publications over periods of two years containing the word ‘‘mul-
tiscale’’ and ‘‘multi-scale’’ in the title only or in the title, abstract, and keywords (found through the
Web of Science). An explosion in the number of publications is observed. However, this search is just
a measure since many of these papers do not really adhere with the definition of multiscale modeling
used here, and others, while truly multiscale, are not accounted for because ‘‘multiscale’’ or ‘‘multi-
scale’’ is not present in their title, abstract, and keywords.
Science in 2004 that have been dedicated to Complex Systems and Multi-scale
Methodology, the forth issue of the 29th volume in Computers in Chemical
Engineering on Multiscale Simulation published in 2005, the Springer-Verlag
IMA edited book on Dispersive Transport Equations and Multiscale Models
resulting from a related workshop, numerous workshops, and a topical con-
ference on Multiscale Analysis in the 2005 AIChE meeting, just to mention a
few.
Multiscale simulation builds on the foundations developed in the 20th cen-
tury of continuum, deterministic and discrete, particle-type models. It attempts
to seamlessly integrate models at various scales, extend existing tools to larger
length and time scales, and develop theoretical connections between tools over
multiple scales. It seems then appropriate to first provide a classification and
an overview of models at various scales before multiscale simulation is more
formally introduced and recent progress is reviewed. Since we have recently
given a review on multiscale simulation in catalysis and reaction engineering
(Raimondeau and Vlachos, 2002a), here a broader overview of multiscale sim-
ulation is given. The multidisciplinary nature of this emerging field makes this a
daunting task. For this reason, I have chosen to mainly focus on the areas of
systems biology and materials growth because these two fields are enticing an
increasing number of chemical engineers. Furthermore, by choosing two areas
one can clearly see unifying multiscale concepts that emerge across chemical
engineering. Some rather introductory examples from statistical mechanics and
reaction systems are also employed to illustrate key points and methods. Fi-
nally, I have tried to include references to some key mathematical pieces of work
and multiscale references from the physics, materials, and hydrodynamics com-
munities I am aware of with the hope of cross-fertilizing various disciplines
without necessarily being exhaustive in coverage (these areas deserve their own
review). For example, a recent, very good review from the mathematics com-
munity has just appeared after the submission of this manuscript that presents
some of the mathematical underpinnings of the algorithms and methods
touched upon below (Givon et al., 2004). While the discussed multiscale ap-
proach and issues are generic and apply to various models and scales, I have
judiciously chosen to mainly focus on the MC method, among other atomistic
or mesoscopic models, and the integration of MC with deterministic, continuum
models as an example of stochastic/continuum hybrid multiscale models. This
naturally provides more coherence to the chapter. Some key references from
other types of multiscale models are also given.
II. Deterministic, Continuum Models
Traditionally, modeling in chemical engineering has invoked continuum de-

scriptions of momentum, mass, and energy conservation (Bird et al., 1960)
where substantial mathematical and computational contributions have been

made over the past decades. Here, the discussion is limited to a brief classi-
fication that introduces the necessary terminology used in the remainder of the
chapter.
A. HIERARCHY OF MODELS
Continuum modeling has often been based on algebraic equations (AEs),

ordinary differential equations (ODEs), partial differential equations (PDEs),
and differential-algebraic equations (DAEs). PDEs provide the most general
description at the continuum level. ODEs typically describe transient, well-
mixed systems, such as the concentrations and temperature in a batch reactor or
in a continuous stirred tank reactor (CSTR), or 1D steady state balances, such
as a plug flow reactor (PFR) model or an axial dispersion model. A distinction
of ODEs entails initial vs. boundary value problems depending on where the
conditions are imposed, namely, only at the entrance or at the entrance and exit,
respectively. The hierarchy of deterministic, continuum models is summarized in
Fig. 3a. Using concepts of dimensional analysis and symmetry, models toward
the bottom of the graph can be thought of as reductions or limits of higher
dimensionality models (found toward the top of the graph).
Chemical master
(a) (b) equation; MC methods
Partial differential
equation based model
Model reduction
noise reduction
Population increase,
τ-leap approximate
Ordinary differential stochastic method
equation based model
Discretization
Chemical Langevin
Algebraic equation equation
based models
Deterministic,
continuum ODEs
FIG. 3. (a) Hierarchy of deterministic, continuum models. Dimensional analysis and symmetry are
powerful concepts in reducing the dimensionality of complex models. (b) Hierarchy of stochastic
models for chemically reacting well-mixed systems.
B. SOLVING DETERMINISTIC, CONTINUUM DIFFERENTIAL EQUATION MODELS:

TECHNIQUES AND STATUS
Substantial advances in computational power (Moore’s law) have had a tre-

mendous impact on the numerical solution of engineering problems. Concom-
itant with increases in computational power, significant advances in problem
solving have resulted through mathematical and/or computational develop-
ments. One example is the introduction of stiff ODE solvers by Gear (1969,
1971) that led to the development of adaptive time step, variable-order methods
that nowadays are available through the web. LSODA, VODE, and RADAU5
are some commonly employed packages for solving stiff ODEs and DDASSL
for DAEs (Brenan et al., 1996; Hairer and Wanner, 1991; Petzold, 1982). As
another example, over the last 30 years the solution of systems of linear equa-
tions has evolved from sparse Gauss elimination, to Gauss–Seidel, to successive
over-relaxation, to conjugate gradient, to multigrid, to parallel multigrid. It was
recently reported by Petzold in NRC (2003b) that such algorithm and software
development has led to four orders of magnitude speedup.
Methods for solving continuum models have advanced to such a point that
they are nowadays considered to be relatively mature. Since several packages
are available based on one of these methods, it becomes a matter of choosing an
appropriate package. Typical CFD and transport packages include Fluent, FI-
DAP, CFX, and Femlab (Femlab is a general purpose finite element (FEM)-
based program with specialized modules for chemical engineering applications).
Simple 1D problems can be solved with widely used teaching software, such as
Matlab and Mathematica. While in the 1980s considerable effort was devoted to
the discretization of PDEs and meshing of a complex domain, this is now a
relatively routine exercise that uses Gambit and the internal mesh generator of
Femlab even for complex geometries. CFD, depicted at the top of the pyramid
in Fig. 1, can be used as process simulator in a multiscale simulation of chemical
engineering. Memory limitations, especially for 3D simulations, robustness in
convergence, speed for complex chemistry in reacting flows, and accuracy are
still issues that need further improvements. Finally, interfacing CFD codes with
complex homemade chemistry codes or finer scale codes from the multiscale
ladder shown in Fig. 1 is also important.
III. Overview of Discrete, Particle Models
Discrete models treat individual atoms, molecules, or particles and can be

deterministic or stochastic. Examples of the former include MD simulations.
Examples of the latter are various MC methods, BD, DPD, DSMC, and LB
simulations. There are different ensembles in which these simulations can be
performed, depending on the quantities that one is interested in computing.
Different techniques are suitable for different tasks. For example, BD focuses
on molecules and particles in solution where the solvent is implicitly lumped
into a friction force. On the other hand, DSMC and LB are typically applied to
various fluid-related problems. MD is the only fundamental, first principles tool
where the equations of motion are solved using as input an interparticle po-
tential. MC methods map the system description into a stochastic Markov-
based framework. MD and MC are often thought of as molecular modeling
tools, whereas the rest are mesoscopic tools (lattice MC is also a mesoscopic
tool).
Many excellent sources on discrete, particle simulations exist (e.g., Allen and
Tildesley, 1989; Binder, 1986; Bird, 1988; Chen and Doolen, 1998; Frenkel and
Smit, 1996; Landau and Binder, 2000; Rastogi and Wagner, 1995; Wolf-Glad-
row, 2000). Volume 28 of Advances in Chemical Engineering, entitled ‘‘Mo-
lecular Modeling and Theory in Chemical Engineering,’’ presents an excellent
collection of molecular-based papers with applications across chemical engi-
neering. A recent overview of the tasks that can be accomplished via molecular
modeling, with special emphasis on MC methods, for irreversible chemical
processes is given in Vlachos (2005).
Obviously, the spectrum of mesoscale, particle-based tools is too vast to be
covered in a single paper. Therefore, in this and the subsequent sections, I
mainly elaborate on MC methods to illustrate various aspects of multiscale
modeling and simulation. Below, the modeling hierarchy for stochastic well-
mixed chemically reacting systems is first outlined, followed by a brief intro-
duction to MC methods.
A. HIERARCHY OF STOCHASTIC MODELS FOR WELL-MIXED, CHEMICALLY REACTING

SYSTEMS
A hierarchy of models can often be derived from a more detailed model under
certain assumptions. This approach was discussed above in the case of deter-
ministic, continuum models (see Fig. 3a). Such hierarchical models can be val-
uable in multiscale modeling. Let us just mention two cases. First, one could use
different models from a hierarchy of models for different situations or length
scales. This approach plays a key role in hybrid multiscale simulation discussed
extensively below. Second, one could easily apply systems tasks to a simpler
model to obtain an approximate solution that is then refined by employing a
more sophisticated, accurate, and expensive model from the hierarchy.
A major difficulty is that such hierarchies of molecular models are not exactly
known. Recent work by Gillespie (2000, 2002) has established such a hierarchy
for stochastic models of chemical reactions in a well-mixed batch reactor. This
hierarchy is depicted in Fig. 3b. In particular, it was shown that the chemical
master equation is deduced to a chemical Langevin equation when the pop-
ulation sizes are relatively large. Finally, the deterministic behavior can be
recovered in the limit of an infinite-size system ðN ! 1Þ. The concept of hi-

erarchical modeling is revisited in the section on systems tasks.
B. SOLVING MASTER EQUATIONS STOCHASTICALLY: MONTE CARLO METHODS
The introduction of the Metropolis MC algorithm in 1953 (Metropolis et al.,

1953) has established a new paradigm in computational statistical mechanics for
computing equilibrium solutions. Starting from a description of the physical
system in terms of a Hamiltonian, MC solves stochastically an underlying
master equation using pseudo-random numbers by constructing the probability
with which the various states of the system have to be weighted according to a
Markov process. The introduction of simulated annealing (Kirkpatrick et al.,
1983) has substantially expanded the scope of the Metropolis MC method to
problems far beyond equilibrium solutions of statistical mechanics. Specifically,
MC has been established as a powerful tool in global optimization in process
engineering, combinatorial materials library development, and reverse engineer-
ing of solid state structure determination (Deem, 2001; Vlachos, 2005). How-
ever, reverse engineering problems related to structure determination of bulk
liquids, solids, nanoparticles, and interfaces, using forward (based on a poten-
tial) and reverse (based on experimental data) modes, are outside the scope of
this paper.
MC is also successful in far from equilibrium processes encountered in the
areas of diffusion and reaction. It is precisely this class of non-equilibrium
reaction/diffusion problems that is of interest here. Chemical engineering ap-
plications of MC include crystal growth (this is probably one of the first areas
where physicists applied MC), catalysis, reaction networks, biology, etc. MC
simulations provide the stochastic solution to a time-dependent master equation
dpa X
¼ ½W ab pb W ba pa (1)
dt b
where paðbÞ is the probability that the system is in configuration aðbÞ and W ab is
the transition probability per unit time of the system going from configuration b
to a. The master equation is deterministic.
Direct solution of the master equation is impractical because of the huge
number of equations needed to describe all possible states (combinations) even
of relatively small-size systems. As one example, for a three-step linear pathway
among 100 molecules, 104 such equations are needed. As another example, in
biological simulation for the tumor suppressor p53, 211 states are estimated for
the monomer and 244 for the tetramer (Rao et al., 2002). Instead of following all
individual states, the MC method is used to follow the evolution of the system.
For chemically reacting systems in a well-mixed environment, the foundations
of stochastic simulation were laid down by Gillespie (1976, 1977). More
recently, Gillespie’s (1992) algorithm was connected with collision theory. His
approach is easily extendable to arbitrary complex computational systems when
individual events have a prescribed transition probability per unit time. This is
often referred to as the kinetic Monte Carlo (KMC) or dynamic Monte Carlo
(DMC) method, and is the tool used herein. In contrast to the classic Me-
tropolis MC algorithm, KMC provides real-time information.
There are two different exact algorithms for stochastic simulation proposed
by Gillespie, namely, the direct simulation method and the first reaction sim-
ulation method. The former is computationally much more efficient and has
been the method of choice (it requires two random numbers per MC event as
compared to Nr random numbers for the latter, where Nr is the number of
reactions). The work of Gibson and Bruck (2000) aims at reducing the com-
putational cost of KMC for complex reaction networks by modifying the first
reaction simulation method of Gillespie. Their approach uses dependency
graphs to minimize the computation time spent on updating the transition
probabilities per unit time (propensities in the terminology of Gillespie). This
idea resembles the lists of neighbors approach used in spatial distributed mo-
lecular models (Allen and Tildesley, 1989; Frenkel and Smit, 1996) and graph
theory used in building complex reaction networks (Broadbelt et al., 1994).
Furthermore, Gibson and Bruck kept the time increments of unaffected reac-
tions, determining when reactions occur, fixed to their current values. As a
result, the number of random numbers needed per MC event is reduced to just
one.
The extension of Gillespie’s algorithm to spatially distributed systems is
straightforward. A lattice is often used to represent binding sites of adsorbates,
which correspond to local minima of the PES. The work of Bortz et al. (1975)
on the n-fold or continuous time MC (CTMC) method is a significant achieve-
ment in computational speedup of the lattice KMC method, which, however,
has been underutilized probably owing to its difficult implementation. In
CTMC, probabilities are computed a priori and each event is successful, in
contrast to null-event algorithms (e.g., Metropolis) whose fraction of unsuc-
cessful (null) events increases considerably at low temperatures and for stiff
problems (Reese et al., 2001; Vlachos et al., 1993). While simulations carried out
early on reported results in MC events or steps and lacked a connection with
real time, calculation of real time by a continuous amount is straightforward, as
demonstrated several years ago (e.g., Fichthorn and Weinberg, 1991; Vlachos et
al., 1990, 1991). Real time can be implemented in both null event and CTMC
methods, and practically the same results are obtained regardless of the algo-
rithm used (Reese et al., 2001). Generalization of the KMC method to treat
arbitrarily complex surface kinetics and comparison of null event KMC and
CTMC have been presented in Reese et al. (2001).
The microscopic processes occurring in a system, along with their corre-
sponding transition probabilities per unit time, are an input to a KMC sim-
ulation. This information can be obtained via the multiscale ladder using DFT,
TST, and/or MD simulations (the choice depends mainly on whether the proc-
ess is activated or not). The creation of a database, a lookup table, or a map of
transition probabilities for use in KMC simulation emerges as a powerful
modeling approach in computational materials science and reaction arenas
(Maroudas, 2001; Raimondeau et al., 2001). This idea parallels tabulation ef-
forts in computationally intensive chemical kinetics simulations (Pope, 1997). In
turn, the KMC technique computes system averages, which are usually of in-
terest, as well as the probability density function (pdf) or higher moments, and
spatiotemporal information in a spatially distributed simulation.
IV. Classification of Multiscale Simulation Approaches
Multiscale simulation enables coupling of phenomena at various scales from

the quantum scale to the molecular, mesoscopic, device, and plant scale (Alkire
and Verhoff, 1998; Christofides, 2001; Lerou and Ng, 1996; Maroudas, 2000;
Raimondeau and Vlachos, 2002a; Villermaux, 1996; Weinan et al., 2003). For
most applications, multiscale modeling has been practiced sequentially. The
smaller (finer) scale model is typically solved first, and information is passed to
the larger (coarser) scale (upscaling), i.e., from the bottom-up. This is a one-way
information traffic paradigm (see also Fig. 1), also termed serial (Maroudas,
2003), and has been practiced successfully in several applications. One such
example includes the development of first principles chemistry via ab initio
methods, statistical mechanics, and kinetic theory. These finer length scale
models parameterize effectively the reaction rate constants that are subsequent-
ly employed in reacting process flow simulations. Laminar flames and chemical
vapor deposition (CVD) are two reaction-engineering applications where this
sequential approach has successfully been used. For gas-phase reacting flows at
low to moderate pressures, the density is so low that the probability of
trimolecular events is negligible. As a result, the PES of two chemical species
describes accurately the reaction coordinate, i.e., the coupling between scales is
practically non-existent. Therefore, one-way coupling is adequate. Another ex-
ample of one-way coupling for crystal growth of GaAs and InP was presented in
Rondanini et al. (2004), where FEM-computed fluxes were passed to a 3D
KMC code.
In most liquid- and solid-phase systems, the dilute approximation is typically
invalid, and, as a result, many body effects play a significant role. Many body
effects are manifested through the effect of solvent or catalyst on reactivity and
through concentration-dependent reaction rate parameters. Under these con-
ditions, the one-way coupling is inadequate, and fully coupled models across
scales are needed, i.e., two-way information traffic exists. This type of modeling
is the most common in chemical sciences and will be of primary interest here-
after. In recent papers the terms multiscale integration hybrid, parallel, dynamic,
and concurrent simulation have been employed, depending on the algorithm

used (Abraham et al., 1998b; Maroudas, 2003; Rudd and Broughton, 2000;
Vlachos, 1997).
Figure 4 depicts the most common multiscale simulation approaches. When
there is a large separation of length scales between phenomena, then a different
model can be used at each scale. This type of multiscale simulation is termed as
hybrid. At the core QM simulations, farther out molecular models (e.g., MD,
KMC), and even farther out continuum mechanics form an onion structure or
nested hierarchy of models (Fig. 5). At the other extreme, in processes whose
phenomena do not exhibit separation of scales, one has two options (see Fig. 4).
The first one is to extend a suitable tool, such as a MD or KMC method, to
large length and time scales to enable comparison with experiments. This is
termed mesoscopic modeling or coarse graining. The second one is to apply a
coarser model over large length and time scales of experimental interest on a
coarse grid and estimate small-scale information for the coarser model from a
finer scale model on a fine grid (Fig. 6). This last method is hereafter termed
multigrid-type hybrid multiscale simulation, but the terms heterogeneous hybrid
simulation (Weinan et al., 2003) and tooth gap have also been used (Gear et al.,
2003) to denote similar ideas.
MULTISCALE MODELING
Solution
strategies
Large
Yes No
separation
of scales? Fit coarse potentials
Solution and/or dynamics to
strategies microscopic
Same Yes predictions
Onion or nested method?
Use non-equilibrium
Iterative, steady state
stat. mech.,
or QSS
HYBRID MODELS No wavelets, etc.
Dynamic
Multigrid or
Parameterize lower heterogeneous COARSE-
scale with a reduced GRAINING OR
model and pass this
to the next scale MESOSCOPIC
model; surface MODELS
response methods
FIG. 4. Types of multiscale modeling and solution strategies. Hybrid models (one model at each
scale) apply well when there is separation of scales (onion or nested-type models). When there is lack
of separation of scales, mesoscale models need to be developed where the same technique (e.g., MD
or MC) is accelerated. Alternatively, multigrid (heterogeneous) hybrid models can be employed
where the unresolved degrees of freedom are determined from a finer scale model and passed to a
coarser scale model.
Continuum
mechanics
Finer scale Coarser scale

subdomain
Overlap
subdomain
region
Quantum
mechanics
Molecular
modeling
FIG. 5. Schematic of onion-type hybrid multiscale simulation. At each scale a different model is
used. Consecutive scale models are simultaneously solved in the overlap region where exchange of
information occurs.
MULTIGRID-TYPE HYBRID MULTISCALE MODELS

Coarse (continuum)
model grid
Fine (discrete, particle)

model grid
Concentration gradient
FIG. 6. Schematic of multigrid-type hybrid simulation with two grids. At the coarse grid a mac-
roscopic model is advanced over large length and time scales. Information is passed to the mac-
roscopic grid/coarse model from a microscopic simulation executed on a fine grid over short length
and time scales. The coarse model is advanced over macroscopic length and time scales and provides
to the microscopic simulation a field for constraint fine scale simulation.
Below, the various types of multiscale simulation are elaborated and various
examples are provided. The presentation on coarse graining is mainly focused on
stochastic (KMC) simulations to provide the underlying foundations and ideas
in some depth. Coarse graining of other atomistic, e.g., MD, and mesoscopic
tools will be covered in a forthcoming communication. Some excellent reviews
on coarse graining in soft-matter physics problems are available (e.g., Kremer
and Muller-Plathe, 2001; Muller-Plathe, 2002, 2003; Nielsen et al., 2004).
V. Hybrid Multiscale Simulation
Hybrid multiscale simulation is the most developed branch of multiscale

simulation and will be covered in this section. The onion-type hybrid simulation
is first covered, followed by a discussion of its application to crystal growth and

then to other areas. The multigrid-type hybrid method is also discussed. Finally,
some of the challenges in hybrid multiscale simulation are elaborated.
A. ONION-TYPE HYBRID MULTISCALE SIMULATIONS AND ALGORITHMS
Consider an example from nucleation and growth of thin films. At least three
length scales can be identified, namely, (a) the fluid phase where the continuum
approximation is often valid (that may not be the case in micro- and nano-
devices), (b) the intermediate scale of the fluid/film interface where a discrete,
particle model may be needed, and (c) the atomistic/QM scale of relevance to
surface processes. Surface processes may include adsorption, desorption, sur-
face reaction, and surface diffusion. Aside from the disparity of length scales,
the time scales of various processes differ dramatically, ranging from picosecond
chemistry to seconds or hours for slow growth processes (Raimondeau and
Vlachos, 2002a, b).
The vastly varying time scales and similar variations in length scales prevent
simply ‘‘brute force’’ molecular simulation of the entire process. No amount of
foreseeable advances in computational power will ever enable such a modeling
approach. For these processes, application of different models and tools at
different scales is essential, resulting in onion-type hybrid multiscale models (see
Fig. 5). For example, a hybrid model for crystal growth may consist of a CFD
model far away from the growing interface, an appropriate molecular or me-
soscopic model (e.g., MD, DSMC, or BD) in the boundary layer, and MC or
MD, with potentials parameterized using DFT, to simulate microscopic proc-
esses on the surface of a growing nanoparticle. Another example of hybrid
simulation entails, the ONIOM method of the software Gaussian (Frisch et al.,
2002). In particular, one treats quantum mechanically the core where chemical
reactivity and high activity is crucial, uses molecular mechanics farther out, and
continuum approximation (e.g., via the dielectric constant) for the solvent even
farther out.
The overall idea of hybrid simulation lies in the domain being decomposed
into subdomains and the application of a different model in each subdomain
(see Fig. 5). This approach is called domain decomposition. In order to improve
the coupling between codes, an overlapping subdomain or interfacial or hand-
shaking region is created within which both models are solved and exchange
information. The size of the interface can be adjusted to ensure proper coupling
between codes (in physics terminology it can be diffuse or sharp, i.e., of zero
thickness). In general, the interface width has to be small enough to minimize
the cost of the finer scale model but sufficiently large to allow proper relaxation
of the macroscopic information in the atomistic domain.
Three algorithms, depicted in Fig. 4, have been proposed to solve onion-type
hybrid multiscale models (Vlachos, 1997). The first applies to steady-state
problems. The solution can be obtained iteratively by passing the steady-state

information back and forth (in each iteration) between the two models until
convergence of the hybrid scheme is achieved (iterative scheme). Given the in-
herent noise of discrete particle simulations, suitable criteria are needed to ensure
steady-state convergence of a stochastic model (Vlachos, 1997). Furthermore,
owing to nonlinear phenomena, there is no guarantee of convergence (a common
observation for realistic systems) (Raimondeau and Vlachos, 2002b).
The second algorithm entails developing an approximate surface (a reduced
model) of the finer length scale model as a function of parameters of the coarser
scale model. For example, one could compute the isotherm in the case of fluid-
surface equilibrium or the reaction rates in the case of a catalytic reaction as a
function of surface temperature and partial pressures of the fluid phase. In
essence, what one does is to map or parameterize the boundary conditions of the
coarser scale model using the finer scale model. This mapping typically entails
some ensemble/spatial averaging technique that reduces the degrees of freedom
of the finer scale model to provide coarse information needed in the next model
up of the multiscale ladder. In mathematical terminology, this step can be
thought of as a restriction or contraction operator that operates onto the micro-
scopic model to provide coarse information. This reduced model is subsequently
coupled with or fed into the coarser scale model. This algorithm is suitable when
steady-state or quasi-steady-state (QSS) solutions are desired. In the latter case it
is tacitly assumed that the finer scale model relaxes fast enough for QSS to be
established. The idea of developing a reduced model (in this case a boundary
condition) using the finer scale model works well as long as the mapping is
accurate. Accuracy, however, is a non-trivial issue to satisfy (see Ludwig and
Vlachos (2004) for an example illustrating the difficulties in DFT/MD coupling).
Obviously, the above algorithms are not suitable when transients of the finer
scale model are involved (Raimondeau and Vlachos, 2000), as, for example,
during startup, shut down, or at a short time after perturbations in macroscopic
variables have occurred. The third coupling algorithm attempts fully dynamic,
simultaneous solution of the two models where one passes information back and
forth at each time step. This method is computationally more intensive, since it
involves continuous calls of the microscopic code but eliminates the need for a
priori development of accurate surfaces. As a result, it does not suffer from the
problem of accuracy as this is taken care of on-the-fly. In dynamic simulation,
one could take advantage of the fast relaxation of a finer (microscopic) model.
What the separation of time scales between finer and coarser scale models
implies is that in each (macroscopic) time step of the coarse model, one could
solve the fine scale model for short (microscopic) time intervals only and pass
the information into the coarse model. These ideas have been discussed for
model systems in Gear and Kevrekidis (2003), Vanden-Eijnden (2003), and
Weinan et al. (2003) but have not been implemented yet in realistic MC sim-
ulations. The term projective method was introduced for a specific implemen-
tation of this approach (Gear and Kevrekidis, 2003).
B. APPLICATION OF ONION-TYPE HYBRID MULTISCALE SIMULATION TO GROWTH OF

MATERIALS
In the area of nanomaterials and thin films, product ‘‘quality’’ is judged from
the sharpness of interfaces, crystallinity, defects, polymorphism, shape, uni-
formity in particle-size distribution, film texture, etc. Engineering product qual-
ity demands linking of phenomena at very different scales and has attracted
considerable interest over the last few years (Alkire and Verhoff, 1998; Christ-
ofides, 2001; Raimondeau and Vlachos, 2002a). A recent review of multiscale
simulation of CVD processes for various materials is given in Dollet (2004).
Several hybrid simulations on crystal growth can be found in recent literature.
Examples include dendritic solidification by coupling finite-different discretizat-
ion of a phase field model to a MC simulation (Plapp and Karma, 2000),
coupling a finite difference for the melt with a cellular automata for the so-
lidification (Grujicic et al., 2001), a DSMC model for the fluid phase with a
Metropolis-based MC for the surface to address cluster deposition onto subst-
rates (Hongo et al., 2002; Mizuseki et al., 2002), a step model for the surface
processes coupled with a CFD simulation of flow (Kwon and Derby, 2001) (two
continuum but different feature scale models), an adaptive FEM CVD model
coupled with a feature scale model (Merchant et al., 2000), and one-way coupled
growth models in plasma systems (Hoekstra et al., 1997). Some specific appli-
cations are discussed in more detail below.
1. Polycrystalline Films
Fabrication of polycrystalline films is an inherently multiscale problem of

substantial technological importance; as a result, several studies have been
conducted recently. For the most part, these have been serial (one-way coupling
or even fully disconnected) simulations. Gilmer and co-workers presented de-
coupled, different scale growth models of sputtering. Specifically, the level-set
method was used at the largest scale for film evolution, the level-set method
coupled with a diffusion model for dissolution of TiN clusters on a surface, and,
finally, a dual lattice KMC model for Al particle growth (see Baumann et al.
(2001) and references therein). In a similar spirit, front-tracking techniques were
employed by our group to delineate the factors controlling zeolitic film texture
fabricated from pre-grown seeds (Bonilla et al., 2001) followed by fundamental
transport/colloids or KMC-based models to elucidate single nanoparticle
growth mechanisms (Nikolakis et al., 1999, 2000, 2003). The work of Srolovitz
et al. on diamond film growth under CVD conditions is yet another example
where film texture is important and where decoupled, different type simulations
were exploited; see overview in Srolovitz et al. (1997). While these models were
applied to different materials and scales (completely decoupled), the insights
gained at different scales have been instrumental for materials design in
different applications. It is expected that future work will aim at linking these
models and phenomena over multiple scales.
2. Physical and Chemical Fluid Deposition of Thin Films (Fluid– Surface

Coupling)
In microelectronics fabrication, small-scale features (e.g., trenches) need to be
conformally covered, i.e., a spatial uniform deposit is desired. In such small
features, the continuum approximation typically breaks down when the mean
free path becomes comparable to the feature size, whereas the continuum ap-
proximation is often fine in the main body of the reactor. This separation of
geometric length scales demands a multiscale approach beyond just mesh ad-
aptation. In CVD, a ballistic growth model was used by Rodgers and Jensen
(1998) to compute an effective sticking coefficient of incoming molecules as a
function of surface topology in the case of non-continuum transport of mol-
ecules in small-scale features (large Knudsen number). This information was
passed to the boundary condition of an FEM code of a CVD reactor. Given the
disparity of time scales, QSS could be assumed at the reactor scale and iterations
could be used where information is passed between the two models to evolve the
growing interface of a trench (a moving boundary problem). This coupling
algorithm is an example of parameterization of boundary conditions method
(surface response approach; see solution strategies at the left of Fig. 4). In order
to account for the sticking probability of molecules, semi-empirical potential-
based MD have been carried out as a function of incident angle and energy, and
this information was incorporated (by suitable integration over necessary de-
grees of freedom) in a line-of-sight transport model to compute local growth
velocities (Hansen et al., 2000). This lower scale model information was then
incorporated into a level-set calculation that is ideal for moving boundary
problems arising in crystal growth or etching.
Next, coupled hybrid stochastic-continuum models are discussed. The stoc-
hastic KMC model is employed to describe surface processes of epitaxy and
account for spatial correlation effects, nucleation, and microstructure evolution.
The continuum model, on the other hand, describes the fluid mechanics and
transport phenomena at the reactor scale. One of the first such hybrid multiscale
growth simulations coupled a fluid-phase PDE model with a pseudo-3D KMC
model to study the transition between growth modes in epitaxial growth of films
(Lam and Vlachos, 2001; Vlachos, 1999). These simulations were performed in
stagnation flow geometry, where a similarity transformation reduces the 3D
fluid-phase problem into a 1D problem, with isoconcentrations being parallel to
the surface (this is an important point, discussed further below in the multigrid
methods, because it requires a single KMC-simulation box). Pseudo-3D KMC
simulations are based on the solid-on-solid (SOS) approximation and are ac-
tually 2D simulations. In particular, each point represents the film height, so the
3D surface can be described by a 2D array of heights; within the SOS
Continuum model
Boundary condition, e.g., External field,

e.g.,
Nk · n = rak – rdk – rRk
concentration,
temperature, etc.
Discrete particle
m odel, e.g ., KMC
FIG. 7. Schematic illustrating the coupling of a fluid-phase mass transfer model with a discrete,
particle model, such as KMC, through the boundary condition. The continuum model passes the
external field and the KMC simulation computes spatial and temporal rates that are needed in the
boundary condition of the continuum model.
approximation, no vacancies and overhangs can form in a film. While the SOS
assumption is not fully realistic, it is commonly employed in KMC simulations
of crystal growth.
In these hybrid simulations, coupling happened through the boundary condi-
tion. In particular, the fluid phase provided the concentration to the KMC
method to update the adsorption transition probability, and the KMC model
computed spatially averaged adsorption and desorption rates, which were sup-
plied to the boundary condition of the continuum model, as depicted in Fig. 7.
The models were solved fully coupled. Note that since surface processes relax
much faster than gas-phase ones, the QSS assumption is typically fulfilled for the
microscopic processes: one could solve for the surface evolution using the KMC
method alone, i.e., in an uncoupled manner, for a combination of fluid-phase
continuum model parameter values to develop a reduced model (see solution
strategies on the left of Fig. 4). Note again that the QSS approach does not hold
at very short (induction) times where the microscopic model evolves considerably.
These multiscale simulations linked, for the first time, macroscopic variables,
such as flow rate, substrate temperature, and composition, with microscopic
features, e.g., surface roughness of a growing film. As an example of such a link,
Fig. 8a depicts a kinetic phase diagram in the growth rate-inverse temperature
plane from simulations where various conditions were varied (Lam and
Vlachos, 2001). At relatively high temperatures and slow growth, the atoms
have sufficient time to reach steps, and step flow is the observed growth mode.
On the other hand, at relatively low temperatures and fast flow, the atoms do
not have sufficient time to reach steps, and small nuclei form between steps,
contributing to roughening. Typical snapshots of film morphologies from the
two modes are also shown in Figs. 8c and d. The solid line in Fig. 8a separates
the desirable step propagation growth mode (higher temperatures and lower
growth rates) from the 2D nucleation mode (lower temperatures and higher
growth rates).
(d)
(c)
104 102
TEG Flux (mmoles/min)

103 2D nucleation
Growth rate [MLS/s]
101
102
Step flow
100
101
(a ) (b)
100 -1
1.1 1.2 1.3 1.4 1.5 10 1.1 1.2 1.3 1.4
1000/T [K-1] 1000/T [K-1]
FIG. 8. (a) Kinetic phase diagram from hybrid, onion-type (stagnation, continuum/KMC) sim-
ulations of homoepitaxy in an atmospheric pressure reactor depicting the transition from step flow
to 2D nucleation as the temperature decreases or the growth increases. Redrawn from Lam and
Vlachos (2001). (b) Analogous experimental data of Kisker et al. (1995) for GaAs in terms of
tetraethyl–gallium flux. (c) and (d): snapshots depicting different growth modes.
There is a limited number of experiments from coupled systems to enable

comparison to experiments, in part because UHV techniques are difficult to
apply to higher pressures where coupling of fluid flow and surface processes are
important. Nonetheless, Fig. 8b shows a similar kinetic phase diagram for
growth of GaAs obtained by X-rays (Kisker et al., 1995). The similarity (e.g.,
Arrhenius type of behavior) of experimental data with the simulations is strik-
ing, but more quantitative comparison with the prototype (simple) model is
meaningless. Owing to the linking of micro-features with macroscopic variables,
multiscale simulation could be used to enable design or control of films and
nanoparticles with certain characteristics, such as a certain surface roughness,
i.e., a top-down realization of multiscale simulation depicted in Fig. 1. This issue
is further discussed in the systems tasks section.
In a similar spirit, Alkire, Braatz and co-workers developed coupled hybrid
continuum-KMC simulations to study the electrodeposition of Cu on flat sur-
faces and in trenches (Drews et al., 2003b, 2004; Pricer et al., 2002a, b). A 3D
KMC simulation accounted for the surface processes as well as diffusion in the
boundary layer next to the surface, whereas a 1D or 2D continuum model (with
adaptive mesh) was applied to simulate the boundary layer farther away. In
their initial implementation, the continuum model passed fluxes to the KMC
and the KMC passed concentrations to the continuum model. Since the length
scales simulated by the KMC method are relatively small, each cell of KMC was
assumed to extend over a certain non-atomic length scale (10–100 nm) to
enable comparison with experiments (an ad hoc coarse-grained KMC). Com-
parison of surface roughness to AFM data for Cu deposition was also done.
These studies have nicely demonstrated that linking multiscale simulation with
experimental results is definitely a reasonable short-term goal. In fact, one could
also use such experiments to parameterize transition probabilities of KMC. This
issue is revisited in the systems tasks’ section.
In coarse graining of KMC on a single monolayer entailing the same mi-
croscopic processes as in growth, the transition probabilities of various proc-
esses are scaled by different factors (Katsoulakis et al., 2003b; Katsoulakis and
Vlachos, 2003); thus, when multiple processes occur (as in the case of growth),
appropriate scaling of the various processes is necessary. Future work in coarse
graining (see section on spatially coarse-grained KMC) related to crystal growth
is essential to further extend the exciting hybrid simulations of Alkire, Braatz
and co-workers.
3. Deposition by Molecular Beam Epitaxy (Uncoupled Fluid– Surface Systems)

Another problem of hybrid multiscale simulation in crystal growth entails
coupling of KMC with a diffusion/reaction, continuum type of model to de-
scribe epitaxial growth of a film. Here, the bulk fluid phase is ignored, i.e., the
model applies to molecular beam epitaxy (MBE) conditions. Crystal growth
may occur on low index crystallographic planes, such as the (1 0 0) surface, or
vicinal surfaces, such as an (h10) plane, consisting of terraces separated by steps
(see Fig. 8c for step–terrace structure). Planes corresponding to small misorien-
tations with respect to a low index surface ðhb1Þ are composed of large terraces
separated by steps that are far apart. These surfaces are impossible to simulate
with currently available KMC simulations even in the absence of flow. Careful
examination of the aforementioned hybrid KMC/flow simulations reveals that
they have all been performed for step distances (or more generally features) that
are too small so they can be handled by a single KMC simulation box.
So how can one handle situations where the steps are far apart? One answer
lies in the pioneering work of Burton et al. (1951), also known as the BCF
model. The BCF model is a continuum PDE that describes adsorption of atoms
to and desorption from terraces along with surface diffusion on terraces [see Eq.
(2) below for a simplified version of the BCF model]. When the concentration of
adatoms is relatively large, nucleation between distant steps is most likely to
occur, because the probability of a diffusing adatom to reach steps before en-
countering another adatom is low. Under these conditions, the BCF model is
inadequate since it does not account for nucleation. Furthermore, the boundary
conditions in the BCF model ignore the discrete nature of steps and treat them
as continuum points where either partial equilibrium or Robin boundary con-

ditions are applied. Robin boundary conditions can account for the adatom
kinetics of attachment to and detachment from steps. Accumulated recent ex-
perimental work from STM and statistical thermodynamic analysis dating back
to the original BCF paper have clearly shown that steps consist of kinks and
straight ledges, and that thermal fluctuations control the structure of steps and,
thus, the velocity by which steps advance on a film. Therefore, a microscopic
resolution of the steps could be important under certain conditions.
As a first step toward overcoming the above problems, a hybrid diffusion–ad-
sorption model for the terrace linked with a KMC model near the steps was
developed (Schulze, 2004; Schulze et al., 2003). This domain decomposition stems
from a natural separation of scales. The continuum terrace model between steps is
@c
¼ Dr2 c þ F (2)
@t
where D is the surface diffusivity, c the adatom concentration, and F the ad-
sorption flux. A KMC simulation is employed near each step to provide the
boundary conditions of Eq. (2).
Note that in this specific model, desorption is neglected, and sites get regen-
erated upon adsorption, so the classic Langmuir blocking of sites is uncommon
for MBE modeling. Furthermore, the diffusion–adsorption model for the ter-
race is only approximate since interactions between molecules are not accounted
for. As a result, this hybrid model cannot handle nucleation between terraces,
and applies only to small supersaturations or high temperatures [note that for
high temperatures, one needs to include desorption in Eq. (2)] where the adatom
concentration on terraces is relatively low.
The rationale of using hybrid simulation here is that a classic diffusion–ad-
sorption type of model, Eq. (2), can efficiently handle large distances between
steps by a finite difference coarse discretization in space. As often happens in
hybrid simulations, an explicit, forward discretization in time was employed. On
the other hand, KMC can properly handle thermal fluctuations at the steps, i.e.,
provide suitable boundary conditions to the continuum model. Initial simula-
tions were done in ð1 þ 1Þ dimensions [a pseudo-2D KMC and a 1D version of
Eq. (2)] and subsequently extended to ð2 þ 1Þ dimensions [a pseudo-3D KMC
and a 2D version of Eq. (2)] (Schulze, 2004; Schulze et al., 2003). Again, the
term pseudo is used as above to imply the SOS approximation. Speedup up to a
factor of 5 was reported in comparison with KMC (Schulze, 2004), which while
important, is not as dramatic, at least for the conditions studied. As pointed out
by Schulze, one would expect improved speedup, as the separation between
steps increases while the KMC region remains relatively fixed in size. At the
same time, implementation is definitely complex because it involves swapping a
microscopic KMC cell with continuum model cells as the steps move on the
surface of a growing film.
C. APPLICATIONS OF ONION-TYPE HYBRID MULTISCALE SIMULATION TO OTHER

AREAS
There have been many hybrid multiscale simulations published recently in

other diverse areas. It appears that the first onion-type hybrid multiscale sim-
ulation that dynamically coupled a spatially distributed 2D KMC for a surface
reaction with a deterministic, continuum ODE CSTR model for the fluid phase
was presented in Vlachos et al. (1990). Extension to 2D KMC coupled with
1D PDE flow model was described in Vlachos (1997) and for complex reaction
networks studied using 2D KMC coupled with a CSTR ODEs model in
Raimondeau and Vlachos (2002a, b, 2003). Other examples from catalytic ap-
plications include Tammaro et al. (1995), Kissel-Osterrieder et al. (1998), Qin et
al. (1998), and Monine et al. (2004). For reviews, see Raimondeau and Vlachos
(2002a) on surface–fluid interactions and chemical reactions, and Li et al. (2004)
for chemical reactors.
In the area of fluids, coupling of MD near a wall with a continuum, de-
terministic description of the Navier–Stokes unidirectional flow farther away
from the wall based on overlapping subdomains of domain decomposition was
presented in O’Connell and Thompson (1995). A nice description of ensuring
continuity of momentum flux was given and the velocity field was made con-
sistent across the interface by using constraint dynamics in MD. See also Nie
et al. (2003) for coupling of MD with a continuum model of flow, and
Hadjiconstantinou and Patera (1997), where MD was again coupled with a
continuum description of the incompressible Navier–Stokes solved using a
spectral element method and the Schwart alternating method with overlapping
subdomains. While the work of Hadjiconstantinou and Patera was applied to
steady-state problems and invoked an iterative scheme to reach convergence
(see solution strategies in Fig. 4), the separation of time scales between mi-
croscopic and continuum models was emphasized as a means of reducing the
computational burden of hybrid schemes. Another example of MD/CFD for a
tethered polymer on a surface in share flow was recently studied (Barsky et al.,
2004) and shown to be in very good agreement with MD simulations, with
significant reduction in CPU. Coupling of continuum mesoscopic or stochastic
models near the Earth’s surface with a fluid model has successfully been ap-
plied to tropical convection in order to study the effect of fluctuations from
unresolved degrees of freedom of fine scales on climatology (Khouider et al.,
2003; Majda and Khouider, 2002).
The materials community has made significant advances in predicting me-
chanical properties of materials and initiation of defects using hybrid multiscale
simulation. This is one of the application areas where multiscale simulation has
advanced the most. Several nice reviews and perspectives have already been
published (Maroudas, 2000, 2003; Miller and Tadmor, 2002; Rudd and
Broughton, 2000). Therefore, it suffices to give only a brief account of the
evolution of multiscale simulation in this area here. One of the earlier and
successful multiscale approaches in solid mechanics is the quasi-continuum

method of Tadmor et al. (1996), Philips (1998), Shenoy et al. (1999), and Miller
and Tadmor (2002), where an adaptive finite element mesh that is refined to
atomistic dimensions at interfaces is used. The energy of each cell is computed
from the underlying Hamiltonian from a single ‘‘representative’’ atom that is
embedded in the cell and subject to the deformation field of the cell. Subse-
quently, the equilibrium configuration at 0 K is determined from an energy
minimization of the total energy of all cells to provide the deformation field. It is
worth noting that the embedding process of the quasi-continuum method (single
atom resolving the energy of a coarse FEM cell) has a stronger parallel to the
multigrid-type hybrid simulation discussed in the next section than the onion-
type simulation, at least in the overlapping regime.
The quasi-continuum approach has been successful in static problems but
limited to equilibrium situations and 0 K. Its extension to dynamic problems has
not been easy, as revealed by subsequent works (Abraham et al., 1998a;
Broughton et al., 1999; Cai et al., 2000; Maroudas, 2000; Weinan and Huang,
2002). The simulations by Abraham et al. (1998a) are one of the first to con-
currently couple quantum mechanics at the core of a dislocation, MD to capture
the atomic motion near the core, and finite elements of continuum elasticity
farther out to simulate defect formation and propagation in materials. In a
similar spirit, application of onion-type hybrid multiscale simulation to oxida-
tion of Si has also been reported (Nakano et al., 2001; Ogata et al., 2001).
Coupling of atomistic MD and continuum FEM models in the overlapping
region can be accomplished by refining the mesh of FEM to atomistic sizes.
However, materials simulations have revealed that this approach causes prob-
lems in some cases. Coarse graining of MD to large scales (Rudd and Brough-
ton, 1998), or combination of FEM refining and MD coarse graining are other
options that may in fact be superior (for a review see Rudd and Broughton
(2000)). The issues of proper coupling in the overlapping subdomain, along with
additional challenges of hybrid simulations, are discussed in detail in section F.
D. MULTIGRID-TYPE HYBRID MULTISCALE SIMULATIONS
The above problems exhibit phenomena with well-separated length scales,

where the coupling between the continuum and discrete models happens at an
interface (the overlapping or handshaking regime). In most published work on
discrete particle/flow distributed systems, the external field (e.g., the concentra-
tion profile) of the continuum model parallel to a surface is either uniform, such
as in an ideal, infinite-size stagnation flow, or exhibits nanometer-scale in-
homogeneities (smaller than the KMC simulation box size). In this situation, a
single discrete particle simulation is adequate to resolve the spatial correlations,
and one could couple it with the deterministic, continuum model.
There is another very important class of problems where no well-defined sep-

aration of length scales and spatial gradients exist over large length scales. There
are numerous examples of such problems. Flow along a long tube is one where
gradients in pressure and velocity fields occur. Growth on a large wafer is another
where flow, concentration, and temperature non-uniformities across the substrate
exist. Nucleation and growth of materials within a thick substrate in the coun-
tercurrent diffusion-reaction configuration (Gummalla et al., 2004) and diffusion
through realistically thick microporous films used for separation or membrane
reactors (Chatterjee et al., 2004a) are two more. These problems exhibit mac-
roscopic gradients (over millimeters to inches), which are beyond the realm of
conventional discrete particle models. Furthermore, with a few exceptions, mainly
in our group (Lam et al., 2001; Snyder et al., 2003; Vlachos and Katsoulakis,
2000), KMC simulations have been limited to situations where the external field
(e.g., pressure) is uniform, and as a result, they were carried out under periodic
boundary conditions. On the other hand, non-equilibrium MD simulations under
a gradient in chemical potential have already been introduced (Cracknell et al.,
1995; Fritzsche et al., 1995; Heffelfinger and van Swol, 1994; MacElroy, 1994;
MacElroy and Suh, 1997; Maginn et al., 1993; Sunderrajan et al., 1996; Xu et al.,
1998). While microscopic models under gradients are now available, they cannot
cope with the large length and time scales of realistic systems.
Recently, there has been strong interest in multigrid-type hybrid multiscale
simulation. As depicted in Fig. 6, a coarse mesh is employed to advance the
macroscopic, continuum variable over macroscopic length and time scales. At
each node of the coarse mesh, a microscopic simulation is performed on a finer
mesh in a simulation box that is much smaller than the coarse mesh disc-
retization size. The microscopic simulation information is averaged (model re-
duction or restriction or contraction) to provide information to the coarser
mesh by interpolation. On the other hand, the coarse mesh determines the
macroscopic variable evolution that can be imposed as a constraint on micro-
scopic simulations. Passing of information between the two meshes enables
dynamic coupling.
The computational advantages of such multigrid methods arise from two key
factors. First, microscopic simulations are carried out over microscopic length
scales instead of the entire domain. For example, if the size of fine grid is 1% of
the coarse grid in each dimension, the computational cost of the hybrid scheme
is reduced by 102d, compared with a microscopic simulation over the entire
domain, where d is the dimensionality of the problem. Second, since relaxation
of the microscopic model is very fast, QSS can be applied at the microscopic
grid while the entire system evolves over macroscopic time scales. In other
words, one needs to perform a microscopic simulation at each macroscopic
node for a much shorter time than the macroscopic time increment, as was the
case for the onion-type hybrid models as well.
The multigrid branch of multiscale simulation is less developed. To my
knowledge, Tammaro and Evans were the first to introduce such multigrid-type
hybrid multiscale simulations for the example of a traveling wave in a catalytic

reaction (Tammaro et al., 1995). In their example, species A diffuses very
quickly, whereas species B diffuses slowly. To cope with the large length scales
and separation of time scales, they advanced over the entire interface of the
traveling wave species A, using the continuity equation based on a finite dif-
ference coarse grid. At each node of the finite difference grid, they carried out a
KMC simulation in which species A was randomly distributed (this approach
copes with the huge disparity in time scales between diffusion of species A and B
and is another type of hybrid simulation) and species B was treated by KMC.
Information was passed back and forth between the two models at the two
grids.
Interesting results from coupling MD with continuum equations using the
multigrid-type hybrid approach were presented by Weinan et al. (2003) for
dislocation dynamics and crack propagation. The method was termed hetero-
geneous multiscale method and is conceptually in the same spirit as the work of
Tammaro et al. (1995). The tooth-gap method is a related technique (Gear et al.,
2003) to deal with these problems. In a different context of fluid flow simu-
lations, coarse levels were modeled with continuum fluid mechanics and fine
levels with discrete particle simulations (the DSMC method) (Garcia et al.,
1999). Multigrid ideas to resolve small-scale information and pass it into large-
scale models for climate predictions have also been discussed by Majda and co-
workers (Majda and Khouider, 2002).
E. AN EXAMPLE OF MULTIGRID-TYPE HYBRID MULTISCALE SIMULATION FOR

GROWTH UNDER LARGE LENGTH SCALE GRADIENTS
An example of the aforementioned multigrid-type hybrid multiscale simula-

tion from crystal growth for simulating nucleation and growth in large length
scale systems is provided following Gummalla et al. (2004). Nucleation and
growth are distributed in space and occur often in relatively localized areas, but
the time and place where this happens is stochastic, i.e., the multiple grids have
to be built as a simulation progresses and remeshing may be necessary as time
evolves. The specific system refers to Pd deposition under CO2 supercritical
conditions within an alumina disk 1 mm thick. Hydrogen and the organome-
tallic precursor are introduced from opposite sides of a countercurrent geometry
to react, leaving behind Pd, as depicted in Fig. 9a. Experimentally, a challenge is
to confine the chemistry within the substrate in such a way that a thin but
continuum Pd film forms, which can be used for hydrogen separation. In this
system, gradients in concentrations of species develop over the entire domain
owing to diffusion and chemical reactions. On the other hand, nucleation occurs
at random locations and times but is limited to the nanometer scale. Nucleation
can be thought of as a noisy term within the governing PDEs, whose closed
form is unknown, rather than in the boundary condition.
Level 1
*
Mesh refinement
Level 2 10 µm
x x x x x x x x x x x x x x x x x x x x x xx x x
Reactant A Reactant B Level 3 1 µm
xoooooooooooooooooooooooo
x
0.1 µm
Level 4
o xo o o
(a) (b) 1 nm
0.25 KHet= 102
KHet = 102 12.0 (cm3 /gmol)0.5/s
Pd density (gm/cm3)
0.20 (cm3/mol)0.5/s
Nucleation rate
= 10 9.0 = 10
0.15 =1 =1
6.0
0.10
Growth rate constant 3.0
0.05
0.0
0.00 3.5 4.0 4.5 5.0 5.5 6.0 0.48 0.52 0.56
Time (min) Depth within Al2O3 (mm)
(c) (d)
FIG. 9. (a) Schematic of opposed flow geometry. (b) Schematic of four-level refinement starting
from coarser (top) and moving toward finer (bottom) meshes. The continuum diffusion–reaction
equations are solved at the top three (coarser) grids with a possible consumption (sink) term de-
termined from the fourth level. At the fourth (finest) level stochastic treatment of nucleation and
front tracking are used to determine whether and where nucleation occurs and the consumption term
of the nucleation precursor fed back to the third level. Effect of growth rate constant on nucleation
rate (fraction of sites on finest grid on which nucleation occurs, per unit time) (c) and density of
deposit when the alumina pores are plugged (d). Fast growth decreases nucleation and leads to films
with less density variation at pore plugging and less Pd used. For these conditions, the Pd film forms
in the middle of the Al2O3 substrate that is 1 mm thick. The effect of fluctuations on deposit density
is apparent in (d). Data redrawn from Gummalla et al. (2004).
This disparity in length scales on the one hand and the stochastic nature of
nucleation on the other underscore the multiscale nature of the problem. The
governing PDEs describing the concentrations of reagents determine the prob-
ability of nucleation and must be solved over large length scales that are far
beyond the realm of microscopic KMC. To overcome the disparity of length
scales, an adaptive mesh refinement strategy has been used with four levels that
enable linking macroscopic scales to the nanometer, as shown schematically in
Fig. 9b (note that these multiple grids differ from the schematic of Fig. 6 to
better fit the problem at hand). One question is: Where and when does one
decide to refine the mesh? This is actually done probabilistically. Since
nucleation has a higher probability of occurrence wherever concentrations are

high, the mesh is refined when and where the probability for nucleation is above
a certain low threshold. The chosen region for mesh refinement typically in-
volves high concentrations of nucleation precursors. Nuclei can also form in
regimes where the probability of nucleation is low, but in our experience, this
does not lead to growth but to a few isolated nucleation events. The use of a
threshold eliminates spurious mesh refinement in ‘‘wrong’’ regions.
The multigrid, hybrid multiscale approach entails solving the continuum
governing (diffusion–reaction) equations in porous media at the three coarser
meshes and a stochastic treatment of nucleation at the finest mesh. In the finest
mesh, a KMC simulation could be employed and linked to a front tracking
technique to follow the evolution of growing clusters. Upon significant growth
of clusters (cluster size4mesh size of level 3), growth could be handled from the
next coarser mesh. In order to accelerate the hybrid scheme, an exponential
distribution was used instead of an actual KMC. KMC simulations in a well-
mixed batch reactor have been compared with the hybrid approach, and good
agreement was found (Gummalla et al., 2004). Thus, at each location of the
finest mesh the probability for nucleation per unit time, Po, which is propor-
tional to the nucleation rate, is computed. The probability for a nucleation event
in a time tnuc after the creation of a previous nucleus is assumed to be
Pðtnuc Þ ¼ 1 exp½tnuc Po (3)
As time evolves by Dt, the continuum model at the third mesh provides con-
centrations that affect the nucleation rate of the stochastic model via Eq. (3). At
every time step, Pðtnuc Þ is computed and compared with a random number
between 0 and 1. When the random number is larger than Pðtnuc Þ, tnuc increases
by Dt, whereas when the random number is less than Pðtnuc Þ, a new nucleus is
seeded and tnuc is set to zero. Nucleation and growth, when occurring at the
finest mesh, consume nucleation precursors, whose rate of consumption is
passed to coarse grids. These hybrid multiscale simulations can provide insights
into the roles of nucleation and growth kinetics in microstructure, defects, film
continuity, etc. that can be directly compared with experiments. An example is
depicted in Figs. 9c and d.
These hybrid approaches have a lot of potential for treating nucleation stoc-
hastically while enabling simulations on large domains. Simulations in higher
dimensionalities and of self-organization phenomena (e.g., Lebedeva et al.,
2004a, b) using multigrid hybrid multiscale models are definitely desirable.
F. CHALLENGES IN HYBRID MULTISCALE SIMULATIONS
The major issue in hybrid multiscale simulation is ‘‘patching’’ of models used

in different subdomains (Nie et al., 2003; Raimondeau and Vlachos, 2002a). In
brief, coupling may lack convergence (Raimondeau and Vlachos, 2002b)

(especially in an iterative scheme) and could result in spurious solutions (Reich,
1999) and violation of conservation laws. One of the best expositions of patch-
ing problems can be found in Weinan and Huang (2002) for problems related to
dislocations, friction, and crack propagation. Matching conditions were devel-
oped between atomistic (MD) and continuum regions to minimize reflection of
phonons of MD at the MD/continuum model interface. While the authors were
successful, they noted that at higher temperatures and in nonlinear situations,
overheating may occur. General solutions to patching in most applications are
still needed. These issues are elaborated below, and recent progress made in
overcoming them is outlined by focusing on crystal growth problems.
A frequent problem in hybrid multiscale simulations is noise-induced nu-
merical instability (Raimondeau and Vlachos, 2002a, b; Rusli et al., 2004). Such
instabilities may occur when the time step of the KMC becomes too large to
violate the numerical stability criterion of the continuum model, or when rare
events happen that create huge variations in the boundary condition or in the
source/sink term of the continuum model. This numerical instability is a result
of the small size of the KMC simulation box (a problem stemming from our
inability to deal with realistically large length scales). Consequently, the KMC
response is considerably noisier than what one would have for realistic length
scales. In order to reduce the noise of KMC passed to the continuum model, in
Vlachos (1999) and Lam and Vlachos (2001) the KMC simulation was run for a
certain number of events before the gas-phase model was solved. This is jus-
tifiable, given that the time step of KMC is typically much smaller than that of
the gas-phase model, i.e., surface processes have a much shorter relaxation time.
Thus, the numerical strategy followed is spatial and temporal averaging in
KMC to compute rates with reduced noise (variance reduction) prior to passing
them to the continuum model. The number of MC events used in temporal
averaging was varied to ensure that the results were unaffected. In cases where
the time scales of the KMC and the continuum model are comparable, one
could use parallel processing by running multiple images of the KMC to create
microscopic-model based rates with reduced noise. Similar problems were
also reported by Drews et al. They used a filtering approach to reduce the noise
in hybrid simulations for improved code robustness (Drews et al., 2004;
Rusli et al., 2004). Note that temporal averaging in a discrete particle model
has the advantage of minimizing the number of continuum model calls; as a
result, it leads to a speedup of a hybrid scheme. System level tasks, such
as filtering, arising from the controls community, was also employed in the
work of Lou and Christofides (2003a, b, 2004) (see also corresponding section
below).
The exposition in Schulze’s (2004) recent paper underscores in an excellent
manner some additional difficulties encountered in hybrid multiscale simulation
(not just of crystal growth problems) when overlapping subdomains are used.
The replacement of KMC on terraces with the continuum model Eq. (2) reduces
the noise of the hybrid scheme compared with the microscopic KMC model,
and alters phenomena controlled by noise such as the time scale for bunching of
steps (a common instability in crystal growth). It was reported that while
bunching of steps occurs under the same conditions as in KMC, the dynamics of
the processes was altered. This is obviously an undesirable situation and results
from over coarse graining the microscopic processes on a terrace, i.e., from
replacing the microscopic KMC method with a continuum model.
The unintentional noise-induced numerical instability in coupled fluid-KMC
codes and the reduced noise in the growth on a terrace model underpin just one
of the problems of hybrid multiscale simulation that stem from the incorrect
(over- or under coarse graining of) noise. We expect that coupling of continuum
models with the recently introduced coarse-grained KMC (CG-KMC) simula-
tions, discussed below, will improve or eliminate this noise-induced numerical
instability. This improvement is expected because the much larger length scales
simulated via CG-KMC will result in (correctly) less noisy signals than those
produced by microscopic KMC simulation. On the other hand, use of
the adaptive coarse-grained-KMC (ACG-KMC) method, also touched upon
below, could completely eliminate the need for hybrid simulation for surface
processes, such as the terrace-step model of Schulze, and overcome the reduc-
tion in noise that in turn affects nucleation. Further work is needed to exploit
these ideas.
There is another subtle but fundamental issue in coupling of hybrid models
that has to do with differences in constitutive relations in various subdomains.
In particular, models at various scales correspond (upon passing to the con-
tinuum limit) to different constitutive relations. For example, in the continuum
model on a terrace, Eq. (2), there are no interactions between molecules. Con-
sequently, Fick’s first law
j ¼ Drc (4)
describes the system adequately. On the other hand, within the KMC subdo-
main, interactions between molecules result in a different underlying mesoscopic
transport equation and constitutive relation, i.e., Fick’s first law does not hold.
We have found out that the specifics of mesoscopic equations and constitutive
relations depend on the microscopic mechanisms of diffusion. For example,
when the activation energy depends only on the energy of the departing site, the
corresponding continuum model (termed Arrhenius dynamics) for the problem
of growth, based on Vlachos and Katsoulakis (2000), is
@c
¼ DrfebJc ½rc bcð1 cÞrJ cg þ F (5)
@t
where D ¼ Do ebU o is the diffusion coefficient, Do the diffusion coefficient at

high (infinite) temperature, J the intermolecular potential of adatom–adatom
R
interactions, and J c ¼ Jðjr r0 jÞcðr0 Þ dr0 a convolution. In this case, the
constitutive relation reads
j ¼ DfebJc ½rc bcð1 cÞrJ cg (6)
As another example, when the activation energy for diffusion depends on the
energy difference between the initial and final locations (termed Metropolis
dynamics), the corresponding continuum model for growth reads
@c
¼ rDo f½rc bcð1 cÞrJ cg þ F (7)
@t
and the constitutive relation is
j ¼ Do f½rc bcð1 cÞrJ c (8)
Equations (6) and (8) reduce to Eq. (4) only when the intermolecular potential J
is zero. These are the proper constitutive relations if the microscopic mecha-
nisms of diffusion are the assumed ones.
What are the implications of different constitutive relations in different re-
gimes? In brief, conservation laws are not that easy to satisfy. For example, in
the presence of interactions, matching the concentration profiles in the over-
lapping regime (a common strategy in domain decomposition) is inadequate
since continuity in concentration and its gradient does not ensure the same flux
at the interface. Matching of chemical potentials is potentially a more rigorous
approach, but different mobility terms do not guarantee continuity in fluxes
across the overlapping region. Furthermore, this is a difficult task to accomplish
because constitutive equations, such as the ones written above, do not exist for
most microscopic models. Matching of fluxes at the interface leads at least to
conservation, but further work is needed to fully understand this point.
Another issue in hybrid multiscale simulation pertains to possible mass con-
servation caused by truncation errors. In particular, mapping discrete molecules
into continuum quantities, e.g., updating the concentration, is easy. However,
the reverse task of mapping continuum changes of concentrations into an in-
teger number of molecules along with their spatial placement is also important
(see Schulze (2004)) for some interesting ideas and a coupling factor that is
iteratively determined to match fluxes).
The discussion above focused on onion-type hybrid multiscale simulation.
Finally, even though there are a limited number of examples published, I expect
that the multigrid-type hybrid simulations share the same problems with onion-
type hybrid multiscale models. In addition, appropriate boundary conditions
for the microscopic grid model need to be developed to increase the accuracy
and robustness of the hybrid scheme. Furthermore, the inverse problem of
mapping coarse-grid information into a microscopic grid is ill posed. Thus, it is
not currently clear what the best way of reconstructing the information on the
fine grid is. Future work will elucidate these issues.
VI. Coarse Graining of Stochastic Models
Hybrid multiscale simulation is currently by far the main multiscale compu-

tational toolkit under development. However, as discussed in the last section,
many problems lack separation of scales, and since a molecular model cannot be
applied to the entire process, coarse graining (upscaling) of molecular models is
an appealing approach, leading to mesoscopic models that can reach larger
length and time scales. These coarse grained or mesoscopic models could be
used as stand-alone models (see examples below) or in hybrid multiscale sim-
ulators (see Fig. 4), e.g., a coarse grained surface simulator is linked with a fluid-
phase model, as in the work of Pricer et al. (2002a, b) and Drews et al. (2003b,
2004). One advantage of stand-alone coarse-grained models over multigrid-type
hybrid simulations is that one does not have to interface multiple models; thus,
one avoids the challenges mentioned in the previous section. Another is that it is
possible to retain the correct noise and thus overcome either numerical insta-
bilities or the alteration of the physics (see discussion above on challenges in
hybrid simulation, and below for the effect of coarse graining on noise).
Next time acceleration is first discussed, followed by space acceleration, and
finally by space-time acceleration of KMC methods. Similar developments are
under way for MD, but this subject is left for a future communication.
A. TEMPORAL UPSCALING OF KMC SIMULATION IN WELL-MIXED SYSTEMS
Separation of time scales is the rule rather than the exception in chemical
kinetics, irrespective of deterministic or stochastic modeling. The disparity of
time scales is easily rationalized by the considerable difference in activation
energies and the strong dependence of reaction rates on activation energies via
the Boltzmann factor. The stiffness of deterministic ODEs is now easily handled
owing to the machinery of implicit, adaptive time step, variable-order solvers.
However, extensions to stochastic systems are far behind. Until recently, KMC
simulations could not deal with separation in time scales. In a conventional
KMC simulation, fast processes with large transition probabilities are fre-
quently sampled, resulting in small simulated times, whereas slow events are
rare and are poorly sampled during a simulation.
Recently, several approaches have been proposed to overcome the disparity
of time scales for certain classes of problems. In order to overcome the problem
of stiffness caused by rapid, partial equilibrated reactions in a living free-radical
polymerization system, a hybrid analytical-KMC method was suggested (He
et al., 1997). In particular, the partial equilibrium (PE) was enforced to elim-
inate the fast processes by adjusting deterministically the concentrations of
species involved in PE, whereas the slow reaction events were treated stochas-
tically. A problem with this technique is that when the separation of time scales
is moderate, PE is not as accurate. Furthermore, PE applies only after some
induction time. Finally, PE requires conversion of real numbers into integers,
and while this can be done so that mass is conserved, it is not clear what the
errors are.
Resat et al. generalized the above idea and implemented a weighted-prob-
ability KMC method (WP-KMC) to overcome the separation of time scales of
stochastic simulation (Resat et al., 2001). The idea of probability weighting
stems from equilibrium MC umbrella sampling simulations introduced in Torrie
and Valleau (1977). The slow reactions determine the long-term dynamics of
system evolution. In WP-KMC, during each slow reaction event, several events
of fast reactions are simultaneously executed, i.e., one moves a number of mol-
ecules (bundles) rather than moving one molecule per time. The rationale for
this method is that over the time scale of slow reactions (rare events), the
transition probabilities of fast reactions and the concentrations of major reac-
tants (large populations) vary slowly. As a result, one may assume that they do
not change as much, and consequently execute a number of events simultane-
ously. A problem with this approach is that the weighting of probabilities am-
plifies the noise, a physically unrealistic situation (see Fig. 10).
The net-event KMC (NE-KMC) or lumping approach has been introduced
by our group. The essence of the technique is that fast reversible events
are lumped into an event with a rate equal to the net, i.e., the difference
between forward and backward transition probabilities per unit time (Vlachos,
1998). The NE-KMC technique has recently been extended to spatially distrib-
uted systems (Snyder et al., 2005), and it was shown that savings are propor-
tional to the separation of time scales between slow and fast events. The method
is applicable to complex systems, and is robust and easy to implement.
Furthermore, the method is self-adjusted, i.e., it behaves like a conven-
tional KMC when there is no separation of time scales or at short times,
and gradually switches to using the net-event construct, resulting in accelera-
tion, only as PE is approached. A disadvantage of the method is that the noise
is reduced.
A comparison of the WP-KMC, NE-KMC, and conventional KMC is shown
in Fig. 10. These acceleration approaches are successful regarding CPU. How-
ever, since the objective is often to study the role of noise, they do not provide
the correct fluctuations. In a similar vein, use of simple rate expressions, such as
the Michaelis–Menten or Hill kinetics, derived via PE and QSS approximations,
are capable of accelerating KMC simulation since fast processes are eliminated.
However, the noise of the resulting simulation, based on a reduced rate ex-
pression that lumps some of the reaction steps, is usually adversely affected
(Bundschuh et al., 2003).
WP-KMC
10
Population size of species C
Conventional KMC
NE-KMC
0 50 100 150 200 250

Time [s]
FIG. 10. Number of molecules of species C vs. time from two time acceleration algorithms and the
microscopic KMC method. The reaction network studied in a constant volume batch reactor is
k1 k2
A þ B Ð C and A þ C Ð D. For the set of conditions picked, a fivefold speedup is obtained using
k1 k2
the WP-KMC. However, the amplification in noise is apparent. On the other hand, the NE-KMC
gives comparable noise at the beginning and speeds up as equilibrium is approached with the same
or better speedup than WP-KMC. The curves have been displaced from each other for better
visualization. The rate constants are k1 ¼ 105 molecules1 s1, k1 ¼ 1 s1 , k2 ¼ 103 mole-
cules1 s1, and k2 ¼ 0:1 and the initial distribution of molecules is NA(t ¼ 0) ¼ 10,000,
N B ðt ¼ 0Þ ¼ 2500, and N C ðt ¼ 0Þ ¼ N D ðt ¼ 0Þ ¼ 0.
Recently, Gillespie (2001) introduced an approximate approach, termed the

t-leap method, for solving stochastic models. The main idea is the same as in the
WP-KMC method. One selects a time increment t that is larger than the mi-
croscopic KMC time increment, and multiple molecular bundles of fast events
occur. However, one now samples how many times each reaction will be
executed from a Poisson rather than a uniform random number distribution.
Prototype examples indicate that the t-leap method provides comparable noise
with the microscopic KMC when the leap condition is satisfied, i.e., the time
increments are such that the populations do not change significantly between
time steps.
Gillespie’s recent work on the t-leap method is a significant advance in ac-
celerating KMC simulation with respect to time constraints. However, some
issues need to be resolved before the method becomes widely used. First, dis-
parity in time scales caused by reaction rate constants rather than concentra-
tions may not be as easy to handle. Second, negative concentrations result with
probability one, i.e., if one runs long enough, since the Poisson distribution is
unbounded and a molecular bundle can be larger than the actual population of
a species. This situation becomes common as the size of the molecular bundle,
and thus the time step, increases. Additional problems are that large jumps in
time can cause incorrect behavior even if the concentrations are non-negative,
and that the magnitude of the noise is increased for substantial coarse graining
in time increments.
In order to overcome the problem of negative concentrations, two versions of
the binomial t-leap method were recently introduced (Chatterjee et al., 2005d;
Tian and Burrage, 2004). While the essence of the techniques is the same, the
method of Tian and Burrage (2004) appears to be limited to reaction networks
whose species are not shared by multiple chemical reactions. The elimination of
negative concentrations enables substantial acceleration of stochastic simulation
of complex biological networks (Chatterjee et al., 2005b). It has also been shown
analytically and numerically that the binomial t-leap method gives a better
approximation of the noise in comparison with the original Poisson-based
t-leap method of Gillespie (Chatterjee et al., 2005d).
The initial criterion proposed to ensure accuracy and avoid negative con-
centrations in simulations of typical length required a small change in the pro-
pensity functions. While an improved criterion was subsequently proposed
(Gillespie and Petzold, 2003), improved and additional criteria should be de-
veloped. Finally, calculation of Poisson random numbers required by the
method is more expensive. We will illustrate some of these issues below in the
context of spatiotemporal CG-KMC. The t-leap method has further been ex-
tended by Petzold, Gillespie, and co-workers (Rathinam et al., 2003) by em-
ploying implicit solvers that could potentially further increase the time step
increments and overcome the problem of stiffness of stochastic systems. With
the implicit t-leap the evolution is captured more accurately for large jumps in
time, even though the noise is now actually reduced. Stability criteria for the
various t-leap methods were recently developed and the variation of noise be-
tween various methods was rationalized (Cao et al., 2004). A trapezoidal t-leap
method was found to provide better noise characteristics. The t-leap method is
revisited in the section on spatial CG-KMC methods.
Noteworthy are some alternative approaches that address the issue of sep-
aration of time scales by starting with the master equation. Rao and Arkin
(2003) have employed the QSS assumption in stochastic simulation, expanding
on ‘‘adiabatic elimination’’ ideas of fast variables from the master equation
discussed in Janssen (1989a, b) and Vlad and Pop (1989). Haseltine and Raw-
lings portioned events into slow and fast (instead of treating species as done in
the work of Rao and Arkin), and treated fast reactions either deterministically
or with Langevin equations, and slow reactions as stochastic events (Haseltine
and Rawlings, 2002). This hybrid type of modeling builds upon the hierarchy of
models depicted in Fig. 3b and is further discussed below in the biological
networks section.
It is clear that this is an exploding branch of multiscale simulation. While

significant progress has already been made, different methods pose different
advantages and disadvantages. The main difficulty with most techniques is their
inability to preserve the noise. In this regard, the t-leap method and its deriv-
atives are promising. I expect more work to be devoted to this rapidly growing
branch of multiscale simulation along with many applications from various
areas. While simple reaction networks have been treated with some success, I
believe that there is a clear need to develop a robust, generic methodology that
overcomes the problem of stiffness of complex reaction networks while preserv-
ing the noise that can be important in some applications.
B. SPATIAL UPSCALING OF DISTRIBUTED (LATTICE) KMC SIMULATION
The problem of coarse graining in space is also very important but has re-
ceived less attention. The overall idea of coarse graining degrees of freedom to
move up in scales comes originally from renormalization group theory. An
interesting idea revolves around coarse graining of the Hamiltonian using
wavelets. This idea has been applied successfully to study critical behavior of
prototype fluids (Ismail et al., 2003a, b) and is being currently extended to
complex polymeric systems (Ismail et al., 2005a, b). Coarse graining of the
Hamiltonian was also presented by Ishikawa and Ogawa (2002), but it can be
shown that the proposed expression does not obey detailed balance.
Recently, the mathematical foundations for spatial CG-KMC have been
introduced for grand canonical and canonical ensemble simulations of
Ising-type systems in Katsoulakis et al. (2003a, b) and Katsoulakis and Vlachos
(2003). This work deserves a review of its own. However, in order to put it
in context with the other multiscale developments, some exciting developments
are briefly summarized. The essence of the method is the creation of a lattice
of coarse cells, each consisting of several microscopic cells. Within a coarse
cell, the local mean field is assumed (a closure at the stochastic level). In this
way, some information (degrees of freedom) is lost during coarse graining.
The potential of interactions, the Hamiltonian, and the transition probabilities
are all coarse-grained using wavelets for projecting the energetics and by en-
suring that the microscopic and macroscopic limits are correctly captured (this
is an essential attribute for the success of the method). Simulations have
demonstrated that when the intermolecular potential is relatively long, CG-
KMC gives results in very close agreement with microscopic KMC in terms of
dynamics and equilibrium states, while retaining the noise and reducing the
CPU by many orders of magnitude. Thus, CG-KMC is an ideal tool for reach-
ing large length scales.
While CG-KMC can reach large scales at reasonable computational cost,
it can lead to substantial errors at boundaries and interfaces where large
gradients exist, and the local mean field assumption is not as accurate. Recent
work has extended coarse graining to adaptive meshes (Chatterjee et al., 2005a,
2004b), in a similar spirit to well-established discretization methods of PDEs.
This method is termed ACG-KMC, and can considerably improve accuracy
with similar or improved computational savings compared with the uniform
mesh CG-KMC simulation. Analytical error estimates of information loss
during coarse graining from finer to coarser scales can be used to design
optimum meshes that ensure high accuracy with minimal computational cost
(Chatterjee et al., 2005a, c).
Next, an example of CG-KMC from pattern formation on surfaces is pre-
sented. Another application to relatively thick membranes was given in Snyder
et al. (2004). In the example considered here, atoms adsorb from a fluid res-
ervoir on a flat surface. Subsequently, they may desorb back to the fluid, diffuse
on the surface, or be annihilated by a first-order surface reaction, as shown in
Fig. 11a. Attractive interactions between atoms trigger a phase transition from a
dilute phase (a low coverage) to a dense phase (a high coverage) (Vlachos et al.,
1991), analogous to van der Waals loops of fluid–vapor coexistence. Surface
reactions limit the extent of phase separation; the competition between micro-
phase separation and reaction leads to nanoscopic patterns by self-organization
under certain conditions (Hildebrand et al., 1998).
A major challenge in simulating such problems is that nucleation occurs at
the nanometer scale whereas self-organization entails competition between nu-
merous pattern blocks for reagents over microns to millimeters. These problems
do not exhibit an obvious separation of length scales. From a different point of
view, the stochasticity is built within the PDE as a source or sink term (if one
were able to write such a PDE). Furthermore, surface diffusion is faster than the
other microscopic processes by many orders of magnitude, but PE cannot be
applied since the actual value of diffusion dictates the presence or absence of
patterns.
Grand canonical
ensemble
Canonical ensemble Time
Surface reaction
(a) (b) Space
FIG. 11. (a) Schematic of microscopic processes for fluid–surface interacting systems. (b) Spa-
tiotemporal evolution of 1D concentration patterns (coarse graining of two sites into each coarse cell
is used). Bifurcation splittings and mergings occur as time evolves. The fast diffusion necessary for
pattern formation (five to six orders of magnitude faster than the rest of the processes) renders
microscopic KMC unsuitable even for small domains.
Instead of using multigrid-type hybrid multiscale simulation discussed above,

the CG-KMC method that retains the noise is employed, and it can thus cor-
rectly capture the effect of fluctuations on nucleation and pattern evolution.
Figure 11b shows an example of such a 1D simulation [see also Chatterjee et al.
(2004a)]. Nucleation happens at short times at different locations and times (not
shown), and patterns evolve in time owing to thermal fluctuations, giving rise to
bifurcation splittings and mergings. This evolution of patterns is driven entirely
by thermal fluctuations.
It is expected that simulations like this as well as various other coarse-grained
fluid-like simulation tools (e.g., LB) will become key players in nanometer scale
design and control in the emerging area of nanotechnology, as well as in in-
tervention for control in biological systems. Examples include pattern forma-
tion, self-assembly of nanoparticles, nucleation and growth of materials, and
computational cell biology. Given that these models are generic, application to
very diverse areas is entirely feasible. For a recent application example to trop-
ical convection, see Khouider et al. (2003).
C. SPATIOTEMPORAL ACCELERATION OF DISTRIBUTED (LATTICE) KMC SIMULATION
Integration of spatial and temporal acceleration methods discussed above to

create a stochastic simulation toolkit that can reach large length and time scales
is entirely possible. The first example of integrating spatial and temporal ac-
celeration methods entails the combination of NE-KMC with ACG-KMC
methods to simulate diffusion through relatively thick (10 mm) membranes
where diffusion becomes rate determining (Snyder et al., 2005). While combi-
nation of time-acceleration methods with lattice KMC is possible, most time
acceleration methods affect noise adversely. Therefore, it appears that integra-
tion of the t-leap method (or a derivative of it) with a spatially distributed
(lattice) KMC simulation is the most promising approach for many applica-
tions. It turns out that the t-leap method developed for well-mixed systems is
fully consistent with the local mean field assumption of the CG-KMC method.
Here, the first example of combining the two methods for the grand canonical
ensemble (adsorption/desorption) is presented. Figure 12a compares the results
of the t-leap CG-KMC method to the CG-KMC ones for a fixed value of the
acceleration parameter e, (see Gillespie (2001) for a precise definition of e). In
this simulation one starts from an empty lattice and monitors the lattice uptake,
i.e., the spatially averaged coverage vs. time, for a fixed value of the fluid
chemical potential. Figure 12b shows the corresponding bundle sizes vs. time. It
is clear that in each adsorption/desorption event molecular bundles greater than
a single molecule occur, leading to acceleration of the simulation. As the ac-
celeration parameter increases, the size of molecular bundles, i.e., the number of
molecules or processes participating in each event, increases. However, for
larger bundle sizes, the possibility of getting negative concentrations, as happens
0.6 103
0.5 βJo = 2 , L =15 Adsorption
Bundle size per cell

0.4 KP = 0.4 102 ε =10-2
Coverage
0.3 Desorption
Adsorption
0.2 101 ε =10-3
Line: CG-KMC
0.1
Symbols: τ-leap CG-KMC, ε=10-3 Desorption
0.0 100
0 5 10 15 20 0 5 10 15 20
(a) Time (b) Time
8 10-4 108
βJo = 2, L =15
ε=10-1 107 CG-KMC: m=200, q=103
6 10-4
ε=10-2 CG-KMC: m=20, q=104
STD in coverage
106
CPU time [s]
4 10-4 105 βJo = 2, L = 1
ε=10-3 104 q=103 θ = 0.5

2 10-4 CG-KMC Symbols:
103
τ-leap CG-KMC q=104
0 100 102
0.0 0.2 0.4 0.6 0.8 1.0 10-4 10-3 10-2
(c) Coverage (d) Time acceleration parameter, ε
FIG. 12. (a) Coverage on an initially empty lattice vs. time from both the CG-KMC simulation
and the Poisson-based t-leap CG-KMC simulation in the grand canonical ensemble. The agreement
is excellent. (b) Corresponding molecular bundles vs. time. (c) Standard deviation (STD) in noise of
coverage vs. coverage (corresponding to different values of the fluid chemical potential) for various
values of the acceleration parameter e. For smaller bundles the noise is nearly exact. However, as one
coarse-grains considerably the time increments, the noise of the t-leap method is slightly increased.
(d) CPU of CG-KMC and of t-leap CG-KMC for two meshes (q is the coarse cell size and m is the
number of coarse cells) and a fixed lattice size of N ¼ mq ¼ 2 105 microscopic sites. Application of
the Poisson-based t-leap can accelerate the lattice CG-KMC by orders of magnitude, especially
when the meshes are coarse to enable large time increments.
with well-mixed systems, limits the application of the t-leap CG-KMC method
(an expected result; not shown). Use of the binomial t-leap overcomes this
problem (Chatterjee and Vlachos, 2005).
Next, two main issues are discussed. First, Fig. 12c compares the noise of CG-
KMC to that obtained from t-leaping as a function of coverage, obtained by
varying the fluid chemical potential, for various values of the acceleration
parameter e. At relatively small bundle sizes compared to the coarse cell size q,
the t-leap CG-KMC method gives very good results in both the expected value
and the standard deviation. However, when larger time increments are at-
tempted, the noise is slightly increased. Finally, Fig. 12d compares the CPU
from the standard CG-KMC simulation and the t-leap CG-KMC simulation
for two meshes as the time increments increase. It is clear that substantially
higher savings than by CG-KMC simulation can be obtained when large time
increments are attempted. This is an exciting result that opens up the possibility
of stochastic simulation of large length and time scales.
One may ask, what are some of the important future directions suggested by
such findings? Developments so far have been focused on examples of prototype
statistical mechanics. There is a need to extend these to realistic, complex sys-
tems such as catalytic reactions, crystal growth, polymers, proteins, self-organ-
ization, etc. Furthermore, benchmark examples from other areas are needed to
further evaluate the success and limitation of various methods. Finally, inte-
gration of the t-leap method with the CG-KMC method, demonstrated here,
holds the greatest promise for enabling stochastic simulation of large length and
time scales.
VII. Multiscale, Stochastic Modeling of Biological Networks
A. SPATIALLY WELL-MIXED SYSTEMS
The need for multiscale modeling of biological networks in zero-dimensional

(well mixed) systems has been emphasized in Rao et al. (2002). The multiscale
nature of stochastic simulation for well-mixed systems arises from separation of
time scales, either disparity in rate constants or population sizes. In particular,
the disparity in species concentrations is commonplace in biological networks.
The disparity in population sizes of biological systems was in fact recognized
early on by Stephanopoulos and Fredrickson (1981). This disparity in time
scales creates slow and fast events. Conventional KMC samples only fast events
and cannot reach long times.
Several methods for speeding up Gillespie’s original algorithm for well-mixed
systems were reviewed above. Among these, the WP-KMC method was recently
employed to study the coupled epidermal growth factor receptor (EGFR) traf-
ficking and transduction (Resat et al., 2003). The binomial t-leap method ap-
plied to the complex MAP kinase cascade (94 signaling species among 296
reactions) demonstrated hundred- to thousand-fold savings in CPU with ex-
cellent accuracy despite the disparity in species populations (Chatterjee et al.,
2005b). An alternative approach to cope with the disparity in population sizes
that has received more attention is hybrid multiscale simulation. In particular,
one treats deterministically species in excess or reactions that involve species in
large populations and stochastically species in low concentration or reactions
invoking species in low population (see hierarchy in Fig. 3b). Other possible
pairs of models depicted in Fig. 3b could be employed and are mentioned below.
The first applications of such hybrid approaches have just emerged. An ex-
ample of coupled deterministic/stochastic ODEs was recently introduced by Zak
et al. (2003) for a relatively large regulatory genetic network (118 reactions, 44
species, 97 parameters). In this case, species in large concentrations (proteins
and transcription factor dimers) were treated as continuum variables and in-
tegrated with the implicit Euler method, whereas species in relatively low
concentrations (promoters and transcripts) were treated as discrete variables
and their corresponding material balances were solved using the KMC method.
The use of an implicit deterministic integrator demands solution of the KMC;
therefore, some reactions describing the effect of small species on large ones
were omitted (their effect was found negligible when carrying out fully deter-
ministic simulations), leading to one-dimensional coupling of the subsystems,
which was found to provide correct solutions.
It is entirely possible that the ensemble average (expected values) of a stoc-
hastic system differs from the deterministic model solution. For example, in the
simulations of Zak et al. (2003), single stochastic trajectories were found to
deviate significantly from the deterministic ones. In particular, the deterministic
solution does not show adaptation, whereas some of the stochastic trajectories
show adaptation and others do not. One question is whether a system composed
of hundreds to thousands of subsystems, i.e., the ensemble average, approaches
the deterministic behavior or not. Simulations were performed for the model of
Zak et al. (2003) using 2300 different random number initializations, and the
simulations were extended to 1000 h (simulated time). To economize simulation
time, only genes in the ‘‘core’’ of the network (genes A, B, C, D, E, and F) were
simulated. However, the simulation results for the core genes are identical to
those that would be obtained if the cascade genes were included in the sim-
ulations. The ensemble mean of the stochastic simulations converged to ap-
proximately the value from the deterministic simulations for only two out of the
six genes. However, for the remaining genes, the ensemble median or mode did
generally converge to values that were close to the deterministic value. Figure 13
shows illustrative results from these ensemble-based hybrid multiscale simula-
tions (corresponding to their Fig. 3). Gene F had the most complex distribution
of the genes, but the ensemble median did match the deterministic result rea-
sonably well. Despite the large number of simulations, ensemble properties,
especially of gene F, exhibit systematic differences from their deterministic pre-
diction.
As another example of hybrid simulation touched upon above, Haseltine and
Rawlings (2002) treated fast reactions either deterministically or with Langevin
equations and slow reactions as stochastic events. Vasudeva and Bhalla (2004)
presented an adaptive, hybrid, deterministic-stochastic simulation scheme of
fixed time step. This scheme automatically switches reactions from one type to
the other based on population size and magnitude of transition probability.
140 80
120 70
100 60
50
Gene B
Gene F
80
40
60
30
40 20
20 10
0 0
0 50 100 150 200 0 50 100 150 20
Time (hours) Time (hours)
120 50
100 40
Deterministic
80 30
Gene B
Gene F
60 Mean
20 Median
40 10
Mode
20 0
0
0 50 100 150 200 0 50 100 150 20
Time (hours) Time (hours)
FIG. 13. Deterministic and hybrid stochastic/deterministic simulations for gene B (left) and gene F
(right). For all plots: bright green thick line, deterministic simulation. For bottom plots: red line,
ensemble mean; magenta line, ensemble median, black line, ensemble mode. Ensemble properties
shown in these plots are computed using all 2300 hybrid simulations. Top plots: Deterministic and
five representative hybrid simulation results. Note that the fluctuations increase dramatically after
80 h simulated time, due to the ‘transient adaptation’ of the genes to the ligand input. Bottom
graphs: Deterministic simulation results and ensemble properties versus time. Note that the mode of
the distribution (black line) shows some degree of variability but remains centered around the
deterministic simulation for gene B and deviates substantially from the deterministic one for gene F.
The median and mean of the ensemble follow the deterministic simulation closely, but deviations for
gene F are apparent despite the large number of realizations averaged. Simulations performed by D.
Zak.
Two prototype reaction examples (reversible first-order and irreversible

second-order kinetics) were discussed to address issues of rounding when
switching from deterministic variables to stochastic (i.e., conversion of real
numbers to integers), as well as the thresholds of population sizes and transition
probabilities to control accuracy in the first two moments of the population
(mean and variance). Other more complex examples were also mentioned. The
BioNetS software was recently published. It can perform various types of sim-
ulations in well-mixed environments (exact stochastic Gillespie method, chem-
ical Langevin model, and deterministic ODE model) as well as hybrid
deterministic simulation for some reactions coupled with exact stochastic sim-
ulation for the rest (Adalsteinsson et al., 2004). The issue of time patching of
hybrid schemes was touched upon, and partitioning of reactions was done based
on the population size (small populations require stochastic treatment, whereas
large populations are treated deterministically). Several examples, including a
dimerization reaction in constant volume as well as cell growth and division, a
chemical oscillator, and a synthetic gene network, were used for model vali-
dation. In a similar spirit, Kiehl et al. (2004) proposed hybrid multiscale sim-
ulation by combining deterministic with exact stochastic simulation. Emphasis
was placed on the time patching between the two types (levels) of models and
the algorithm was applied to the lambda phage switch model system.
As another example of hybrid multiscale simulation, recent work combined
the Poisson-based t-leap method of Gillespie with the next reaction method of
Gibson and Bruck (2000) for reactions invoking large and small populations
(Puchalka and Kierzek, 2004). This two-level method, termed the maximal time
step method, is an interesting hybrid multiscale simulation where large disparity
in populations can be handled efficiently while the noise is nearly exact. Fur-
thermore, partitioning of reaction sets between the two algorithms is easy to
automate. The method was applied to the simulation of glucose, lactose, and
glycerol metabolism in Escherichia coli. Partitioning reactions as jump and
continuous Markov processes, and handling them using the next reaction
method and Langevin method, respectively, were also proposed by Salis and
Kaznessis (2005). In their approach, reactions modeled using the Langevin
method were defined as those that have a large transition probability (occur
many times in the time scale of slow reactions) and slightly change the pop-
ulations of reactants and products.
Burrage et al. (2004) provided an overview of the various methods used for
modeling of chemical kinetics with emphasis on SODEs. Hybrid schemes,
building on the hierarchy depicted in Fig. 3b, were again developed by com-
bining the exact stochastic simulation method of Gillespie, the t-leap method,
and the chemical Langevin equation (first three levels of Fig. 3b). A departure
from other recent hybrid simulations mentioned above is that the authors par-
titioned the reactions into three levels, namely, slow, intermediate, and fast. This
partitioning was based not only on propensities but also on population sizes.
Furthermore, they emphasized that semi-implicit or implicit solvers should be
used for SODEs to cope with possible stiffness, a very reasonable proposal that
was also followed by Zak et al. (2003). The constraint on the t-leap time in-
crement of the original method being sufficiently small to avoid negative con-
centrations (see section on temporal acceleration of KMC methods) was also
brought up, and it appears that it limited, at least in part, the computational
speed up to less than a factor of 2. Their hybrid simulation was applied to the
expression and activity of LacZ and LacY proteins in E. coli consisting of 22

reactions among 23 species. This study highlighted the point made earlier about
the original t-leap method. It would be interesting to study these examples using
the new binomial t-leap method of Chatterjee et al. (2005d). Furthermore, it
becomes clear that while the proposed partitioning was successful, automatic,
generic criteria allowing one to partition on-the-fly reactions into the various
levels of models of Fig. 3b are needed. Furthermore, the adverse effect of hybrid
schemes on fluctuations of shared species (Takahashi et al., 2004) needs to be
addressed.
As temporal upscaling methods for acceleration of KMC simulation become
mature and more robust, I expect that they will have a significant impact on the
modeling of biological reaction networks.
B. SPATIALLY DISTRIBUTED SYSTEMS
Spatially realistic models are important because most systems, while, not
being well mixed, still comprise a small number of molecules deeming stochasti-
city important. The ramifications of spatial non-uniformity can be substantial.
As an example, spatial variations in the receptor concentration on the mem-
brane surface of a living cell, i.e., receptor clustering, can have important effects
on downstream signaling (Duke and Bray, 1999; Goldman et al., 2002; Shea
et al., 1997). In their review, Meng et al. (2004) made the comment that spa-
tiotemporal modeling of biological systems is still infeasible. However, some
papers have started to emerge. One of the first examples of spatial KMC in
biological systems entails the spatial clustering of membrane receptors in bac-
terial chemotaxis that may lead to collective activity (Shimizu et al., 2003). This
is basically an Ising type of model with first nearest–neighbor interactions that
trigger local spatial organization of receptors close to and below the critical
temperature. Goldman and co-workers have conducted the first off-and-on lat-
tice simulations of a single type of EGF receptor by allowing dimerization,
cluster–cluster collisions, and diffusion of all cluster sizes (Goldman et al., 2002;
Gullick et al., 2002). As discussed in Goldman et al. (2002), off lattice simu-
lations are very slow even when there is no significant separation of time scales
to enable a molecular-level simulation of cell receptor dynamics. Very interest-
ing lattice KMC simulations of diffusion and dimerization events leading to
spatial self-organization of the G-protein-coupled receptor family have been
carried out by Woolf and Linderman (2003, 2004). Other spatially distributed
biological systems modeling using KMC simulation include that of Saxton
(1995, 2001) and Shea et al. (1997). But none of these papers have really ad-
dressed multiscale issues.
Examination of reported values of diffusion and reaction rate constants point
to the inherent multiscale challenges encountered in spatiotemporal modeling of
realistic systems using molecular models (Mayawala et al., 2005). What are the
multiscale challenges? First, there is a huge disparity in time scales between
various transport and reaction events. For example, this disparity is estimated
to be at least eight orders of magnitude in the dimerization and auto-
phosphorylation events of EFGR on the cell membrane. Second, a microscopic
KMC simulation box (100 100 nm2 in 2D or 30 30 30 nm3 in 3D) is far
too small to be applied to an entire cell of 10 mm diameter, and periodic
boundary conditions may be inadequate for such systems owing to the
extremely low density of several features such as surface pits. The disparity in
length scales is further attenuated owing to the low density of molecules, which
makes the probability of collision too low. This, in turn, results in inefficient
sampling and extremely long simulations that currently cannot reach experi-
mental time scales. As a result, only simulations with judiciously chosen pa-
rameters, i.e., probabilities of similar magnitude (see Goldman et al. (2002)), in
model systems have been carried out.
Two CG-KMC simulations for diffusion of non-interacting molecules and
simple reaction mechanisms have been proposed apparently independently (Elf
et al., 2003; Stundzia and Lumsden, 1996) as the first multiscale spatial models
for biological application. These models discretize the space into cells or ele-
ments (the latter term is more suitable here, to avoid confusion with biological
cells), in exactly the same way as in the CG-KMC simulation described above.
Within each element, the local mean field is assumed, and thus, the connection
with the Gillespie algorithm is straightforward. A major difference between the
aforementioned CG-KMC simulations of Katsoulakis et al. (2003a, b) and
Katsoulakis and Vlachos (2003) and the work of Stundzia and Lumsden and Elf
et al. is that in the latter, there is neither an exclusion principle nor interactions
between molecules. When the chemistry is nonlinear, strong spatial correlations
between molecules usually arise rendering the local mean field approximation
inaccurate. Therefore, these CG-KMC models are reasonable only when the
diffusion is relatively fast compared with reactions to locally homogenize the
concentrations of species and establish local mean field conditions (see Chat-
terjee et al., 2004a for an example). Under such conditions, these are local mean
field models, but they include noise and are thus suitable for small populations.
Obviously, this is an area to which significant multiscale efforts are expected to
be devoted in future work, in order to enable spatiotemporal modeling of bi-
ological systems.
VIII. Systems Tasks
The widespread use of multiscale modeling necessitates the concomitant de-

velopment of system-level tasks (see Fig. 1) for designing suitable experiments,
estimation of important parameters, reconstruction of entire reaction networks

(reverse engineering), metabolic path optimization and control, model-based
optimization and control of nanomaterials, to mention a few. A main difference
from previous systems tasks work is that one has to extract information from
spatiotemporal data of unprecedented resolution that are more noisy. At the
same time, one is interested in estimating and controlling additional variables
such as population size, intermolecular forces, and spatial correlations, tasks
that were inconceivable a few years ago. The systems branch of multiscale
simulation is at the embryonic stage, but is expected to grow rapidly. After all,
the benefit of multiscale simulation is not only to provide insights into complex
systems, but also to enable tasks that lead to improved performance and
control. Here, a brief outline of some recent progress on systems-level tasks is
provided. Perspectives on systems tasks have appeared recently (Braatz et al.,
2004; Christofides, 2001; Kevrekidis et al., 2004).
A. SENSITIVITY AND IDENTIFIABILITY ANALYSES
One of the most important tools in complex systems modeling and analysis is
sensitivity analysis (SA) (Tomlin et al., 1997; Varma et al., 1999). In order to
carry out SA, system responses (R) have to be defined first. These responses are
system specific, and some may be experimentally measurable whereas others may
not. For example, in developing gas-phase combustion mechanisms, the ignition
delay time, the flame speed, flammability limits, and concentrations of major and
some radical species are common responses. These responses are actually ob-
tained with different types of experiments, such as shock tubes, flame-propa-
gation experiments in wide tubes, continuous flow jet stirred reactor or one-
dimensional burner flame species measurements. In a typical catalytic reactor on
the other hand, conversion and selectivity are two key responses. Pollutant mole
fractions, explosion limits for safety, and hot spots are just some additional
responses. Lastly, in advanced materials and pharmaceuticals growth, the pri-
mary concerns are different: making the right material (e.g., the correct poly-
morph) with specific particle morphology (habit), high growth rate, appropriate
roughness, controllable fraction of defects, etc. is the main goal.
I expect that SA of stochastic and multiscale models will be important in
traditional tasks such as the identification of rate-determining steps and pa-
rameter estimation. I propose that SA will also be a key tool in controlling
errors in information passing between scales. For example, within a multiscale
framework, one could identify what features of a coarse-level model are affected
from a finer scale model and need higher-level theory to improve accuracy of the
overall multiscale simulation. Next a brief overview of SA for deterministic
systems is given followed by recent work on SA of stochastic and multiscale
systems.
1. Sensitivity Analysis of Deterministic Systems

SA determines the change in a response R as a result of a perturbation in one
of the parameters P of the model. Parameters of a model can be any conceivable
ones. For example, in a MD simulation, parameters could be all factors ap-
pearing in the intermolecular potential. Since the magnitude of various param-
eters can be very different, it is common to compute a normalized sensitivity
coefficient (NSC) defined as
d ln Ri Pi dRj Pi Rj ðPi þ DPi Þ Rj ðPi Þ

NSCij ¼ ¼ (9)
d ln Pj Rj dPi Rj DPi
The last approximate equality is simply a forward finite difference approxima-

tion of the derivative (higher order, more accurate approximations can obvi-
ously be used). The partial derivative simply indicates the slope, i.e., the change
in a response for an infinitesimal change in a parameter. When the partial
derivative is computed by differentiation of the response function, the SA is
termed local SA. On the other hand, when the parameter is changed and the
response is recomputed by solving the entire problem, the SA is termed brute
force. Brute force SA is straightforward. However, a perturbation must be per-
formed for each parameter, leading to prohibitive computational costs, espe-
cially when the number of parameters is large and the simulation is expensive.
This is typically the case with multiscale codes.
2. Sensitivity Analysis of Stochastic and Multiscale Models

SA of SODEs describing chemically reacting systems was introduced early
on, in the case of white noise added to an ODE (Dacol and Rabitz, 1984). In
addition to expected values (time or ensemble average quantities), SA of var-
iances or other correlation functions, or even the entire pdf, may also be of
interest. In other words, in stochastic or multiscale systems one may also be
interested in identifying model parameters that mostly affect the variance of
different responses. In many experimental systems, the noise is due to multiple
sources; as a result, comparison with model-based SA for parameter estimation
needs identification of the sources of experimental noise for meaningful con-
clusions.
One of the difficulties in performing SA of stochastic or more generally mul-
tiscale models is that a closed form equation does not often exist. As a result,
brute force SA has so far been the method of choice, which, while possible, is
computationally intensive. As suggested in Raimondeau et al. (2003), since the
response obtained is noisy, one has to introduce relatively large perturbations to
ensure that the responses are ‘‘reliable,’’ so that meaningful SA results are
obtained. For most complex systems, local SA may not be feasible. However, I
do not see this being an impediment since SA is typically used to rank-order the
importance of model parameters (see section on parameter estimation for a

different use of SA).
In our group we have used SA in lattice 2D and 3D KMC in order to identify
key parameters for parameter estimation from experimental data (see corre-
sponding section below). Finite difference approximations of NSC were em-
ployed (Raimondeau et al., 2003; Snyder and Vlachos, 2004). Drews et al.
(2003a) motivated by extraction of parameters for Cu electrodeposition, ob-
tained an expression for the sensitivity coefficient, analogous to Eq. (9), that
minimizes the effect of noise on the NSC assuming that the variance of the
stochastic correction is unaffected by the perturbation.
In order to elucidate some of the issues in SA of stochastic systems, the gene-
expression model proposed in Thattai and van Oudenaarden (2001) and
Ozbudak et al. (2002) for transcription and translation, shown schematically in
150
kr=0.01 s-1, b=10
100
Protein copies
50
γp
Protein
8
kp
6 γr
mRNA
mRNA
4
kr
2 DNA
0
0 1 2 3 4 5 6 7 8
Time [h]
FIG. 14. Simulation of gene expression depicting the number of protein copies and mRNA as well
as the corresponding pdf. While a normal distribution describes reasonably the population of pro-
tein, the pdf of the mRNA, whose population is very low, is far from Gaussian (solid lines on the
right panels). The parameters are: transcript initiation rate kr ¼ 0.01 s1, decay rates of gr ¼ 0:1 s1
and gp ¼ 0:002 s1 , and b ¼ 10. The deterministic steady-state values are hri ¼ 0:1; hpi ¼ 50. The
inset is a schematic of the gene expression process.
the inset of Fig. 14, is studied. At the deterministic, continuum level, the con-
centrations of mRNA (r) and protein (p) are given by the following ODEs:
dr
¼ k r gr r (10)
dt
dp
¼ k p r gp p (11)
dt
where g denotes the rate of decay (lnð2Þ=g is the half-life time) and k is the rate
constant for transcription or translation. The steady-state solution is hri ¼ kr =gr
and hpi ¼ kp kr =ðgp gr Þ ¼ bkr =gp , where b ¼ kp =gr is the average number of
proteins produced per transcript. For this linear system, it is relatively easy to
show that the variance over the mean (termed the Fano factor) is given by
hdp2 i
¼ 1 þ kp =ðgp þ gr Þ (12)
hpi
Typically, gp =gr is small (the mRNA is unstable compared to the protein).
Single stochastic trajectories obtained using Gillespie’s KMC algorithm are
shown in Fig. 14 for the protein and the mRNA levels vs. time. The corre-
sponding pdfs are also shown. Figure 15 shows the dependence of copies of
mRNA and proteins along with the variance of proteins on kr . Using the steady
solution of deterministic equations, the NSC with respect to kr (chosen pa-
rameter for illustration) can be easily computed to be 1. In order to exploit the
accuracy of computed NSCs, central second-order and forward or backward
0.2 70
Solid symbols: 800
65
Variance in protein copies
time averaged KMC KMC: Time averaged

Analytical
60
Protein copies
700
mRNA
55
0.1 600
50
45 500
40 Nominal case
Solid lines: Analytical 400
0.0 35
8.0 10-3 1.0 10-2 1.2 10-2 8.0 10-3 1.0 10-2 1.2 10-2
(a) kr (b) kr
FIG. 15. (a) Expected values of mRNA and protein from a KMC simulation computed as time
averaged (filled symbols), along with the corresponding values from the deterministic gene-expres-
sion model (solid lines) as a function of kr. (b) Corresponding variance in protein copies. Time-
averaged population and variance are in very good agreement with the analytical solution for this
linear model despite the low number of protein and mRNA copies. The rest of the parameters are
those of Fig. 14.
20
10
protein, 2nd-order FD
% change in kr
0 protein, 1st-order FD
Deterministic
mRNA, 1st-order FD
-5
-10
-20
0.0 0.2 0.4 0.6 0.8 1.0 1.2

NSC
FIG 16. NSC for protein and mRNA copies for different size perturbations in kr when time-
averaged concentrations are used. First-order (forward or backward) and central finite difference
approximations are used. The deterministic NSC for both species is 1 (vertical dashed line).
first-order finite difference approximations were employed, and the results are
depicted in Fig. 16. Owing to the inherent noise, the different methods make
little difference in the value of NSC, at least using steady-state data from a
single KMC trajectory and the perturbation sizes indicated (expected values are
typically computed from a total of 106–109 MC events to ensure very good
statistics). That is, there is no specific, clear-cut trend regarding accuracy in
computing NSCs. From a fundamental point of view, it is obviously desirable to
understand how many simulations and what simulation sizes are needed to
improve accuracy. Future work should explore this issue in detail.
Doyle and co-workers have used sensitivity and identifiability analyses in a
complex genetic regulatory network to determine practically identifiable pa-
rameters (Zak et al., 2003), i.e., parameters that can be extracted from exper-
iments with a certain confidence interval, e.g., 95%. The data used for analyses
were based on simulation of their genetic network. Different perturbations (e.g.,
step, pulse) were exploited, and an identifiability analysis was performed. An
important outcome of their analysis is that the best type of perturbations for
maximizing the information content from hybrid multiscale simulations differs
from that of the deterministic, continuum counterpart model. The implication
of this interesting finding is that noise may play a role in systems-level tasks.
3. Hierarchical Approaches to Sensitivity Analysis

A hierarchical approach could be an efficient way of reducing the CPU re-
quirements for performing systems-level tasks. In my experience, mean field or
continuum models, while not as accurate, provide a qualitatively picture of

sensitivity when the model itself is qualitatively correct (Raimondeau et al.,
2003; Snyder and Vlachos, 2004). Mean field-based SA could be used in various
ways. First, SA of a mean field model could identify key processes controlling a
system. Second, optimization of parameters of a stochastic model can be done
by first optimizing the parameters of the mean field model. These optimized
parameters could serve as a good initial guess for optimizing the parameters of a
KMC or multiscale model. The advantage of this hierarchical approach is that it
narrows down the parameter space where an optimum parameter search has to
be conducted for a stochastic or multiscale model. Obviously, this idea is based
on the premise that the optimum set of parameters of the stochastic or
multiscale model is close to that of the corresponding continuum model. This
looks to be the case in several examples explored by our group so far. However,
it is expected to fail when mean field models are qualitatively different from
stochastic or multiscale models.
The evolution equation of the sensitivity of the chemical master equation,
along with a first-order deterministic approximation of the sensitivity of the
mean, was recently derived in Haseltine (2005). In a similar spirit, the SA ev-
olution equations in MD simulations were developed in Stefanovic and Pant-
elides (2001). Haseltine et al. suggested computing the mean using the KMC
method and the sensitivity evolution of the mean through the deterministic first-
order approximation. This is an improvement over simply using the SA of the
mean field model mentioned above. The advantage of this hierarchical, hybrid
approach is that the sensitivity is a smooth function of time, an important aspect
for many systems tasks, and less expensive to compute. For the examples con-
sidered in a batch reactor by Haseltine (2005), the deterministic approximation
gave a smooth evolution of the sensitivity that works very well for linear systems,
and shows moderate deviations from the exact sensitivity for nonlinear systems.
From the prolegomena it is clear that hierarchical methods have significant
potential for accomplishing systems tasks at reduced computational cost.
B. PARAMETER ESTIMATION FROM EXPERIMENTAL DATA AND FINER SCALE MODELS
One question that arises is: if one uses multiscale simulation to predict sys-
tems behavior from first principles, then why does one need to carry out pa-
rameter estimation from experimental data? The fact is that model predictions
using even the most accurate QM techniques have errors. In the foreseeable
future, one would have to refine parameters from experiments to create a fully
quantitative multiscale model. Furthermore, for complex systems, QM tech-
niques may be too expensive to carry out in a reasonable time frame. As a result,
one may rely on estimating parameters from experimental data. Finally, an
important, new class of problems arises when one has to estimate parameters of
a coarser scale model in order to minimize the difference in predictions (in some
proper measure) from the ones of the finer scale model.
Parameter estimation is a very mature subject for continuum, deterministic
models and is an integral part of reverse engineering. In general, a cost function,
such as the Euclidean distance between modeling results and experimental ob-
servables, is formulated. For deterministic systems, this is sufficient. For mul-
tiscale models, however, a cost function for expected values may not be enough.
For example, in a stochastic simulation one would like to fit the entire pdf,
described by a Fokker–Planck equation. However, the pdf is known only for a
limited number of rather trivial problems, and its calculation is computationally
impractical. Practically, one may extend the approach of deterministic systems
to stochastic or multiscale models by including, aside from expected values,
additional quantities such as variances. While this is possible, including variance
into the objective function requires hundreds to thousands of trajectories, ren-
dering parameter estimation very time-consuming (Fullana and Rossi, 2002). In
passing, I should note an overview of parameter estimation of stochastic dif-
ferential equations based on maximum approximate likelihood ideas given in
Nielsen et al. (2000). Some additional complications in parameter estimation are
due to the large number of parameters of multiscale models and the noisy results
arising from molecular models. The introduction of high throughput exper-
iments or combinatorial methods opens up the possibility of creating massive
data sets for parameter estimation. However, one may not be able to extract
useful information for all parameters. It is important that the relevant param-
eters get extracted.
Rawlings and co-workers proposed to carry out parameter estimation using
Newton’s method, where the gradient can be cast in terms of the sensitivity of
the mean (Haseltine, 2005). Estimation of one parameter in kinetic, well-mixed
models showed that convergence was attained within a few iterations. As ex-
pected, the parameter values fluctuate around some average values once con-
vergence has been reached. Finally, since control problems can also be
formulated as minimization of a cost function over a control horizon, it was also
suggested to use Newton’s method with relatively smooth sensitivities to ac-
complish this task. The proposed method results in short computational times,
and if local optimization is desired, it could be very useful.
Since complex systems most probably exhibit complicated surfaces with mul-
tiple minima, convergence may not be obtained using local searching tech-
niques, and the probability of obtaining the global optimum with local
optimizers is low. Alternatively, one can employ global-type optimization
methods, such as simulated annealing and genetic algorithms. While these
techniques are often successful in determining the global minimum, they require
hundreds of thousands of function evaluations, i.e., KMC or multiscale sim-
ulations. Such a task is impractical. To overcome this challenge we have pro-
posed to develop reduced models or surfaces approximated by low-order degree
polynomials using solution mapping or surface response methods typically
employed in design of experiments. This idea leads to a relatively smooth sur-

face that can be used in optimization (Raimondeau et al., 2003). Hierarchical
parameterization where the mean field model parameters are estimated first and
serve as an initial guess for optimization of the molecular or multiscale model
parameters offer an attractive approach for constructing accurate surfaces.
The first application of hierarchical SA for parameter estimation included
refinement of the pre-exponentials in a surface kinetics mechanism of CO
oxidation on Pt (a lattice KMC model with 6 parameters) (Raimondeau et al.,
2003). A second example entailed parameter estimation of a dual site 3D lattice
KMC model for the benzene/faujasite zeolite system where benzene–benzene
interactions, equilibrium constants for adsorption/desorption of benzene on
different types of sites, and diffusion parameters of benzene (a total of 15
parameters) were determined (Snyder and Vlachos, 2004). While this approach
appears promising, the development of accurate but inexpensive surfaces (re-
duced models) deserves further attention to fully understand its success and
limitation.
C. MODEL REDUCTION AND CONTROL
Online multiscale model-based control is beyond current computer capabil-

ities owing to the computational intensity of multiscale simulation. Two ap-
proaches are proposed to enable control at the nanometer scale using multiscale
simulation. The first entails a suitable model reduction, where the full multiscale
model is effectively mapped into an approximate surface that is subsequently
used in process design and control. Toward this goal, proper-orthogonal de-
composition was explored to derive a small number of modes (space dimen-
sion), i.e., spatially global eigenfunctions, to form a basis that captures
spatiotemporal computer data. While this is indeed possible, we have found that
the noise of KMC or multiscale simulations renders model reduction challeng-
ing Raimondeau and Vlachos (2000). In a way, microscopic surface processes
have slower dissipation mechanisms than fluid-phase processes, and as a result,
noise-induced phenomena, such as nucleation, demand many modes for accu-
rate reduction. It appears then that a second strategy based on optimum design,
where one designs the system using a multiscale model to behave in a desirable
manner, may be more suitable than online control. This is also underscored by
the current lack of easy implementation of sensors and actuators operating at
the nanoscopic scale the way their macroscopic counterparts work at the large
scale. While materials engineering could possibly overcome this problem in the
future, at least in part, manipulating a few input and output coarse variables
may still remain the only viable way for many processes. A similar view is
shared by Braatz and co-workers (2004). However, further work is needed to
delineate the necessity of online control and the suitability of various model
reduction tools for this task.
Aside from proper orthogonal decomposition, alternative model reduction

strategies have also been explored. For a simple reaction network with a species
in QSS, a reduced description of the master equation has been successful by
applying the projection operator formalism (Shibata, 2003), and subsequently
applied to a simple gene expression network. An advantage of this theoretical
study is that it provides insight into how the noise of the eliminated species
affects the population of the other species. However, extension of such theo-
retical analysis to complex reaction networks is not straightforward. Reduction
of the master equation was also carried out, and the reduced model was used to
determine open-loop temperature profiles for epitaxial growth (Gallivan and
Atwater, 2004; Gallivan and Murray, 2003, 2004).
One of the objectives of model reduction is the possibility of carrying
out model-based control. Some initial, promising efforts along this direction
have already appeared. Control of surface roughness and growth rate in hybrid
KMC/stagnation flow simulations of epitaxial growth mentioned above
was demonstrated by employing integral control in Lou and Christofides
(2003a, b, 2004). In particular, the overall approach employed real-time
estimators from KMC (using multiple, small KMC simulation boxes),
filters to reduce the noise of KMC, and error compensators followed by
feedback controllers. In an alternative approach, a time stepper method was
used to derive an optimal control policy for reactions (modeled by the
LB technique or a well-mixed KMC model to stabilize an unstable open-
loop state) or to derive a local linearization of a stochastic model that was
subsequently employed in linear control theory (Armaou et al., 2004; Siettos
et al., 2003).
Systems approach borrowed from the optimization and control communities
can be used to achieve various other tasks of interest in multiscale simulation.
For example, Hurst and Wen (2005) have recently considered shear viscosity as
a scalar input/output map from shear stress to shear strain rate, and estimated
the viscosity from the frequency response of the system by performing short,
non-equilibrium MD. Multiscale model reduction, along with optimal control
and design strategies, offers substantial promise for engineering systems. In-
tensive work on this topic is therefore expected in the near future.
D. BIFURCATION
Many systems exhibit nonlinear behavior. This is another systems-level

task that is computationally very demanding. Application of bifurcation
analysis to simple and complex chemistry hybrid stochastic (KMC)-deterministic
(ODE) models has been presented by our group (Raimondeau and Vlachos, 2002b,
2003; Vlachos et al., 1990) for various catalytic surface reactions.
Prototype hybrid continuum-stochastic models that exhibit bifurcations
were recently explored by Katsoulakis et al. (2004). It was found that mesoscopic
models based on the stochastic averaging principle are excellent approximations of

fully stochastic models when there is disparity in relaxation times of microscopic
(fast) and flow (slow) processes. Multiple states in tropical convection model pre-
diction were reported in Majda and Khouider (2002). Kevrekidis and co-workers
have been successful in applying time steppers in constructing bifurcation diagrams
of stochastic simulation such as KMC (e.g., Makeev et al., 2002). One of the
advantages of their method is that unstable branches and bifurcation points can be
computed, a task that is difficult with direct KMC simulation. On the other hand, it
is expected that the stabilization has an adverse effect on understanding metasta-
bility and transitions between states of small systems.
IX. Outlook
The multiscale simulation framework presented here is generic and can be

applied across multiple disciplines and problems of chemical sciences. Obvi-
ously, specific scientific problems may be amenable to special twists. While
substantial progress in multiscale analysis has already been achieved, the
emerging field is still at an embryonic stage; many exciting developments are
expected in the next decade. The area of systems tasks is by far the least de-
veloped. However, the significantly increasing number of presentations at the
AIChE meeting (in area 10d) is an indicator of the explosion of the new field
and the exciting contributions of the systems community to the design and
control of complex systems via multiscale modeling and simulation. A central
theme in multiscale modeling, as one moves from finer to coarser scales, is
model reduction. While universal approaches to model reduction may not exist
or even be desirable, robust reduction methodologies along with methods of
assessing the resulting errors of coarse graining for various types of multiscale
simulation are needed.
I believe that growth in a number of critical areas of technological importance
to the nation, such as nanotechnology, biotechnology, and microengineering,
will be accelerated and catalyzed by the new multiscale modeling and compu-
tational paradigm. In all these and other areas of chemical sciences, as alluded
to in the introduction, multiscale analysis could have the most significant en-
gineering impact in top-down and reverse engineering modes. While multiscale
analysis research is multidisciplinary and is currently conducted, in many cases,
in a collaborative manner, training of future undergraduate and graduate stu-
dents on these topics is also important. Graduate and undergraduate education
is at a crossroads (Cussler et al., 2002; Dudukovic, 2003), and modern and in
many cases undeveloped tools need to be taught in efficient ways for preparing
students for the modeling challenges arising from the new technologies. To
achieve this goal, there is a clear need for revision of core courses to incorporate
elements of multiscale analysis and for development of new multiscale modeling

and simulation courses.
ACKNOWLEDGMENTS
The author’s multiscale simulation work has been supported over the years by
NSF grants (Career Award no. CTS-9702615, CTS-9811088, ITR-0219211, and
CTS-0312117). The ideas on spatial coarse graining of the lattice MC method
have been developed in collaboration with Prof. Markos A. Katsoulakis. The
research of Robert Lam, Dr. Stephanie Raimondeau, Dr. Mallika Gummalla,
Abhijit Chatterjee, and Mark Snyder as ex- or current graduates students of
DGV has contributed substantially to the author’s understanding on multiscale
simulation and comprises the main body of the results presented herein. The
author is indebted to Dan Zak, who performed additional simulations to clarify
differences between deterministic and stochastic simulation for his system.
Proofreading by graduate students (Mark Snyder, Jeff Ludwig, Kapil May-
awala, and Dan Zak) of the University of Delaware is also greatly appreciated.
REFERENCES
Abraham, F. F., Broughton, J. Q., Bernstein, N., and Kaxiras, E. Europhys. Lett. 44(6), 783 (1998a).
Abraham, F. F., Broughton, J. Q., Bernstein, N., and Kaxiras, E. Comput. Phys. 12(6), 538 (1998b).
Adalsteinsson, D., McMillen, D., and Elston, T. C. BMC Bioinformatics 5(24), 1–21 (2004).
Alkire, R., and Verhoff, M. Electrochim. Acta 43(19–20), 2733 (1998).
Allen, M. P., and Tildesley, D. J., ‘‘Computer Simulation of Liquids’’. Oxford Science Publications,
Oxford (1989).
Armaou, A., Siettos, C. I., and Kevrekidis, L. G. Int. J. Robust Nonlinear Control 14(2), 89 (2004).
Barsky, S., Delgado-Buscalioni, R., and Coveney, P. V. J. Chem. Phys. 121(5), 2403 (2004).
Baumann, F. H., Chopp, D. L., de la Rubia, T. D., Gilmer, G. H., Greene, J. E., Huang, H.,
Kodambaka, S., O’Sullivan, P., and Petrov, I. MRS Bull. 26(3), 182 (2001).
Binder, K., Ed., ‘‘Monte Carlo Methods in Statistical Physics’’. Springer, Berlin (1986).
Bird, G. A. Comm. Appl. Num. Methods 4, 165 (1988).
Bird, R. B., Stewart, W. E., and Lightfoot, E. N., ‘‘Transport Phenomena’’. Wiley, New York
(1960).
Bonilla, G., Vlachos, D. G., and Tsapatsis, M. Micro. Meso. Mat. 42(2–3), 191 (2001).
Bortz, A. B., Kalos, M. H., and Lebowitz, J. L. J. Comp. Phys. 17, 10 (1975).
Braatz, R. D., Alkire, R. C., Seebauer, E., Rusli, E., Gunawan, R., Drews, T. O., Li, X., He, Y.
‘‘Perspectives on the Design and Control of Multiscale Systems’’. Proceedings of the 7th
International Symposium on Dynamics and Control of Process Systems, Cambridge, MA,
paper 96, July 5–7.
Brenan, K. E., Campbell, S. L., and Petzold, L. R., ‘‘Numerical Solution of Initial-value Problems in
Differential-algebraic Equations’’. SIAM, Philadelphia (1996).
Broadbelt, L., Stark, S. M., and Klein, M. T. Ind. Eng. Chem. Res 33(4), 790 (1994).
Broughton, J. Q., Abraham, F. F., Bernstein, N., and Kaxiras, E. Phys. Rev. B 60(4), 2391 (1999).
Bundschuh, R., Hayot, F., and Jayaprakash, C. Biophys. J. 84, 1606–1615 (2003).
Burrage, K., Tian, T. H., and Burrage, P. Progr. Biophys. Mol. Biol. 85(2–3), 217 (2004).
Burton, W. K., Cabrera, N., and Frank, F. C. Proc. Roy. Soc. Lond. A 243, 299 (1951).
Cai, W., de Koning, M., Bulatov, V. V., and Yip, S. Phys. Rev. Lett. 85(15), 3213 (2000).
Cao, Y., Petzold, L. R., Rathinam, M., and Gillespie, D. T. J. Chem. Phys. 121(24), 12169 (2004).
Chatterjee, A., Katsoulakis, M. A., and Vlachos, D. G. Rev. E 71, 0267021 (2005a).
Chatterjee, A., Mayawala, K., Edwards, J. S., and Vlachos, D. G. Bioinformatics 21(9), 2136
(2005a).
Chatterjee, A., Snyder, M. A., and Vlachos, D. G. Chem. Eng. Sci. 59, 5559–5567 (2004a).
Chatterjee, A., and Vlachos, D. G. J. Comp. Phys. accepted (2005).
Chatterjee, A., Vlachos, D.G., and Katsoulakis, M. Int. J. Multiscale Comp. Eng. in press (2005c).
Chatterjee, A., Vlachos, D. G., and Katsoulakis, M. A. J. Chem. Phys 121(22), 11420 (2004b).
Chatterjee, A., Vlachos, D. G., and Katsoulakis, M. A. J. Chem. Phys 122, 024112 (2005d).
Chen, S., and Doolen, G. D. Annu. Rev. Fluid Mech. 30, 329 (1998).
Christofides, P. D. AIChE J. 47(3), 514 (2001).
Cracknell, R. F., Nicholson, D., and Quirke, N. Phys. Rev. Lett. 74(13), 2463 (1995).
Cussler, E. L., Savage, D. W., Middelberg, A. P. J., and Kind, M. Chem. Eng. Prog. 98(1), 26S
(2002).
Cussler, E. L., and Wei, J. AIChE J. 49(5), 1072 (2003).
Dacol, D. K., and Rabitz, H. J. Math. Phys. 25(9), 2716 (1984).
Deem, M.W., ‘‘A statistical mechanical approach to combinatorial chemistry, in (A.K. Chakr-
aborty, Ed.), ‘‘Molecular Modeling and Theory in Chemical Engineering’’, vol. 28, p. 81.
Academic Press, New York, (2001).
Dollet, A. Surf. Coat. Technol. 177–178, 245–251 (2004).
Drews, T. O., Braatz, R. D., and Alkire, R. C. J. Electrochem. Soc. 150(11), C807 (2003a).
Drews, T. O., Ganley, J. C., and Alkire, R. C. J. Electrochem. Soc. 150(5), C325 (2003b).
Drews, T. O., Webb, E. G., Ma, D. L., Alameda, J., Braatz, R. D., and Alkire, R. -C. AIChE J.
50(1), 226 (2004).
Dudukovic, M. P. Chem. Eng. News May12, 3 (2003).
Duke, T. A. J., and Bray, D. Proc. Natl. Acad. Sci. USA 96(18), 10104 (1999).
Elf, J., Doncic, A., and Ehrenberg, M. Mesoscopic reaction–diffusion in intracellular signaling, in (S.
M. Bezrukov, Hans Frauenfelder, Frank Moss Eds.), ‘‘Fluctuations and Noise in Biological,
Biophysical, and Biomedical Systems’’. Proc. SPIE, 5110, 114–124, SPIE Digital Library,
http://spie.org/app/Publications/, Santa Fe, NM, USA (2003).
Fichthorn, F. A., and Weinberg, W. H. J. Chem. Phys. 95(2), 1090 (1991).
Frenkel, D., and Smit, B., ‘‘Understanding Molecular Simulation: From Algorithms to Applica-
tions’’. Academic Press, New York (1996).
Frisch, M. J., Trucks, G. W., Schlegel, H. B., Scuseria, G. E., Robb, M. A., Cheeseman, J. R.,
Montgomery, Jr., J. A., Vreven, T., Kudin, K. N., Burant, J. C., Millam, J. M., Iyengar, S. S.,
Tomasi, J., Barone, V., Mennucci, B., Cossi, M., Scalmani, G., Rega, N., Petersson, G. A.,
Nakatsuji, H., Hada, M., Ehara, M., Toyota, K., Fukuda, R., Hasegawa, J., Ishida, M.,
Nakajima, T., Honda, Y., Kitao, O., Nakai, H., Klene, M., Li, X., Knox, J. E., Hratchian, H.
P., and V. Bakken, J. B. C., Adamo, C., Jaramillo, J., Gomperts, R., Stratmann, R. E.,
Yazyev, O., Austin, A. J., Cammi, R., Pomelli, C., Ochterski, J. W., Ayala, P. Y., Mo-
rokuma, K., Voth, G. A., Salvador, P., Dannenberg, J. J., Zakrzewski, V. G., Dapprich, S.,
Daniels, A. D., Strain, M. C., Farkas, O., Malick, D. K., Rabuck, A. D., Raghavachari, K.,
Foresman, J. B., Ortiz, J. V., Cui, Q., Baboul, A. G., Cli_ord, S., Cioslowski, J., Stefanov, B.
B., Liu, G., Liashenko, A., Piskorz, P., Komaromi, I., Martin, R. L., Fox, D. J., Keith, T.,
Al-Laham, M. A., Peng, C. Y., Nanayakkara, A., Challacombe, M., Gill, P. M. W., Johnson,
B., Chen, W., Wong, M. W., Gonzalez, C., and Pople, J. A. Gaussian, 03, Revision, C.02,
(2002).
Fritzsche, S., Haberlandt, R., and Karger, J. Z. Phys. Chem. 189, 211 (1995).
Fullana, J. M., and Rossi, M. Phys. Rev. E 65(3) (2002).

Gallivan, M. A., and Atwater, H. A. J. Appl. Phys. 95(2), 483 (2004).
Gallivan, M. A., and Murray, R. M. ‘‘Proceedings of the American Control Conference’’. Denver,
CO (2003).
Gallivan, M. A., and Murray, R. M. Int. J. Robust Nonlinear Control 14(2), 113 (2004).
Garcia, A. L., Bell, J. B., Crutchfield, W. Y., and Alder, B. J. J. Comp. Phys. 154(1), 134 (1999).
Gear, C. W. The automatic integration of stiff ordinary differential equations, Proceedings of the
IP68 Conference, North-Holland, Amsterdam (1969).
Gear, C. W., ‘‘Numerical Initial Value Problems in Ordinary Differential Equations’’. Prentice-Hall,
Inc., Englewood Cliffs (1971).
Gear, C. W., and Kevrekidis, I. G. SIAM J. Sci. Comput. 24(4), 1091 (2003).
Gear, C. W., Li, J., and Kevrekidis, I. G. Phys. Lett. A 316(3–4), 190 (2003).
Gibson, M. A., and Bruck, J. J. Phys. Chem. A 104, 1876 (2000).
Gillespie, D. T. J. Comp. Phys. 22, 403 (1976).
Gillespie, D. T. J. Phys. Chem. 81, 2340 (1977).
Gillespie, D. T. Physica A 188, 404 (1992).
Gillespie, D. T. J. Chem. Phys. 113(1), 297 (2000).
Gillespie, D. T. J. Chem. Phys. 115(4), 1716 (2001).
Gillespie, D. T. J. Phys. Chem. A 106(20), 5063 (2002).
Gillespie, D. T., and Petzold, L. R. J. Chem. Phys. 119(16), 8229 (2003).
Givon, D., Kupferman, R., and Stuart, A. Nonlinearity 17, R55–R127 (2004).
Goldman, J. P., Gullick, W. J., Bray, D., and Johnson, C. G. Individual-based simulation of the
clustering behaviour of epidermal growth factor receptors, ‘‘Proceedings of the ACM Sym-
posium on Applied Computing (SAC), March 10–14, 2002’’. ACM, Madrid, Spain (2002).
Grujicic, M., Gao, G., and Figliola, R. S. Appl. Surf. Sci. 183, 43 (2001).
Gullick, W. J., Goldman, J. P., Johnson, C. G., and Hayes, N. L. V. Br. J. Cancer 86, S10 (2002).
Gummalla, M., Tsapatsis, M., Watkins, J. J., and Vlachos, D. G. AIChE J. 50(3), 684 (2004).
Hadjiconstantinou, N. G., and Patera, A. T. Int. J. Mod. Phys. C 8(4), 967 (1997).
Hairer, E., and Wanner, G., ‘‘Solving Ordinary Differential Equations II. Stiff and Differential
Algebraic Problems’’. Springer, Berlin (1991).
Hansen, U., Rodgers, S., and Jensen, K. F. Phys. Rev. B 62(4), 2869 (2000).
Haseltine, E. L. ‘‘Systems Analysis of Stochastic and Population Balance Models for Chemical
Reactions’’. Department of Chemical Engineering, Madison, University of Wisconsin (2005).
Haseltine, E. L., and Rawlings, J. B. J. Chem. Phys. 117(15), 6959 (2002).
He, J., Zhang, H., Chen, J., and Yang, Y. Macromolecules 30, 8010 (1997).
Heffelfinger, G. S., and van Swol, F. J. Chem. Phys. 100(10), 7548 (1994).
Hildebrand, M., Mikhailov, A. S., and Ertl, G. Phys. Rev. Lett. 81(12), 2602 (1998).
Hoekstra, R., Grapperhaus, M., and Kushner, M. J. Vac. Sci. Technol. A 15(4), 1913 (1997).
Hongo, K., Mizuseki, H., Kawazoe, Y., and Wille, L. T. J. Cryst. Growth 236(1–3), 429 (2002).
Hurst, J. L., Wen, J. T. ‘‘Computation of Shear Viscosity: A Systems Approach, American Control
Conference’’. Portland, OR (2005).
Ishikawa, A., and Ogawa, T. Phys. Rev. E 65(2), 026131 (2002).
Ismail, A. E., Rutledge, G. C., and Stephanopoulos, G. J. Chem. Phys. 118(10), 4414 (2003a).
Ismail, A. E., Rutledge, G. C., and Stephanopoulos, G. J. Chem. Phys 122, 234901 (2005).
Ismail, A. E., Stephanopoulos, G., and Rutledge, G. C. J. Chem. Phys. 118(10), 4424 (2003b).
Ismail, A. E., Stephanopoulos, G., and Rutledge, G. C. J. Chem. Phys. 122, 234902 (2005).
Janssen, J. A. M. J. Stat. Phys. 57(1–2), 157 (1989a).
Janssen, J. A. M. J. Stat. Phys. 57(1–2), 171 (1989b).
Katsoulakis, M., Majda, A. J., and Vlachos, D. G. Proc. Natl. Acad. Sci. 100(3), 782 (2003a).
Katsoulakis, M. A., Majda, A. J., and Sopasakis, A. Com. Math. Sci. 2, 255 (2004).
Katsoulakis, M. A., Majda, A. J., and Vlachos, D. G. J. Comp. Phys. 186, 250 (2003b).
Katsoulakis, M. A., and Vlachos, D. G. J. Chem. Phys. 119(18), 9412 (2003).
Kevrekidis, I. G., Gear, C. W., and Hummer, G. AIChE J. 50(7), 1346 (2004).
Khouider, B., Majda, A. J., and Katsoulakis, M. Proc. Natl. Acad. Sci. 100(14), 11941–11946 (2003).
Kiehl, T. R., Mattheyses, R. M., and Simmons, M. K. Bioinformatics 20(3), 316 (2004).
Kirkpatrick, S., Gelatt, C. D. J., and Vecchi, M. P. Science 220(4598), 671 (1983).
Kisker, D. W., Stephenson, G. B., Fuoss, P. H., and Brennan, S. J. Cryst. Growth 146, 104 (1995).
Kissel-Osterrieder, R., Behrendt, F., and Warnatz, J. ‘‘Twenty-Seventh Symposium (International)
on Combustion, The Combustion Institute’’, p. 2267 (1998).
Kremer, K., and Muller-Plathe, F. MRS Bulletin 26(3), 205 (2001).
Kwon, Y. I., and Derby, J. J. J. Cryst. Growth 230(1–2), 328 (2001).
Lam, R., Basak, T., Vlachos, D. G., and Katsoulakis, M. A. J. Chem. Phys. 115(24), 11278 (2001).
Lam, R., and Vlachos, D. G. Phys. Rev. B 64(3), 5401 (2001).
Landau, D. P., and Binder, K., ‘‘A Guide to Monte Carlo Simulations in Statistical Physics’’.
Cambridge University Press, Cambridge (2000).
Lebedeva, M. I., Vlachos, D. G., and Tsapatsis, M. Ind. Eng. Chem. Res. 43, 3073 (2004a).
Lebedeva, M. I., Vlachos, D. G., and Tsapatsis, M. Phys. Rev. Lett. 92(8), 088301 (2004b).
Lerou, J. J., and Ng, K. M. Chem. Eng. Sci. 51(10), 1595 (1996).
Li, J., Zhang, J., Ge, W., and Liu, X. Chem. Eng. Sci. 59(8–9), 1687 (2004).
Lou, Y., and Christofides, P. D. Chem. Eng. Sci. 58(14), 3115 (2003a).
Lou, Y., and Christofides, P. D. AIChE J. 49(8), 2099 (2003b).
Lou, Y., and Christofides, P. D. Comp. Chem. Eng. 29, 225–241 (2004).
Ludwig, J., and Vlachos, D. G. Mol. Sim. 30(11–12), 765 (2004).
MacElroy, J. M. D. J. Chem. Phys. 101, 5274 (1994).
MacElroy, J. M. D., and Suh, S. -H. J. Chem. Phys. 106(20), 8595 (1997).
Maginn, E. J., Bell, A. T., and Theodorou, D. N. J. Phys. Chem. 97(16), 4173 (1993).
Majda, A. J., and Khouider, B. Proc. Natl. Acad. Sci. 99(3), 1123 (2002).
Makeev, A. G., Maroudas, D., Panagiotopoulos, A. Z., and Kevrekidis, I. G. J. Chem. Phys.
117(18), 8229 (2002).
Maroudas, D. AIChE J. 46(5), 878 (2000).
Maroudas, D., Modeling of radical-surface interactions in the plasma-enhanced chemical vapor
deposition of silicon thin films, in (A.K. Chakraborty, Ed.), ‘‘Molecular Modeling and The-
ory in Chemical Engineering’’, vol. 28, p. 252. Academic Press, New York (2001).
Maroudas, D. ‘‘Multiscale modeling, Challenges for the chemical sciences in the 21st century:
Information and communications report’’, National Academies, Washington, DC. p. 133.
(2003).
Mayawala, K., Vlachos, D. G., and Edwards, J. S. in preparation (2005).
Meng, T. C., Somani, S., and Dhar, P. In Silico Biol. 4(0024), 1 (2004).
Merchant, T. P., Gobbert, M. K., Cale, T. S., and Borucki, L. J. Thin Solid Films 365(2), 368 (2000).
Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., and Teller, E. J. Chem. Phys.
21, 1087 (1953).
Miller, R. E., and Tadmor, E. B. J. Comput-Aided Mater. Des. 9(3), 203 (2002).
Mizuseki, H., Hongo, K., Kawazoe, Y., and Wille, L. T. Comput. Mat. Sci. 24(1–2), 88 (2002).
Monine, M. I., Pismen, L. M., and Imbihl, R. J. Chem. Phys. 121(22), 11332 (2004).
Muller-Plathe, F. Chem. Phys. Chem. 3, 754 (2002).
Muller-Plathe, F. Soft Mater. 1(1), 1 (2003).
Nakano, A., Bachlechner, M. E., Kalia, R. K., Lidorikis, E., Vashishta, P., Voyiadjis, G. Z.,
Campbell, T. J., Ogata, S., and Shimojo, F. Comput. Sci. Eng. 3(4), 56 (2001).
Nie, X., Chen, S., Weinan, E., and Robbins, M. Phys. Fluids to be submitted.
Nielsen, J. N., Madsen, H., and Young, P. C. Ann. Rev. Control 24, 83 (2000).
Nielsen, S. O., Lopez, C. F., Srinivas, G., and Klein, M. L. J. Phys. Condens. Matter 16, R481–R512
(2004).
Nikolakis, V., Kokkoli, E., Tirrell, M., Tsapatsis, M., and Vlachos, D. G. Chem. Mater. 12(3), 845
(2000).
Nikolakis, V., Tsapatsis, M., and Vlachos, D. G. Langmuir 19, 4619 (2003).
Nikolakis, V., Vlachos, D. G., and Tsapatsis, M. J. Chem. Phys. 111(5), 2143 (1999).
NRC ‘‘Beyond the Molecular Frontier: Challenges for Chemistry and Chemical Engineering’’. Na-
tional Research Council, The National Academy Press, BCST, www.nap.edu publication
(2003a).
NRC ‘‘Challenges for the chemical sciences in the 21st century: Information and communications
report’’. National Research Council, The National Academy Press, Washington, DC (2003b).
O’Connell, S. T., and Thompson, P. A. Phys. Rev. E 52(6), 5792 (1995).
Ogata, S., Lidorikis, E., Shimojo, F., Nakano, A., Vashishta, P., and Kalia, R. K. Comp. Phys.
Commun. 138(2), 143 (2001).
Ozbudak, E. M., Thattai, M., Kurtser, I., Grossman, A. D., and van Oudenaarden, A. Nat. Genet.
31, 69 (2002).
Petzold, L. R. DDASSL: A differential/algebraic system solver. Albuquerque, NM, Sandia National
Laboratories Report No. SAND82-8637 (1982).
Philips, R. Curr. Opin. Solid State Mater. Sci. 3, 526 (1998).
Phillips, C. H. ‘‘Modeling in Engineering-The challenge of multiple scales’’. ENGenious, Division of
Engineering and Applied Science, California Institute of Technology (2002).
Plapp, M., and Karma, A. J. Comp. Phys. 165, 592 (2000).
Pope, S. B. Combust. Theory Model. 1, 1 (1997).
Pricer, T. J., Kushner, M. J., and Alkire, R. C. J. Electrochem. Soc. 149(8), C396 (2002a).
Pricer, T. J., Kushner, M. J., and Alkire, R. C. J. Electrochem. Soc. 149(8), C406 (2002b).
Puchalka, J., and Kierzek, A. M. Biophys. J. 86, 1357–1372 (2004).
Qin, F., Tagliabule, L., Pivesan, L., and Wolf, E. E. Chem. Eng. Sci. 53(5), 919 (1998).
Raimondeau, S., Aghalayam, P., Mhadeshwar, A. B., and Vlachos, D. G. Ind. Eng. Chem. Res. 42,
1174 (2003).
Raimondeau, S., Aghalayam, P., Vlachos, D.G., and Katsoulakis, M. Bridging the gap of multiple
scales: From microscopic, to mesoscopic, to macroscopic models, ‘‘Foundations of Molecular
Modeling and Simulation, AIChE Symposium Series No, 325’’, vol. 97, pp. 155–158 (2001).
Raimondeau, S., and Vlachos, D. G. J. Comp. Phys. 160, 564 (2000).
Raimondeau, S., and Vlachos, D. G. Chem. Eng. J. 90(1–2), 3 (2002a).
Raimondeau, S., and Vlachos, D. G. Comput. Chem. Eng. 26(7–8), 965 (2002b).
Raimondeau, S., and Vlachos, D. G. Chem. Eng. Sci. 58, 657 (2003).
Rao, C. V., and Arkin, A. P. J. Chem. Phys. 118(11), 4999 (2003).
Rao, C. V., Wolf, D. M., and Arkin, A. P. Nature 420(6912), 231 (2002).
Rastogi, S. R., and Wagner, N. J. Comp. Chem. Eng. 19, 693 (1995).
Rathinam, M., Petzold, L. R., Cao, Y., and Gillespie, D. T. J. Chem. Phys. 119(24), 12784 (2003).
Reese, J. S., Raimondeau, S., and Vlachos, D. G. J. Comp. Phys. 173(1), 302 (2001).
Reich, S. J. Comp. Phys. 151(1), 49 (1999).
Resat, H., Ewald, J. A., Dixon, D. A., and Wiley, H. S. Biophys. J. 85, 730 (2003).
Resat, H., Wiley, H. S., and Dixon, D. A. J. Chem. Phys. 105, 11026 (2001).
Rodgers, S. T., and Jensen, K. F. J. Appl. Phys. 83(1), 524 (1998).
Rondanini, M., Cavallotti, C., Moscatelli, D., Masi, M., and Carra, S. J. Cryst. Growth 272, 52–58
(2004).
Rudd, R. E., and Broughton, J. Q. Phys. Rev. B 58(10), R5893 (1998).
Rudd, R. E., and Broughton, J. Q. Phys. Stat. Sol. 217, 251 (2000).
Rusli, E., Drews, T. O., and Braatz, R. D. Chem. Eng. Sci. 59, 5607 (2004).
Salis, H., and Kaznessis, Y. J. Chem. Phys. 122, 054103 (2005).
Saxton, M. J. Biophys. J. 69, 389 (1995).
Saxton, M. J. Biophys. J. 81, 2226 (2001).
Schulze, T. P. J. Cryst. Growth 263(1–4), 605 (2004).
Schulze, T. P., Smereka, P., and Weinan, E. J. Comp. Phys. 189(1), 197 (2003).
Shea, L. D., Omann, G. M., and Linderman, J. J. Biophys. J. 73, 2949 (1997).
Shenoy, V. B., Miller, R., Tadmor, E. B., Rodney, D., Phillips, R., and Ortiz, M. J. Mech. Phys.
Solids 47(3), 611 (1999).
Shibata, T. J. Chem. Phys. 119(13), 6629 (2003).
Shimizu, T. S., Aksenov, S. V., and Bray, D. J. Mol. Biol. 329(2), 291 (2003).
Siettos, C. I., Armaou, A., Makeev, A. G., and Kevrekidis, I. G. AIChE J. 49(7), 1922 (2003).
Snyder, M. A., Chatterjee, A., and Vlachos, D. G. Comput. Chem. Eng 29(4), 701 (2005).
Snyder, M. A., and Vlachos, D. G. Mol. Sim. 30(9), 561 (2004).
Snyder, M. A., Vlachos, D. G., and Katsoulakis, M. A. Chem. Eng. Sci. 58, 895 (2003).
Srolovitz, D. J., Dandy, D. S., Butler, J. E., Battaile, C. C., and Paritosh J. Metals 49(9), 42 (1997).
Stefanovic, J., and Pantelides, C. C. Mol. Sim. 26(3), 167 (2001).
Stephanopoulos, G., and Fredrickson, A. G. Bull. Math. Biol. 43(2), 165 (1981).
Stundzia, A. B., and Lumsden, C. J. J. Comp. Phys. 127, 196 (1996).
Sunderrajan, S., Hall, C. K., and Freeman, B. D. J. Chem. Phys. 105(4), 1620 (1996).
Tadmor, E. B., Ortiz, M., and Phillips, R. Philos. Mag. A 73(6), 1529 (1996).
Takahashi, K., Kaizu, K., Hu, B., and Tomita, M. Bioinformatics 20(4), 538–546 (2004).
Tammaro, M., Sabella, M., and Evans, J. W. J. Chem. Phys. 103(23), 10277 (1995).
Thattai, M., and van Oudenaarden, A. Proc. Natl. Acad. Sci. 98(15), 8614 (2001).
Thompson, T. B. ‘‘DOE workshop roadmap for computational chemistry’’. http://itri.loyola.edu/
molmodel (1999).
Tian, T., and Burrage, K. J. Chem. Phys. 121(21), 10356 (2004).
Tomlin, A. S., Turanyi, T., and Pilling, M. J. Elsevier. Sci. J. 35, 293 (1997).
Torrie, G. M., and Valleau, J. P. J. Comp. Phys. 23(2), 187 (1977).
Vanden-Eijnden, E. Comm. Math. Sci. 1(2), 385 (2003).
Varma, A., Morbidelli, M., and Wu, H., ‘‘Parametric Sensitivity in Chemical Systems’’. Cambridge
University Press, New York (1999).
Vasudeva, K., and Bhalla, U. S. Bioinformatics 20(1), 78 (2004).
Villermaux, J. (1996). ‘‘New horizons in Chemical Engineering, Plenary lecture in Fifth World
Congress of Chemical Engineering’’, July 15. San Diego, CA (1996).
Vlachos, D. G. AIChE J. 43(11), 3031 (1997).
Vlachos, D. G. Chem. Eng. Sci. 53(1), 157 (1998).
Vlachos, D. G. Appl. Phys. Lett. 74(19), 2797 (1999).
Vlachos, D.G., Molecular modeling for non-equilibrium chemical processes, in (S. Lee, Ed.), ‘‘En-
cyclopedia of Chemical Processing’’, Marcel Dekker, Inc., New York (2004).
Vlachos, D. G., and Katsoulakis, M. A. Phys. Rev. Lett. 85(18), 3898 (2000).
Vlachos, D. G., Schmidt, L. D., and Aris, R. J. Chem. Phys. 93, 8306 (1990).
Vlachos, D. G., Schmidt, L. D., and Aris, R. Surf. Sci. 249, 248 (1991).
Vlachos, D. G., Schmidt, L. D., and Aris, R. Phys. Rev. B 47, 4896 (1993).
Vlad, M. O., and Pop, A. Physica A: Stat. Theor. Phys. 155(2), 276 (1989).
Weinan, E., Engquist, B., and Huang, Z. Y. Phys. Rev. B 67(9), 092101 (2003).
Weinan, E., and Huang, Z. Y. J. Comp. Phys. 182(1), 234 (2002).
Wolf-Gladrow, D. A. ‘‘Lattice-gas cellular automata and lattice Boltzmann models: An introduc-
tion. ‘‘Lecture Notes in Mathematics’’’’. vol. 1725Springer, Berlin (2000).
Woolf, P. J., and Linderman, J. J. Biophys. Chem. 104, 217–227 (2003).
Woolf, P. J., and Linderman, J. J. J. Theor. Biol. 229, 157–168 (2004).
Xu, L., Sedigh, M. G., Sahimi, M., and Tsotsis, T. T. Phys. Rev. Lett. 80(16), 3511 (1998).
Zak, D. E., Gonye, G. E., Schwaber, J. S., and Doyle, F. J. Genome Res. 13, 2396 (2003).
QUANTIFYING PHYSICS AND CHEMISTRY AT MULTIPLE
LENGTH-SCALES USING MAGNETIC RESONANCE
TECHNIQUES
Lynn F. Gladden, Michael D. Mantle and Andrew J. Sederman
Department of Chemical Engineering, University of Cambridge, Pembroke Street,

Cambridge CB2 3RA, UK
I. Introduction 64
II. Principles of MR Measurements 68
A. Spatially Unresolved And Spatially Resolved
Experiments 69
B. Nuclear Spin Relaxation Times 73
C. Transport 76
D. Temperature 83
E. The k-space Raster 83
F. Fast Data Acquisition 87
Engineering Research 92
A. ‘‘Ultra-fast’’ Imaging of Velocity Fields 93
B. Multiple Images From a Single Excitation 96
C. Imaging Rotating Systems 98
D. ‘‘Ultra-fast’’ Diffusion Measurement 99
E. Gas-phase MR 100
IV. Reaction Engineering: From Catalyst to Reactor 103
A. MR Spectroscopy of Catalysts 103
B. Micro-imaging and Molecular Diffusion Studies
of Formed Catalyst Pellets 105
C. Single-Phase Flow in Fixed-Bed Reactors 110
D. Measuring Chemical Composition and Mass Transfer
in Fixed-Bed Reactors: In Situ Studies of Reactions 115
E. Two-Phase Flow in Fixed-Bed Reactors 119
F. Hydrodynamic Transitions in Fixed-Bed Reactors 123
V. Future Prospects 128
Acknowledgments 130
References 131
E-mail: Gladden@cheng.cam.ac.uk
63
DOI 10.1016/S0065-2377(05)30002-0
64 LYNN F. GLADDEN ET AL.
Abstract
Magnetic resonance (MR) is finding increasing use in chemical engi-

neering research. The real power of MR techniques is that by bringing
together spectroscopy, diffusion, micro-imaging and flow imaging, we
have a non-invasive, chemically-specific measurement technique that can
characterise a system over length-scales ranging from Å to the cm-scale.
The aims of this chapter are two-fold: first, to outline the principles of
MR measurements such that they are presented as an integrated set of
measurements clearly based on the same physicochemical phenomena;
and second, to highlight the recent advances in the field, with a focus on
the development of measurement techniques with immediate application
to chemical engineering research. The power of bringing together the full
range of MR measurements to address phenomena occurring over mul-
tiple length-scales is illustrated using examples taken from the field of
chemical reaction engineering.
I. Introduction
In recent years there have been notable developments in measurement science

and technology, particularly in applications of non-invasive measurement tech-
niques to the chemical and process industries. This interest in implementing
modern metrology for the study of multi-component, multi-phase systems is
likely to be central to the development of many aspects of chemical engineering
research in the coming years. Techniques already established in use include
g- and X-ray absorption, ultrasound attenuation, laser and phase Doppler an-
emometry, particle-imaging velocimetry and capacitance tomography (Boyer
et al., 2002). Of course, there remains significant opportunity to identify new
applications of such techniques and to assess and improve the accuracy and
spatial and temporal resolution of these measurements, as well as to use these
new data to aid in the development and validation of new theoretical or nu-
merical approaches. This chapter focuses on recent developments in magnetic
resonance (MR) applications in chemical engineering research, the majority of
the examples being drawn from ongoing work in our own research laboratory.
MR is a measurement technique that is particularly well suited to providing
insight at multiple length-scales. Until recently, each ‘‘family’’ of MR tech-
niques tended to be found in different fields of research. For example, MR
spectroscopy (usually referred to as nuclear magnetic resonance, or NMR,
spectroscopy) was a tool in common use by chemists and to some extent phys-
icists, while MR imaging (or MRI) was the domain of the medical physicist.
QUANTIFYING PHYSICS AND CHEMISTRY AT MULTIPLE LENGTH-SCALES 65
MR diffusion measurements, usually referred to as pulsed field gradient (PFG)

or pulsed gradient spin echo (PGSE) techniques, tended to be a niche area of
activity undertaken by particular research groups in physics and chemistry lab-
oratories around the world. Clearly, given that these techniques are all based on
the phenomenon of the resonant excitation of the nuclear spin system, they can
all be applied to a given sample within the same sample environment. In many
cases, combinations of MR measurements can be integrated into a single MR
experiment. For example, why not image (i.e. spatially resolve) a spectroscopic
or diffusion measurement? It is when the whole family of MR techniques is
brought together that MR has its greatest impact on chemical engineering re-
search, particularly at the level of Å- to cm-scale physical and chemical proc-
esses. It is therefore timely that an article is dedicated to the recent
developments made in this field and their proven and potential applications.
In addition to the ability of MR to probe a hierarchy of length-scales in a
range of systems, the field has been given additional impetus by the development
and implementation of ‘‘fast’’ data acquisition techniques. These will be dis-
cussed in detail in this chapter. The key development has been to reduce data
acquisition times from tens of minutes to tens of milliseconds, making study of
the unsteady state possible while retaining all the attributes of chemical
specificity, non-invasiveness and transport measurement capability inherent to
MR methods. Two major themes are covered in this chapter. First, the recent
technical developments in fast data acquisition will be explained and illustrated.
Second, an applications focus will be developed: we will look at one case study
in particular—that of a fixed-bed catalytic reactor—and review the nature of
information that can be obtained over the range of length-scales characterising
the system.
Before introducing the principles of MR methods it will be useful to illustrate
the multi-scale nature of MR measurements with examples of relevance to
chemical engineering that fall outside the scope of the main case study of re-
action engineering. The first of these is the field of Rheo–NMR. The current
state of the art integrates conventional rheological measurements with MR
experimentation. Commercially available sample environments can be pur-
chased such that, for example, a cone-and-plate rheology measurement can be
performed in situ within the MR magnet, thereby enabling an image of the
velocity field within the sample to be acquired during shear. This new type of
Rheo–NMR measurement has already produced some important insights. For
example, Britton and Callaghan (1997) reported visualisations of anomalous
behaviour such as apparent slip, shear banding and fracture during shear
measurement in worm-like surfactants (Fig. 1). Quite apart from the interesting
rheology under study, this result demonstrates how MR imaging can give us
new insights into the limitations of our day-to-day ‘‘macroscopic’’ materials
characterisation techniques—in this case, the cone-and-plate rheometer. Obvi-
ous questions are: Is the rheometric characterisation provided by the conven-
tional cone-and-plate device giving us all the relevant information about the
FIG. 1. (a) Velocity image and (b) shear rate map for a worm-like surfactant in a 41 cone gap, at a
shear rate of 16 s1 (above critical shear rate). A distinct deviation from a linear velocity gradient
and shear banding are observed. The velocity scale lies between the limits 712 mm/s and the shear
rate between 757 s1. Reproduced with permission from Britton and Callaghan (1997).
in-use properties of our material? How can we use this new information in
optimising product and process performance? In situ Rheo–NMR studies can
also be combined with in situ MR spectroscopy, which then allows us to char-
acterise molecular re-arrangements occurring during shear; such data sets have
enormous potential value in testing and validating theoretical models of pol-
ymer rheology (e.g. Cormier et al., 2001; Cormier and Callaghan, 2002). At
larger length-scales, process applications of Rheo–NMR have been demon-
strated with, for example, routine in-line monitoring of shear viscosity–shear
rate data (Arola et al., 1997; Powell et al., 1994). The effect of shear on the
structural evolution of multi-phase systems during processing operations can
also be probed (Altobelli et al., 1997).
A different area of application of MR is that of solids flow. Visualisation of
the distribution of solids within vessels by conventional means is limited by
opacity; therefore, MR provides one of the few means to visualise structure in
particulate arrays and the flow of solids. MR studies of solids are limited to the
study of materials that give a detectable MR signal under the relevant exper-
imental conditions—this usually restricts studies to particles that have a liquid-
like (i.e. relatively long spin–spin relaxation time—to be discussed later in Sec-
tion II.B) core. Typical systems are oil-filled plastic beads or seeds which have a
naturally high oil content. Early studies focussed on imaging the particle dis-
tribution at rest within the container, following perturbations to the system.
Such studies have been able to probe convection phenomena within granular
systems resulting from vibrations of granular beds (Ehrichs et al., 1995) and to
follow the evolution of segregation in rotating granular flow (Hill et al., 1997).
Imaging measurements have also been reported while motion is occurring. For
example, Seymour et al. (2000) have imaged spatial distributions of collisional
correlation times during three-dimensional (3-D) granular flow in a horizontal
rotating cylinder. A further example has been reported by Metcalfe et al. (1999),
who reported results of experiments designed to investigate axial transport and
core formation in granular systems. Figure 2 shows some images reported from
this study—the detail in the internal structure of the bed is clearly seen. MR
(a) (b) (c)
FIG. 2. Central slice of a rotating tube 75% full, with an aspect ratio of 3, comprising mustard
seeds (MR active) and polystyrene beads (MR inactive). The mustard seeds are identified as the light
grey pixels. The polystyrene beads give no signal intensity. In images (a)–(c) the mustard seeds and
polystyrene beads are seen to segregate rapidly but retain a non-mixing core region. Reproduced
with permission from Metcalfe et al. (1999).
imaging has also been used to visualise formation of bands of high shear within
vibrating granular beds (Caprihan et al., 1997), and to quantify the time-av-
eraged density variations and the random motion of granular particles in the
presence of gas flow in a model gas-fluidised bed reactor (Savelsberg et al.,
2002). Although not performing a spatially resolved (imaging) experiment,
Yang et al. (2002) have used a pulsed field gradient stimulated echo technique to
measure the short time (3-D) displacement, density and granular temperature of
mustard seeds vibrated at 15 g in a vertical column of internal diameter 9.0 mm.
The behaviour of the dense lower region of the sample was adequately described
by inelastic hard-sphere hydrodynamics. However, in the upper layers of the
sample, where the mean free path is long, the vertical and horizontal velocities
became decoupled, and these observations could not be predicted by current
theory.
Hopefully, these examples have given the reader not previously acquainted
with MR methods a feel for the potential uses of this field of measurement
science in their own research. However, to appreciate the wealth of information
that can be obtained from MR measurements, knowledge of the basic principles
of MR measurements is required. Therefore, the structure of this chapter is as
follows. In Section II, the principles of MR measurements will be described.
Sections III and IV report two areas of MR research in which developments of
immediate relevance to chemical engineering research have been made. Section
III addresses the technical developments made in the field of rapid data acqui-
sition, and examples of data acquired using these new techniques are given.
Areas of application are illustrated, examples being chosen from outside the
field of reaction engineering to broaden the scope of the chapter. Section IV is
dedicated to a case study of reaction engineering studied by MR over a range of
length-scales from the Å- and nm-scale behaviour of reactant molecules within
the catalyst and the nature of the active site and pore structure of the catalyst, to
the macro-scale hydrodynamics characterising single- and two-phase flow with-

in a fixed-bed reactor packed with catalyst pellets. It is hoped that the reader
will see that the same combination of MR methods can be applied to many
areas of chemical engineering research including the optimisation of oil recovery
processes and the design and control of the manufacture of controlled release
pharmaceutical and agrochemical delivery systems, to name just two.
Before proceeding further, it is important that the limitations of MR meas-
urements are identified. In general, we assume that ferromagnetic and para-
magnetic systems cannot be studied. However, study of samples and sample
environments containing such materials (in small amounts) is sometimes pos-
sible. Clearly, from a practical point of view, large ferromagnetic objects cannot
be handled within and close to a superconducting magnet. However, sample
environments comprising modest amounts of aluminium and brass can be used.
With respect to the sample itself, the ability to study a given system is very
material-specific. Ferromagnetic and paramagnetic particles, present even at
parts per million levels, act to distort the local magnetic fields and relaxation
times within the sample, thereby making all studies based upon quantitative
analysis extremely difficult. However, depending on the nature of the informa-
tion that is of interest, systems containing such species can be addressed. MR is
also considered to be inherently insensitive compared with other spectroscopic
techniques; i.e. it takes 1015 nuclear spins to generate a detectable signal. For
this reason, gas-phase studies have been a rather specialist area. However, as we
will see later (Section III.E), gas-phase studies are now very much a reality with
respect to both direct gas-phase imaging and the use of gases as probe molecules
to characterise species adsorbed at interfaces.
II. Principles of MR Measurements
There are a large number of different MR techniques that, at first glance,

seem quite unrelated. Perhaps the simplest way to look at MR methods is to
recognise that they will fall into one of two categories—those that do not re-
quire any spatial resolution or positional encoding within the data, and those
that do. If you need to encode spatial information in some way, then in addition
to a basic MR experiment, which is typically performed in a large, supercon-
ducting magnetic field, additional smaller linear gradients in the magnetic field
will also be applied at some point during the measurement. This section is set
out as follows. In Section II.A, the concepts behind spatially unresolved and
spatially resolved experiments are introduced. The simplest MR measurement
records a signal proportional to the absolute number of nuclear spins within the
system; careful calibration of signal intensity against a reference sample con-
taining a known number of nuclear spins allows the absolute number of spins
present in the sample to be quantified which, in turn, yields the number of
species of interest present. Extensions to the MR measurement allow the signal

obtained to give quantitative information on the physical–chemical environ-
ment of those spins as well as on the temperature and transport processes
associated with them—the principles of these measurements are presented in
Sections II.B–II.D. In Section II.E, the actual ‘‘working’’ formalism for under-
standing imaging experiments (known as ‘‘pulse sequences’’) is introduced, and
this is then used, in Section II.F, to introduce the principles of ‘‘fast’’ MR
measurements. There is no attempt here to give a detailed introduction to the
principles of MR techniques; the interested reader should refer to excellent texts
by Callaghan (1991) and Kimmich (1997). A very basic introduction to under-
standing MR spectroscopy measurements and simple imaging sequences
(Gladden, 1994) and a more recent, detailed review of fast imaging sequences
of relevance to chemical engineering (Mantle and Sederman, 2003) may also be
of interest.
A. SPATIALLY UNRESOLVED AND SPATIALLY RESOLVED EXPERIMENTS
1. Spatially Unresolved Measurements

A spatially unresolved experiment is usually taken to imply a spectroscopy
experiment, although relaxation time studies (Section II.B), transport (Section
II.C) and temperature (Section II.D) measurements are also often measured as
‘‘bulk’’ properties. The spatially unresolved spectroscopy experiment is the
simplest type of MR experiment, and the principles upon which this measure-
ment is based underpins all types of MR measurement, be they spatially un-
resolved or spatially resolved.
When a nucleus of non-zero nuclear spin quantum number is placed in an
external magnetic field (typically a superconducting magnetic field of 2–10 T), its
nuclear spin energy levels become non-degenerate. As a result of this, at the
equilibrium state of the spin system, there exists a net magnetisation vector
aligned parallel to the direction of the external magnetic field, assumed to be
along the z-direction. By exposing the system to electromagnetic energy of ap-
propriate frequency (radio-frequency (r.f.)), a resonant absorption occurs be-
tween these nuclear spin energy levels. The specific frequency at which this
resonance occurs is called the resonance (or Larmor) frequency, and is pro-
portional to the strength of the external magnetic field, B0, used in the exper-
iment. The precise energy-level splitting is specific to a given isotope of an
element, and the resonance frequency (o0) is given by
o0 ¼ gB0 (1)
where g is the gyromagnetic ratio, which is an isotope-specific property. The

precise energy-level splitting is slightly modified by the electronic environment
of the nucleus under study; thus, o0 is also modified and becomes specific to
individual molecules containing the element of interest. We can therefore take a

spectrum of a mixture of chemical species and identify the presence of particular
molecular species in that mixture (i.e. a conventional NMR or MR spectroscopy
experiment). In principle (see Section II.B), the measurement is quantitative; i.e.
following calibration we know exactly how much of each chemical species is
present. A standard way of representing the basic MR measurement is shown in
Fig. 3. Initially, the net magnetisation vector, M, is aligned along the direction
of the magnetic field. The action of the excitation pulse, in this case a pulse of
r.f. applied at right angles (along x0 ) to the direction of the superconducting field
is to rotate M about the x0 -axis. In this example, the r.f. excitation is applied for
sufficient time so that M is rotated to lie along the y0 -axis in the x0 –y0 plane. If
this condition is met, the r.f. pulse is referred to as a p=2 (or 901) pulse; i.e. it has
rotated M through p=2 radians. These processes actually occur in the ‘‘rotating
frame’’ of reference (hence the primed symbol) which, in the laboratory frame,
precesses about the z-axis (i.e. about B0) at the Larmor frequency. This con-
z' z' z' z'

M(t=0)
(i)
y' y' y' y'
x' x' x' x'
(/2)x'
r.f.
(ii)
time
My'
time (iii)
FT
*
1/(T2) (iv)
frequency
FIG. 3. The behaviour of the magnetisation vector (i) is shown in response to the application of a
single p=2 r.f. pulse along the x0 -direction, (ii). The decay of the magnetisation vector in the x0 –y0
plane yields the received time-domain signal, called the FID, shown in (iii). The result of a digital
Fourier transform of the FID is the spectrum shown in (iv). For a liquid-like sample, the full-width
at half-maximum-height of the spectral signal is 1=pT 2 (see Section II.B).
vention is adopted to simplify the representation of the action of the r.f. pulses.
In this rotating frame representation, the MR time-domain signal, following r.f.
excitation, is measured by acquiring the signal (i.e. magnitude of the magnet-
isation vector) aligned along y0 as a function of time; this signal will decay with
time owing to the recovery of the magnetisation along z0 and, at shorter time-
scales, owing to the loss of phase coherence of the spin isochromats comprising
the net magnetisation vector along the y0 -axis. These decay processes are termed
the spin–lattice and spin–spin relaxation processes and will be discussed further
in Section II.B. The decay of the magnetisation along the y0 -axis is recorded as a
decaying voltage in a receiver coil. Fourier transform of this time-domain
signal, usually referred to as the free induction decay (FID) yields the frequency
domain spectral response in which the area under the spectral peak, following
appropriate calibration, gives a quantitative measure of the number of nuclear
spins associated with that spectral frequency (i.e. a quantitative measure
of the number of molecules of a given molecular species present). Thus, MR
is an intrinsically chemical-specific, quantitative measurement. This is the esse-
ntial attribute that makes it such a powerful tool in science and engineering
research.
2. Spatially Resolved Measurements

In the context of this chapter, spatial resolution refers to any MR measure-
ment that requires identification of the spatial location of nuclear spins. As
such, spatially resolved measurements include both imaging and transport
measurements. To achieve a spatially resolved measurement, the same physical
principles and experimental ‘‘excitation-acquire’’ strategies required for the
spatially unresolved measurements still apply. The basic spatially unresolved
experiment is still performed, but by applying a spatially varying magnetic field,
in addition to the large static field B0, the resonance frequency of species within
the sample becomes a function of their position along z and the strength of the
applied gradient. Thus, for a magnetic field gradient applied along the
zdirection, Gz:
oz ¼ gðB0 þ G z zÞ (2)
Clearly, this is the basis of an imaging experiment; the measurement can be

calibrated such that the relationship between resonance frequency and spatial
position is known. Figure 4 illustrates the basic principles of an imaging ex-
periment. Without application of the linear gradient in the magnetic field we
perform a spatially unresolved experiment; i.e. the water in both test tubes
resonates at the same frequency. Therefore, we see only one MR signal, which is
a quantitative measure of the total amount of water in the two test tubes. Upon
application of the field gradient, the water at every spatial location along the
direction of that gradient has a different resonance frequency; therefore, we
acquire an FID that represents, after Fourier transformation, a 1-D spatial
field gradient
no with
gradient gradient
FT spatial separation, ∆z FT
ω
frequency, ω
frequency separation, ∆ω
FIG. 4. Consider two test tubes containing different amounts of water. Spatial resolution is ob-
tained by applying a linear gradient in the magnetic field, which makes the resonant frequency of the
nucleus of interest a function of its position in real space. Without the presence of the field gradient,
the water within the two tubes resonates at the same frequency and a single-peak spectrum is
obtained, the area under it being a quantitative measure of the total amount of water in the two
tubes. Upon application of the field gradient, the resonant frequency of the water molecules becomes
a function of their position along the direction of the applied field gradient. Fourier transform of the
acquired signal yields a 1-D profile of the amount of water present. The area under each ‘‘peak’’
gives the amount of water in each tube.
projection (along the direction of the applied gradient) of the amount of

water in the two tubes. 2-D and 3-D images are acquired by applying gradients
in 2 and 3 orthogonal directions, respectively. However, just as importantly, we
see from Eq. (2) that if the application of a magnetic field gradient encodes a
given nucleus by its spatial position, then by developing a pulse sequence which
allows us to measure the change in position in a given time interval, we have the
basis of a MR measurement of transport. This is indeed the principle of the
majority of MR transport measurement techniques, e.g. PGSE, often called
PFG, measurements of molecular diffusion and MR flow-imaging measure-
ments. The detailed principles of transport measurements are described in Sec-
tion II.C.
In addition to measurements of ‘‘how much’’ and ‘‘what type’’ of chemical
species is present, modification of the MR experiment allows us to quantify the
physical state of that species (e.g. gas, liquid, gel, solid), temperature and, as has
just been alluded to, any incoherent or coherent transport processes within the
system of interest. By integrating any of these measurements into an imaging
experiment, we can spatially map these quantities or exploit the effect of these
characteristics on the magnitude or frequency of the MR signal to preferentially
observe sub-populations of spins within the system. In this latter application we
are exploiting so-called ‘‘contrast’’ mechanisms in the image acquisition. These
concepts will be illustrated in Sections II.B–II.D.
B. NUCLEAR SPIN RELAXATION TIMES
In the following section, the principles of nuclear spin relaxation processes

will be summarised and their use in data acquisition discussed. Following the
application of the r.f. excitation pulse, the nuclear spin system has excess energy.
To return to equilibrium the spin system has to recover its initial energy and
entropy states a process known as ‘‘relaxation’’. A number of different relax-
ation times characterise the mechanisms for the processes involved. The most
important are the spin–lattice relaxation (T1) and spin–spin relaxation (T2) time
constants. These time constants characterise the physicochemical environment
of the molecules being studied. T1, as the name suggests, characterises the en-
ergy exchange between the excited spin and the surrounding physical environ-
ment (i.e. the lattice), while T2 characterises the loss of phase coherence between
nuclear spins within the nuclear spin ensemble. If a system is characterised by a
very small T2 (e.g. many solids) it may not be possible to study it using MR; this
is the major limitation in imaging the solid state. Each chemical species will have
its own T 1 =T 2 characteristics, and these will vary depending on the physical
state in which that species exists.
1. Spin– Lattice Relaxation, T1

As shown in Fig. 5, before application of the r.f. excitation pulse the net
magnetisation vector associated with the nuclear spin system is aligned along
the direction of the static magnetic field. It is the magnitude of this vector that
provides the quantitative measurement of the number of nuclear spins excited
within the sample. After excitation by a p/2 r.f. pulse applied along the x0 -axis,
the magnetisation vector is rotated through p/2 to lie along the y0 -axis. As soon
as the excitation stops, the system acts to return to equilibrium; this corresponds
to a monotonic increase in the magnitude of the magnetisation vector back
along z0 (z) as a function of time. If we wait a short time, only a fraction of the
magnetisation will have been re-established along z0 . If we wait 5–7 T1, the full
magnitude of the magnetisation will have recovered along z0 . The magnitude of
the magnetisation vector along z0 , Mz0 , as a function of the ‘‘waiting’’ time, t,
can be written down analytically for any specific r.f. pulse sequence. Equation
(3) describes the recovery of the magnetisation back along z for a saturation
recovery pulse sequence:
M z0 ðtÞ ¼ M 0 ½1 expðt=T 1 Þ (3)
By recording Mz0 for a number of t values and fitting these data to Eq. (3), both
the T1 characterising the system and the value of M0 (which quantifies the
number of initially excited spins) are obtained. In a spatially resolved ‘‘relax-
ometry’’ experiment, images are acquired at different values of t, and a fit of Eq.
(3) to the intensity as a function of t, for the equivalent pixel, i, in each image
allows a complete map of M 0i and T 1i to be obtained. Thus, spatial variation in
M0 z'
y'
x'
a)
M0 z'
y'
x'
b)
intensity
signal
time
(c)
FIG. 5. (a) As described in Fig. 3, the action of the p=2 pulse (applied along the x0 direction) is to
rotate the magnetisation vector into the x0 –y0 plane, along the y0 -direction. The individual spin
isochromats then dephase in the x0 –y0 plane, as shown by the increasing size of the shaded region
with time. (b) At timescales longer than T2, the magnetisation recovers back along the direction of
the magnetic field B0, with a characteristic time constant T1. (c) Two different species within the
same sample may have different characteristic T1 values. In this example, the species associated with
the black arrows has a shorter T1 than the species associated with the grey arrows; the arrows
indicate the magnitude of the acquired signal intensity following the initial r.f. excitation. If data are
acquired at long times after r.f. excitation, equal signal intensity will be acquired from both species.
However, if data are acquired very soon after the excitation pulse, the acquired signal will be
predominantly associated with the species characterised by the shorter T1. This illustrates the prin-
ciple of relaxation contrast.
T1 can be mapped throughout the image. Figure 5 also demonstrates that the
magnitude of the signal we will acquire will depend on the time at which we
acquire the signal. Thus, if we have two species with different T1 characteristics,
by careful selection of the delay time between excitation and acquisition of the
resulting signal, a signal can be acquired preferentially from one of the com-
ponents.
2. Spin– Spin Relaxation, T2

On timescales of less than or equal to that of T1, spin–spin relaxation (T2)
processes occur. T2 characterises the loss of phase coherence of the individual
spin isochromats within the spin ensemble comprising the total magnetisation
vector M0. A spin isochromat represents a group of spins that experiences the
same homogeneous magnetic field and which, therefore, behaves in the same
way following the excitation pulse. During the period following excitation, the
individual isochromats will lose phase coherence with each other as a result of
spin–spin interactions and local variations in B0. The decay of the coherent
magnetisation aligned along y0 , due to spin–spin interactions but not magnetic
field heterogeneities, is characterised by the time constant T2 and measured
using a ‘‘spin-echo’’ pulse sequence, as shown in Fig. 6. T2 is defined as follows:
M y0 ðtÞ ¼ M 0 expðt=T 2 Þ (4)
With reference to Fig. 6, t ¼ 2td in Eq. (4). The p pulse acts to reverse the de-
phasing effects due to the local heterogeneities in B0 such that the final acquired
signal (the ‘‘echo’’) suffers attenuation due to spin–spin interactions only. The
spin-echo shown in Fig. 6, or rather ‘‘echoes’’ in general (since they can be
produced by actions other than a p pulse) have widespread use in MR methods,
far beyond simple measurement of T2. In short, by using an echo sequence,
instead of exciting the system and then allowing the magnetisation to decay to
zero as in Fig. 3, the majority of the magnetisation can be recovered for use in
subsequent measurements. The simple ‘‘echo’’ sequence shown in Fig. 6 is a
common feature of MR imaging pulse sequences (e.g. Fig. 12).
An additional and important relaxation time constant is T 2 , which charac-
terises a faster decay of the magnetisation along y0 and accounts for, in par-
ticular, the additional effects of magnetic field heterogeneities on the loss of
phase coherence of the magnetisation. Thus, the simple pulse-acquire sequence
(with no re-focusing) shown in Fig. 3, will give a response in which the envelope
(/2)x' ()y' echo
r.f.
0 time
td 2td
z'
1 2 5 4
3 3
y' 4
5 1 2
x' M(2td)
(a) (b) (c) (d)
FIG. 6. A spin-echo pulse sequence used to determine T2. (a) A (p/2)x0 pulse puts M0 into the y0 -
direction, and (b) the spin isochromats dephase with time. At a time td later, a p pulse is applied
along the y0 -axis causing the spins to rotate through p radians (c) such that they ‘‘refocus’’ along the
y0 -axis to form an ‘‘echo’’ at time. 2td.(d) The decrease in magnitude of the magnetisation vector
between stages (a) and (d) provides a measure of T2 [Eq. (4)]. All ‘‘reversible’’ contributions to the
spin-spin relaxation process are removed by the application of the p pulse.
of the decay in the time domain and hence the width of the frequency-domain
signal is characterised by T 2 .
In summary, the reasons that the nuclear spin relaxation times, and in par-
ticular T1 and T2, are so important are:
(i) Each molecule in a given physicochemical environment is characterised by

specific values of T1 and T2. Hence, by knowing the T1 and T2 charac-
teristics of the species in each state of matter or for each chemical species in
a mixture we are able to obtain a wealth of information about a multi-
component, multi-phase system.
(ii) The timings of the MR experiment (pulse sequence) allow us to control the
extent to which we acquire a signal from the entire spin population. To
acquire spectra or image data that are fully quantitative in terms of the
signal intensity yielding a true measure of the number of species of interest
present, the relaxation times of that system must be measured and the pulse
sequences optimised such that any reduction in signal intensity leading to
loss of quantitation is avoided. If, as in some cases, it is impossible to
implement a pulse sequence without the effects of relaxation ‘‘contrast’’ on
the signal intensity, it is possible (but often non-trivial) to correct the signal
intensities to their true values as long as the accurate values of the appro-
priate relaxation times are known.
(iii) In a multi-component or multi-phase system, each component/phase will be
characterised by different relaxation times. Therefore, the timings in the
MR pulse sequence can be set so that the signal is preferentially acquired
from one component/phase. This approach has been exploited in, for ex-
ample, studies of the separation of an oil–water emulsion in which the T1
characteristics of the oil and water phases are significantly different (Kau-
ten et al., 1991). Relaxation contrast can also be used to discriminate be-
tween liquid and solid phase of the same material during a crystallisation
process (e.g. Simoneau et al., 1991).
C. TRANSPORT
Transport processes are measured by means of applying pulsed magnetic field

gradients to the system, in addition to the normal r.f. pulses. PFG techniques
measure molecular displacement as a function of time without the need for
introducing tracers into the experiment. The principle of the experiments is easy
to understand although the detailed implementation of the experiments is
somewhat more challenging. The application of a pulsed field gradient at the
beginning of an experiment (i.e. immediately after r.f. excitation) encodes a
given spin with a ‘‘label’’ describing its position along the direction of that
applied field gradient. At a time D later, referred to as the observation time, a
second pulsed field gradient is applied. The net effect of applying these two
gradients separated by time D is that we can monitor the distance travelled

during a known time and hence quantify the transport process of interest.
Figure 7 shows the principles of transport measurement. When considering
the application of pulsed magnetic field gradients to measure transport proc-
esses we use the lower-case symbol g, as opposed to G, which is reserved for use
in describing the imaging gradients [see Eq. (2) and Section II.E]. Perhaps the
most important point to appreciate with respect to MR measurements of
transport is that the same measurement methodology is used to quantify in-
coherent (e.g. diffusion, dispersion) and coherent (e.g. flow) processes occurring
within the same system. The basic principle derives directly from Eq. (2). If the
magnetic field gradient is applied for a short time period (ie. a ‘‘pulse’’), as
opposed to ‘‘continuously’’ during which time data are acquired, instead of
imposing a time-independent modified resonance frequency on a nucleus as
determined by its spatial position, the nuclear spin is given phase offset (say f1 )
time
g
(i) δ (ii) (iii) (iv) (v)
z no net phase shift

no motion
flow
net phase shift
coherent motion
net reduction in
amplitude
random diffusional motion
FIG. 7. The principle of transport measurements using the ‘‘phase shift’’ approach. Two pulsed
magnetic field gradients (of magnitude g and duration d) are applied a time D apart. The cases of no
motion, coherent motion (i.e. constant velocity) and random diffusional motion are shown. The
schematics show the relative phase offsets of the spin isochromats initially at different positions in z
along the length of the sample. (i) Initially all the spins are aligned in the rotating frame. (ii) The first
gradient pulse applies a phase offset to the spin isochromats depending on their position along the z-
direction. (iii) The position of the spin isochromats after the systems has evolved for the time D. (iv)
The orientation of the spin isochromats after the action of the second, equal and opposite polarity
gradient pulse. (v) The magnitude and phase shift of the net magnetisation vector after application of
this bipolar gradient pair (i.e. equal and opposite) pulse sequence.
after application of the pulse characteristic of its spatial position when the pulse
was applied. In the rotating frame of the spin system, this phase offset, f1 , is
equal to ggdz1 , where d is the duration of the applied gradient, z1 the position of
the spin and g the magnitude of the magnetic field gradient along the z-direc-
tion. Although many variations on the theme exist, the basic concept under-
pinning the vast majority of transport measurements is that after an observation
time, D, an equal but opposite polarity magnetic field gradient pulse is applied,
which gives the spins a further phase offset, f2 , such that the total phase offset is
f1 þ f2 ¼ ggdðz1 z2 Þ. Clearly, if the molecule (i.e. spin) has not moved during
the time D, it will experience a net phase shift of f1 þ f2 ¼ 0; i.e. the mag-
netisation vector will again be aligned along the y0 -axis, as it was immediately
after application of the initial excitation pulse. However, if the molecule has
moved during the time D, i.e. z1 az2 , then f1 þ f2 a0, and observation of the
magnetisation will show a phase shift that is proportional to the distance moved
(z1z2). Since g, g and d are known, the displacement or average velocity over
the timescale D is obtained. A typical transport measurement would proceed by
making several measurements at differing values of d or g and recording the
resulting phase shift and amplitude of the signal.
Let us now consider, in detail, the effect on the acquired signal of a coherent
transport process (i.e. the molecules move with a velocity v in the direction of
the applied pulsed gradients). With reference to Fig. 7, we see that the effect of
the second pulsed gradient is to re-align the spin isochromats with each other,
but at an increasing angle (phase offset) with respect to the y0 -axis. As d or g
increases, the net magnetisation will rotate through the x0 –y0 plane of the ro-
tating frame. This manifests itself as a continuously increasing phase shift, while
the magnitude of the magnetisation vector (i.e. signal amplitude) remains con-
stant (ignoring relaxation effects). If only the real component of the complex
signal is recorded (i.e. we observe the magnitude of the magnetisation vector
projected along y0 ), an oscillatory signal is recorded as a function of dg, and the
period of the oscillation is directly related to the velocity of the moving spins. In
the case of an incoherent transport process, the random molecular displacements
cause a random distribution of phase shifts of the individual spins and the
acquired signal is a vector sum of these phase shifts. As d or g increases, the
magnitude of the acquired signal decreases monotonically. An interesting ex-
tension to this is when the diffusion occurs within a confined geometry (e.g. an
emulsion droplet). In this case, the distance travelled is constrained to a max-
imum value. Therefore, by taking measurements at increasing values of D, a
value of D is reached, above which no further signal attenuation is measured—
this value of D quantifies the typical dimension of the discrete phase. When
pulsed magnetic field gradients are applied to study diffusive processes, the MR
technique is often referred to as PFG or PGSE-MR. Application of PGSE-MR
techniques to quantify molecular diffusion was pioneered by Stejskal and Tan-
ner (Stejskal, 1965; Stejskal and Tanner, 1965), and the techniques typically
probe molecular displacements of 106–105 m over timescales of order
103–1 s. An overview of applications of PGSE-MR in chemical engineering is

given elsewhere (Gladden, 1994).
Transport measurements performed using pulsed magnetic field gradients are
more clearly understood using a more mathematical framework. It follows from
Eq. (2) that the phase shift (i.e. the instantaneous phase offset in resonance
frequency) fðtÞ acquired in the rotating frame following application of a mag-
netic field gradient, g, along the z-direction, will be
Z t
fðtÞ ¼ g gðtÞzðtÞ dt (5)
0
We also know that the change of position with time of a ‘‘spin’’ or its associated
magnetic moment can be written as
zðtÞ ¼ z0 þ vt þ 12 at2 þ (6)
where z0 is initial position, v the velocity and a the acceleration in the direction
of the gradient. Substituting Eq. (6) into Eq. (2) gives:

oðtÞ ¼ g B0 þ g z0 þ vt þ 12 at2 þ (7)
The total relative phase of the MR signal is then calculated by considering the
time integrals of the individual terms on the right-hand side of Eq. (7). These
integrals are the moments of the magnetic field gradient and the zeroth, first, and
second are proportional to
Z
zeroth moment : z0 gðtÞ dt (8)
Z
first moment : v gðtÞt dt (9)
Z
1
second moment : a gðtÞt2 dt (10)
2
Let us now consider the action of the two equal and opposite pulse gradients
(referred to as a bipolar pair) shown in Fig. 7 of amplitude 7g and length d,
separated by time D. The first pulse, +g, will, in the absence of relaxation, cause
a phase shift according to the integral defined by the zeroth moment:
Z d
gðtÞ dt ¼ ½gtd0 ¼ gd (11)
0
The second, equal and opposite, gradient pulse will have a zeroth moment given
by
Z dþD
gðtÞ dt ¼ ½gtdþD
D ¼ ðgd gDÞ ðgDÞ ¼ gd (12)
D
Addition of Eqs. (11) and (12) gives the total relative phase shift which, for the
case of no motion, is clearly zero.
Now consider the evaluation of the first moment when the magnetic moment
undergoes motion at a constant velocity v in the direction of the gradient. The
total first moment is given by
Z d Z Dþd
1
gðtÞt dt þ gðtÞt dt ¼ g d2 0 D2 2dD d2 þ D2 ¼ gdD (13)
0 D 2
Therefore, the residual phase shift of the MR signal for a magnetic moment
undergoing uniform motion with velocity v for a set of bipolar gradients, 7g, of
duration d separated by time D is gvgdD, i.e. the measured phase shift is
linearly proportional to the velocity.
In practice, we may wish to measure only one of the moments (Eqs. (8)–(10)),
thereby removing the sensitivity of the measurement to position, velocity or
acceleration. This is done by modifying the basic transport measurement pulse
sequence (Fig. 7) so that the integrals are zero for all moments except the one
that is to be measured. These so-called ‘‘compensated’’ pulse sequences have
been reviewed in detail by Pope and Yao (1993a). Figure 8 shows the effect of
gradient nulling on a system of mustard seeds in a rotating drum (Fukushima,
1999). Figure 8a is taken with a sequence in which only the zeroth moment of
(a) (b) (c)
FIG. 8. Mustard seeds flowing in a drum. (a) Only the zeroth moment is nulled, and the image
shows significant signal attenuation (visible as the lighter shades in the sliding layer) due to the
velocity distribution in each pixel and higher-order contributions. (b) The first moment is nulled.
There is no signal loss due to the velocity distribution; the remaining signal attenuation reflects
velocity fluctuation and diffusion in the presence of background gradients. (c) Diffusion image. The
darkness of the image is proportional to the diffusion coefficient D perpendicular to the plane of the
image, calculated from signal attenuation in the presence of a gradient applied along the cylinder
axis. All images were acquired at a 1H frequency of 80 MHz. Reprinted with permission from
Fukushima (1999), copyright 1999, Annual Reviews www.annualreviews.org.
the magnetic field gradients is zero. The signal loss along the exposed edge,
visible as lighter shades in the image, is due to the velocity distribution as well as
higher order motions. Figure 8b was recorded with an imaging sequence in
which the first moment was made zero. The signal is much stronger because
there is no signal loss due to the velocity distribution. Figure 8c shows an image
that has been encoded for diffusive motion along the cylinder axis; darker
regions correspond to a higher diffusion coefficient.
Another type of experiment commonly used when wishing to characterise
transport phenomena is the propagator measurement. The propagator gives a
statistical description of the evolution of motion characterising the system; it
provides a complete description of the random (e.g. diffusion) as well as co-
herent motions. The propagator, Ps(r|r0 ,t), gives the probability of finding a spin
initially at r at time t ¼ 0, and at r0 after a time t. If the propagator only depends
on the displacement R ¼ r0 r, we can define the average propagator:
Z
P̄s ðR; tÞ ¼ Ps ðrjr þ R; tÞrðrÞ dr (14)
where r(r) is the spin density (i.e. number density of MR active nuclei) at
position r. Let us consider the same pair of gradients as shown in Fig. 7. In the
case that a molecule moves from z0 to z0 during the time D, the net phase shift,
F, following the second gradient pulse is determined by the zeroth moment,
provided that the displacement of the molecule is small during the time the pulse
is applied (d):
F ¼ gdgðz0 z0 Þ (15)
Defining a dynamic displacement Z ¼ z0 z0 , and the average displacement

propagator Ps ðZ; DÞ as the average probability that any molecule in the sample
will move by a displacement Z over time D, the acquired signal (relative to that
acquired when no magnetic field gradients are used) for a population of spins
characterised by a range of displacements is given by
Z
Ps ðZ; DÞ expði2pqZ Þ dZ (16)
where q ¼ ð1=2pÞgdg is the reciprocal displacement vector (Callaghan, 1991).

The average displacement propagator distribution, Ps ðZ; DÞ, is obtained by
Fourier inversion of the acquired MR signal. The propagator measurement is
equivalent to a tracer measurement in which the tracer is introduced into the
flow and the average distribution of tracer from its location determined in a
completely non-invasive manner. Figure 9 shows propagators determined for
water flow within a packed bed of spheres. The major features of propagator
measurement are clearly seen. As the observation time increases, the peak in the
propagator occurs at a greater displacement, and the width of the propagator
Ps (Z,∆) (a.u.)
2
−1 0 1 2 3 4
Displacement, Z (mm)
FIG. 9. Displacement propagators recorded for flow of water through a packed bed of 1 mm
diameter glass beads packed within a 10 mm diameter column. The average flow velocity was
0.77 mm/s, corresponding to Pe and Re of 350 and 0.77, respectively. Propagators are shown for
observation times, D ¼ 0:3 s (—), 1 s (- - - -) and 2 s ( ).
distribution increases, reflecting the magnitude of molecular diffusion and dis-

persion phenomena occurring within the bed.
For obvious reasons, the methods just described are termed ‘‘phase shift’’
measurements of transport. These methods are considered the most robust and
quantitative, and are therefore the most commonly used. Another approach is
time-of-flight (TOF) imaging. TOF or ‘‘spin tagging’’ methods were first re-
ported by Singer (1959), and their use has been widespread since then, partic-
ularly with respect to velocity measurement, although the same measurement
also probes other transport processes. At its simplest, the TOF approach mon-
itors velocity by the attenuation observed in the acquired image—no absolute,
direct measure of velocity using pulsed gradients is employed. The principle is
that a set of spins are given an initial excitation pulse—signal will only be
acquired from these excited spins at a given time later. Thus, if we excite a set of
spins in a plane and then acquire a signal from that plane a time, D, later, the
signal will be reduced in the positions at which the fluid has moved fastest, i.e.
the excited spins will have moved out of the image plane to be replaced by fast
moving spins that have moved into the image plane during D; these spins will
not have received the initial excitation and therefore will not give any signal
upon data acquisition. There are many variants of this approach (e.g. Pope and
Yao, 1993b) but the principle remains the same. Because these methods rely on
image intensity to determine the fluid velocity, calibration is necessarily relax-
ation dependent and quantification can be difficult. An extension of this ap-
proach is the Delays Alternating with Nutations for Tailored Excitation
(DANTE) method, in which a grid of spins is excited in the imaging plane and
(a) (b)
FIG. 10. Example of a DANTE-type velocity image. 1H 2-D image of a 2.0 mm thick longitudinal
slice at (a) zero flowrate of water and (b) a flowrate of 486 cm3/min. Flow is from left to right.
the motion of the spins observed at a later time (Mosher and Smith, 1990). An
example of a DANTE-TOF image of laminar flow in a pipe is shown in Fig. 10.
D. TEMPERATURE
Temperature can be mapped by its effect on nuclear spin relaxation (Doran et

al., 1994; Jezzard et al., 1992), resonance frequency (usually termed ‘‘chemical
shift’’) (Bertsch et al., 1998; Hall et al., 2001) and diffusion (Le Bihan et al.,
1989). Because temperature influences so many MR characteristics of a system,
great care must be taken to ensure that the effect on the MR signal that we
assign to temperature is influenced only by temperature or that other influences
can be quantified and hence ‘‘deconvolved’’ from the temperature measurement.
For example, if a chemical reaction occurred within the sample of interest, the
chemical shifts of the individual chemical species may be modified slightly by
the change in mixture composition as well as by temperature variation within
the sample. However, careful studies have been performed, and MR offers the
opportunity for non-invasive temperature measurement. A recent review of this
field has been given by Nott and Hall (1999). Figure 11 shows an example of
temperature mapping reported by Bows et al. (2001) that compares 3-D tem-
perature maps with maps of structural heterogeneity within a jar containing
soup and meat balls.
E. THE K-SPACE RASTER
While the simple schematic of Fig. 4 allows us to appreciate the concept of

obtaining spatial resolution in the measurement, it is almost impossible to un-
derstand and design MRI pulse sequences using this approach. Instead, the
approach used is that of the so-called k-space raster, introduced by Mansfield
(1977).
(a)
(b)
FIG. 11. 3-D (a) temperature and (b) spatial heterogeneity images of a 500 cm3 jar containing soup
with meat balls that has been immersed into a water bath between 95 and 98 1C for 11 min, displayed
as a set of 2-D sections. Spatial resolution is 0.78 mm (along cylindrical axis) 3.13 mm 3.13 mm.
Reproduced with permission from Bows et al. (2001).
Re-writing Eq. (2) for the general case of the variation of resonance frequency
with spatial position r:
oðrÞ ¼ gðB0 þ G rÞ (17)
and neglecting the influence of relaxation on signal intensity, the transverse
magnetization—and therefore the acquired signal, dS—in an element of volume
dr at position r with spin density r(r) is given by
dSðG; tÞ ¼ rðrÞ exp½ioðrÞt dr (18)
Inserting Eq. (17) into Eq. (18) gives
dSðG; tÞ ¼ rðrÞ exp½iðgB0 þ gG rÞt dr (19)
A transformation into the rotating frame of reference allows us to re-write Eq.
(19) as
ZZZ
S ðtÞ ¼ rðrÞ exp½igG rt dr (20)
Mansfield and Grannell (1973) simplified the interpretation of Eq. (20) and the
development of imaging pulse sequences by introducing the concept of k-space,
where the k-space vector is defined as k ¼ ðgGt=2pÞ. It follows that Eq. (20) can
now be written in terms of the k-space vector as
ZZZ
S ðkÞ ¼ rðrÞ exp½i2pk r dr (21)
and the spatial distribution of spins is given by the inverse 3-D Fourier trans-
form:
ZZZ
rð r Þ ¼ S ðkÞ exp½i2pgk r dk (22)
Thus, the imaging experiment is seen as acquisition of data in the time domain
and the sampling of the k-space raster, followed by Fourier transformation to
the frequency domain, which in turn is directly related to real space. An imaging
sequence can now be understood. Figure 12 shows a schematic of a simple 2-D
imaging sequence. In this case let us assume that the sample is cylindrical and
oriented along the z-axis, and an xy image is to be recorded. The first com-
ponent of the pulse sequence is the so-called ‘‘slice selection’’ phase. This com-
prises the application of a narrow band r.f. excitation simultaneously with a
magnetic field gradient imposed along the direction in which the 2-D image is to
be taken (i.e. along the z-direction). The effect of this is that the only spins that
will be excited will be those that resonate within the bandwidth Do of the r.f.
pulse, and therefore only those spins that lie within a certain ‘‘image slice
thickness’’ Dz. The rest of the sequence acquires data along a separate row of
/2
/2 S(t)
ky(phase)
r.f.
Gz
Gx
kx(read)
Gy
(a) repeat N times time (b)
FIG. 12. (a) Schematic of a simple slice selective 2-D spin-echo pulse sequence. In this pulse
sequence the magnetic field gradient is varied for successive acquisitions of different rows of the k-
space raster. (b) The corresponding k-space raster used to show how we interpret the pulse sequence.
Following a sufficient T1-relaxation period the sequence is repeated to acquire a second row of the k-
space raster. Acquisition of each row of k-space requires a separate r.f. excitation and application of
a Gy -gradient of different magnitude.
the k-space raster for successive r.f. excitations. With reference to Fig. 12, a
magnetic field gradient is first applied in the x-direction, simultaneously with the
maximum magnitude negative magnetic field gradient in the y-direction. A slice
selective p ‘‘re-focusing’’ pulse is then applied; this is represented on the k-space
raster as a move from kx;max , ky;max to kx;max , ky;max . A second gradient is
applied along the x-direction while data, typically 128 or 256 complex data
points, are acquired at a specified digitisation rate. The signal, S(t), that is ac-
quired during application of the second x-gradient is said to be frequency-en-
coded, since the signal is acquired in the presence of a magnetic field gradient.
This gradient along the x-direction is therefore referred to as the frequency-
encoding gradient and is also termed the ‘‘read’’ gradient. The acquisition of
complex data points in the presence of a constant linear ‘‘read’’ gradient yields a
straight line k-space data trajectory whose direction is defined by the Cartesian
orientation of the gradient. A straight, equally spaced k-space trajectory will
always result so long as the read amplitude gradient is kept constant and the
digitisation (acquisition) rate of the complex data is fixed. The spin system is then
allowed to return to equilibrium, via T1 relaxation, and the pulse sequence re-
peated, this time with the second-largest negative y-gradient being applied, hence
‘‘reading’’ the next row of k-space. This process is repeated until the entire raster
has been sampled. In this example, the gradient applied in the y-direction is
referred to as the ‘‘phase encoding’’ gradient. Phase encoding refers to the action
of an applied gradient that is responsible for moving the acquisition through the
k-space raster. In this case, the action of the x-gradient is the same in each
implementation of the pulse sequence, and it is the y-gradient that enables
successive rows of k-space to be sampled. Therefore, if M complex points are
acquired along each row of k-space and N rows of k-space are sampled (i.e. there
are N phase-encoding steps), the final data matrix will consist of M N points on
a rectilinear grid. A 2-D Fourier transform of these data followed by modulus
correction gives a 2-D spin density map. 2-D images are acquired in typically a
few minutes using this approach. While this might be considered slow, in that
only pseudo-steady-state processes can be studied using this pulse sequence, it is
robust in use and straightforward to implement. It is also easy to minimise, or at
least account for, relaxation contrast effects within the acquired image.
F. FAST DATA ACQUISITION
At the heart of recent developments in applying MR in chemical engineering

research has been the implementation and further development of fast MR
spatially unresolved and spatially resolved measurements. Recent implementa-
tions designed for specific applications in chemical engineering are described,
with examples, in Section III. In this section, the principles of the three main
strategies for fast MR imaging are described. We should note at this point that
fast imaging (the term fast is used interchangeably with the terms ultra-fast and
rapid both in this chapter and in the wider literature) is considered here to refer
to the acquisition of, say, a 128 128 2-D image in less than 1 s. This section is
not intended as a detailed review of fast-imaging strategies, but focuses spe-
cifically on those that are finding application in chemical engineering research
rather than in the homeland of fast imaging, the field of medical imaging. While
the data collection strategies will be similar in the medical and engineering fields,
the physical and chemical nature of the samples to be studied and the nature of
the data required are quite different; therefore, the details of the implementation
will differ in terms of the hardware and the pulse sequences themselves. In using
a fast sequence, we will often have to relax our desire for high spatial resolution
(15–30 mm) and take great care, if quantitative data are required, to account
for relaxation contrast effects in the final image. Despite these additional con-
siderations, in the hands of skilled users, these techniques are opening up a
wealth of opportunities in chemical engineering research.
As we have seen, conventional spin echo imaging (Section II.E) typically
takes approximately a few minutes because an independent r.f. excitation is
required for acquisition of each row of k-space data. Hence, sampling of the
complete raster is limited by the repetition/recycle time of the pulse sequence
used, which in turn is governed by the inherent T1 relaxation time(s) of the
system under study. In general, the acquisition speed of an MR image may be
improved by two basic methods:
(i) The sampling of more than one line of k-space for each r.f. excitation of the
spin system
(ii) The use of rapid multiple r.f. excitations (and subsequent acquisitions).
Three sampling strategies will now be introduced: echo planar imaging (EPI),
rapid acquisition with relaxation enhancement (RARE) and low excitation an-
gle imaging (e.g. flash low angle shot (FLASH). The first two are based on the
sampling of more than one line of k-space for each r.f. excitation, while the third
uses rapid multiple r.f. excitations.
1. Echo Planar Imaging

The first example of rapid k-space imaging was demonstrated by Mansfield
(1977), who realised that a complete image could be formed by the acquisition
of multiple lines in k-space following a single r.f. excitation, in a technique today
known as EPI. There are now many EPI-based sequences in use, mostly in the
medical field. In chemical engineering research, while these methods acquire
data faster than any other sampling strategy, their successful implementation is
particularly sensitive to variations in magnetic susceptibility (i.e. gas/liquid/solid
interfaces) within the sample. A common EPI sequence, known as Modulus
Blipped Echo planar Single-pulse Technique (MBEST-EPI) or Blipped EPI
(Howseman et al., 1988), is shown in Fig. 13. Using this sequence, a 128 128
image would typically take 100–130 ms to acquire. In this sequence, a single r.f.
excitation is used to sample the entire k-space raster. Following r.f. excitation,
gradients simultaneously applied in the x- and y-direction take you to the bot-
tom row of k-space, this row of data is acquired, and then a small (or ‘‘blipped’’)
y-gradient takes you to the next row of k-space which is then read in the
opposite direction and so on until the whole raster has been ‘‘read’’. Post-
acquisition data processing techniques address the fact that alternate lines of
k-space are read in opposite directions. The significant step forward made
by Mansfield was that he saw it was possible to ‘‘refocus’’ the decaying
/2
/2 ky(phase)
r.f.
Gz
Gx kx (read)
Gy
S(t)
continue N times
(a) time (b)
FIG. 13. MBEST-EPI: (a) slice selective pulse sequence and (b) corresponding k-space raster.
FIG. 14. 2-D slice sections through 3-D MR images of water distribution within an initially water-
saturated packing of 500 mm glass spheres. Voxel resolution is 94 mm 94 mm 94 mm. Data are
shown before drying commences and at 8 time intervals during the drying process. Only the water
within the inter-particle space of the bead pack is imaged (white pixels). No signal is obtained from
the solid and gas phases present.
magnetisation following r.f. excitation, using the concept of an ‘‘echo’’, so that

it could be used to sample further lines of k-space (see Section II.B). Many
variants of EPI exist; most of these have been developed with the aim of re-
ducing potential artefacts in the image. Particular pulse sequences include radial
EPI or REPI (Silva et al., 1998), p-EPI or PEPI (Guilfoyle et al., 1992) and
Gradient and Spin Echo (GRASE) (Oshio and Feinberg, 1991). Manz et al.
(1999b) have implemented an EPI-based sequence, which was sufficiently robust
to spatially resolve drying within a packed bed of 100 mm glass spheres. Typical
data are shown in Fig. 14; despite the drying process being characterised by a
uniform water mass loss with time, the MR images clearly show the spatial
heterogeneity in the drying process. This has been used to develop an inva-
sion–percolation modelling strategy that predicts both the drying rate and the
heterogeneity of the drying process (Gladden et al., 2004).
/2
/2
r.f. ky (phase)
Gz
Gx
Gy kx (read)
(4)
S(t) (3)
(2)
(1) (2) (3) (4) (1)
continue R times
continue N/R times
(a) (b)
FIG. 15. k-space trajectory of a RARE sequence. The order of the phase encoding is shown by the
numbers to the left of the raster. After each line in k-space is read the spins are returned to the same
point on the kxðreadÞ axis prior to the application of the refocusing pulse shown by the dashed line and
arrow.
2. Rapid Acquisition with Relaxation Enhancement

This type of image acquisition was first realised by Hennig et al. (1986), and
variants such as turbo spin echo (TSE) and fast spin echo (FSE) exist. The
k-space sampling strategy appropriate to RARE is shown in Fig. 15, and at first
it might not seem different from the EPI approach; i.e. after acquisition of a
single line in k-space the coherent signal is refocused and used to acquire further
lines in k-space (or averages of the same line). However, there are important
differences that make RARE much more robust (i.e. artefact-free) in applica-
tion to systems characterised by heterogeneous magnetic susceptibility typical of
those found in chemical engineering. Further, the decay of the MR signal in a
RARE experiment is dependent on T2 (and, of course, to some extent T1) and
not T 2 , as is the case for EPI. In the magnetically heterogeneous systems of
interest to us, T 2 is considerably shorter than T2; therefore, RARE allows us to
acquire data of higher signal-to-noise over longer timescales, which is important
if we wish to investigate the time–evolution of a system. An example of the
application of RARE for rapid image acquisition is shown in Fig. 16, in which a
single frame is shown from a series of 2-D images of an oscillatory chemical
reaction occurring within a fixed bed. Relaxation contrast is used to discrim-
inate between the reaction products Mn2+ and Mn3+ (Britton et al., 2005). In
this example, MR offers the opportunity to map the detailed structure of the
fixed bed and the product distribution within it. This pulse sequence has also
been recently applied to obtain quantitative images of the evolution of a lys-
ozyme–urea separation within a chromatography column (Holland et al., 2004).
FIG. 16. 2-D MR image of an oscillating chemical reaction occurring within a bed of diameter
15 mm, packed with glass spheres of diameter 1 mm. In-plane resolution is 195 mm 195 mm, and the
image slice thickness is 1 mm. A single image was acquired in 1 s. Chemical waves are imaged as a
result of the oscillatory production of Mn2+ and Mn3+ species; the location of Mn2+ and Mn3+
bands are identified as dark and light bands, respectively.
T2- and T1-relaxation contrast in the RARE acquisition is exploited such that
the images are made selective to lysozyme and urea, respectively.
3. Low Excitation Angle Imaging

The important feature of this technique is that the r.f. excitations are char-
acterised by a low flip angle, y, typically 5–101, in contrast to the p/2 pulse
discussed in Section II.A. For this reason, this technique has been termed
FLASH imaging or SNAPSHOT imaging (Haase et al., 1986). The signal
resulting from the small flip angle y is proportional to sin y, while the
longitudinal (z-axis) magnetisation that remains after the excitation is
proportional to cos y. Fractions of this remaining magnetisation are then used

to sample successive lines of k-space. A 128 128 image based on a repetition
time of 3 ms takes approximately 380 ms to acquire. The disadvantage of this
approach is that by using only a proportion of the available magnetisation,
signal-to-noise ratio in the image is significantly reduced. However, the impor-
tant advantage is that the acquired signal is not strongly influenced by relax-
ation contrast effects or artefacts associated with molecular motion and
magnetic susceptibility. In short, one is sacrificing temporal and, potentially,
spatial resolution (i.e. signal-to-noise) for increased robustness in implementa-
tion.
There is no ‘‘rule’’ as to which fast-imaging technique should be used in a
given application, and the following guidelines should be treated with caution.
In short, the best approach is to implement each method and see which works
best in a given application. EPI-based techniques are the fastest and can in-
corporate measurements of velocity vectors, but are also the most sensitive to
variations in magnetic susceptibility within the sample. They are the most dif-
ficult to implement and in many circumstances cannot be used, as the resulting
image contains obvious artefacts. FLASH/SNAPSHOT and RARE are, in
general, slightly slower than EPI, but they are more robust, thereby making
them more useful in day-to-day chemical engineering applications. The major
differences between them are that they are influenced predominantly by differ-
ent contrast mechanisms: RARE by T2 and FLASH/SNAPSHOT by T 2 . Thus,
using the same argument as that used in comparing EPI and RARE previously,
RARE will usually be the technique of choice. Further, RARE offers substan-
tial signal-to-noise gains over FLASH/SNAPSHOT.

Engineering Research
What constitutes an advance in any field will always be subjective. However,

the combination of the inherent ability of MR methods to probe the internal
structure and transport processes from the Å- to cm-scale phenomena non-
invasively, quantitatively and with chemical resolution, and with the ability to
acquire these data sufficiently fast so that unsteady state processes can be stud-
ied is undoubtedly going to open up new avenues of research and allow us to
investigate many phenomena for the first time. This section summarises five
recent developments in the field of MR in chemical engineering. The first four
sub-sections (Sections III.A–III.D) report developments of fast MR measure-
ment pulse sequences, which have recently been implemented for application in
chemical engineering research. The final sub-section (Section III.E) addresses a
new and different field of research, that of gas-phase imaging.
A. ‘‘ULTRA-FAST’’ IMAGING OF VELOCITY FIELDS
Imaging unsteady state or turbulent flow fields is a subject of long-standing

interest in chemical engineering. Many experimental approaches have been
used, including particle imaging velocimetry (PIV) and laser Doppler an-
emometry (LDA) techniques. The motivation for implementing MR flow meas-
urement in this application is not only to allow us to study flow fields within
optically opaque fluids, but also to extend the measurement to multi-phase,
multi-component flows. In principle, as long as the MR pulse sequence can be
made robust with respect to variations in magnetic susceptibility across phase
boundaries, the measurement should be quantitative, regardless of the void
fraction of any phase. In contrast to other methods, an image reconstruction
algorithm is not required to take us from the acquired raw data to the final
image: the only processing operation is a Fourier transform of the time-domain
data. This should make MR much more quantitative in the sense that no cor-
rections need to be considered for the scattering of probe radiation from in-
terfaces within the sample. Fast MR techniques are now being applied to
various topics in the field of fluid mechanics, including mapping spatio-temporal
structures formed in Taylor Couette flow (Kose, 1994), and the velocity patterns
within a falling liquid film (Heine et al., 2002) and a free-falling drop (Han et al.,
2001a); this final example is particularly elegant (see Fig. 17). Sufficient signal-
to-noise cannot be acquired from a measurement of a single falling drop; hence,
data from multiple falling drops are acquired. This is achieved by releasing
successive ‘‘identical’’ drops and gating the data acquisition such that data are
acquired as each drop falls through the field of view of the imaging experiment.
Measurements of turbulent flows have been addressed previously by two MR
methods: by measurement of signal attenuation and time-averaged velocity
profiles. Kuethe (1989) reported the first imaging study of turbulent flow in
which the flow was generated by a jet of water passing through a nozzle. In this
study, a measure of turbulent diffusivity was obtained from a measurement of
the signal attenuation observed in time-averaged images, employing the ap-
proach described in Section II.C. Similar approaches were later used by Gate-
nby and Gore (1994), employing a spatially unresolved diffusion measurement
sequence to study pipe flow up to Reynolds numbers as high as 6270, and
Kuethe and Gao (1995) to study pipe flows up to higher Reynolds numbers of
12000–58000. The alternative approach is that of Li et al. (1994), who reported
time-averaged velocity measurements of flow in a cylindrical pipe; 1-D velocity
profiles were measured for Reynolds numbers up to 9000. The results showed a
correlation between the pixel intensity of the time-averaged flow image and the
local turbulence intensity, with the latter showing good agreement with earlier
data recorded using a hot-wire anemometer technique (Laufer, 1954). More
recently, Han et al. (2001b) used an approach similar to that of Li et al. (1994)
to investigate the non-Newtonian flow of blood in a pipe at Reynolds numbers
up to 3500.
FIG. 17. Velocity components vz , vx are shown within a drop of pure water falling at 2.0 m/s as a
function of position along the zx plane, averaged over the y dimension of the drop. The vz com-
ponent is scaled relative to the average falling velocity of 2.0 m/s. Reprinted with permission from
Han et al., (2001a). Copyright (2001), American Physical Society.
Neither of the two aforementioned methods directly measures the actual ve-
locity vectors describing the flow field. In both cases even if an ‘‘excitation-
acquire’’ sequence is fast, the data then have to be signal-averaged to provide
adequate signal-to-noise in the final data set. In contrast, EPI (Section II.F) can
be integrated with a transport measurement sequence (e.g. Fig. 7) to provide
images of the flow field at timescales of 30 ms or less. Kose (1991a,b, 1992)
reported the first EPI-MR images of turbulent flow in which individual velocity
vectors characterising the flow were resolved. In particular, the velocity distri-
bution in a cross-sectional plane perpendicular to the direction of superficial
flow at a Reynolds number of 2250 was imaged; clear visualisation of a tur-
bulent ‘‘puff’’ was reported. More recently, the work of Kose has been extended
to acquire images of three orthogonal velocity components from a single ex-
citation over a timescale of 60 ms, with each velocity component itself being
acquired in o20 ms. Alternatively, up to 16 velocity images can be acquired, at
time intervals of 20 ms, from a single excitation. The maximum number of
images acquired is limited by the residence time of the moving fluid within the
imaging coil. These images were acquired in a 2-D data matrix of size 64 32
(Sederman et al., 2004a). This is fast enough for the evolution of turbulent
eddies in the flow to be followed. Acquisition of successive velocity images
allows the acceleration field to be calculated. These data were acquired over a
π/2 π π
r.f.
Gr
Gp
Gs
δ
g
∆
repeat unit
time
FIG. 18. Schematic of the GERVAIS pulse sequence, identifying all pulse and delay timings. The
pulse sequence shown is that for acquisition of successive velocity images. The magnitude and
orientation of the g-gradient are changed as determined by the velocity or acceleration vector that is
to be measured.
wide range of Reynolds numbers encompassing the transition from pure

laminar flow into the turbulent flow regime (1250oReo5000). This pulse
sequence has been named the gradient echo rapid velocity and imaging sequence
(GERVAIS) and is based on the MBEST-EPI, or Blipped EPI, sequence
(Section II.F), with the modification that after the initial excitation pulse, each
image is preceded by a velocity-encoding gradient pair, separated by a
p-refocusing pulse. The GERVAIS pulse sequence is shown in Fig. 18, and
the basic combination of transport measurement gradient pair and EPI
sequence is clearly seen with reference to Figs. 7 and 13. With reference to
Fig. 18, velocities in the x-, y- and z-directions are acquired by changing the
direction of the velocity measurement gradients (g). Figure 19 shows samples of
these three Cartesian component velocity images for six different liquid veloc-
ities. The velocity component in the superficial direction (vz ) is given by the
colour scale bar and the magnitude and direction of the transverse (vx , vy )
velocity components are indicated by the vector arrows present on each of the
images. More recently, the GERVAIS pulse sequence has been applied
successfully to image unsteady-state flows in narrow fixed-bed reactors (Sains
et al., 2005).
FIG. 19. Imaging unsteady state and turbulent flow of water within a 29 mm diameter pipe. Three
orthogonal component velocity images acquired at increasing Re of (a) 1250, (b) 1700, (c) 2500, (d)
3300, (e) 4200 and (f) 5000 are shown. The colour scale identifies the magnitude of the z-velocity, and
the flow velocity in the plane of the image (i.e. x–y) is shown by the vectors on each image. The
vector scale bar on each image corresponds to 1 cm/s. Reprinted from Sederman et al. (2004a) with
permission from Elsevier. Copyright (2004).
B. MULTIPLE IMAGES FROM A SINGLE EXCITATION
As discussed in Section II.F, the RARE method is a particularly robust

method for fast data acquisition in chemical engineering systems, which are
often characterised by a range of magnetic susceptibilities. The resulting short
nuclear spin relaxation times combined with rapid timescales over which a
system may change impose far greater requirements on the imaging pulse se-
quence than are encountered when studying medical systems. Recently, a var-
iant of RARE has been implemented (Sederman et al., 2003) that yields multiple
images from full k-space data at multiple echo times following a single r.f.
excitation, not just a single complete image from a single excitation. This pulse
sequence has been named single excitation multiple image RARE (SEMI-
RARE) and is an extension of a standard RARE experiment. Implementation
of SEMI-RARE allows the acquisition of up to 120 images, in immediate
succession, from a single excitation pulse. The actual number of images ac-
quired from a single excitation is determined by the characteristic nuclear spin
relaxation times of the system under study. A fine demonstration of the power
of this technique is the visualisation of gas–liquid flow within the parallel
channels of a ceramic monolith. To our knowledge MR is the only technique
that has been able to image non-invasively within such a ceramic monolith. This
is also a good example of the robustness of the RARE methodology. Not only
do we have varying magnetic susceptibility across the gas–liquid interface, we
also have the interface of both with the walls of the ceramic channels. In this
system, it is not practical to reduce the magnetic susceptibility variations be-
tween the gas and liquid by introducing chemical species that modify the mag-
netic susceptibility of the liquid phase, because such species will be adsorbed
from the liquid onto the porous walls of the monolith. Thus, RARE is probably
the only MR method that can be used in this application.
Figure 20 shows images, recorded in real-time using the SEMI–RARE pulse
sequence, of gas flow through stagnant liquid within parallel-channel ceramic
monoliths (Gladden, 2003; Gladden et al., 2003b). The monoliths were of di-
ameter 48 mm and length 0.15 m, contained within a Perspex column of inner
diameter 50 mm. Gas (compressed air) was introduced at the base of the liquid
(water) flooded column at flow rates of between 50 cm3/min and 300 cm3/min. In
this experiment, four images were acquired in immediate succession from a
single excitation. A single frame took 146 ms to acquire. The images are taken
over a field-of-view of 50.6 mm 50.6 mm, with an in-plane pixel resolution of
391 mm(x) 781 mm(z). The centre of the imaging section was positioned at the
centre of the monolith, i.e. 7.5 cm from the monolith outlet. The thickness of the
image slice is 600 mm. Figure 20 shows the first (a) and second (b) images of a
sequence of four taken after a single excitation. The orientation of the slice was
selected to image just a single row of parallel channels within the monolith. The
imaging resolution allows easy identification of the individual channels and
several bubbles (zero signal intensity) can clearly be identified. From these im-
ages, gas-phase volume fractions and distributions of gas-bubble length and
z
(a) (b)
FIG. 20. Two successive 2-D xz images of two-phase flow through the parallel channels of a
ceramic monolith rated at 400 cpsi, for a gas flow rate of 200 cm3/min: (a) 74 ms after excitation (b)
220 ms after r.f. excitation. In-plane image resolution is 393 mm (x) 783 mm (z). Reprinted from
Gladden, (2003), with kind permission of Springer Science and Business Media.
velocity as a function of gas flow rate and channel size are obtained directly. In
this particular example, monoliths rated at 300 and 400 channels per square
inch (cpsi) were studied. Increasing the gas flow rate was seen to increase the
number of large bubbles and the average bubble velocity. A bimodal distribu-
tion in the bubble velocities was observed for flow within the larger channel size
(300 cpsi) in contrast to a broad unimodal distribution characterising two-phase
flow within the smaller channel size (400 cpsi).
C. IMAGING ROTATING SYSTEMS
Albeit a technique with a niche application, this variant on the RARE meth-
odology demonstrates the flexibility of MR to achieve specific measurements.
The particular application for which this pulse sequence was developed was to
avoid blurring of images recorded for systems rotating at a known, constant
angular frequency. This RARE sequence with rotational compensation is
shown in Fig. 21. It works by reorienting the direction of the imaging gradients
between each cycle or echo acquisition, as shown schematically in Fig. 21b, to
ensure that the k-space raster remains aligned with a chosen rotational velocity,
o. This means that the directions of the read and phase gradients rotate after
each cycle or echo acquisition, but remain perpendicular to one another, their
angle with the x-axis being a and a þ p=2, respectively. a increases with t, the
time between successive echo acquisitions, by an amount ot. With reference to
Fig. 21b, note that unlike conventional RARE (see Fig. 15), the sequence begins
and ends at the origin of k-space between each cycle or echo acquisition. This
increases the robustness of the pulse sequence since the origin of k-space is the
only point unaffected by the gradient rotation. In the laboratory frame this
kx1
/2 kx2
r.f.
ky1
Gx read
Gz slice
Gy phase
ky2
N times time
(a) (b)
FIG. 21. (a) The ‘‘rotating RARE’’ pulse sequence, and (b) the associated k-space raster. The
pulse sequence is a variant of the RARE pulse sequence with the read and phase directions rotated
with respect to the laboratory-frame after each echo acquisition such that the orientation of k-space
relative to the rotating object remains constant. After each echo acquisition, the system is returned to
the centre of k-space, which is not affected by sample rotation.
corresponds to an irregularly sampled k-space raster; for a sample rotating at

the pre-defined rotation rate this will correspond to the sampling of a rectilinear
k-space raster.
Figure 22 shows a 2-D image of a paddle wheel rotating within a cylindrical
vessel (Sederman et al., 2004b). The paddle is rotating at an angular frequency
of 1.6 rev/s and therefore appears blurred when acquiring the 2-D image with an
acquisition time of 300 ms. By applying the rotational compensation, blurring is
prevented. The particular application for which this pulse sequence was devel-
oped was to image accurate droplet shapes in situ during exposure of an emul-
sion to a shear field within a Couette cell. By allowing a given droplet to attain
its equilibrium radial position within the Couette cell during shear, the rota-
tional compensation is optimised for that particular angular frequency. Figure
23 shows the deformation of a water droplet in a 1000 cSt silicone oil solution
under conditions of increasing shear within a wide-gap Couette cell. The droplet
has zero intensity in this image since the image has been acquired under con-
ditions of T2 contrast such that signal is acquired only from the silicone oil. By
measuring the short and long axis of the droplet as a function of rotation rate,
standard analysis (Taylor, 1934) is used to obtain an in situ estimate of the
interfacial tension characterising the system.
D. ‘‘ULTRA-FAST’’ DIFFUSION MEASUREMENT
The diffusion train (DIFFTRAIN) pulse sequence (Stamps et al., 2001) is an

elegant technique for fast measurement of PFG-MR (PGSE-MR) data, and
finds application in the rapid measurement of, for example, emulsion droplet-
size distributions or propagators. DIFFTRAIN speeds up the conventional
(‘‘slow’’) diffusion measurement experiment by using successive stimulated ech-
oes from a single excitation pulse, such that a portion of the available mag-
netisation is recovered for each echo (cf. FLASH in Section II.F). Recently,
DIFFTRAIN has been demonstrated in application to the determination of
emulsion droplet-size distributions and propagators of water transport through
a desalting column packed with Sephadex matrix (Buckley et al., 2003). In both
cases, the data were in quantitative agreement with data acquired using con-
ventional ‘‘slow’’ techniques. In the case of the droplet-size distributions, data
acquisition times were reduced from 10–20 min to 4 s, while in the propagator
determinations, fast acquisition took less than 10 min, an order of magnitude
faster than is the case with conventional PFG techniques. The reduction in
measurement time for the emulsion droplet-size distributions means that it is
now possible to follow the evolution of droplet size in situ during a mixing
process. As long as signal-to-noise requirements are satisfied, there is no reason
why these measurements cannot be integrated into an imaging strategy to ob-
tain spatially resolved measurements of droplet size.
(a) (b)
FIG. 22. (a) Conventional RARE and (b) ‘‘rotating RARE’’ images of a paddle rotating in water
at 1.6 rev/s. Reprinted from Sederman et al. (2004b), with permission from Elsevier. Copyright
(2004).
(a) (b) (c)
FIG. 23. Images acquired of 2 ml water droplets of 1% Tween 60 (w/w) with 0.005 M MnSO4 in
1000 cSt silicone oil under shear rates of (a) 0.63 s1 (b) 1.67 s1 and (c) 3.25 s1. Water inside the
rotating inner cylinder also appears in the images; it is off centre due to its chemical shift relative to
that of the silicone oil. The figures have been rotated to show the droplets in the same position in the
Couette. Reprinted from Sederman et al. (2004b), with permission from Elsevier. Copyright (2004).
E. GAS-PHASE MR
Both thermal and hyperpolarized gas-phase MR studies have until now pre-
dominantly addressed medical applications and, in particular, lung imaging.
However, non-medical applications are now being reported. Conventional MR
measurements rely on the ‘‘thermal’’ polarisation of the atomic nuclei within the
magnetic field. Since the interaction energy of the nuclear ‘‘magnets’’ with
the applied magnetic field is about 104 times smaller than the thermal energies of
the system, the resulting polarisation of the nuclear magnets is very weak
and the inherent sensitivity of MR measurements is therefore low compared to
other spectroscopic techniques. This characteristic of MR measurements is
particularly limiting in studying systems of low nuclear spin density, such as

gases. More recently, methods that enhance the signal available from gas-phase
experiments by exposing the gas (usually 3He and 129Xe) to laser optical pump-
ing have been developed. The population distribution between adjacent energy
levels is thus modified, resulting in an increase in the polarisation of the nuclear
spin system by 3–4 orders of magnitude, and a state of ‘‘hyperpolarisation’’ is
reached. First, we will briefly review the current state of the art in using ther-
mally polarised gases and then discuss the latest developments in using hyper-
polarisation methods, focussing on non-medical applications.
Thermally polarised gas MR measurements have been successfully performed
using 129Xe and 19F observation (for the case of the 19F nucleus, the molecules
CF4, C2F6 and SF6 are typically studied), and have been shown to be a powerful
probe of inorganic porous media (e.g. rock cores), giving insights both to the
structure of the porous medium and the transport of the gas within this struc-
ture (Bencsik and Ramanathan, 2001; Mair et al., 2002). As discussed by Mair
et al. (2002), 3He and 129Xe are particularly well suited for such studies given
their rapid diffusion, inert nature, low surface interactions (which reduces re-
laxation time effects) and the ability to tailor the diffusion coefficient by altering
the gas pressure in the sample. In particular, the faster diffusivities of gas-phase
species compared with liquid-phase probes of the pore space have made it
possible to extend the length-scales that can be probed in porous media by more
than one order of magnitude. However, in addition to the low signal-to-noise
ratio associated with the low polarisation of the spin system, a further challenge
in using gas-phase MR is that the T2 of the gases is often very short. Kuethe
et al. (2000) have made advances in using thermally polarised gases by devel-
oping hardware that recovers from the transmitter pulse rapidly enough to
detect gases characterised by rapid T2 relaxation times as short as 1 ms. This is
a significant advance, since the combined effect of decreased signal loss due to
the rapid switching times and the ability to perform rapid signal averaging now
allows images of good quality to be obtained. Although originally developed for
medical applications, the method has been successfully applied to the imaging of
porous glasses and ceramics. Particularly noteworthy has been the extension
of the method to map the specific surface area and trends in the local adsorption
energy within these porous materials. This is done by exposing the porous
sample to a gas at a range of gas pressures, and acquiring a spin density image
at each pressure condition. In this way a spatially resolved Brunauer-
Emmett-Teller (BET) isotherm has been generated (Beyea et al., 2002).
More recent studies of gas-phase imaging with more direct relevance to
chemical processing and reaction engineering have involved the examination of
thermally polarized gas (and liquid) flow in monolithic catalysts. Koptyug et al.
(2000a) have obtained quantitative, spatially resolved velocity maps for the flow
of thermally polarized acetylene, propane, butane and water flowing through
the channels of alumina monoliths with an in-plane spatial resolution of 400 mm.
The monoliths had a channel cross-section of 4.0 mm2 and a wall channel
thickness of 1.0 mm. Axial gas-phase velocities of up to 0.93 m/s (Re ¼ 570)
were studied. The flow maps showed a highly non-uniform distribution of shear
rates within the individual monolith channels. In a follow-up study, Koptyug
et al. (2001) reported images of both liquid and gas flow and mass transport
phenomena in two different cylindrical monolith catalysts (one with triangular
channels, the other with square channels) at different axial locations within the
monolith.
The advent of hyperpolarized 3He and 129Xe for use in MR studies has given
further impetus to gas-phase studies. Brunner et al. (1999) have described a
continuous flow system that circulates laser-polarised 129Xe through the sample,
yielding an enhancement of signal intensity of 3–4 orders of magnitude com-
pared to the equilibrium 129Xe signal. Gas flow displacement profiles of 129Xe
flowing through polyurethane foams of different densities and pore sizes were
reported. Another exciting area of potential development is the use of laser-
polarised gases to provide enhanced sensitivity to study species at surfaces
(Pietrab et al., 1998). Exploitation of laser-polarised gases is also finding ap-
plication in so-called ‘‘remotely detected MRI’’. A good description of this has
been given by Seeley et al. (2004). The principle behind the measurement is easy
to understand although the concepts are hard to grasp for those not well ac-
quainted with MR. The key feature of remote detection is the spatial separation
of the MR encoding and detection steps, which allows for their separate opt-
imisation by providing the most suitable conditions for encoding without com-
promising detection quality. This is achieved by employing a ‘‘signal carrier’’ to
encode MR information indicative of its environment. Time-domain (or
k-space) information is stored point-by-point as spin polarization in the en-
coding location, each point being subsequently transported to the detection
location. Remote detection is an indirect detection technique that exploits the
principles of phase encoding. The resulting phase acquired after each signal
acquisition provides one point in the indirect signal. In such a ‘‘remote’’ ex-
periment, the indirect dimension provides information about the encoding en-
vironment, while the direct dimension gives the signal of the sensor in the
detection region. In principle, this methodology can be extended to the ‘‘signal
carrier’’ being water and oil—of particular relevance to, say, understanding oil
recovery processes; the T1 of these fluids currently limits their use in this ap-
plication. In contrast, laser-polarised 129Xe has a range of particular attributes,
the most obvious being that it is a highly polarisable and chemically inert noble
gas with a long T1, and therefore acts as a very efficient carrier of the spin
polarisation. Further, its wide chemical shift range makes it a powerful sensor of
its local environment and therefore provides a wealth of physicochemical and
biomedical information. To date, clear demonstration of enhanced signal-to-
noise for indirect over direct detection has been demonstrated for a model
porous medium comprising a 14-mm-diameter ‘‘honeycomb’’ phantom with
1.4-mm-diameter pores. There remains much work to do in optimising this
experiment and in quantifying the effect of contrast mechanisms on the acquired
image, but the approach may offer opportunities to study systems inaccessible
to more conventional MR measurement.
IV. Reaction Engineering: From Catalyst to Reactor
The following section does not dwell on the details of any of the MR tech-
niques used, but brings together the various ways that MR can characterise and
give insight into a particular field of chemical engineering–in this case, heter-
ogeneous catalysis. As will be demonstrated, MR makes contributions to our
understanding of a system from the size scale of the active site of the catalyst to
the macroscopic phase distributions within a fixed-bed reactor. Inevitably, the
examples are not exhaustive, and for the sake of continuity between examples,
the majority are drawn from research in our own group. It is hoped that the
reader, armed with knowledge of the examples cited in earlier sections that draw
on applications in various fields of chemical engineering, will be able to identify
applications of MR over various length-scales in their own field of interest.
A. MR SPECTROSCOPY OF CATALYSTS
In most cases reported in the literature, studies of the catalyst are performed
on the as-prepared catalyst in powder form. It is important to remember that
such characteristics may be modified when the catalyst is formed into the pellet
for process operation. The motivation for current studies in the field of MR
spectroscopy of catalysts is to characterise the active catalytic site and, in some
cases, to study how reactant/product molecules interact with it. For these rea-
sons this research remains predominantly in the area of mainstream chemistry
research. However, it is quite likely that as chemical engineers engage with MR
methods, there will be increasing interest in probing the catalyst structure and
chemical processes occurring within the catalyst as part of understanding and
optimising the overall catalyst/reactor system. An area that has already been
taken up by chemical engineers is the study of molecular diffusion within cat-
alysts. This will be described briefly in this section and then considered further
in Section IV.B, where we discuss studies of molecular diffusion within catalysts
at the length-scale of the formed catalyst typical of those manufactured for
supported-metal catalysis.
With respect to in situ spectroscopy, there is no attempt here to give a detailed
overview of the ongoing activities in this field. The interested reader is directed
to excellent articles describing applications of MR spectroscopy in catalysis that
have been published elsewhere (Baba and Ono, 1999; Dybowski et al., 1991;
Fraissard, 1999; Haw, 1999; Hunger and Weitkamp, 2001; Packer, 1996; Parker
2000; Roe et al., 1998; van der Klink, 2000). Instead, we focus on the nature of
the information obtained and cite some illustrative examples in the respective
fields.
1. Characterisation of Surface Chemistry

MR spectroscopy is widely used to study the surface chemistry of catalysts
and the interaction of the surface with reactant and product species, as well as
investigating the role of surface modifiers to be exploited in catalytic reactions.
For example, 27Al and 31P solid-state MR spectroscopy have been used to study
the acidity and surface structure of aluminas subjected to a fluorination treat-
ment (Chupas and Grey, 2004). In this study, 27Al was used to show the pref-
erential removal of 5-coordinate aluminium sites during fluorination, suggesting
that they are predominantly localised near the surface. The sorption of basic
phosphines as probe molecules for acid sites coupled with 31P-MR was used to
follow the changes in acidity after fluorination of the surface. Zhang et al. (2000)
have used MR to probe the adsorption of SO3 onto oxide surfaces.
2. Characterisation of Coke
13
C-MR spectroscopy and related techniques has been established as a tech-
nique for characterising the carbon-to-hydrogen ratio of carbonaceous residues
on the catalyst surface (Duncan et al. 1985, Egiebor et al., 1989; Groten et al.
1992). MR spectroscopy continues to be used in this way, but is more often used
in combination with other characterisation techniques such as X-ray diffraction,
electron spin resonance, infra-red spectroscopy and temperature-programmed
oxidation studies, to gain information on the structure, carbon-to-hydrogen
ratio and combustion behaviour of the coke (Martin et al., 2004; Wang et al.,
2001).
3. MR Spectroscopy Studies of In Situ Reactions

A review of this field has been given by Haw (1999). Reactions can be fol-
lowed either in sealed glass ampoules or flow-through cells constructed within
the spectrometer. The formation of intermediates can be studied in real time. An
elegant example of this was shown in an early study of methanol to gasoline
conversion over HZSM-5 zeolites. As a result of the shape selectivity of the
catalyst, spectroscopic evidence of reaction intermediates, which were not seen
as reaction products, was observed (Anderson and Klinowski, 1990).
4. Molecular Diffusion within Catalysts

Most studies of molecular diffusion within catalysts have been performed on
the as-prepared (i.e. powder form) of the catalyst and focus predominantly on
zeolites and related catalytic materials. A recent review of this field has been
given by Karger and Freude (2002). In situ studies of molecular diffusion during
reaction include 13C-PFG-MR studies of diffusion and reaction of isopropanol
in zeolite CsNaX (Schwarz et al., 1995), intracrystalline diffusion during is-

opropanol conversion in X-type zeolites (LiNaX, CsNaX and NaX) (Schwarz
et al., 1997) and molecular diffusion during catalytic reaction of cylopropane to
propene in zeolite X (Hong et al., 1992).
B. MICRO-IMAGING AND MOLECULAR DIFFUSION STUDIES OF FORMED CATALYST

PELLETS
The majority of the early MR imaging studies specific to catalysis addressed

the heterogeneity in structure and transport within catalyst pellets. In-plane
spatial resolution was 30 mm, and the pellets themselves were typically 1–5 mm
in size. In the majority of cases, studies have addressed the pure, usually oxide,
support so that the quantitative nature of the data obtained is not sacrificed by
the presence of metal, which can introduce an unknown extent of nuclear spin-
relaxation time contrast into the images.
1. Structure-Transport Relationships
By working with typical spatial resolutions of 30–50 mm, individual pores
within the material are not resolved. Further, the transport characteristics
within small pores (e.g. nanopores) may not contribute to the measurement
owing to the very short relaxation times of fluid contained within such pores.
However, a wealth of information can be obtained at this lower resolution.
Typical data are shown in Fig. 24, which shows images or maps of spin density,
nuclear spin-lattice relaxation time (T1) and self-diffusivity of water within the
porous pellet (Hollewand and Gladden, 1993). In-plane spatial resolution is
45 mm 45 mm, and the image slice thickness is 0.3 mm. The spin density map is
a quantitative measure of the amount of water present within the porous pellet,
i.e. it is a spatially resolved map of voidage. Estimates of voidage obtained agree
to within 5% of those obtained by gravimetric analysis. The spin-lattice relax-
ation time map (as discussed in Section II.B) yields information on the spatial
distribution of mean pore size within a given image pixel. Lighter shades in the
image correspond to larger mean pore size. Even at this course resolution, these
data give us additional insight to that which may be obtained from a 1-D pore-
size distribution obtained by, for example, mercury porosimetry or nitrogen
adsorption measurement. Thus, by using MR, we can now probe the spatial
heterogeneity in porosity within a catalyst pellet that will have been introduced
during the manufacturing process. Figures 24a and b allow us to discriminate
between a region of given voidage comprising many small pores or a single large
pore. The structure–transport relationship characteristic of the catalyst pellet is
seen in comparing Fig. 24b and c; i.e. the spatial heterogeneity in variation of D
values is much more consistent with the heterogeneity in the intensity seen in the
T1 maps as opposed to the spin density maps. Thus, we conclude that it is the
(a) (b) (c)
FIG. 24. Spin density, T1 and water diffusion images for a 2.2 mm diameter, spherical silica
catalyst support. In-plane pixel resolution is 45 mm 45 mm; image slice thickness is 0.3 mm. (a) Spin
density map; lighter shades indicate higher liquid content. (b) T1 map (150–400 ms); lighter shades
indicate longer values of T1. (c) Diffusivity map (0–1.5 109 m2/s); lighter shades indicate higher
values of water diffusivity within the pellet.
spatial variation of local pore size that has the dominant influence on molecular
transport within the pellet. There have been a number of studies exploiting this
type of MR measurement. For example, water spin density imaging has been
used to explore the 3-D structure of activated alumina spheres. The spheres were
observed to exhibit a uniform ball structure comprising spherically layered
structures and voids (Timonen et al., 1995).
Simple spin-density imaging has also been used to characterise the tortuosity
of catalyst pellets manufactured in different ways (Rigby et al., 1996). This is
achieved by initially impregnating the catalyst with deuterated water (invisible
to a 1H-MR experiment) and then immersing it in normal, protonated water.
The diffusive exchange of 1H2O with 2H2O within the pellet is followed by 1H
imaging. The resulting time-resolved 1H2O concentration profiles are then fitted
to a standard diffusion equation and the effective water diffusivity and, hence,
catalyst tortuosity, are obtained. Measurements of this type are straightforward
to perform and give immediate insight into transport anisotropies within the
catalyst resulting from manufacturing process parameters.
When originally recorded, the types of images shown in Fig. 24 suggested a
potential dilemma in using this type of information to aid the development of
modelling the structure of catalyst pellets and the transport processes occurring
within these porous structures. Each pellet that is imaged, even if taken from the
same batch, will yield a different image. However, we know that a given batch
of pellets is expected to have consistent (i.e. ‘‘typical’’) properties ‘‘in-use’’. In
studying the spin density (voidage) and spin-lattice relaxation time maps of
many pellets it was found that the heterogeneity in pore size, as characterised by
the fractal dimension of the T1 map, was consistent between images of pellets
drawn from the same batch (Gladden et al., 1995). The fractal dimension of
these images actually identifies a constant perimeter–area relationship for clus-
ters of pixels of similar intensity (i.e. pore size) in the image. The most obvious
physical interpretation of this is that a given manufacturing process imposes a

particular meso-scale (0.1–1 mm) structure characterising the spatial distribu-
tion of pore sizes within the pellet.
Numerical modelling strategies employing simulation lattices constructed to
have the same fractal dimension as that obtained from the MR images have had
some success in predicting mass transfer characteristics in porous catalyst sup-
ports (Rigby and Gladden, 1998). Current strategies include combining the
characterisation of meso-scale structure obtained from MR with characterisa-
tion of the micro-pore structure obtained by, say, nitrogen adsorption or mer-
cury porosimetry to construct hierarchical models of the micro- and meso-
length scales that control transport processes (Rigby and Daut, 2002; Rigby and
Gladden, 1999).
2. Catalyst Preparation
The ability to image liquid distribution in catalyst support pellets suggests
immediate applications in investigating transport processes occurring during
catalyst manufacture and, in particular, wet impregnation methods. Aspects of
catalyst preparation that MR has been used to address include both liquid and
ion migration process. Early work demonstrated the ability of 2-D MR images
to discriminate between so-called uniform and capillary-controlled drying mod-
els (Hollewand and Gladden, 1994). The dominant drying mechanism was
shown to be dependent on the timescale over which drying was performed.
Further, for those systems for which a capillary-controlled drying mechanism
was identified as appropriate, a simple numerical simulation of the drying
process provided an estimate of the mean co-ordination number associated with
a random pore network describing the micro-pore structure of the material.
Network connectivities estimated in this fashion were consistent with those
estimated from mercury porosimetry analysis.
Detailed studies of 1-D 1H imaging of alumina and titania cylinders (of length
1.2 cm and diameter 3.6 mm) have been reported by Koptyug and co-workers
(Koptyug et al. 1998, 2000b) in which consideration of relaxation time effects on
the water concentration profiles are discussed. These porous catalyst supports
were impregnated with acetone, benzene, cyclohexane and water, and the drying
of these liquids was monitored under different drying conditions, i.e. different
nitrogen flow rates. The 1-D profiles were able to discriminate fast- and slow-
drying regimes. The detail of the water concentration profiles as a function of
time were found to be sensitive to the pore-size distribution of the catalyst
support. A numerical analysis of the quantitative liquid concentration profiles
was performed by fitting the profiles to the diffusion equation, allowing for a
concentration-dependent diffusivity. It was shown that for liquids characterised
by low surface tension, such as acetone, benzene and cyclohexane, transforma-
tions of the concentration profiles could be adequately modelled assuming a
liquid-independent diffusivity. In contrast, the diffusivity of water in both ti-

tania and alumina pellets substantially decreased as a function of water content.
MR imaging has also been demonstrated to be a potentially useful tool for
characterising metal-ion distribution within catalyst pellets, both in a static state
and as an in situ measurement during catalyst synthesis. Khitrina et al. (2000)
have exploited the effect of relaxation time contrast within images to map out
the metal-ion distribution. It is non-trivial, and it has not yet been attempted to
make such measurements quantitative, but the method does provide a non-
destructive way of identifying the position of metal bands within the catalyst
that is consistent with the position identified by an electron probe micro-an-
alyser. The approach uses 1H micro-imaging to follow the distribution and spin-
lattice relaxation time of cyclohexane within the catalyst. Cyclohexane species
interacting with the hexachloroplatinate dianion during the timescale of the
measurement are associated with an increased spin-lattice relaxation time;
therefore, T1-maps of the sample identify the location of the hexachloroplat-
inate dianion within the porous catalyst support. This methodology has been
used to follow the dynamics of hexachloroplatinate dianion re-distribution
during a competitive impregnation of the support with an aqueous solution of
H2PtCl6+H2C2O4. It is seen that after 5 h of impregnation, the egg-white-type
profile transforms into a classic egg-yolk profile.
At a slightly larger length-scale we can also follow in real time the water loss
during drying from liquid-saturated arrays of particles, as was shown in Fig. 14.
Although this work was originally undertaken to investigate product hetero-
geneity arising from a different particle-drying process, this experimental con-
figuration has immediate relevance to catalyst synthesis in the context of the
drying of a batch of catalyst pellets following a wet impregnation synthesis
(Gladden et al., 2004).
3. Reaction
This is an area in which there is likely to be considerable future interest. Of
course, apart from the challenge of being able to measure species-dependent
profiles within the pellet, we need to think carefully about how these data will
be used. To date, images of liquid distribution during chemical reaction within
a cylindrical Pt/g-Al2O3 catalyst pellet (diameter and height 4.7 mm) have been
reported (Koptyug et al., 2002). In-plane spatial resolution was 230 mm -
140 mm, with an image slice thickness of 2 mm. A stream of hydrogen gas heated
to 67–69 1C and saturated with a-methylstyrene vapour was supplied to the
pellet at a flow rate of 18.5 cm3/s. 1H images of the evolution of the liquid
distribution within the pellet were recorded during a-methylstyrene evaporation
accompanied by its vapour-phase hydrogenation. To date there have not been
any reports of chemically resolved images within individual supported-metal
catalyst pellete—the 1H signal in the previous example did not differentiate
between reactants and product species. This is because of the severe line
broadening that occurs in the 1H-MR resonances associated with individual

molecular species when these molecules interact with the surface of the catalyst,
causing a reduction in the relaxation times of those species which, in turn, is
manifested by an increased line width in the frequency spectrum. However, this
is not necessarily an insurmountable problem. Judicious selection of reaction
systems may identify specific systems in which the relevant 1H resonances can be
discriminated. Alternatively, imaging of nuclei other than 1H may provide the
solution—13C is an obvious choice. While the sensitivity of a 13C measurement
is substantially less than that of a 1H measurement, the existence of fewer, well-
separated 13C lines in a typical 13C spectrum of a reaction mixture might make
such measurements possible (see Section IV.D).
4. Coke Deposition
A natural extension of the studies of single pellet micro-imaging is to explore
the effect of pore structure on the spatial distribution of coke deposition within
the pellet. MR images have clearly shown that even while visual observations of
pellets as a function of coking might be consistent with the shrinking-core model
commonly employed in chemical engineering (i.e. coke deposition proceeds to-
wards the centre of the pellet at a uniform rate such that the catalytically active
region of the catalyst takes the form of a shrinking core of un-coked catalyst),
coke deposition actually proceeds in a much more heterogeneous fashion with
regions of coke deposition (of a length-scale corresponding to that of the length-
scale of heterogeneities in the pore structure) occurring towards the interior of
the pellet ahead of a higher concentration of coke deposition which is well
described by the shrinking core concept (Cheah et al., 1994). A more recent
study by Bonardet et al. (1999) employs a different strategy. These workers
imaged the 1H spin density associated with 2,3-dimethylpentane adsorbed on
pellets of HY zeolites coked to levels of 7.5 and 10% (w/w). The 1-D image
profiles of the 7.5% (w/w) coked sample showed a heterogeneous coke distri-
bution in the pellet at the macroscopic level. The most heavily coked region is
that which was first exposed to reaction, and the heterogeneity in coking was
again found to be associated with structural heterogeneities arising from the
pellet manufacturing process. Bonardet et al. also concluded that the coke is
heterogeneous in quality throughout the pellet. By varying the echo time in the
r.f. pulse sequence it was found that the T2 of the probe molecule, which is a
function of the aromaticity of the coke, varies within the pellet. In particular,
the more heavily coked zones were associated with coke characterised by the
shortest T2 values, indicative of coke of the most graphitic content. The more
highly coked sample had a more homogeneous coke distribution, with the coke
being of a homogeneous graphitic content throughout. A third approach to
addressing coke deposition has recently been demonstrated by Bar et al. (2002),
in a study of the location of coke deposits in industrial HZSM-5 pellets, of
diameter 5 mm, contained in a small fixed-bed reactor. The coke distribution
was detected directly using the Single Point Ramped Imaging with T1 En-
hancement (SPRITE) technique, a special MR imaging sequence for detecting
materials with short T2 relaxation times. In the example given, SPRITE was
used to produce a 1-D coke profile along the axis of the model fixed bed of inner
diameter 3 cm, containing two layers of coked pellets (20.5 wt% coke) separated
by a 3.3-cm layer of fresh pellets. A spatial resolution of 0.5 cm was obtained,
this being limited by the rapid nuclear spin relaxation times of the sample. These
workers also used an approach similar to that of Bonardet et al. (1999), in which
the presence of carbonaceous deposits is observed indirectly by imaging the 1H
spin density of propane loaded within the sample. The different adsorption
strengths of propane on the fresh and coked HZSM-5 cause T4 to vary between
the fresh and coked regions, hence allowing the regions of coked and fresh
catalyst to be identified. The effect of coke deposition upon pore structure and
molecular diffusion within supported-metal catalysts has also been studied
(Wood and Gladden, 2003).
C. SINGLE-PHASE FLOW IN FIXED-BED REACTORS
There exist a wealth of studies of spatially unresolved measurements of single-

phase flow in model porous structures such as bead packs (e.g. Manz et al.,
1999a; Seymour and Callaghan, 1997; Stapf et al., 1998). The motivation for
many such studies has been to understand flow within the porous structure of
rock cores. In particular, PGSE-MR techniques have been used to measure
diffusion and dispersion within such systems. However, it was not until rela-
tively recently that the relevance of these measurements to the study of flow in
fixed-bed reactors was identified. Indeed, fixed beds of column-to-particle di-
ameter ratio 10–20 are an ideal system in which to apply MR flow imaging
experiments since the flow field within the inter-particle space can be well re-
solved. While many industrial reactors are characterised by column-to-particle
diameter ratios 420, the results obtained for these narrower beds are still rel-
evant. Importantly we have shown that the major feature of the flow field
observed in narrow beds (i.e. channelling throughout the entire bed, not just at
the walls) is still observed for beds of column-to-particle diameter ratio of 40
(which is the largest bed we have studied). It should also be remembered that
studies of beds of column-to-particle diameter ratio p20 will be directly rel-
evant to reactor designs used for reactions characterised by high exothermicity
and relatively poor heat transfer, such as the synthesis of methyl-isobutyl ketone
(Mariani et al., 2001) and the conversion of natural gas into transportation fuels
(Sie et al., 1991). Even when the ultimate objective is in understanding the
operation of larger column-to-particle diameter ratio beds, the MR imaging
data are still valuable since they provide a wealth of data that can be used for
the development and validation of numerical codes, which can then be used in
subsequent scale-up studies. Further, MR imaging may also be used to identify
a specific phenomenon that influences process operation; once identified it may

then be possible to develop a cheap, robust probe, such as pressure drop meas-
urement, to monitor that phenomenon at plant-scale.
Early MR flow imaging studies of single-phase flow in fixed beds were re-
ported for packings of non-porous particles. Figure 25 shows 2-D sections
through 3-D volume images of the x, y and z components of flow within a fixed
bed of non-porous spherical particles. The map of the z-component of the flow
velocity is the most interesting; the +z-direction is the direction of superficial
flow in the reactor. In this particular example, the superficial flow velocity was
0.56 mm/s, corresponding to a Reynolds number of 2.8. Hence, flow in much of
the bed is dominated by viscous forces, associated with flow velocities less than,
or of the order of, the superficial velocity. The most striking characteristic of
these images is the extent of heterogeneity in the flow field; a relatively small
fraction of the inter-particle space carries a high percentage of the liquid flow
(Sederman et al., 1997, 1998). Such regions of the bed are associated with high
fluid velocities, and inertial effects increasingly influence the flow profile (Johns
et al., 2000). On the basis of these images, it is clear that any theoretical analysis
of the flow within such a reactor must account for distinct populations of fast-
and slow-moving liquid, as channelling does not just occur at the walls of the
bed. The contact time between feed and catalyst will differ very significantly
(a) (b) (c)

yz yz yz
xy xy xy
xz xz xz
y
−7.2 mm s−1 Vx, Vy 7.2 mm s−1
x
z −2.7 mm s−1 Vz 9.0 mm s−1
FIG. 25. MR visualisation of water flowing within a fixed bed of spherical glass beads; the beads
have no MR signal intensity associated with them and are identified as black voxels. Flow velocities
in the (a) z- (b) x- and (c) y-directions are shown with slices taken in the xy, yz and xz planes for each
of the velocity components. For each image the positions at which the slices in the other two
directions have been taken are identified. Voxel resolution is 195 mm 195 mm 195 mm. The glass
beads are of diameter 5 mm and are packed within a column of internal diameter 46 mm. Typically
40% of the flow is carried by 20% of the inter-particle space within any 2-D slice section through
the bed, perpendicular to the direction of superficial flow. Regions of high and low flow velocity in
the direction of superficial flow are highlighted in (a). Reprinted from Sederman and Gladden,
(2001a), with permission from Elsevier. Copyright (2001).
across the bed, i.e. by up to at least one order of magnitude in regions of the bed
characterised by the highest and lowest flow velocities, and this will introduce
spatially varying mass transfer characteristics within the bed. The spatial res-
olution of the flow field is such that the flow profile between individual particles
can be extracted from the data. This is clearly illustrated in Fig. 26 in which it is
seen that at low velocities (or more precisely, at low local Reynolds numbers)
parabolic, laminar flow is observed, whereas at higher local velocities the flow
profile flattens such that it is much more characteristic of plug flow. MR data of
this type are useful for identifying how catalyst size, shape and method of
loading into the reactor influence heterogeneities in hydrodynamics while, at a
more academic level, the combination of 3-D MR images of the bed structure
with flow visualisation allows us to explore how the geometry and inter-
connectivity of the inter-particle space determine the local flow characteristics
within the bed. These insights increase our generic understanding of fluid
transport in porous materials and are equally relevant to understanding fluid
transport processes in rocks and soils, with immediate application to oil recov-
ery and groundwater remediation processes, respectively.
Figures 27 and 28 show how the combined application of MR imaging and
flow visualisation allows us to study the deposition of fines within fixed beds.
This is a common problem in reaction engineering. For example, during process
1.0
vz (normalised)
0.5
0.0
0.0 1.0 2.0
d (mm)
FIG. 26. The velocity profile for flow of water through two different regions (highlighted in Fig.
25a) of the inter-particle space within a fixed bed of spherical glass beads. The velocity profiles are
measured across the inter-particle space between two packing elements. Profiles are shown for local
regions associated with fast and slow flow velocities, characterised by a local Reynolds number of 50
and 12, respectively. At low Re number, a parabolic flow profile typical of Poiseuille flow is seen
(____). At the higher Re number, inertial effects in the flow are evident and the flow profile ap-
proaches that of plug flow (——). For this particular bed, the Re number based on bed diameter is
of Re
15. Designs based on such a value may well not adequately describe the true hydrodynamics of
the system. Re is defined as rvA1=2 =m where r and m are the density and viscosity of the fluid
phase, and v and A1/2 are a characteristic fluid velocity and length-scale, respectively, of the system
under study. Reproduced with permission from Johns et al. (2000). Copyright (2000), A.I.Ch.E.
FIG. 27. 2-D-MR image of the deposition of 80 mm particles (‘‘fines’’) within a bed packed with
5 mm spherical glass beads. The water flowrate was 300 ml/min. All images were acquired in 3-D
with isotropic spatial resolution of 188 mm 188 mm 188 mm. Two local regions associated with a
build-up of fines are highlighted, and are identified by the low apparent 1H spin density from these
regions resulting from low voidage and relaxation time effects. Flow is in the +z-direction. Re-
printed from Gladden (2003), with kind permission of Springer Science and Business Media.
(a) (b) (c)

yz yz yz
xy xy xy
xz xz xz
y
-5.8 s−1 (exz, exy) 5.8 s−1
x
z 0.0 s−1 |ez| 10.0 s−1
FIG. 28. Images of the liquid stress tensor derived from the data shown in Fig. 25. Data are shown
for (a) exz , (b) exy and (c) jez j with slices taken in the xy, yz and xz planes for each of the shear
components. Regions of high shear are particularly evident in (c), i.e. along the direction of su-
perficial flow, z. Reprinted from Sederman and Gladden (2001a), with permission from Elsevier.
Copyright (2001).
operation, attrition of the catalyst will occur, and the resulting ‘‘fines’’ will
deposit throughout the bed. These deposits will then influence the flow paths
through the bed and this, in turn, can influence the operating conditions within
the bed (e.g. pressure drop) and chemical conversion. Exactly the same exper-
imental strategy would allow us to understand and optimise filtration processes.
Figure 27 shows an image of water concentration (or, more precisely, 1H spin
density) within a 2-D slice section through a fixed bed of glass beads. The beads
are seen as black on the colour scale and the free water as white. Regions in
which fines are deposited are readily identified as the grey contrast level; when
fines are deposited, they pack to produce regions of low voidage, and are hence
observed as regions of reduced water (1H) content. The images shown in Fig. 28
explain how the MR flow images shown in Fig. 25 may be used to understand
the phenomena leading to deposition. Figure 28 shows maps of the shear stress
in the liquid as it moves through the bed. These maps of shear stress within the
fluid have been calculated directly from the flow visualisations shown in Fig. 25
(Sederman and Gladden, 2001a). From these maps, we see precisely where
regions of high liquid shear stress (lighter shades) exist within the bed; such
regions will be associated with particle erosion but are not likely to be regions in
which significant fines deposition will occur. In contrast, in regions of low shear
stress little particle erosion will occur, but we will expect these regions to be
particularly susceptible to fines deposition. As regions of particle deposition
extend within the bed, the hydrodynamics and pressure drop characteristics of
the bed will also change. Such images can be used directly in optimising the
hydrodynamic characteristics of the bed to promote or minimise the existence of
particular flow phenomena.
High-resolution images of the type shown in Fig. 25 combined with 3-D high-
resolution images of the structure of the bed can also be used to study the more
fundamental aspects of structure–transport relationships characterising the
fixed bed itself. In a series of papers (Baldwin et al., 1996; Johns et al., 2000;
Sederman et al., 1997; Sederman et al., 2001) it has been shown that in addition
to extracting data such as radial distribution functions of either the void space
or packing elements, image-analysis techniques can be used to partition the void
space characterising the bed into individual elements—referred to as ‘‘pores’’—
which are characterised by their volume, surface area and connectivity to each
other. By combining this information with the images of flow occurring within
the same bed, the influence of pore size, shape and connectivity on the flow
within the void space of the bed can be investigated. This field is now receiving
increasing interest both at an academic level and also from catalyst manufac-
turers interested in investigating how subtle changes in pellet size and shape can
influence the spatial distribution of porosity and, hence, the hydrodynamics
within the reactor (e.g. Gotz et al., 2002). As in the case of the micro-imaging
studies mentioned earlier (Section IV.B), a real advantage of employing MR
techniques is the ability to quantify the extent and nature of heterogeneity in
both structure and transport characteristics characterising the system under
study.
The combination of techniques and research methodology described above is
being applied to other fields. Examples include assessing the performance of
filtration processes (e.g. Dirckx et al., 2000) and biofilm reactors (e.g. Beuling
et al., 1998; Nott et al., 2001). Both these applications exploit the capability of
measuring the flow field within the process unit in combination with imaging the
internal structure of the system so as to understand the inter-relationship of the
rate of formation and structure of deposited material with the local and macro-
scale hydrodynamics. Other applications include visualisation of flow within
hollow-fibre bioreactors (Heath et al., 1990), and heat exchangers (Pangrle
et al., 1992; Sun and Hall, 2001; Wang et al., 1999). Extensive studies of chro-
matographic column design and operation have been published by Tallarek,
Van As and co-workers addressing issues such as the quantification of the mass
transfer between the intra-particle pore network of the porous particles and the
inter-particle void space comprising the column, as well as dispersion and
electroosmotic perfusion phenomena (Tallerek et al., 1996, 1999, 2001).
D. MEASURING CHEMICAL COMPOSITION AND MASS TRANSFER IN FIXED-BED

REACTORS: IN SITU STUDIES OF REACTIONS
Recently, there has been a surge of interest in this area, but it is important to
remember that early studies were reported as far back as 1978, albeit in rel-
atively simple systems. In 1978, Heink et al. (1978) used 1-D profiling to study
the time-resolved concentration profile of butane and water in packings of
NaCaA and NaX crystallites, respectively. Other early studies include the use of
19
F imaging to follow the intercalation of AsF5 into highly oriented pyrolytic
graphite at room temperature (Chingas et al., 1986). Since then, interest has
focussed mainly on oscillations and travelling waves during chemical reaction
(Balcom et al., 1992; Menzinger et al., 1992; Ra’bai et al., 1990; Scott, 1987;
Tzalmona et al., 1990, 1992). Butler et al. (1992) employed 1H-MR imaging
techniques to investigate the extent of reaction in a single crystal of 4-
bromobenzoic acid during exposure to ammonia gas. A second case study used
MR imaging to follow the reaction of a deep bed of powdered toluic powder
with ammonia gas flowing over it. The apparent reaction rate constant and the
effective ammonia diffusion coefficient (into the toluic bed, perpendicular to
the direction of flow) were obtained from a fit of a diffusion–reaction model
to the experimental data. In principle, all the methods of MR spectroscopy can
be integrated into an MR imaging sequence; hence, spatially resolved meas-
urements of chemical conversion should be possible. However, the practical
challenges in achieving this are substantial. In particular, as with any measure-
ment, achieving adequate signal-to-noise is the key to a successful measurement.
This is why in the vast majority of MR imaging experiments, signal from the 1H
nucleus is acquired. The 1H nucleus has high MR sensitivity and exists in 100%
natural abundance. The problem is that in any reaction mixture there is likely to
be a wealth of 1H resonances many of which will overlap, making it impossible
to follow the change in concentration of a particular species. This problem is
increased because of the interaction of the fluid phase with the solid-phase
catalyst pellets, which causes the relaxation times of the fluid species to decrease.
This further reduces the available signal-to-noise in the experiment and also
broadens the spectral resonances, further increasing the overlap of individual
resonances. For this reason 1H studies are likely to be limited to the study of
simple reactions in which the 1H spectral peaks are readily resolved. Alterna-
tively, we must also consider 13C, 31P and 19F imaging, if possible. 31P and 19F
are commonly used in the medical field and have high abundance and sensi-
tivity. 13C is the most likely candidate species to be studied in heterogeneous
catalytic processes, but successful implementation of the technique is not
straightforward. Despite these difficulties, initial 1H experiments and, most re-
cently, the first report of 13C observation in this application suggest that sub-
stantial new insights into the coupling of hydrodynamics and chemical
conversion can be obtained by developing the appropriate MR techniques.
Albeit in a simple reaction, the nature of the information that can be obtained
regarding chemical mapping and quantification of mass transfer processes has
been demonstrated using the liquid-phase esterification of methanol and acetic
acid catalysed within a fixed bed of H+-ion exchange resin (particle size
600–850 mm). The purpose of this initial experiment was to confirm the link
between the heterogeneity in hydrodynamics identified in Section IV.C and
possible heterogeneity in conversion within the bed. This study is a good ex-
ample of how the attributes of MR spectroscopy and imaging can be combined.
In principle, there is no reason why the chemical mapping of the experiment
cannot be performed at the spatial resolution achieved in the imaging of the
structure of the bed. However, in the present example, spectra were acquired
from relatively large volumes within the bed (1.5 mm 1.5 mm 0.5 mm) to
ensure high signal-to-noise data for subsequent analysis. Using this approach,
we were able to study quantitatively the extent of conversion at various loca-
tions within transverse sections through the bed and at several positions along
the length of the bed. The experiment is reported in detail elsewhere (Yuen et al.,
2002). The technique of volume-selective spectroscopy (Kimmich and Hoepfel,
1987) is used to obtain, non-invasively, MR spectra from the well-defined vol-
ume elements with the bed; this is achieved by the application of magnetic field
gradients to selectively excite spins only within the volume of interest. The
chemical conversion within each volume is obtained directly from the positions
of the spectral peaks within the 1H spectrum. The measurement of chemical
composition, and hence conversion, within the bed is entirely non-invasive,
thereby overcoming the problem of introducing sampling points within the bed,
which will disturb the local hydrodynamics and potentially influence the local
conversion. It is worth summarizing the principle of the measurement since it is
a way of overcoming the problem of not being able to resolve molecule-specific
resonances (or ‘peaks’) in the 1H spectrum. Quantification of chemical com-
position is achieved by exploiting the phenomenon of fast exchange of the 1H
species associated with hydroxyl groups within the acetic acid, methanol and
water species comprising the reaction mixture:
CH3 OH þ CH3 COOH Ð CH3 COOCH3 þ H2 O (23)

As a result of 1H fast exchange, the observed chemical shift (i.e. resonance

frequency) for the 1H species associated with OH groups present in the reaction
mixture (dobserved ) is determined by an average of the component chemical shifts,
weighted by the amount of each molecular species present, as follows from Eq.
(23). Thus, the concentration of acetic acid present at any time and, hence, the
extent of conversion at that time is determined directly from the value of
dobserved identified in the volume selective spectrum. Figure 29 shows a 2-D slice
section through a 3-D RARE image of the bed; the image voxels are of size
97.7 mm 97.7 mm 97.7 mm. 1H-MR spectra have been recorded from regions
of in-plane dimension 1.5 mm 1.5 mm, with a slice thickness of 500 mm in the
direction of superficial flow. From each spectrum it is possible to determine,
directly and quantitatively, the extent of conversion within a given local volume,
as shown in Figure 30. All images in Fig. 30 show that there is significant
heterogeneity in conversion within a single transverse section through the bed,
perpendicular to the direction of superficial flow; fractional variations in con-
version of up to 20% are typical under steady-state operating conditions.
Although such studies are in their early stages, we clearly have the measurement
tools to study the complex interaction of hydrodynamics and chemical kinetics
in the complex porous medium represented by a fixed bed.
While the esterification experiment described in the previous paragraph has
provided the most detailed in situ visualisation of a heterogeneous catalytic
reaction to date, the approach to measuring conversion cannot be used rou-
tinely for other reactions. This is because we are monitoring quite small var-
iations in chemical shift to calculate conversion, which requires a relatively
simple 1H spectrum that can be interpreted unambiguously. Two other studies
have been reported recently. In the first of these, Koptyug et al. (2004) used 1H-
NMR to produce spatially resolved spectra within a 2-D slice section along the
axial direction of a fixed bed of Pd/Al2O3 (1%, (w/w)) catalyst pellets. The
reaction considered was that of the hydrogenation of a-methylstyrene to
cumene. The spectra show clear evidence of changes in chemical composition
along the length of the bed, but conversion is not quantified, most likely owing
to problems in deconvolving the 1H resonances from the reactant and product
species. The most recent study reports the successful implementation of 13C
imaging to study reaction in situ (Akpa et al., 2005), illustrated by the com-
petitive etherification and hydration reactions of 2-methyl-2-butene occurring
within a fixed bed of H+-ion exchange resin. By exploiting polarisation transfer
techniques, sufficient signal-to-noise is achieved for spatial resolution of chem-
ical composition, and hence conversion, within the bed to be obtained without
the need for 13C isotopic enrichment. Although much work remains to be done,
the ability to spatially map 13C signals at natural isotopic abundance opens
up significant new opportunities for widespread studies exploring the interaction
of hydrodynamics, mass transfer and chemical kinetics within catalytic
reactors. Our ultimate goal must be to gain enough understanding to design
an integrated, optimised (scaled-up) catalyst reactor system that can deliver the
FIG. 29. 2-D slice through a 3-D RARE image of a fixed bed of ion exchange resin. The image has
an isotropic resolution of 97.7 mm 97.7 mm 97.7 mm. The image slice in which the local volumes
are located for the volume-selective spectroscopy study is identified. The image was acquired by
saturating the bed with pure methanol. The image is acquired employing T2 contrast such that signal
is acquired only from the methanol in the inter-particle space. The arrow indicates the direction of
superficial flow. Reprinted from Gladden (2003) with kind permission of Springer Science and
Business Media.
high activities and selectivities that the catalyst can achieve at small scale in
powder form.
Further extensions of chemical mapping within a fixed bed lie in quantifying
mass transfer between the intra-pellet and inter-pellet pore space. One approach
to measuring mass transfer processes is to use displacement propagator meas-
urements (see Section II.C). The data acquired for the esterification reaction
described previously are shown in Fig. 31. In this propagator measurement, the
total propagator measured for the system has been separated into two
(a) (b) (c)
10% X 54%
FIG. 30. Visualisation of mean conversion, X, within selected volumes located within the slice
section identified in Fig. 29. The local volumes have in-plane dimensions of 1.5 mm 1.5 mm, and
have a depth (image slice thickness) of 500 mm. Data are shown for three feed flow rates: (a) 0.025,
(b) 0.05 and (c) 0.1 ml/min. As flow rate increases, so the residence time characterising the bed
decreases and it therefore follows that conversion will decrease. Reprinted from Gladden (2003),
with kind permission of Springer Science and Business Media.
component propagators by combining the transport measurement pulse sequence

with a spin-lattice relaxation time experiment. Since the relaxation time of liquid
molecules that have existed only within the inter-particle space during the ob-
servation time (100 ms) of the experiment is significantly longer than that of
molecules that have moved between the inter- and intra-particle space, independ-
ent propagators characterising these two populations of liquid molecules can be
obtained (Gladden et al., 2003a). In Fig. 31 the broader propagator, which shows
two peaks, is that associated with the liquid in the inter-particle space. The two
peaks are consistent with there being populations of very slow-moving and much
faster moving fluid within the bed—note that these observations confirm that the
flow heterogeneity observed in beds of low column-to-particle diameter (10—
Fig. 25) is also observed for beds of significantly higher column-to-particle ratios
(20). The full-width at half maximum of the (narrower) ‘‘exchange’’ propagator
provides an estimate of the effective diffusion coefficient for water molecules
moving between the pore space of the catalyst and the inter-particle space of the
bed of 2 109 m2/s, which gives a lower limit to the value for the mass transfer
coefficient of 4 106 m/s. This value is obtained by defining a mass transfer
coefficient as D=d, where d is a typical distance travelled to the surface of the
catalyst that we estimate as half a typical bead dimension (500 mm). Such a mass
transfer coefficient would give rise to a rate of reaction of 2 103s1, con-
sistent with the reaction occurring under conditions of kinetic control.
E. TWO-PHASE FLOW IN FIXED-BED REACTORS
The next level of complexity in implementing MR to study fixed-bed proc-

esses is to study two-phase flow phenomena. Initial studies focussed less on fully
1.0
0.8
0.6
Ps(Z,T1) (arbitrary units)
0.4
0.2
0.0
−0.2
−0.1 −0.05 0.0 0.05 0.1 0.15 0.2
Z (mm)
FIG. 31. T1-resolved propagators for water flowing within the inter-particle space of a bed packed
with ion exchange resin () and for water exchanging between inter- and intraparticle environments
(’) during the timescale of the transport measurement. Data are shown for a volumetric feed flow
rate of 2 ml/min, to a column of internal diameter 20 mm.
resolving the liquid flow field and more on the macroscopic gas–liquid distri-
bution within the bed, with the specific aim of measuring liquid holdup (w) and
surface wetting. The latter is particularly important since MR provides the first,
direct, non-invasive measure of this quantity. The first reported MR imaging
study of liquid holdup and wetting in two-phase flow within a fixed bed was
performed on a column of internal diameter 40 mm, packed with 5-mm-diam-
eter non-porous glass spheres (Sederman and Gladden, 2001b). This system was
considered a model system for a trickle bed, and it was demonstrated that
measurements of both liquid holdup in quantitative agreement with gravimetric
data, and surface wetting could be obtained directly. The typical spatial res-
olution obtained in these measurements was 300 mm; therefore, the absolute
thickness of wetted films on the surface of the packing elements is not imaged.
Instead, by selecting a suitable gating level on the image intensity that discrim-
inates between image pixels that are part-filled with liquid (i.e. containing a
solid/gas/liquid interface) and those that have zero liquid content, it is possible
to identify the presence of liquid films on the surfaces of the glass spheres
unambiguously. Liquid holdup and wetting efficiency (ws ), as defined below,
were determined:
w.—The fraction of (inter-particle) void space pixels containing some liquid
provides an upper estimate of liquid saturation, from which values of liquid
holdup are obtained. By extrapolation of the data to zero liquid superficial
velocity, the static liquid holdup is identified, and the dynamic liquid holdup,
wdynamic , is obtained.
ws .—The wetting efficiency is obtained by calculating the fraction of the pixels

identifying the surface of the packing that are in contact with liquid during
gas–liquid flow. Liquid-containing voxels adjacent to the wall of the column,
and the internal surface of the porous packing elements are not considered in
the analysis.
Figure 32 shows the nature of the data obtained. In Fig. 32b the total liquid
holdup in the image slice has been segmented in to what we call rivulets using
the pore space segmentation algorithm referred to in Section IV.C. Application
of this algorithm is merely an objective, reproducible procedure to characterise
the liquid distribution within the bed. By recording these data in 3-D, a range of
statistics can be produced, such as the number of ‘‘pores’’ within the bed as-
sociated with different levels of fractional filling and the number distribution of
the fractional surface wetting of the packing elements.
Extension of this methodology to packings of porous packing elements (e.g.
catalyst support pellets) is not straightforward. The challenge arises because the
signal we wish to measure is associated with the water in the bed. However, the
signal intensity we acquire from a specific region of water will depend on its
local environment, because the nuclear spin relaxation times of water in dif-
ferent physical environments will be different. In this system, the different en-
vironments will be (i) free water in the bulk of the inter-pellet space, (ii) water
(a) (b) (c)
FIG. 32. Identification of rivulets and surface wetting in a packing of 5 mm diameter glass spheres
contained within a column of internal diameter 40 mm. The data were acquired in a 3-D array with
an isotropic voxel resolution of 328 mm 328 mm 328 mm. (a) The original image of trickle flow was
first binary-gated, so that only the liquid distribution within the image is seen (white); gas-filled
pixels and pixels containing glass spheres show up as zero intensity (black). (b) The liquid distri-
bution is broken up into individual liquid rivulets, each identified by a different shade on a grey
scale. (c) Pixels containing any liquid–solid interface are then identified using image analysis tech-
niques and ‘‘images’’ of surface wetting are produced. Data are shown for liquid and gas superficial
flow velocities of 3 and 66 mm/s, respectively. Reprinted from Sederman and Gladden (2001b), with
permission from Elsevier. Copyright (2001).
within the intra-pellet pore space and (iii) water existing in films on the surface
of the pellets but not part of a rivulet within the inter-pellet space. Initial studies
of holdup and wetting during trickle flow in a fixed bed packed with cylindrical,
porous alumina extrudate have been reported. Data were presented for two
packings: (i) packing of diameter 1.5 mm and a distribution of lengths in the
range 572 mm; (ii) packing of diameter and length equal to 3 mm. A constant
gas superficial velocity of 31.3 mm/s was used, with liquid superficial velocities
in the range 0.1–6 mm/s. 2-D visualisations of liquid distribution within trans-
verse sections, of thickness 1 mm, were acquired. The total data acquisition time
for each image was 25 min. Data were recorded with a field-of-view of
40 mm 40 mm and a data array size of 256 256, thereby yielding an in-plane
spatial resolution of 156 mm 156 mm (Gladden et al., 2003b).
Figure 33 shows plots of dynamic liquid holdup and wetting efficiency against
liquid superficial velocity for a constant gas velocity of 31.3 mm/s (Gladden
et al., 2003b). It is clearly seen that dynamic liquid holdup increases more
rapidly as a function of liquid superficial velocity within the 1.5 mm packing,
and values of holdup and wetting efficiency are always greater, for a given liquid
velocity, for the 1.5-mm-diameter packing relative to the 3-mm-diameter pack-
ing. The line through the dynamic liquid holdup data is the best fit of the
percolation-based model described by Crine et al. (1992). The form of the ex-
pression for the dynamic liquid holdup is
2=3
1=3 Q
wdynamic ¼ ðKQÞ (24)
Q þ Qmin
0.4 0.6
0.5
0.3
dynamic
0.4
s
0.2
3 mm pellets 0.3 3 mm pellets
1.5 mm pellets 1.5 mm pellets
0.1 0.2
0 1 2 3 4 0 1 2 3 4
(a) Liquid flow rate (kg m−2 s−1) (b) Liquid flow rate (kg m−2 s−1)
FIG. 33. (a) Dynamic liquid hold-up, and (b) wetting efficiency as a function of liquid superficial
velocity for 1.5 and 3 mm cylinders. Gas flow rate is constant at 66 mm/s. The line shows the best fit
of the percolation model of Crine et al. (1992). Reprinted from Gladden et al. (2003b), with per-
mission from Elsevier. Copyright (2003).
where Q is the liquid superficial velocity and Qmin the minimum liquid
superficial velocity. K ¼ kmL a2 =rL g, where k is a proportionality factor de-
pending on the fluid and packing properties, mL the liquid dynamic viscosity, rL
the liquid mass density, a the specific surface of the packing and g the accel-
eration due to gravity. Equation (24) is fitted to the experimental data, with
Qmin and K (i.e. k) as variables in the fit. As seen from Fig. 33, the fit of Eq. (24)
to the data is good. The values of Qmin obtained are 3.56 104 m/s and
12.5 104 m/s for the 1.5- and 3-mm-diameter cylinders, respectively. Follow-
ing the argument of Toye et al. (1996), Qmin characterises solid-phase (i.e.
packing) wettability such that smaller values of Qmin are associated with better
packing wettability. This, as discussed earlier, is clearly supported by inspection
of Fig. 33. These early studies of trickle flow within beds of porous packing
elements identified three general results: (i) values of liquid holdup compare
well, typically to within 5% of gravimetric data; (ii) the general trends in both
holdup and wetting data are consistent with the predictions of existing models in
the literature; (iii) the absolute value of surface wetting tended to be lower than
that previously reported in the literature. Considering (iii), there are, of course,
likely to be significant errors in the values of wetting obtained from the range of
indirect methods used in the earlier works. However, an underestimation of
surface wetting using data obtained from MR imaging can be explained as a
consequence of not ‘‘seeing’’ liquid layers on the surface when they are not
associated with a larger scale liquid rivulet. As a result of this observation,
improved data-acquisition and image-analysis strategies have been developed.
The key modification to the overall methodology is that we apply an image-
analysis algorithm that applies a local gating level: this is an objective procedure
that accounts for the fact that the gating level we choose must discriminate
between intra-pellet water and water bound to the surface as a wetted film. The
particular gating level needed to achieve this will be very sensitive to the signal
from the intra-pellet water, and since this varies with the characteristics of
individual pellets, a global gating level will always introduce inaccuracies into
the measurement. Using this improved algorithm, MR measurements of both
holdup and wetting are in good agreement with the predictions of the neural
network analysis of Larachi et al. (1999).
F. HYDRODYNAMIC TRANSITIONS IN FIXED-BED REACTORS
Trickle- and pulse-flow regimes are the contacting patterns most commonly
encountered in commercial-scale trickle beds; therefore, understanding the na-
ture and characteristics of the hydrodynamics in these flow regimes and the
transitions between them are subjects of long-standing interest (e.g. Blok et al.,
1983; Boelhouwer et al., 2002; Dankworth et al., 1990; Grosser et al., 1988;
Holub et al., 1992; Larachi et al. 1999; Ng, 1986; Reinecke and Mewes, 1997;
Sicardi and Hofmann, 1980). The flow pattern will significantly influence the
performance of a given reactor through characteristics such as phase holdups,
power consumption and mass-transfer fluxes, so successful modelling of trickle-

bed reactors requires precise tools for the identification of the flow pattern
expected for a specified set of operation conditions. An accurate prediction of
the trickle-to-pulse transition is also of key importance in this regard (Larachi
et al., 1999). Extending MR to study unsteady-state flows requires rapid data
acquisition times. In this case, FLASH techniques (see Section II.F) were em-
ployed, and spatial resolution was reduced such that sufficient signal-to-noise
was achieved. This is the subject of ongoing work, but some interesting new
observations have already been made. This research also highlights that every
time we implement a new type of data acquisition we have to think carefully
about how to process the data so that the important information is captured.
Images were acquired as a data array of size 32 16 (in-plane spatial resolution
1.4 mm 2.8 mm), with an acquisition time of 20 ms for a 2-mm slice thickness.
Images were acquired in immediate succession, and frame rates of 50 frames per
second (f.p.s.) were achieved. The maximum number of images acquired in a
single series was 540, this number being limited by hardware considerations.
High-resolution images of 2-D slice sections through the bed were also acquired
to provide accurate identification of the position of packing elements within the
bed. These were acquired using a standard spin-echo imaging sequence with an
in-plane resolution of 175 mm 175 mm for a slice thickness of 1 mm. The
methods we have chosen are outlined below:
Signal intensity in the images is acquired only from the liquid phase. The MR
imaging data obtained have been analysed using the following two procedures:
(i) Standard deviation maps: The simplest way to assess the stability of the
gas–liquid distribution is to calculate a map of the standard deviation of the
pixel intensities for each pixel in a time series of images. Thus, for a series of
n images, the standard deviation of the intensity associated with pixel i, si , is
calculated as follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
ðxi x̄i Þ2
si ¼ (25)
n
where xi is the signal intensity of pixel i, and x̄i the average intensity of
pixel i in the series of n images, the summation being taken over n images.
(ii) Temporal autocorrelation function ðRI ðtÞÞ plots of effective liquid holdup:
These are calculated from the signal intensity (i.e. effective liquid holdup)
data as follows:

IðtÞIðt þ tÞ
RI ðtÞ ¼
2 (26)
I
where IðtÞ is the total signal intensity within the region of interest acquired
at time t. This analysis extracts correlations in the signal intensity for all
time separations t and highlights any periodicity in the holdup data and the
timescales over which they exist.
Figure 34 shows three time series of 2-D magnetic resonance images of liquid
distribution (Lim et al., 2004). The three sets of operating conditions character-
ising these series correspond to operation in the trickle-, transition- and pulse-
flow regimes. The gas velocity is constant at 112 mm/s, data in rows (a), (b) and
(c) correspond to liquid velocities of 0.8, 6.8 and 10.6 mm/s, respectively. The
final image in each series is the standard deviation map calculated from a series
of 512 images acquired in immediate succession over a period of 10 s. It is
clearly seen that the liquid distribution in each of the trickle-flow images
(Fig. 34a) appears constant; therefore, the standard deviation map has pixel
intensities approaching the noise level consistent with a given pixel containing the
same phase or combination of phases (liquid, solid, gas) in each successive image.
FIG. 34. Evolution of gas–liquid distribution in the (a) trickle-, (b) transition, and (c) pulsing-flow
regimes. The gas velocity is 112 mm/s; liquid velocities are (a) 0.8, (b) 6.8 and (c) 10.6 mm/s. Four
images are shown acquired within the series of 540; each image took 20 ms to acquire. High intensity
(white) corresponds to high liquid content. Signal intensity is associated only with the liquid phase.
In-plane spatial resolution is 1.4 mm 2.8 mm, and the image slice thickness is 2 mm. All images are
on the same intensity scale. The final image in each row is the SD map calculated from 512 con-
secutive images taken from the complete time-series acquired. The three standard deviation maps are
plotted on the same scale with white (highest) and black (lowest) values. Reprinted from Lim et al.
(2004), with permission from Elsevier. Copyright (2004).
Fig. 34c corresponds to conditions of pulsing flow during which the content (i.e.
gas or liquid) of a given pixel changes with time. Thus, the resulting standard
deviation map has high intensity values (lighter shades). There also exist regions
within this standard deviation map that appear to suggest stable gas–liquid dis-
tribution. In many cases, such features in the standard deviation maps actually
correspond to regions where the given image pixel contains (mostly) solid catalyst
pellets. Since the pellets do not move, and remain liquid filled at all times, the
signal intensity associated with that pixel, even if such a pellet is at times sur-
rounded by either gas or liquid, will change very little. The series of images shown
in Fig. 34b are taken for a set of conditions that exist at an operating condition
within the transition from trickle to pulsing flow. Inspection of this row of
images shows a constant gas–liquid distribution in most regions of the field
of view. However, some regions clearly show a time-varying liquid distribution.
The regions of stable and unstable liquid content are clearly identified in the
associated standard deviation map and provide strong evidence in support of the
‘‘microscopic’’ models of the trickle-to-pulse transition. Movies of the 2-D images
recorded in the trickle-, transition- and pulse-flow regimes can be found at the
website http://www.cheng.cam.ac.uk/groups/mri/aichej.htm. Figure 35 explores
the nature of these local pulsing events in more detail by overlaying the standard
deviation maps with the high-resolution image of the structure of the bed; the
pixels associated with pellets have been gated out and are depicted in black for
clarity (Lim et al., 2004). It is clearly seen that the initial local pulses occur at the
size scale of the packing elements within the bed.
Further analysis of the standard deviation maps is extended by applying a
binary gate to these data sets to identify pixels that are associated with values of
standard deviation significantly above the noise level, thereby clearly identifying
the position of local pulses within the bed. Figure 36a shows a binary-gated
version of the standard deviation map shown in Fig. 34b. The value at which the
binary gate is applied is chosen to be 5s, where s is the average standard
deviation value of the noise calculated from the images of a completely liquid-
filled bed. Figure 36b shows the temporal autocorrelation function data cal-
culated from the two regions identified by the binary gate. The regions identified
by black pixels, which identify pixels associated with a constant gas–liquid
distribution, are characterised by a constant value of the temporal autocorre-
lation function of 1. The pixels associated with time-varying gas–liquid con-
tent are seen to have a temporal autocorrelation function containing several
distinct periodicities. In Fig. 36c, local temporal autocorrelation functions are
calculated for the local regions highlighted in Fig. 36a. Local periodicities of
0.4, 2 and 10 s are identified. The fluctuations in the liquid signal occurring
with a period of 0.4 s are of much smaller magnitude. A possible explanation
for this is that these fluctuations in liquid distribution are associated with fluc-
tuating liquid films on the surface of the packing. This hypothesis is further
investigated in Fig. 37, which shows the standard deviation map and temporal
autocorrelation data for images recorded at a liquid and gas velocity of 2.0 and
(a)
(b)
FIG. 35. Identification of location and size of local pulses within the trickle bed. A high spatial
resolution image (in-plane spatial resolution 175 mm 175 mm; slice thickness 1 mm) is overlayed
with a standard deviation map calculated from images acquired at a spatial resolution of in-plane
spatial resolution 1.4 mm 2.8 mm, and slice thickness 2 mm. The standard deviation maps have
been linearly interpolated to the same in-plane spatial resolution as the high resolution data. Images
are shown for a constant gas velocity of 112 mm/s; (a) increasing liquid velocity, and (b) decreasing
liquid velocity. The liquid velocities increase left to right: 2.8, 3.7, 6.1 and 7.6 mm/s. Reprinted from
Lim et al. (2004), with permission from Elsevier. Copyright (2004).
1.003
1.4 1.04
(iii) 1.002
(i)
1.2 1.02
τ) (a.u.)
τ) (a.u.)
1.001
1.0 1.00
RI(τ)
RI(τ)
1.000
(ii) 0.8 0.98
0.999
0.6 0.96
0.998
0 1 2 3 4 5 0 1 2 3 4 5
(a) (b) τ (s) (c) τ (s)
FIG. 36. (a) The binary-gated map derived from the SD map shown in Fig. 34b. The pixels
identified as being associated with local pulsing are identified as white; pixels associated with ‘‘con-
stant’’ gas–liquid distribution are identified as black. (b) The temporal autocorrelation function
calculated for white pixels (_____) and black pixels (- - - - -). (c) Local autocorrelation functions
calculated from the identified regions (i, ii, iii) identified in (a) correspond to periods of oscillation of
10, 2 and 0.4 s, respectively. The solid and dashed lines take the values on the left-hand y-axis;
the much smaller scale intensity fluctuations are associated with the much smaller values of standard
deviation on the right-hand y-axis. Reprinted from Lim et al. (2004), with permission from Elsevier.
Copyright (2004).
275 mm/s, respectively. This condition lies in the trickle regime before the onset
of the transition to pulsing flow. The standard deviation map is not shown on
the same scale of intensity as those shown in Fig. 34. Instead, the scale varies
between the highest (white) and lowest (black) values of standard deviation
(a)
FIG. 37. (a) standard deviation map, and (b) temporal autocorrelation functions for data recorded
at liquid and gas velocities of 2.0 and 275 mm/s, respectively. The grey scale varies between lowest
(black) and highest (white) standard deviation values calculated. The temporal autocorrelation
functions are shown for regions (i) and (ii) by solid and dashed lines, respectively. Reprinted from
Lim et al. (2004), with permission from Elsevier. Copyright (2004).
calculated from this specific time series of liquid-distribution images. All the
standard deviation values in Fig. 37a fall below the gating level selected for the
data shown in Fig. 36a. Therefore, all the fluctuations in liquid distribution are
much smaller than those observed when macroscopic (size of the packing
elements) liquid pulses occur. Short time-scale fluctuations in the local liquid
distribution that occur with well-defined periods of 0.15 and 0.3 s (see Fig.
37b) can be observed. This work has recently been extended to acquire 3-D
images of the formation and evolution of local pulsing events through the
trickle-to-pulse transition (Anadon et al., in press; Gladden et al., in press).
V. Future Prospects
The role of MR in chemical engineering research is changing. It is now es-

tablished as a quantitative measurement tool in chemical engineering, and as its
use becomes more widespread the ‘‘language’’ of MR will appear less daunting.
There remains significant scope to develop new MR methods that address spe-
cific measurement needs in chemical engineering. To this end we are likely to see
continued interest in the development and implementation of fast measurement
methods and increasing use of MR beyond 1H to observation of nuclei such as
13
C, 31P, 23Na and 19F. We are also likely to see MR systems dedicated to
specific MR measurements set up for use by the non-expert. Precedents already
exist in clinical medicine, downhole tools in oil exploration and food quality
control. In each of these applications, MR hardware and software are designed
to perform a restricted set of measurements such that the resulting data are
readily interpreted in terms of specific characteristics of the system of interest.
In addition to this simplified user interface, the design of hardware for a single
specific function can also reduce hardware costs considerably. There is also
potential for using dedicated, single-function MR hardware for process control
applications. For example, droplet-size distributions in emulsions can be char-
acterised in situ under flowing conditions, thereby enabling MR to be used as
part of a process control strategy. This measurement technique also avoids the
introduction of errors into the droplet-size determination caused by extracting
samples from a process line for ex situ analysis (Johns and Gladden, 2002).
Further opportunities lie in combining MR with other tomographic and sensor
technologies. For example, X-ray micro-tomography (XMT) can probe struc-
tures at a higher spatial resolution than MR, but does not readily give infor-
mation on transport processes. By combining MR and XMT, a greater range of
spatio-temporal correlations may be accessed. Opportunities also exist in
using MR to identify and understand, at the laboratory scale, the occurrence of
particular phenomena that may have a detrimental effect on product quality or
process performance, but then using a cheap, robust measurement on plant
to identify when such phenomena occur in the real process. For example,
in the context of trickle-bed operation, the detailed nature of transitions in
hydrodynamics between different flow regimes may be determined and
their signature in terms of variation in pressure drop measurements iden-
tified. The pressure drop measurements may then be used on the full-size op-
erating unit to provide a much more accurate assessment of trickle-bed
operation.
Having introduced and begun to validate the use of MR measurements, the
integration of these measurements into numerical modelling and theoretical
research will also be an area of significant opportunity and development. MR
data can be used (i) to guide the description of the problem to be solved. This
can be done by using the MR image directly, in pixelated form, as the sim-
ulation lattice, or by using the image in a more qualitative way to identify the
structural features that need to be represented within the simulation, and (ii) to
explore the accuracy of the numerical prediction. This is an area of research in
which work in the medical and engineering fields is at a similar stage of de-
velopment, which is not surprising since both research communities require each
other’s skills. A good example of this research methodology is found in the work
of Wang and Li (1998), which used MR images to develop a 3-D simulation
lattice upon which a numerical simulation of the mass transport processes rel-
evant to drug delivery to a brain tumour could be based. From this simulation it
was possible to predict the optimal location of a controlled-release drug delivery
implantation. One might ask, why is MR imaging so much more useful in this
type of application compared with other tomographic techniques? Of course,
the ability to image an optically opaque subject is key, but just as important is
the fact that MR can also give a 3-D mapping of the transport processes oc-
curring within the sample as well as mapping chemical composition—i.e. if one
needs the real diffusion coefficients of species within the sample to input to the
model, one can measure them directly. This is enormously powerful because we
can develop models in which the simulation lattice, boundary conditions and
many of the transport parameters required by such models are not free variables
but parameters measured on the real system.
More generally, the ability of MR to provide 3-D data sets for the validation
and development of numerical codes (e.g. Manz et al., 1999c) will be enor-
mously useful. As we look to the future, many of us expect that we will be able
to exploit developments made in computational physics so as to provide an
alternative strategy to using scale-up rules and correlations in process design. If
this goal is to be achieved, these ‘‘predictive’’ codes must be validated rigor-
ously, and MR has an important role to play in this. The key attribute of MR is
its ability to provide spatially resolved information on structure, transport and,
as appropriate, chemistry, within the system of interest. If the numerical code is
able to predict, at high spatial resolution, the experimentally determined chem-
ical composition, diffusion, dispersion and flow behaviour then there is com-
pelling evidence that the code is capturing the correct chemistry and physics,
and hence should be able to predict the performance of the scaled-up system.
Of course, strategies will still be required for modelling heterogeneities in mac-
ro-scale systems on length-scales greater than those probed in the MR exper-
iments.
In conclusion, this chapter has attempted to bring together the measurement
capabilities that can be united under the heading of ‘‘MR techniques’’.
Each of these can be a research field in their own right, but from the perspective
of the chemical engineer it is the combination of these methods—particularly in
the hierarchy of length-scales they probe—that makes MR methods so
powerful.
ACKNOWLEDGMENTS
We wish to thank Dr. P. Alexander for his help in the preparation of this
chapter.
REFERENCES
Akpa, B. S., Mantle, M. D., Sederman, A. J., and Gladden, L. F. J. Chem. Soc., Chem. Commun.
2741 (2005).
Altobelli, S. A., Fukushima, E., and Monday, L. A. J. Rheol. 41, 1105 (1997).
Anadon, L. D., Lim, M. H. M., Sederman, A. J., and Gladden, L. F. Magn. Reson. Imaging 23, 291
(2005).
Anderson, M. W., and Klinowski, J. Chem. Phys. Letts. 172, 275 (1990).
Arola, D. F., Barrall, G. A., Powell, R. L., McCarthy, K. L., and McCarthy, M. J. Chem. Eng. Sci.
52, 2049 (1997).
Baba, T., and Ono, Y. Annu. Rev. NMR Spectroscopy 38, 355 (1999).
Balcom, B. J., Carpenter, T. A., and Hall, L. D. Macromolecules 25, 6818 (1992).
Baldwin, C. A., Sederman, A. J., Mantle, M. D., P. Alexander, P., and Gladden, L. F. J. Colloid
Interface Sci. 181, 79 (1996).
Bar, N. -K., Bauer, F., Ruthven, D. M., and Balcom, B. J. Catal. 208, 224 (2002).
Bencsik, M., and Ramanathan, C. Magn. Reson. Imaging 19, 379 (2001).
Bertsch, F., Mattner, J., Stehling, M. K., Muller-Lisse, U., Peller, M., Loeffler, R., Weber, J.,
Messmer, K., Wilmanns, W., Issels, R., and Reiser, M. Magn. Reson. Imaging 16, 393 (1998).
Beuling, E. E., van Dusschoten, D., Lens, P., van den Heuvel, J. C., Van As, H., and Ottengraf, S. P.
P. Biotechnol. Bioeng. 60, 283 (1998).
Beyea, S. D., Caprihan, A., Clewett, C. F. M., and Glass, S. J. Appl. Magn. Reson. 22, 175 (2002).
Blok, J. R., Varkevisser, J., and Drinkenburg, A. A. H. Chem. Eng. Sci. 38, 687 (1983).
Boelhouwer, J. G., Piepers, H. W., and Drinkenburg, A. A. H. Chem. Eng. Sci. 57, 4865 (2002).
Bonardet, J. -L., Domeniconi, T., N’Gokoli-Kékélé, P., Spinguel-Huet, M. -A., and Fraissard, J.
Langmuir 15, 5836 (1999).
Bows, J. R., Patrick, M. L., Nott, K. P., and Hall, L. D. Int. J. Food Sci. Technol. 36, 243 (2001).
Boyer, C., Duquenne, A. -M., and Wild, G. Chem. Eng. Sci. 57, 3185 (2002).
Britton, M. M., and Callaghan, P. T. J. Rheol. 41, 1365 (1997).
Britton, M. M., Sederman, A. J., Taylor, A. F., Scott, S. K., and Gladden, L. F. J. Phys. Chem.
(submitted, 2005).
Brunner, E., Haake, M., Kaiser, L., Pines, A., and Reimer, J. A. J. Magn. Reson. 138, 155 (1999).
Buckley, C., Hollingsworth, K. G., Sederman, A. J., Holland, D. J., Johns, M. L., and Gladden, L.
F. J. Magn. Reson. 161, 112 (2003).
Butler, L. G., Cory, D. G., Dooley, K. M., Miller, J. B., and Garroway, A. N. J. Am. Chem. Soc.
114, 125 (1992).
Callaghan, P. T., ‘‘Principles of Nuclear Magnetic Resonance Microscopy’’. Clarendon, Oxford
(1991).
Caprihan, A., Fukushima, E., Rosato, A. D., and Kos, M. Rev. Sci. Instrum. 68, 4217 (1997).
Cheah, K. Y., Chiaranussati, N., Hollewand, M. P., and Gladden, L. F. Appl. Catal. A 115, 147
(1994).
Chingas, G. C., Miller, J. B., and Garroway, A. N. J. Magn. Reson. 66, 530 (1986).
Chupas, P. J., and Grey, C. P. J. Catal. 224, 69 (2004).
Cormier, R. J., Kilfoil, M. L., and Callaghan, P. T. Phys. Rev. E 64 Art. No. 051809 (2001).
Cormier, R. J., and Callaghan, P. T. J. Chem. Phys. 116, 10020 (2002).
Crine, M., Marchot, P., Lekhlif, B., and L’Homme, G. Chem. Eng. Sci. 47, 2263 (1992).
Dankworth, D. C., Kevrekidis, I. G., and Sundaresan, S. A.I.Ch.E. J. 36, 605 (1990).
Dirckx, C. J., Clark, S. A., Hall, L. D., Antalek, B., Tooma, J., Hewitt, J. M., and Kawaoka, K.
A.I.Ch.E. J. 46, 6 (2000).
Doran, S. J., Carpenter, T. A., and Hall, L. D. Rev. Sci. Instrum. 65, 2231 (1994).
Duncan, T. M., Winslow, P., and Bell, A. T. J. Catal. 93, 1 (1985).
Dybowski, C., Bansal, N., and Duncan, T. M. Annu. Rev. Phys. Chem. 42, 433 (1991).
Egiebor, N. O., Gray, M. R., and Cyr, N. Chem. Eng. Commun. 77, 125 (1989).
Ehrichs, E. E., Jaeger, H. M., Karczmar, G. S., Knight, J. B., Kuperman, V. Y., and Nagel, S. R.
Science 276, 1632 (1995).
Fraissard, F. Catal. Today 51, 481 (1999).
Fukushima, E. Ann. Rev. Fluid Mech. 31, 95 (1999).
Gatenby, J. C., and Gore, J. C. J. Magn. Reson. A 110, 26 (1994).
Gladden, L. F. Chem. Eng. Sci. 49, 3339 (1994).
Gladden, L. F. Top. Catal. 24, 19 (2003).
Gladden, L. F., Alexander, P., and Hollewand, M. P. A.I.Ch.E. J. 41, 894 (1995).
Gladden, L. F., Alexander, P., Britton, M. M., Mantle, M. D., Sederman, A. J., and Yuen, E. H. L.
Magn. Reson. Imaging 21, 213 (2003a).
Gladden, L. F., Lim, M. H. M., Mantle, M. D., Sederman, A. J., and Stitt, E. H. Catal. Today 79,
203 (2003b).
Gladden, L. F., Buckley, C., Chow, P. S., Davidson, J. F., Mantle, M. D., and Sederman, A. J. Curr.
Appl. Phys. 4, 93 (2004).
Gladden, L.F., Anadon, L.D., Lim, M.H.M., Sederman, A.J., and Stitt, E.H. Ind. Eng. Chem. Res.
(in press).
Gotz, J., Zick, K., Heinen, C., and Konig, T. Chem. Eng. Process. 41, 611 (2002).
Grosser, K., Carbonell, R. G., and Sundaresan, S. A.I.Ch.E. J. 34, 1850 (1988).
Groten, W. A., Wojciechowski, B. W., and Hunter, B. K. J. Catal. 138, 343 (1992).
Guilfoyle, D. N., Mansfield, P., and Packer, K. J. J. Magn. Reson. 97, 342 (1992).
Haase, A., Frahm, J., Matthaei, D., Hanicke, W., and Merboldt, K. D. J. Magn. Reson. 67, 258
(1986).
Hall, L. D., Amin, M. H. G., Evans, S., Nott, K. P., and Sun, L. J. Electron. Imaging 10, 601 (2001).
Han, S. I., Stapf, S., and Blumich, B. Phys. Rev. Lett. 87 Art. No. 144501 (2001a).
Han, S. I., Marseille, O., Gehlen, C., and Blümich, B. J. Magn. Reson. 152, 87 (2001b).
Haw, J. F. Top. Catal. 8, 81 (1999).
Heath, C. A., Belfort, G., Hammer, B. E., Mirer, S. D., and Pimbley, J. M. A.I.Ch.E. J. 36, 547
(1990).
Heine, C., Kupferschlager, K., Stapf, S., and Blümich, B. J. Magn. Reson. 154, 311 (2002).
Heink, W., Kärger, J., and Pfeifer, H. Chem. Eng. Sci. 33, 1019 (1978).
Hennig, J., Nauerth, A., and Friedburg, H. Magn. Reson. Med. 3, 823 (1986).
Hill, K. M., Caprihan, A., and Kakalios, J. Phys. Rev. Lett. 78, 50 (1997).
Holland, D. J., Sederman, A. J., Mantle, M. D., Gladden, L. F., and Middelberg, A. P. J.
J. Chromatogr. A 1033, 311 (2004).
Hollewand, M. P., and Gladden, L. F. J. Catal. 144, 254 (1993).
Hollewand, M. P., and Gladden, L. F. Magn. Reson. Imaging 12, 291 (1994).
Holub, R. A., Duduković, M. P., and Ramachandran, P. A. Chem. Eng. Sci. 47, 2343 (1992).
Hong, U., Karger, J., Hunger, B., Feoktistova, N. N., and Zhdanov, S. P. J. Catal. 137, 243 (1992).
Howseman, A. M., Stehling, M. K., Chapman, B., Coxon, R., Turner, R., Ordidge, R. J., Cawley,
M. G., Glover, P., Mansfield, P., and Coupland, R. E. Br. J. Radiol. 61, 822 (1988).
Hunger, M., and Weitkamp, J. Angew. Chem. Int. Ed. 40, 2954 (2001).
Jezzard, P., Carpenter, T. A., Hall, L. D., Clayden, N. J., and Jackson, P. J. Polym. Sci. Pt. B-
Polym. Phys. 30, 1423 (1992).
Johns, M. L., and Gladden, L. F. J. Magn. Reson. 154, 142 (2002).
Johns, M. L., Sederman, A. J., Bramley, A. S., Alexander, P., and Gladden, L. F. A.I.Ch.E. J. 46,
2151 (2000).
Karger, J., and Freude, D. Chem. Eng. Technol. 25, 769 (2002).
Kauten, R. J., Maneval, J. E., and McCarthy, M. J. J. Food Sci. 56, 799 (1991).
Khitrina, L. Y., Koptyug, I. V., Pakhomov, N. A., Sagdeev, R. Z., and Parmon, V. N. J. Phys.
Chem. B, 104, 1966 (2000).
Kimmich, R., ‘‘NMR Tomography, Diffusometry, Relaxometry’’. Springer, Berlin Heidelberg

(1997).
Kimmich, R., and Hoepfel, D. J. Magn. Reson. 72, 379 (1987).
Koptyug, I. V., Fenelonov, V. B., Khitrina, L. Y., Sagdeev, R. Z., and Parmon, V. N. J. Phys. Chem.
B, 102, 3090 (1998).
Koptyug, I. V., Altobelli, S. A., Fukushima, E., Matveev, A. V., and Sagdeev, R. Z., J, Magn. Reson.
147, 36 (2000a).
Koptyug, I. V., Khitrina, L. Y., Arsitov, Y. I., Tokarev, M. M., Iskakov, K. T., Parmon, V. N., and
Sagdeev, R. Z. J. Phys. Chem. B, 104, 1695 (2000b).
Koptyug, I. V., Ilyina, L. Y., Matveev, A. V., Sagdeev, R. Z., Parmon, V. N., and Altobelli, S. A.
Catal. Today 69, 385 (2001).
Koptyug, I. V., Kulikov, A. V., Lysova, A. A., Kirillov, V. A., Parmon, V. N., and Sagdeev, R. Z. J.
Am. Chem. Soc. 124, 9684 (2002).
Koptyug, I. V., Lyosova, A. A., Kulikov, A. V., Krlov, V. A., Parmon, V. N., and Sagdeev, R. Z.
Appl. Catal. A 267, 143 (2004).
Kose, K. Phys. Rev. A 44, 2495 (1991a).
Kose, K. J. Magn. Reson. 92, 631 (1991b).
Kose, K. J. Magn. Reson. 96, 596 (1992).
Kose, K. Phys. Rev. Lett. 72, 1467 (1994).
Kuethe, D. O. Phys. Rev. A 40, 4542 (1989).
Kuethe, D. O., and Gao, J. H. Phys. Rev. E 51, 3252 (1995).
Kuethe, D. O., Caprihan, A., Gach, H. M., Lowe, I. J., and Fukushima, E. J. Appl. Physiol. 88, 2279
(2000).
Larachi, F., Iliuta, I., Chen, M., and Grandjean, B. P. A. Can. J. Chem. Eng. 77, 751 (1999).
Laufer, J. NACA Rep. 1174, 1 (1954).
Le Bihan, D., Delannoy, J., and Levin, R. L. Radiology 171, 853 (1989).
Li, T. Q., Seymour, J. D., Powell, R. L., McCarthy, K. L., Ödberg, L., and McCarthy, M. J. Magn.
Reson. Imaging 12, 923 (1994).
Lim, M. H. M., Sederman, A. J., Gladden, L. F., and Stitt, E. H. Chem. Eng. Sci. 59, 5403 (2004).
Mair, R. W., Sen, P. N., Hürlimann, M. D., Patz, S., Cory, D. G., and Walsworth, R. L. J. Magn.
Reson. 156, 202 (2002).
Mansfield, P., and Grannell, P. K. J. Phys. C, 6, L422 (1973).
Mansfield, P. J. Phys. C, 10, L55 (1977).
Mantle, M. D., and Sederman, A. J. Prog. Nucl. Magn. Reson. Spectr. 43, 3 (2003).
Manz, B., Alexander, P., and Gladden, L. F. Phys. Fluids, 11, 259 (1999a).
Manz, B., Chow, P. S., and Gladden, L. F. J. Magn. Reson. 136, 226 (1999b).
Manz, B., Gladden, L. F., and Warren, P. B. A.I.C.hE. J. 45, 1845 (1999c).
Mariani, N. J., Martı́nez, O. M., and Barreto, G. F. Chem. Eng. Sci. 56, 5995 (2001).
Martin, N., Viniegra, M., Lima, E., and Espinosa, G. Ind. Eng. Chem. Res. 43, 1206 (2004).
Menzinger, M., Tzalmona, A., Armstrong, R. L., Cross, A., and Lemaire, C. J. Phys. Chem. 96, 4725
(1992).
Metcalfe, G., Graham, L., Zhou, J., and Liffman, K. Chaos 9, 581 (1999).
Mosher, T. J., and Smith, M. B. Magn. Reson. Med. 15, 334 (1990).
Ng, K. M. A.I.Ch.E. J. 32, 115 (1986).
Nott, K. P., and Hall, L. D. Trends Food Sci. Technol. 10, 366 (1999).
Nott, K. P., Paterson-Beedle, M., Macaskie, L. E., and Hall, L. D. Biotech. Lett. 23, 1749 (2001).
Oshio, K., and Feinberg, D. A. Magn. Reson. Med. 20, 344 (1991).
Packer, K. J. Top. Catal. 3, 249 (1996).
Pangrle, B. J., Walsh, E. G., Moore, S. C., and DiBasio, D. Chem. Eng. Sci. 47, 517 (1992).
Parker, W. O. Comment Inorg. Chem. Part A 22, 31 (2000).
PietraX, T., Seydoux, R., and Pines, A. J. Magn. Reson. 133, 299 (1998).
Pope, J. M., and Yao, S. Concepts Magn. Reson. 5, 281 (1993a).
Pope, J. M., and Yao, S. Magn. Reson. Imaging, 11, 585 (1993b).
Powell, R. L., Maneval, J. E., Seymour, J. D., McCarthy, K. L., and McCarthy, M. J. J. Rheol. 38,
1465 (1994).
Ra’bai, G., Orba’n, M., and Epstein, I. R. Acc. Chem. Res. 23, 258 (1990).
Reinecke, N., and Mewes, D. Chem. Eng. Sci. 52, 2111 (1997).
Rigby, S. P., Cheah, K. Y., and Gladden, L. F. Appl. Catal. A 144, 377 (1996).
Rigby, S. P., and Gladden, L. F. J. Catal. 173, 484 (1998).
Rigby, S. P., and Gladden, L. F. Chem. Eng. Sci. 54, 3503 (1999).
Rigby, S. P., and Daut, S. Adv. Colloid Interface Sci. 98, 87 (2002).
Roe, D. C., Kating, P. M., Krusic, P. J., and Smart, B. E. Top. Catal. 5, 133 (1998).
Sains, M. C., El-Bachir, M. S., Sederman, A. J., and Gladden, L. F. Magn. Reson. Imaging 23, 291
(2005).
Savelsberg, R., Demco, D. E., Blumich, B., and Stapf, S. Phys. Rev. E. 56 Art. No. 020301(R)
(2002).
Schwarz, H. B., Ernst, H., Ernst, S., Karger, J., Roser, T., Snurr, R. Q., and Weitkamp, J. Appl.
Catal. A 130, 227 (1995).
Schwarz, H.B., Ernst, S., Karger, J., Knorr, B., Seiffert, G., Snurr, R.Q., Staudte, B., and We-
itkamp, J.J. Catal. 167, 248 (1997).
Scott, S. K. Acc. Chem. Res. 20, 186 (1987).
Sederman, A. J., Johns, M. L., Bramley, A. S., Alexander, P., and Gladden, L. F. Chem. Eng. Sci.
52, 2239 (1997).
Sederman, A. J., Johns, M. L., Alexander, P., and Gladden, L. F. Chem. Eng. Sci. 53, 2117 (1998).
Sederman, A. J., Alexander, P., and Gladden, L. F. Powder Technol. 117, 255 (2001).
Sederman, A. J., and Gladden, L. F. Magn. Reson. Imaging 19, 339 (2001a).
Sederman, A. J., and Gladden, L. F. Chem. Eng. Sci. 56, 2615 (2001b).
Sederman, A. J., Mantle, M. D., and Gladden, L. F. J. Magn. Reson. 161, 15 (2003).
Sederman, A. J., Mantle, M. D., Buckley, C., and Gladden, L. F. J. Magn. Reson. 166, 182 (2004a).
Sederman, A. J., Hollingsworth, K. G., Johns, M. L., and Gladden, L. F., J, Magn. Reson. 171, 118
(2004b).
Seymour, J. D., and Callaghan, P. T. A.I.Ch.E. J. 43, 2096 (1997).
Seeley, J. A., Han, S. I., and Pines, A. J. Magn. Reson. 167, 282 (2004).
Seymour, J. D., Caprihan, A., Altobelli, S. A., and Fukushima, E. Phys. Rev. Lett. 84, 266 (2000).
Sicardi, S., and Hofmann, H. Chem. Eng. J. 20, 251 (1980).
Sie, S. T., Senden, M. M. G., and van Wechem, H. M. H. Catal. Today, 8, 371 (1991).
Silva, A. C., Barbier, E. L., Lowe, I. J., and Koretsky, A. P. J. Magn. Reson. 135, 242 (1998).
Simoneau, C., McCarthy, M. J., Kauten, R. J., and German, J. B. J. Am. Oil Chem. Soc. 68, 481
(1991).
Singer, J. R. Science 130, 1652 (1959).
Stamps, J. P., Ottink, B., Visser, J. M., van Duynhoven, J. P. M., and Hulst, R. J. Magn. Reson. 151,
28 (2001).
Stapf, S., Packer, K. J., Graham, R. G., Thovert, J. F., and Adler, P. M. Phys. Rev. E 58, 6206
(1998).
Stejskal, E. O. J. Chem. Phys. 43, 3597 (1965).
Stejskal, E. O., and Tanner, J. E. J. Chem. Phys. 42, 288 (1965).
Sun, L., and Hall, L. D. Intl. Commun. Heat Mass Transfer, 28, 461 (2001).
Tallerek, U., Albert, K., Bayer, E., and Guiochon, G. A.I.Ch.E. J. 42, 3041 (1996).
Tallerek, U., Vergeldt, F. J., and Van As, H. J. Phys. Chem. B, 103, 7654 (1999).
Tallerek, U., Rapp, E., Van As, H., and Bayer, E. Angew. Chem. Int. Ed. 40, 1684 (2001).
Taylor, G. I. Proc. Roy. Soc. (London), 146A, 501 (1934).
Timonen, J., Alvila, L., Hirva, P., Pakkanen, T. T., Gross, D., and Lehmann, V. Appl. Catal. A, 129,
117 (1995).
Toye, D., Marchot, P., Crine, M., and L’Homme, G. Meas. Sci. Technol. 7, 436 (1996).
Tzalmona, A., Armstrong, R. L., Menzinger, M., Cross, A., and Lemaire, C. Chem. Phys. Lett. 174,
19 (1990).
Tzalmona, A., Armstrong, R. L., Menzinger, M., Cross, A., and Lemaire, C. Chem. Phys. Lett. 188,
457 (1992).
van der Klink, J. J. Adv. Catal. 44, 1 (2000).
Wang, C. -H., and Li, J. Chem. Eng. Sci. 53, 3579 (1998).
Wang, J. A., Chen, L. F., Li, C. L., and Navaro, O. Stud. Surf. Sci. Catal. 139, 53 (2001).
Wang, W., Walton, J. H., and McCarthy, J. L. J. Food Proc. Eng. 22, 11 (1999).
Wood, J., and L F Gladden, L. F. Appl. Catal. A 249, 241 (2003).
Yang, X. Y., Huan, C., Candela, D., Mair, R. W., and Walsworth, R. L. Phys. Rev. Lett. 88, 44301
(2002).
Yuen, E. H. L., Sederman, A. J., and Gladden, L. F. Appl. Catal. A 232, 29 (2002).
Zhang, J. H., Nicholas, J. B., and Haw, J. F. Angew. Chem. Intl. Ed. 39, 3302 (2000).
MODELING OF TRANSPORT AND TRANSFORMATION
PROCESSES IN POROUS AND MULTIPHASE BODIES
Juraj Kosek1, František Štěpánek2 and Miloš Marek1,

1
Department of Chemical Engineering, Prague Institute of Chemical Technology, Czech
Republic
2
Department of Chemical Engineering, Imperial College London, UK
I. Introduction 138
II. Methodology 140
A. Representation of Multiphase Media 140
B. Structure Acquisition 142
C. Morphological Characterization 143
D. Digital Reconstruction of Multiphase Media 145
E. Calculation of Effective Properties 151
F. Effective-scale Transport Models 159
III. Transformations 160
A. Skeletonization 161
B. Phase Transitions 161
C. Chemically Reactive Systems 170
IV. Applications 175
A. Multi-scale Reconstruction of a Catalyst Pellet 175
B. Reconstruction of Closed-cell Polymer Foam Structure 179
C. Polymer Particle Morphogenesis 182
D. Granulation and Dissolution 189
E. Simulation of CO Oxidation on Reconstructed
Catalytic Washcoat 192
V. Outlooks 195
A. Biological Systems 195
B. Materials Design 196
VI. Conclusions 197
Acknowledgments 198
References 198
Abstract
A methodology for computer representation of the structure of spa-

tially complex multiphase media and for the modeling of reaction, trans-
port, and structure-transformation processes in those media, is reviewed.
Tel.: +420 220 443 104; Fax: +420 233 337 335; E-mail: milos.marek@vscht.cz
137
DOI 10.1016/S0065-2377(05)30003-2
138 JURAJ KOSEK ET AL.
The methodology is demonstrated via several examples including phase

transition and structure evolution in porous and granular media, the
morphogenesis of polymer particles, and heterogeneous catalysis. Several
future potential applications of the methodology are identified.
I. Introduction
Chemical engineering has always been concerned with describing phenomena

occurring over a range of length-scales, from the molecular (e.g., adsorption
equilibrium and reaction kinetics), through the mesoscopic (e.g., agglomera-
tion), to continuum description of transport phenomena (e.g., momentum, heat,
and mass transfer), up to the unit operation and process-systems levels. It has
also been the aim to pass information from models at one length-scale to those
at a different hierarchical level. For example, a fixed-bed catalytic reactor model
may require the effectiveness factor as one of its input parameters, which can be
obtained as a solution of a diffusion problem at the length-scale of an individual
catalyst pellet (Sahimi et al., 1990). The values of a diffusion coefficient or
adsorption equilibrium constant (Davies and Seaton, 2000), which are required
as input to the particle-scale model can, in turn, be obtained from molecular-
level simulations.
Several length-scales have to be considered in a number of applications. For
example, in a typical monolith reactor used as automobile exhaust catalytic
converter the reactor length and diameter are on the order of decimeters, the
monolith channel dimension is on the order of 1 mm, the thickness of the cat-
alytic washcoat layer is on the order of tens of micrometers, the dimension of
the pores in the washcoat is on the order of 1 mm, the diameter of active noble
metal catalyst particles can be on the order of nanometers, and the reacting
molecules are on the order of ångstroms; cf. Fig. 1. The modeling of such
reactors is a typical multiscale problem (Hoebink and Marin, 1998). Electron
microscopy accompanied by other techniques can provide information on par-
ticle size, shape, and chemical composition. Local composition and particle size
of dispersed nanoparticles in the porous structure of the catalyst affect catalytic
activity and selectivity (Bell, 2003).
Simulation techniques suitable for the description of phenomena at each
length-scale are now relatively well established: Monte Carlo (MC) and Mo-
lecular Dynamics (MD) methods at the molecular length-scale, various ‘‘me-
soscopic’’ simulation methods such as Dissipative Particle Dynamics (Groot
and Warren, 1997), Brownian Dynamics, or Lattice Boltzmann in the colloidal
domain, Computational Fluid Dynamics at the continuum length-scale, and
sequential-modular or equation-based methods at the unit operation/process-
systems level.
MODELING OF TRANSPORT AND TRANSFORMATION PROCESSES 139
(a) (b)
(c) (d)
FIG. 1. The multiple scales in the catalytic monolith reactor: (a) catalytic monolith (10 cm), (b)
channel with catalyst washcoat on the walls (1 mm), (c) SEM image of the washcoat layer (10 mm),
(d) TEM image of meso-porous g-Al2O3 with dispersed Pt (200 nm).
If in all situations, parameters could be transferred from models at one hi-

erarchical level to another in the form of precalculated coefficients or closed-
form expressions, such as an equation of state or a correlation for a mass-
transfer coefficient, there would be no need for ‘‘multiscale modeling’’ as a
distinct methodology. However, there are many situations where phenomena
occurring at different length-scales are so integrated that their modeling cannot
be simply de-coupled (Ingram et al., 2004). As an example, one can consider
gas-phase catalytic polymerization of olefins, where the resulting polymer
properties (chain-length distribution, crystallinity) depend simultaneously on
phenomena at several length-scales: from polymer growth around a catalyst-
support fragment, through monomer diffusion and heat transfer within a
growing polymer particle, to population dynamics and fluid-particle flow at the
length-scale of the reactor (Kosek et al., 2001a, b).
As the focus of chemical engineering research shifts from bulk and petro-
chemicals to specialty materials and formulated products, ‘‘microstructure’’ is
becoming an increasingly important product attribute controlling end-use
properties. By microstructure we mean the spatial distribution of components,

or phases, in the product. The distribution of amorphous and crystalline do-
mains within a polymer particle, for example, has a significant influence on its
further processability. The distribution of primary solid particles, binder, and
porosity within a granule has an influence on its dissolution and release char-
acteristics (Stepanek, 2004) and can control attributes such as the bioavailability
of a drug.
In this chapter, we review modeling methods specifically aimed at describing
transport and reaction phenomena that occur within, or lead to the formation
of, complex porous or multiphase microstructures. The chapter is structured as
follows. First, the methodology of computer representation, characterization,
and reconstruction of multiphase and porous media is described. Techniques for
computational determination of effective transport properties in digitally re-
constructed media are then described, as well as methods for the modeling of
transient transport/transformation processes in those media. The methods are
then illustrated by selected examples from areas of interest of the authors:
polymer particle morphogenesis, granule formation and dissolution, and hete-
rogeneous catalysis. Other emerging applications are then mentioned in Section
V, followed by concluding remarks in Section VI. Finally, a list of key literature
references is provided.
II. Methodology
A. REPRESENTATION OF MULTIPHASE MEDIA
The structure of a multiphase medium can be specified by the spatial dis-

tribution of the phases that form the medium. This spatial distribution can be
generally represented by the phase function or, in specific situations, by the
equivalent pore-network diagram, by the spatial distribution of particles or
other constituents, and by the probability density function. Examples of the
representation of the porous media are shown in Fig. 2.
In the general case, the spatial distribution of the phases can be formally
represented by the so-called phase function f i : R3 ! f0; 1g for each phase i. The
phase function is defined as (Adler, 1992, 1994)

1 if r 2 phase i
f i ðrÞ ¼ (1)
0 otherwise
By definition only one phase can be present at any point r 2 R3 . It is further

required that the set Pi R3 , Pi ¼ fr : f i ðrÞ ¼ 1g be a compact set, i.e., that the
inter-phase boundaries are smooth in the mathematical sense. In a discrete
form, the phase function fi becomes the phase volume function fî , which assigns
FIG. 2. Types of computer representation of porous media—network diagrams (left) and solid-
phase function of reconstructed porous media (right).
each finite volume element of space (voxel) a value from the interval h0; 1i, i.e.,
the volume fraction of phase i in that element. The sum of phase volume func-
tions in each voxel must give unity,
X
fî ¼ 1 (2)
i
and so for an N-phase medium only N1 phase volume functions need to be
specified in each voxel. Unless stated otherwise, we will further work only with
phase volume functions on a discrete grid of voxels and leave out the hat
notation from fî .
In a practical implementation, the domain on which the phase volume func-
tions are specified is typically a cubic grid of N x N y N z voxels, which cor-
responds to real dimensions of Lx ¼ hN x , Ly ¼ hN y , and Lz ¼ hN z , where h is
the voxel size. We will further call this region of real space the computational
unit cell. The relationship between the unit cell and the multiphase medium of
interest depends on the absolute dimensions of the medium and on the spatial
resolution at which the medium is represented (feature dimensions). The unit
cell can either contain the entire medium and some void space surrounding it, as
in the case of virtual granules described in Section IV.D below, or be a sample of
a much larger (theoretically infinite) medium, as in the case of transport prop-
erties calculation, described in Section II.E below.
In the latter case, the dimensions of the unit cell must be such that the unit cell
is statistically representative of the entire medium. For spatially periodic regular
media, the unit cell will coincide with one lattice unit. For random media, let
Lm ¼ maxi ðLi Þ be the maximum characteristic length-scale of all the phases
forming the medium, where Li is defined by Eq. (4) below. The unit cell di-
mensions ðLx ; Ly ; Lz Þ should be much larger than the maximum characteristic
length-scale Lm for transport properties calculated on the unit cell to become
cell size-independent, and thus, representative of the entire medium (Adler,
1992).
In situations where the spatial features of the medium occur over a wide range
of length-scales, e.g., in porous media with bi-modal pore size distribution
(Salejova et al., 2004), the voxel size h would have to be small enough to capture
the smallest spatial feature but at the same time the unit cell size must be large
enough for the sample to be statistically representative. This would require an
infeasibly large number of voxels if standard fixed-grid encoding was used. In
that case, one can use a spatially adaptive mesh (Perré, 2004; Sahimi et al., 2004;
Sapoval, 2001), or transform the porous medium into an equivalent pore-
network model (Blunt et al., 2002; El-Nafaty and Mann, 2001; Liang et al.,
2000a; Lin and Miller, 2000; Thompson, 2002).
As already mentioned, the phase function is only one of several possible ways
of representing the structure of porous/multiphase media. Other alternatives
include: (i) pore-network diagrams, (ii) density probability functions, and (iii)
aggregate of interacting micro-elements. The mapping between those different
types of spatially three dimensional (3D) representations is often required.
However, the original structural data will nowadays almost always be in the
form of a 2D or 3D digital image; therefore, their translation into a phase
function is trivial as it involves just a spatial discretization. Some relevant
structure acquisition techniques are now briefly reviewed.
B. STRUCTURE ACQUISITION
Information about the spatial distribution of phases in a specimen of a mul-

tiphase medium can be obtained experimentally by means of one of several
techniques. Optical or scanning electron microscopy (SEM) yields spatially 2D
images of surfaces or cross-sections of the material of interest. Modern electron
microscopes provide resolution down to the atomistic level (e.g., Arenas-
Alatorre et al., 2002). When information about the chemical composition of a
multiphase medium is required, spectroscopic imaging techniques such as IR
imaging (Gupper et al., 2002), energy dispersive X-ray (EDX) analysis, and
X-ray photoelectron spectroscopy (XPS) can be used. Atomic force microscopy
(AFM) can also be used for probing the microstructure of materials, e.g., the
crystallinity in polymers (Hosier et al., 2004) as well as a range of surface
features. In the case of optical microscopy, standard staining can be used to
further distinguish between phases, as is commonly done when imaging bio-

logical tissue samples (Mantalaris et al., 2001).
While standard optical and electron microscopy is limited to 2D imaging,
confocal laser scanning microscopy (CLSM) can be used to probe the material
to a certain depth and thus generate 3D information about its microstructure
(Kanamori et al., 2004). Several techniques can generate fully 3D images of
porous or multiphase media. These include nuclear magnetic resonance (NMR)
imaging (Barrie, 2000; Rigby and Daut, 2002), and tomographic reconstruction
based on X-ray (Blacher et al., 2004) or transmission electron microscopy
(TEM) visualization (Weyland et al., 2001). X-ray computed micro-tomo-
graphy, in particular, has gained in popularity recently, being used for non-
destructive 3D visualization of the microstructure of porous materials such as
solid foams (Elmoutaouakkil et al., 2002), textile (Ramaswamy et al., 2004), or
pharmaceutical granules (Farber et al., 2003), as well as a range of biological
materials. 3D TEM has been used for the direct visualization of metal
nanocrystallites (Ichikawa et al., 2003) and the pore structure of zeolites (Jans-
sen et al., 2001).
Once a spatially 2D or 3D image of the multiphase medium of interest has
been obtained, it is desirable to characterize the image by a set of morphological
descriptors, which can then be correlated with effective properties of the me-
dium or their evolution followed in time when a structure-transformation
process (e.g., dissolution) takes place in the medium. Let us now review some
morphological descriptors most commonly used for the characterization of po-
rous and multiphase media.
C. MORPHOLOGICAL CHARACTERIZATION
The composition of the medium is expressed by the phase volume fraction fi

of each phase i, which can be calculated as the spatial average (denoted by an
overline) of its phase function fi according to
Z
1
fi ¼ f i ðrÞ ¼ f ðrÞ dV (3)
V V i
where V ¼ Lx Ly Lz is the volume of the unit cell. Specifically, porosity ¼ fg is

the volume fraction of the gas phase in an unsaturated porous medium. The
next quantity of interest is the characteristic length-scale of each phase, which is
a measure of its dispersion in the multiphase medium. Assuming non-fractal
media, let Si be the internal surface area of phase i (i.e., the area of its interface
with all other phases). A characteristic length-scale Li can be defined as
V i fi V
Li ¼ ¼ (4)
Si Si
Of particular interest in porous media is the so-called equivalent hydraulic

diameter d e 4V g =S g ¼ 4Lg , which is important for permeability scaling
(Martys and Garboczi, 1992).
In spatially evolving multiphase media (e.g., during dissolution of a porous
medium, or phase separation in a polymer blend), the mean curvature of the
interface between two phases is of interest. Curvature is a sensitive indicator of
morphological transitions such as the transition from spherical to rod-like
micelles in an emulsion, or the degree of sintering in a porous ceramic material.
Furthermore, important physicochemical parameters such as capillary pressure
(from the Young–Laplace equation) are curvature-dependent. The local value
of the mean curvature K i ¼ 12ð1=R1 þ 1=R2 Þ of an interface of phase i with
principal radii of curvature R1 and R2 can be calculated as the divergence of the
interface normal vector ni
K i ¼ r ni (5)
where ni is calculated as the gradient of the phase function
rf i
ni ¼ (6)
jjrf i jj
with the minus sign indicating normal vector orientation from the phase i to the
surrounding phases. Higher-order approximations should be used for numerical
evaluation of the gradient rfi (Kothe et al., 1996; Rider and Kothe, 1998;
Scardovelli and Zaleski, 1999). The mean value of Ki is then obtained by
averaging over the entire surface @i of the phase i,
Z
1
Ki ¼ K i dA (7)
S i @i
The above-mentioned phase volume fractions, internal surface area, and mean
curvature are instances of a more general class of integral measures called
Minkowski functionals (Arns et al., 2001, 2004).
Another frequently used characteristic of random media is the two-point
autocorrelation function (Adler, 1992, 1994; Thovert et al., 1993)
ðf i ðrÞ fi Þðf i ðr þ uÞ fi Þ
Ri ðuÞ ¼ (8)
ðf i ðrÞ fi Þ2
where fi is the phase volume fraction, and the overline denotes spatial average
as defined by Eq. (3). In isotropic media, the correlation function does not
depend on the direction, but only on the magnitude u ¼ jjujj of the displacement
vector. The zeroth moment of the correlation function,
Z 1
Lc;i ¼ Ri ðuÞ du (9)
u¼0
is the so-called correlation length, an alternative measure of the characteristic

length-scale of each phase in the multiphase medium. The characteristic length-
scale of a phase in the medium can also be measured by the so-called chord-length
distribution function (e.g., Roberts and Torquato, 1999), defined as the distribu-
tion of line segments passing through a given phase when the medium is randomly
sampled by lines. The measure of characteristic length-scales is important from the
physicochemical point of view, e.g., for measuring the mean pore diameter, which
influences phenomena such as capillary condensation or Knudsen diffusivity.
In the context of diffusion in porous media, the tortuosity parameter t has
traditionally been used. Tortuosity is defined geometrically as the ratio of the
shortest distance through the pore space between two randomly positioned points
to their Euclidean distance, averaged over a large number of points. Tortuosity of
the pore space (or any other percolating phase in a multiphase medium) can be
evaluated directly from its phase function. The pore space distance of two points
can be calculated using an algorithm based on the propagation of a reaction–
diffusion front through the medium (Stepanek et al., 2000). Topological prop-
erties of a percolating phase can further be characterized by reducing the phase
into a skeleton (Liang et al., 2000a) by applying a thinning algorithm. Skeleton-
ization is useful for establishing the pore connectivity, or for distinguishing be-
tween the conducting and the dead-end pores in a porous medium.
Experimental techniques commonly used to measure pore size distribution,
such as mercury porosimetry or BET analysis (Gregg and Sing, 1982), yield
pore size distribution data that are not uniquely related to the pore space mor-
phology. They are generated by interpreting mercury intrusion–extrusion or
sorption hysteresis curves on the basis of an equivalent cylindrical pore as-
sumption. To make direct comparison with digitally reconstructed porous
media possible, morphology characterization methods based on simulated
mercury porosimetry or simulated capillary condensation (Stepanek et al., 1999)
should be used.
D. DIGITAL RECONSTRUCTION OF MULTIPHASE MEDIA
The reconstruction of the porous/multiphase medium is the process starting

from electron microscopy or other image of the medium, followed by the eval-
uation of suitable morphological descriptors of the image, concluding with the
generation of a spatially 3D porous/multiphase medium with the same
morphological characteristics as those of the original image. The reconstruct-
ed porous/multiphase medium is then used as the input for the calculation of
effective transport, mechanical or electrical properties of the medium (Tor-
quato, 2002), or as the input to the modeling of various reactions and other
transformation processes.
In mathematical terms, given the porosity and suitable morphological char-
acteristics, we want to construct the replica of the porous/multiphase medium
SEM image Binarized image
µm µm
sem 140
1
R
0.8 Rx
Ry
0.6
Autocorrelation
0.4 function R(u)
0.2
0
43µm 0 5 10 15
µm
FIG. 3. Procedure of the reconstruction of porous media: SEM image of the porous silica-sup-
ported catalyst particle, selection of rectangular box, binarization of SEM image, calculation of the
autocorrelation function R(u), and reconstructed porous medium.
represented by the phase function fi(r) defined on the discrete grid of voxels with
coordinates r; cf. Fig. 3. Here, we present the principles of three different
algorithms of the stochastic reconstruction of the porous/multiphase media:
(i) simulated annealing, (ii) Poissonian generation of polydisperse spheres, and
(iii) thresholding of correlated random fields. Then we briefly survey the re-
construction of porous media by mechanical diagenesis.
1. Simulated Annealing
Stochastic reconstruction by the simulated annealing algorithm is one of the
most versatile classes of reconstruction methods, although not very efficient in
terms of the computational demand. The increasing popularity of the simulated
annealing algorithm (Hazlett, 1997; Manwart et al., 2000) has been caused by
ever-growing computational power and straightforward implementation.
Let us consider an isotropic porous medium under the reconstruction
described by a pore phase function fg(k)(r) in the kth iteration step of the sim-
ulated annealing algorithm and let the actual statistical characteristics of this
phase function, i.e., the two-point correlation function, be Rg(k)(u). The distance
of Rg(k)(u) from the target morphological characteristics Rgtarget(u) of the
microscopy image of the porous medium is
X
umax
E ðkÞ
g ¼ ðRðkÞ target
g ðuÞ Rg ðuÞÞ2 (10)
u¼0
The simulated annealing algorithm typically starts from the random phase
function f ð0Þ ð0Þ
g ðrÞ having the required porosity ¼ f g ðrÞ. In the kth iteration step,
we interchange the values of two voxels f g ðr1 Þ and f ðkÞ
ðkÞ
g ðr2 Þ at randomly chosen
positions r1 and r2, where one voxel is from the pore and the other is from the
solid phase, i.e., f ðkÞ ðkÞ
g ðr1 Þ þ f g ðr2 Þ ¼ 1. Thus, a new phase configuration f g
ðkþ1Þ
ðrÞ
is proposed and its statistical characteristics Rðkþ1Þ g ðuÞ as well as the distance
E ðkþ1Þ
g from the target function are calculated according to Eq. (10). The pro-
posed configuration f gðkþ1Þ ðrÞ is accepted with the probability p given by the
Metropolis rule
8
<1 if E gðkþ1Þ pE ðkÞ
g
p¼ ðkÞ ðkþ1Þ (11)
:e g gðE E Þ=T
if E g 4E ðkÞ
ðkþ1Þ
g
Otherwise, the old configuration is restored, i.e., f ðkþ1Þ

g ðrÞ ¼ f ðkÞ
g ðrÞ. In Eq. (11),
the symbol T represents the fictitious temperature of the system. As the tem-
perature decreases, more changes that increase the distance E ðkÞ g from the target
function are rejected. The reason for not rejecting all proposed changes when
E ðkþ1Þ
g 4E ðkÞ
g is to allow the system to evolve to the desired target configuration
without being trapped in a local minimum of the distance from the target
function give by Eq. (10); cf. Yeong and Torquato (1998a). Manwart et al.
(2002) suggested implementing the exponentially decreasing fictitious temper-
ature T ¼ T 0 ek=b , where T0 and b are adjustable parameters. The reconstruc-
tion ends when the target characteristic is reached with a certain tolerance, or
when a large number of successive proposed changes of f ðkÞ g ðrÞ are rejected.
The selection of interfacial voxels (which have at least one neighbor that
contains a different phase) as candidates r1 and r2 of an interchange significantly
improves the rate of convergence of the simulated annealing algorithm (Roz-
man and Utz, 2001). The algorithm is versatile enough to allow the recon-
struction of any multiphase media. Figure 4 illustrates the simulated annealing
algorithm on the section of a reconstructed, spatially 3D cube.
2. Poissonian Generation of Polydisperse Spheres

Poissonian generation of polydisperse spheres is another technique for the
reconstruction of porous media (Thovert et al., 2001). This technique is useful
for the reconstruction of media that result from the deposition of grains
followed by their compaction, but it also allows the reconstruction of non-
granular porous structures.
0.40 target
t two-point CF
0.35 actual two-point CF
0.30
S2(u)
0.25
0.20
0.15
0.10
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
u u u
FIG. 4. Simulated annealing algorithm of the cube consisting of 65 65 65 voxels starting from
random initial condition (left), after 14 106 iterations (middle) and after 17 106 iterations (right).
The two-point correlation function S2 ðuÞ ¼ RðuÞð 2 Þ þ 2 is compared with the target correlation
function (CF) during the reconstruction.
For a porous medium represented by the phase function f g ðrÞ, we define the
covering radius rc(r) of the solid phase as the radius of the largest sphere (or disk
in 2D porous media) placed entirely into the solid phase and covering the point
with coordinates r. The value of rc is zero inside the pores. The covering radius
rc can be found by the morphological operation called ‘‘opening,’’ i.e., the
erosion followed by the dilatation. The dilatation of the solid-phase domain
A by the spherical element Br with radius r is the set A B r covered by all
translations of Br centered in A,
[
A Br ¼ ðr þ sÞ (12)
r2A; s2B r
The erosion is the dual operation, corresponding to all points in A not covered
by a sphere Br centered out of A,
A B r ¼ ðA0 B r Þ0 (13)
where the prime stands for the complement. The opening FðA; rÞ is the result of
an erosion followed by a dilatation
FðA; rÞ ¼ ðA B r Þ B r (14)
The important property of the opening FðA; rÞ of the solid-phase domain A by

the sphere/disk Br of radius r is that it contains only points that can be covered
by Br, because the components that were too small to contain Br were removed
by an erosion. The covering radius can thus be defined as
rc ðrÞ ¼ supfr : r 2 FðA; rÞg (15)
The integral distribution of the covering radius G(rc) is defined as
Gðrc Þ ¼ ðVolume fraction of the setÞ fr 2 A; rc ðrÞorg (16)
and the limiting values of G(rc) are Gð0Þ ¼ 0 and Gðþ1Þ ¼ 1, respectively.
Graphically, the distribution G(rc) is the volume fraction of the solid phase,
which cannot be covered by spheres (disks) of radius rc.
The algorithm of the Poissonian generation of polydisperse spheres is simple
(Thovert et al., 2001). Spheres representing the solid phase are generated and
randomly placed into the reconstruction domain until the desired porosity e is
reached. The distribution of the radii of generated spheres r corresponds to the
spatially 3D distribution of the covering radius G3D(rc) given by Eq. (16). In
practice, the distribution of the covering radii can be evaluated either from the
spatially 2D microscopy images (i.e., G2D(rc)) or from the spatially 3D solid-
phase function obtained by computed micro-tomography or successive TEM
imaging of sections of the porous medium (i.e., G3D(rc)). Clearly, the distribu-
tions G2D(rc) and G3D(rc) are generally different even for the same porous me-
dium, and there is no unique mapping from G2D(rc) into G3D(rc) for general
porous media.
If the spatially 3D distribution of the covering radii G3D(rc) is not available,
one has to reconstruct the porous medium from 2D distribution of covering
radii G2D(rc) evaluated from statistical analysis of microscopy images. The re-
construction algorithm then has to search for such a 3D distribution G3D(rc) so
that 2D distributions G2D(rc) calculated from 2D sections of the reconstructed
porous media match the G2D(rc) obtained from microscopy images. As an ex-
ample, the 2D phase function (white and black color states for solid phase and
pores, respectively) and the corresponding calculated distribution of covering
radii are shown in Figure 5.
3. Thresholding of Correlated Random Fields

Let us generate a spatially 3D isotropic porous medium fg(r) with a given
porosity ¼ f g ðrÞ and a given correlation function Rg(u). The porous medium
fg(r) is considered to be discretized into the grid of N c N c N c voxels, each
voxel of the same size h. The reconstructed cube of the porous medium is usually
considered to have periodic boundary conditions.
The random and discretized field fg(r) can be devised from a Gaussian field
X(r) discretized on the same N c N c N c grid, which is the starting point of
SEM image of low porosity medium

black color = pore
white color = solid
1.0
0.8
Distribution G
Grain-size
0.6
(solid-size)
0.4 distribution
0.2
0.0
0 5 10 15 20 25
Covering radius rC
Covering by
circles of radius rC
FIG. 5. Distribution of covering radius rc. The solid phase of the porous medium is covered by
circles of radius rc and the color of circles depends on circle radius. The integral distribution of
covering radius G(rc) is then displayed.
the algorithm and which is successively passed through the linear and nonlinear
filters. The random variables in X(r) are assumed to be normally distributed and
are independent (i.e., not correlated). A linear operator can be defined by an
array of coefficients a(u0 ), where u0 belongs to a finite cube, u0 2 ½0; Lc 3 , and
Lc oðN c hÞ. A new random field Y(r) can be expressed as a linear combination of
the random variables X(r),
X
Y ðrÞ ¼ aðu0 ÞX ððr þ u0 ÞmodðN c hÞÞ (17)
u0 2½0;Lc 3
where the correlated random field Y(r) is also discretized on the N 3c grid. Its
correlation function Ry(u) is calculated from the given correlation function
Rg(u) and from the porosity e. The coefficients a(u0 ) are obtained from the
correlation function Ry(u); cf. the details of the algorithm described by Adler
and Thovert (1998). The particular choice of a(u0 ) resulting in the Gaussian
correlated random field is
!
0 pu0
aðu Þ ¼ exp 2 (18)
Ly
where Ly is the correlation distance (Mourzenko et al., 2001).

Although the random field Y(r) is correlated, it takes its values in the space of
real numbers R, while the porous medium has to be represented by a discrete-
valued field fg(r). In order to create such a field from Y(r), one applies a non-
linear filter fg(r) ¼ G(Y(r)), i.e., the random variable fg(r) is the deterministic
function of Y(r). The nonlinear filter is realized by the introduction of the

threshold te,
(
1; Y ðrÞpt
f g ðrÞ ¼ (19)
0; Y ðrÞ4t
where the value of te is adjusted so that the given porosity is satisfied, i.e.,
f g ðrÞ ¼ .
The thresholding of correlated random fields was generalized for the case of
multiphase media by Losic et al. (1997). However, this reconstruction technique
is limited only to cases where the third phase is located at the interface between
two phases or within a single phase.
4. Reconstruction Based on Mechanical Diagenesis
An alternative to stochastic reconstruction of multiphase media is the recon-
struction based on the direct simulation of processes by which the medium is
physically formed, e.g., phase separation or agglomeration and sintering of
particles to form a porous matrix. An advantage of this approach is that apart
from generating a medium for the purpose of further computational experi-
ments, the reconstruction procedure also yields information about the sequence
of transformation steps and the processing conditions required in order to form
the medium physically. It is thereby ensured that only physically realizable
structures are generated, which is not necessarily the case when a stochastic
reconstruction method such as simulated annealing is employed.
The mechanical routes of diagenesis usually represent particle packing in the
force field, e.g., gravitational or electrostatic; cf. Fig. 6. Diagenesis-based re-
construction techniques have been used, e.g., by Kim et al. (2000) for the cre-
ation of model ceramic agglomerates by particle packing. Similarly,
Kainourgiakis et al. (2002) have generated model porous media by ballistic
deposition of spheres. Thomson and Gubbins (2000) generated model porous
carbon structure by a reverse MC method. Formation of wet particle assemblies
has also been represented by mechanical diagenesis (Kohout et al., 2004).
There is of course a trade-off between the level of rigorosity of the diagenesis
model and the computational cost (as well as the number of required input
parameters). A qualitatively correct model is often sufficient for the purpose of
porous/multiphase media reconstruction; a rigorous diagenesis model is re-
quired if the results are to be used, e.g., for process optimization.
E. CALCULATION OF EFFECTIVE PROPERTIES
1. Calculation of Effective Transport Properties on Reconstructed Porous Media

Calculation of the effective transport properties—thermal or electrical
conductivity, effective diffusivity, and permeability—as a function of the
FIG. 6. Formation of the granular structure of the cylindrical pellet by the mechanism of se-
quential deposition of individual grains, the so-called ballistic packing.
microstructure of porous and multiphase media is of key interest for many

chemical engineering applications as well as for other areas, such as geology
(e.g., Dodds and Rothman, 2000; Sahimi, 1993). The general methodology for
the calculation of effective transport properties from digitally reconstructed
media is based on the local solution of transport equations followed by volume
averaging (Quintard et al., 1997; Whitaker, 1999), and can be summarized in the
following steps:
(i) Microstructure characterization, as described in Section II.C.

(ii) Microstructure realization, as described in Section II.D.
(iii) Local solution of transport equations on the reconstructed medium with an
imposed macroscopic gradient of temperature (for conductivity), concen-
tration (for diffusivity), or pressure (for permeability).
(iv) Evaluation of the macroscopic transport parameter of interest (thermal

conductivity, effective diffusivity, permeability) from converged temperature,
concentration, or velocity fields according to the macroscopic transport laws.
These steps can then be iterated while a given morphological property of the
medium (e.g., porosity or mean pore diameter) is systematically varied in order
to generate structure–property correlations which can then be used in higher-
level simulations either in the form of a look-up table, or fitted by an empirical
correlation or theoretically founded function. The background to steps (i) and
(ii) has already been discussed, so let us now consider steps (iii) and (iv) in detail.
For an alternative analysis of the effective transport properties of disordered
media based on the exit-time analysis, see Giona et al. (1995). It has to be noted
that the four-step procedure outlined above is the direct computational analogy
of experimental determination of transport properties. Step (ii) corresponds to
the physical preparation of a sample, step (i) corresponds to its physicochemical
characterization such as measurement of porosity, step (iii) corresponds to the
physical measurement of transport rate (e.g., diffusion in a Wicke–Kahlenbach
cell or heat flux in a guarded plate thermal conductivity meter), and step (iv)
corresponds to the evaluation of the transport property of interest from ‘‘raw’’
mass- or heat-flux data.
The underlying problem of the calculation of transport properties is the nu-
merical solution of the Poisson equation on a complex 3D domain defined by
the phase functions. Several methods can be used, e.g., finite-difference disc-
retization and the solution of the resulting linear system by the successive over-
relaxation method (Kohout et al., in press), the conjugate gradient method, or
methods specifically suitable for solving large, sparse systems of linear algebraic
equations. Heat conduction at steady state is governed by Fourier’s law
r ðli rTÞ ¼ 0 (20)
where local values of thermal conductivity li of each phase are used. At phase
boundaries, the continuity of heat flux is required. A macroscopic temperature
gradient ðT 2 T 1 Þ=Lz is imposed on the unit cell in a chosen direction (for
isotropic media the choice of direction is not important), and the effective
conductivity le is obtained from the converged temperature field using mac-
roscopic Fourier’s law,
Q Lz
le ¼ (21)
A T2 T1
where Q is the net steady-state heat flow rate obtained as a solution of Eq. (20),
A ¼ Lx Ly the cross-sectional area of the unit cell perpendicular to the direction
of the temperature gradient, and Lz the dimension of the unit cell in the di-
rection of the temperature gradient (here z). The simulation setup is illustrated
in Fig. 7.
.
Qout
.
T2 Qout
L
T1 .
Qin
FIG. 7. Unit cell containing a wet particle packing for calculating the effective thermal conduc-
tivity. A cross-section showing the local heat fluxes at steady state (from Kohout et al., 2004).
The dependence of conductivity on the composition and microstructure of

porous media has been investigated, e.g., by Argento and Bouvard (1996) for
sphere packing, or Roberts and Knackstedt (1996) for a range of model micro-
structures, including model solid foams. Effective thermal conductivity of two-
and three-phase systems (solid-particle packing partially or fully saturated by a
liquid phase) has recently been determined by Kohout et al. (2004) both com-
putationally and experimentally. For a two-component system, consisting of
particles of phase B dispersed in a continuous phase A, le as function of the
phase volume fraction fB was confirmed to observe Archie’s-like power law
le ¼ ðlB lA ÞfcB þ lA (22)
where lA and lB are component conductivities, and the exponent c is a function

of the ratio lB/lA and of the microstructure. Simulation results for the case of
spherical particles are plotted in Fig. 8. For spherical particles, it was found that
c ¼ 0:136 logðlB =lA Þ þ 1:27 and for cubic particles c ¼ 0:239 logðlB =lA Þ þ 1:10
(Kohout et al., 2004). These correlations are valid in the range 2olB =lA o100
and 0:5ofB o0:7 for spheres, and 0:0ofB o1:0 for cubes.
Calculation of the effective diffusivity is analogous to that of conductivity.
The steady-state Fick’s law
r ðDi rcÞ ¼ 0 (23)
is solved in the unit cell, where Di is the diffusion coefficient in phase i and c is
the concentration. In a multiphase medium, the diffusivities in some of the
phases may be zero. A macroscopic concentration gradient (c2 – c1)/Lz is im-
posed on the unit cell in a chosen direction and Eq. (23) is solved. The effective
diffusivity is then evaluated from the macroscopic Fick’s law
J Lz
De ¼ (24)
A c2 c1
0.8
B:A = 10:1
e [Wm−1K−1]
100:1
0.6
5:1
0.4
50:1
0.2 2:1
20:1
0 10:1
0 0.2 0.4 0.6 0.8 1
xB [−]
FIG. 8. Dependence of effective thermal conductivity of a two-phase medium A–B (randomly

positioned overlapping spheres of phase B in continuous phase A) on the ratio of component
conductivities (from Kohout et al., 2004).
where J/A is the steady-state diffusive flux obtained as a solution of Eq. (23). If
diffusion occurs through only one phase with diffusion coefficient Di, the di-
mensionless ratio De/Di, sometimes also called the formation factor, can be
correlated with the phase volume fraction of the phase, fi, and the tortuosity
factor t of the porous medium can also be evaluated from De ¼ Di fi =t. Recent
examples of effective diffusivity studies in porous media or porous media filled
with two fluid phases are provided, e.g., by Kainourgiakis et al. (2002), Nam
and Kaviany (2003), and Galani et al. (2004).
In microporous materials where Knudsen diffusion prevails, De cannot be
calculated by solving Fick’s law. The use of a discrete particle simulation
method such as dynamic MC is appropriate in such cases (Coppens and Malek,
2003; Zalc et al., 2003, 2004). In the Knudsen regime, relatively few gas mol-
ecules collide with each other compared with the number of collisions between
molecules and pore walls. One of the fundamental assumptions of the Knudsen
diffusion is that the direction in which a molecule rebounds from a pore wall is
independent of the direction in which it approaches the wall, and is governed by
the cosine law: the probability ds that a molecule leaves the surface in the solid
angle do forming an angle y with the normal to the surface is
ds ¼ ðdo=pÞ cos y (25)
The theoretical basis for the cosine distribution of velocities remained unclear
for a long time. Feres and Yablonsky (2004) introduced a simple model that
considers the surface of the pores to have some micro-roughness, and employ
the random billiard model for individual molecules. In this model, gas molecules
move freely inside a region of space until they collide elastically with a rough
surface and instantaneously rebound according to the usual law of equal angles
of incidence and reflection. Geometrical irregularities of the surface naturally
make the angle of reflection sensitive to the precise point at which a collision
takes place. As the scale of surface irregularities is reduced, the deterministic but
highly unstable angle of reflection becomes effectively random. Its probability
law, however, is still a function of the (scale-independent characteristics of the)
surface geometry. Feres and Yablonsky (2004) show that the Knudsen cosine
law (25) is a stationary probability distribution of post-collision scattered di-
rections (velocities) of gas molecules.
Permeability is calculated by imposing a macroscopic pressure gradient across
the unit cell and solving the Stokes and continuity equations in the pore space
rp ¼ Zr2 u; ru¼0 (26)
where p is the pressure, Z the viscosity, and u the velocity vector, with no-slip
boundary conditions on the solid walls. This procedure can be regarded as the
computational equivalent of gas permeation experiments (Capek et al., 2001).
In partially saturated porous media (e.g., water–air and water–oil), each liquid
phase can be considered stationary and the relative permeability of the
other phase calculated as a function of its relative pore-space saturation, or a
two-phase flow problem can be solved and the relative permeabilities of the
two phases determined simultaneously. The permeability k0 is obtained from
converged velocity profile (Fig. 9) using the macroscopic Darcy’s law
U Lz
k0 ¼ Z (27)
A p2 p1
FIG. 9. Streamlines showing steady-state velocity profile during single-phase flow in a recon-
structed porous medium.
0.01 spheres
crystals hydrophobic
crystals hydrophilic
0.008 fit
Carman-Kozeny
κ / de2 [−]
0.006
0.004
0.002
0
0 0.1 0.2 0.3 0.4 0.5
porosity xG [−]
FIG. 10. Dependence of relative permeability of gas phase on the volume fraction of gas phase for
dry and partially water-saturated porous media of various morphology, fitted by a power-law
function and compared with the Carman–Kozeny equation (from Kohout et al., in press).
where U is the imposed macroscopic volume flowrate (U/A is superficial

velocity), and (p2 – p1) the pressure drop across the unit cell. For the purpose of
structure–property correlations, permeability is typically scaled by the square
of the equivalent hydraulic diameter d 2e defined following Eq. (4). An example of
calculated dependence of the relative permeability of the gas phase on porosity
of random packings of spherical and rectangular particles partially saturated by
wetting and non-wetting liquid is shown in Fig. 10 (Kohout et al., in press).
2. Calculation of Tensile Properties

The relationship between the structure of the disordered heterogeneous ma-
terial (e.g., composite and porous media) and the effective physical properties
(e.g., elastic moduli, thermal expansion coefficient, and failure characteristics)
can also be addressed by the concept of the reconstructed porous/multiphase
media (Torquato, 2000). For example, it is of great practical interest to under-
stand how spatial variability in the microstructure of composites affects the
failure characteristics of heterogeneous materials. The determination of the de-
formation under the stress of the porous material is important in porous pack-
ing of beds, mechanical properties of membranes (where the pressure applied in
membrane separations is often large), mechanical properties of foams and gels,
etc. Let us restrict our discussion to equilibrium mechanical properties in static
deformations, e.g., effective Young’s modulus and Poisson’s ratio. The calcu-
lation of the impact resistance and other dynamic mechanical properties can be
addressed by discrete element models (Thornton et al., 1999, 2004).
The stress tensor r in the perfectly elastic and isotropic solid phase of the
porous medium is described by the constitutive equation
r ¼ lðr dÞI þ mðrd þ ðrdÞT Þ (28)

where l and m are the Lamé coefficients, and d the displacement that charac-
terizes the deformation of the solid phase. The condition of elastostatic equi-
librium in the absence of external body forces r r ¼ 0 can be formulated with
the above definition of the stress tensor as
ðl þ mÞrðr dÞ þ mr2 d ¼ 0 (29)
The boundary condition at the solid–pore interphase Sp has to be supplemented

as r n ¼ T, that is
½mðrd þ ðrdÞT Þ þ lðr dÞI n ¼ T on Sp (30)
where n is the unit normal of the solid phase and T the surface traction.
Let us consider the spatially periodic porous medium consisting of an infinite
number of identical unit cells. The spatially periodic medium is subjected to a
macroscopic deformation described by the tensor of deformation A, and the
local displacement d ¼ A x þ d~ can be decomposed into a macroscopic defor-
mation A x and a microscopic spatially periodic displacement d; ~ cf. Poulet
et al. (1996). This decomposition introduced into elastostatic Equations (29) and
(30) yields
~ þ mr2 d~ ¼ 0
ðl þ mÞrðr dÞ (31)
½mðrd~ þ ðrdÞ ~ n
~ T Þ þ lðr dÞI
¼ mðA þ AT Þ n lðtr AÞn þ T on Sp ð32Þ
Hence, the macroscopic deformation A plays the role of an external ‘‘driving

force.’’ The macroscopic stress hri can be defined as the average local stress over
the solid portion Vs of the unit cell
Z
1
hri ¼ s dV (33)
V Vs
where V is the total volume of the unit cell. When the porous medium is iso-
tropic, the macroscopic stress is given by
hri ¼ le tr E I þ 2me E (34)
where le and me are effective Lamé coefficients of the medium and

E ¼ 12ðA þ AT Þ the macroscopic strain. The effective Young’s modulus Ee and
the effective Poisson’s ratio ne are related to effective Lamé coefficients as
E e ¼ me ð3le þ 2me Þ=ðle þ me Þ and ne ¼ le =ð2ðle þ me ÞÞ, respectively.
F. EFFECTIVE-SCALE TRANSPORT MODELS
Effective-scale models make no specific assumption about the geometry of the

pore space, and the reaction-transport equations are formulated with the con-
centration field averaged over a certain effective spatial scale. All the effective-
scale models use in general the so-called tortuosity t as a fitting parameter. The
tortuosity can be related to the geometry of the pore space, cf. Section II.C, but
it lumps too much information to be used as a single characteristic of the porous
medium, because it depends also on pressure and on the actual set of diffusing
gases (Aris, 1975; Bhatia, 1986; Jackson, 1977).
The constitutive equations of transport in porous media comprise both
physical properties of components and pairs of components and simplifying
assumptions about the geometrical characteristics of the porous medium. Two
advanced effective-scale (i.e., space-averaged) models are commonly applied for
description of combined bulk diffusion, Knudsen diffusion and permeation
transport of multicomponent gas mixtures—Mean Transport-Pore Model
(MTPM)—and Dusty Gas Model (DGM); cf. Mason and Malinauskas
(1983), Schneider and Gelbin (1984), and Krishna and Wesseling (1997). The
molar flux intensity of the ith component Ni is the sum of the diffusion N di and
permeation N pi contributions,
N i ¼ N di þ N pi ; i ¼ 1; . . . ; n (35)
In the DGM, the solid phase is modeled as giant dust molecules held motionless
in space with which the diffusing gas molecules collide. The constitutive equa-
tions governing the diffusion molar flux intensities N di for both MTPM and
DGM are the generalized Maxwell–Stefan equations
(
N di X n yj N di yi N dj cT ryi MTPM
þ ¼ (36)
DK
i i¼1;jai Deff
ij
rðcT yi Þ DGM
for i ¼ 1,y,n, where yi is the mole fraction of the ith component, cT the total
concentration, Deff
ij ¼ cDij the effective binary diffusion coefficient, Dij the bi-
nary bulk diffusion coefficient, and c the geometrical factor formally defined as
the ratio of porosity e and tortuosity t, c ¼ =t. Parameter DK i ¼ chriKi is the
effective Knudsenp diffusivityffi of the ith component, where the Knudsen coef-
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ficient is Ki ¼ 23 8RT=ðpM i Þ, hri the mean pore radius, R the gas constant, T
the temperature, and Mi the molecular weight of the ith component.
Both MTPM and DGM share the Darcy’s constitutive equation governing
the permeation molar flux intensity of the ith component N pi ,
yi ki
N pi ¼ yi ki rcT ¼ rp; i ¼ 1; . . . ; n (37)
RT
where ki is the effective permeability coefficient of the ith component. In the

case of DGM, all effective permeability coefficients are identical, i.e.,
k ¼ k 1 ¼ . . . ¼ kn :
k ¼ k0 p=Z (38)
where Z is the dynamic viscosity and k0 the Darcy’s permeability. For cylindrical
pores with mean square pore radius hr2 i, the parameter k0 can be formally
replaced by k0 ¼ chr2 i=8. A number of correlations are available to estimate k0,
e.g., the popular Carman–Kozeny equation applicable for the aggregated bed of
spheres. MTPM assumes that the decisive part of the gas transport takes place
in transport pores that are considered to be cylindrical capillaries distributed
around the mean value hri. The width of this distribution is characterized by the
mean value of the squared transport-pore radii hr2 i. Effective permeability k can
be in the case of the single component permeation described by MTPM ap-
proximated by the Weber equation
k ¼ hricK þ hr2 icp=ð8ZÞ (39)
where K is the Knudsen number.

The DGM by Mason and Malinauskas (1983) is the frequently used alter-
native to effective Fick’s diffusion for the calculation of the multicomponent
diffusion and convection in the porous media. A number of special features of
multicomponent diffusion and convection in the pore space have been outlined
by Krishna (1993).
III. Transformations
The microstructure of the multiphase media is often the product of phase

transitions, e.g. (i) capillary condensation in the porous media, (ii) phase sep-
aration in polymer/polymer and polymer/solvent systems, (iii) nucleation and
growth of bubbles in the porous media, (iv) solidification of the melt with a
temporal three-phase microstructure (solid, melt, gas), and (v) dissolution,
crystallization or precipitation. The subject of our interest is not only the to-
pology of the resulting microstructured media, but also the dynamics of its
evolution involving the formation and/or growth of new phases.
The structure of the porous media can also be shaped by mechanical and
electrostatic forces, e.g., sheeting of granules at conveyor walls (Yao et al.,
2004), deposition of solid grains in the porous media, and fragmentation of
catalyst carriers. Simple examples of mechanical diagenesis of porous media
were already illustrated in Section II.D.4. In this section, we first describe
the morphological operation of skeletonization and the methods used for the
simulation of various phase transitions, and then discuss the modeling of

chemically reactive systems.
A. SKELETONIZATION
Let us first discuss the mapping from the phase function defined in Section
II.A into the pore-network diagram. The analysis of the network connectivity of
the pore (or solid) phase can be based on the construction of skeleton its
(Thovert et al., 1993). In the skeletonization of the porous phase, the elements of
the object are removed until only the skeleton remains. This process is called
‘‘thinning’’ in the terminology of mathematical morphology. The constructed
skeleton of the pore phase allows the analysis of its topological characteristics,
and the determination of the connected and percolating components of the po-
rous phase, e.g., transport or dead-end pores. Liang et al. (2000b) extracted the
skeleton of pore space using an improved parallel thinning algorithm introduced
by Ma and Sonka (1996). Once the skeleton is obtained, its vertices and edges
can be identified as well as the minima in the cross-sectional area of pores along
the skeleton branches, which correspond to bottlenecks of the pore space.
We applied a skeletonization method not to the pore space, but to the solid
phase in order to find the bottlenecks of the skeleton of the solid phase (Grof
et al., 2003). An illustrative example of the object and its skeleton is shown in Fig.
11. We employed the fully parallel 3D thinning algorithm described by Ma and
Sonka (1996) that preserves the original connectivity of the processed phase.
During the thinning, an object element is removed if its neighborhood satisfies any
of the deleting templates, and the thinning ends when no additional object element
can be removed. The elements of a resulting skeleton can then be classified as
(1) Line-end elements if they have only one neighboring element.

(2) Edge elements if they have two neighboring elements.
(3) Vertices elements if they have three or more neighboring elements.
The cross-sectional area of the solid phase perpendicular to the skeleton edge
is calculated along all skeleton branches of the solid phase. The solid phase is
then disconnected at the points corresponding to the absolute minima of that
cross-sectional area for every skeleton edge (branch); cf. Grof et al. (2003). Ex-
amples of solid-phase fragments of the porous media formed by the disconnec-
tion of the solid-phase skeleton at weak points are shown in Figures 11 and 12.
B. PHASE TRANSITIONS
1. Free Surface and Interfacial Flows

Simulation of free-surface and interfacial flows is a topic with many practical
applications, e.g., the formation of droplet clouds or sprays from liquid jets,
Porous medium Skeleton of the Fragmentation of

(cubes represent solid phase the solid phase
solid phase)
FIG. 11. Schema of skeletonization of porous media by the algorithm of conditional thinning and
disconnection of the skeleton at the weak points to predict the location of fractures. Classification of
skeleton elements according to number of neighboring elements: line-ends (1 neighboring element),
edges (2), vertices (3 and more) (from Grof et al., 2003).
Size of cube
158 fragments
64x64x64
Catalyst
fragments in 3D
Skeleton of the solid phase
FIG. 12. Skeleton of the the solid phase of spatially 3D reconstructed porous medium obtained by
conditional thinning and disconnection of the skeleton at the weak point (from Grof et al., 2003).
collision of liquid droplets, capillary condensation in pores, dissolution of po-

rous granules, and drying (Kohout et al., in press). Major difficulties in sim-
ulations are associated with the change of the interface topology that occurs
during the coalescence or breakup of droplets.
Simulation methods for problems with free surfaces governed by Navi-
er–Stokes equations were reviewed by Scardovelli and Zaleski (1999). The spe-
cific problems of these simulations are the location of the interface and the
choice of the spatial discretization:
In fixed-grid methods, there is a predefined grid that does not move with the
interface. The interface has to somehow cut across this structured or un-
structured fixed grid. The popular Volume of Fluid (VOF), Level Set (Sethian,
1996) or cellular automata methods are examples of a fixed-grid approach.
In moving-grid methods, the interface is a boundary between two subdomains
of the grid. The interface then identifies, at some order of approximation, with
boundaries of elements into which the phases are discretized. When the in-
terface undergoes large deformations, the grid has to be remeshed.
In particle-technique methods, no grids are needed. The liquid is considered to
be discretized into number of particles with force interacting among individ-
ual particles (Heyes et al., 2004).
Let us introduce the VOF method in more detail. The phase volume function
f^1 ðrÞ is defined by f^1 ¼ 1 in phase 1 and f^1 ¼ 0 in phase 2; cf. Eq. (1). A discrete
analog of the phase volume function f^1 is the scalar field f1 known as the
volume fraction (or area fraction in spatially 2D cases). It represents the fraction
of the volume of the voxel with size h and the coordinate of its center r filled
with the phase 1,
Z
fb ¼ 1 fb1 ðr þ uÞ du (40)
1
h3 u2½h=2;h=23
where a cubic mesh with a constant grid spacing h is considered. We have

0of b o1 in cells cut by the interface S and f b ¼ 0 or 1 away from it. In
1 1
the incompressible flow, mass conservation is equivalent to conservation of
the volume and hence of the characteristic function f b . In the VOF method, the
1
algorithmic scheme is local in the sense that only the f b values, pressure, and
1
velocity field of the neighboring cells are needed to update the fb ðrÞ value in the
1
voxel with coordinates r. For this reason, it is relatively simple to implement
these algorithms in parallel and, in particular, within the framework of domain
decomposition techniques.
The central problem of the VOF method is the reconstruction and represen-
tation of the interface. For example, in two spatial dimensions, the interface is
considered to be a continuous, piecewise smooth line; the problem of its re-
construction is that of finding an approximation to the section of the interface in
each cut cell by knowing only the volume fraction f b in that cell and in the
1
neighboring ones.
2. Capillary Condensation in Porous Media

Operations like pressure swing adsorption involve the condensation of the
liquid in the porous medium. Several researchers developed predictive models of
the configuration of liquid phase in the wet, unsaturated, porous media. The
method developed by Silverstein and Fort (2000) is based on simulated an-
nealing with random swapping of gas and liquid elements in the system to
achieve a global energy minimum defined by
X
G st ¼ Ai gi (41)
i
where i represents the interface, G st the total interfacial free energy of the system,
Ai the area of the interface, and gi the interfacial free energy of the interface. As
a result of interfacial tension, the system tends to minimize surface area A for a
given volume of condensate V, and this minimization is subjected to the Kelvin
equilibrium and wetting angle constraints. The Kelvin equation relates the in-
terface curvature radius r and the equilibrium vapor pressure above it p*,

2g V m
p ¼ ps exp (42)
r RT
Here ps is the saturated vapor pressure at temperature T, g the surface tension,
Vm the molar volume of the liquid, and the curvature radius r is conventionally
taken as negative for concave interfaces. Kelvin equation for a non-ideal mul-
ticomponent mixture was derived by Shapiro and Stenby (1997).
Stepanek et al. (1999) presented an algorithm for tracking the evolution of
liquid–vapor interfaces moving with curvature-dependent velocity so as to reach
an equilibrium shape, which represents the local minimum of a global Gibbs
energy. Their algorithm is derived from the level-set approach (Sethian, 1996),
and was applied to the modeling of capillary condensation hysteresis cycles in
reconstructed porous media. Concave interfaces tend to evaporate spontane-
ously, while either condensation or evaporation can occur on a convex interface,
depending on its curvature and the total pressure and composition of the vapor
phase. Interface displacements are curvature-dependent; the curvature radius r
is calculated by locally approximating the interface by an elliptic paraboloid
oriented so that its axis is identified with the liquid–gas interface normal vector
given by Eq. (6).
In the algorithm of Stepanek et al. (1999), the simulated domain is discreti-
zed into a set of cubic volume elements; the phase-encoding alphabet is
A ¼ fs; l; f; gg (meaning solid, liquid, interface, and gas, respectively). In addi-
tion, a value f b 2 ð0; 1Þ expressing the volume fraction of condensate in the ith
i
discretization element (i.e., its saturation) is assigned to every interface element.
The driving force of the interface growth in terms of the partial pressure,
jjp p
jjpp , was used as a convergence criterion of the algorithm. The pa-
rameter p has a physical meaning of the potential barrier to phase transition,
which reveals itself by the existence of oversaturated vapors and metastable
liquids.
3. Formation of Gas Cavities in Liquid-saturated Porous Medium

The formation of gas bubbles in a porous medium affects, e.g., the perform-
ance of anaerobic granular sludge particles containing entrapped gas (waste-
water treatment), or the performance of electrochemical reactors where a gas
(hydrogen, oxygen, chlorine) evolves inside a porous electrode. A related prob-
lem is that of bubble nucleation on structure surfaces, which can be as varied as
specially designed surfaces for enhanced nucleate boiling heat transfer, or the
wall of a champagne glass (Liger-Belair et al., 2002). In porous media, patterns

and rates of gas evolution are different than in the bulk phase because of the
heterogeneous nucleation. Moreover, the solute transport, capillary and viscous
forces affect the growth of gas bubbles in the porous media (Dominquez et al.,
2000). The first model of the bubble growth in the porous medium driven by a
constant or a time-varying supersaturation in the far field was presented by Li
and Yortsos (1995). The porous medium was represented in their study by an
equivalent network of nodes (pore bodies) and bonds (pore throats) of distrib-
uted sizes.
According to the Young–Laplace equation, a pressure difference Dp is re-
quired to support a stable, curved interface with curvature radii r1 and r2 (by
convention, positive for convex interfaces and negative for concave ones),

1 1
Dp ¼ pg pl ¼ s þ (43)
r1 r2
Another condition that has to be satisfied by a stable gas cavity existing in the
pore space is that the solid wall and the gas–liquid phase boundary must join at
a certain contact angle W, which depends on the wettability of the solid phase.
Stepanek et al. (2001b) developed an algorithm for finding a configuration of
gaseous and liquid phases in a porous medium, satisfying certain curvature at all
gas–liquid interfaces and a predefined contact angle at all gas–liquid–solid
contact lines; cf. Fig. 13. Their algorithm is based on a modified level-set
Initial condition
Equilibrium state
FIG. 13. Simulation of bubble growth and coarsening in a water-saturated porous medium (from
Stepanek et al., 2001b).
method for the tracking of moving interfaces propagating with curvature-

dependent velocity (Stepanek et al., 1999). The saturation of gas–liquid inter-
face voxels described by the saturation function gi 2 ð0; 1Þ is updated according
to dgi ¼ aðp1 =ZÞð1 p
1 Þ dt, where a is a numerical parameter determining
the rate of convergence and stability Z the viscosity of the liquid, and
p
1 ¼ p0g s0 ð1=r01 þ 1=r02 Þ, with p0g ¼ pg =p1 using the gas over-pressure, s0 ¼
s=ðapl Þ a dimensionless capillary pressure, and r01 and r02 are the estimates of the
interface curvature radii.
4. Solidification of Melts
Microstructure evolution in solids formed from a melt is important in proc-
esses such as prilling or casting. For example, in aluminum alloy shape casting
the final microstructure depends directly on the as-cast microstructure since the
only post-casting process is the heat treatment. The micro-porosity formed
owing to the combined effects of the volumetric shrinkage upon the solidifi-
cation of the melt and the precipitation of the dissolved hydrogen affects the
final properties of the aluminum alloy.
A solidification model of the aluminum melt was proposed by Atwood and
Lee (2003) and by Lee et al. (2004), and consists essentially of the grain growth
and pore growth sub-models. The nucleation and growth of the solid phase and
gas bubbles simulated by the cellular automata method, whereas the multi-
component diffusion and partitioning in a three-phase system (liquid melt, solid
grains, and gas) are simulated by a finite difference method. The driving force in
the cellular automata algorithm for both the nucleation and growth of the solid
phase is the local under-cooling. Hence, the heat balance equations at the mac-
ro-scale are required. When the amount of the solid in a voxel increases, the
dissolved solutes are partitioned between the solid and liquid phases. The
change in the solute concentration affects the degree of under-cooling and,
hence, the nucleation and growth processes. Similarly, a pore may form in a
voxel by a nucleation event, and a pore may expand into a neighboring cell if the
concentration of hydrogen in the melt favors it.
5. Spinodal Decomposition
Nauman and He (2001) considered the modeling of the formation of two-
phase morphologies by the spinodal decomposition of the initially homogene-
ous polymer/polymer or polymer/solvent mixture that was thermally quenched
into a two-phase region. Let us outline their modeling approach.
Fick’s diffusion cannot be applied to the description of the evolution of
multiphase systems because it acts to smooth concentration differences and
drives the system toward the uniform concentration profiles of species. Instead
of Fick’s diffusion, where the concentration gradient acts as a driving force, the
intensity of the diffusion flux N A of the species A can be considered to be
proportional to the gradient of the chemical potential
fA
N A ¼ D rmA (44)
RT
where D is the diffusion coefficient, fA the mole (or volume) fraction, and mA
the chemical potential of component A. Let gmix be the free energy of mixing of
the binary mixture. Then the chemical potential of component A is
mA ¼ gmix þ ð1 fA Þð@gmix =@fA Þ. The local balance of species A correspond-
ing to Eq. (44) is
" #
@fA fA ð1 fA Þ @2 gmix
¼r D rfA ¼ r ðDeff rfA Þ (45)
@t RT @f2A
where Deff is the effective diffusion coefficient. Let us assume that the Gibbs free
energy gmix of the polymer/polymer or polymer/solvent system is described by
the Flory–Huggins equation
gmix fA ln fA ð1 fA Þ lnð1 fA Þ
¼ þ þ wfA ð1 fA Þ (46)
RT NA NB
where NA and NB are dimensionless numbers chosen to reflect relative molec-

ular size and w the interaction parameter. Substituting from Eq. (46) into Eq.
(45) gives the effective diffusion coefficient

eff fA N A þ ð1 fA ÞN B
D ¼D 2fA ð1 fA Þw (47)
N AN B
The effective diffusion coefficient becomes negative for some range of compo-
sition when w is sufficiently large. The separation of phases requires an up-hill
diffusion (characterized by negative Deff) in which the material moves against
concentration gradients. The kinetics of the separation and the morphology of
separated phases depend on the details of such diffusion.
For a system to exhibit two-phase behavior, it is necessary that gmix ðfA Þ
has two minima within 0ofA o1. The points near the minima of gmix ðfA Þ
are called the binodal points, flA ofuA , and are given by the condition
g0mix ðflA Þ ¼ g0mix ðfuA Þ, where the derivative g0mix ¼ @gmix =@fA . The condition of
two minima of gmix ðfA Þ implies g00mix ðfA Þ ¼ @2 gmix =@f2A o0 for some interval
within 0ofA o1. The points fslA and fsu 00
A , where gmix ðfA Þ ¼ 0, are the spinodal
points, and the region between these points is thermodynamically unstable. The
effective diffusivity Deff is negative in this region, cf. Eqs. (45) and (47). An
initially homogeneous mixture quenched to lie within this region will undergo
spontaneous phase separation by the process of spinodal decomposition. How-
ever, Eq. (45) does not account for a surface energy term in the free energy
expression.
The remedy is to assume that the free energy of mixing of the system DG mix
depends both on the composition at a point fA as well as on the composition
gradient rfA
Z h i
k
DG mix ¼ gmix þ ðrfA Þ2 dV (48)
V 2
where k is the gradient energy parameter that is related to the surface tension s.
Equilibrium is achieved when the above integral is minimized subject to the
material balance constraint, i.e.,
Z
1
fA ¼ fA dV (49)
V V
where fA is the mean concentration of species A. Let us consider fA within the

two-phase region and the system governed by Eq. (48). Then phase separation
would lower the free energy through the gmix term but would necessarily cause a
gradient, which would raise the free energy through the k term. A consequence
of this trade-off between bulk and surface energies is that very small systems will
not bifurcate into two phases but will remain in the single phase.
Eq. (48) suggests that the gradient of composition be introduced into the
generalized definition of chemical potential
k
mA ¼ gmix ðrfA Þ2 þ ð1 fA Þðg0mix kr2 fA Þ (50)
2
The local balance equation (45) can then be reformulated with the generalized
(or gradient-dependent) chemical potential defined by Eq. (50) into

@fA f k
¼ r D A r gmix ðrfA Þ2 þ ð1 þ fA Þðg0mix kr2 fA Þ
@t RT 2
fA ð1 fA Þ 00
¼rD ðgmix rfA kr3 fA Þ ð51Þ
RT
Spinodal decomposition simulated with Eq. (51) does not generate stationary
equilibrium configurations. Instead, it gives non-equilibrium structures that
continue to evolve in time even in the late stages of the simulation. There are
two mechanisms for domain growth in such quiescent systems. The first is
Ostwald ripening, where diffusion of individual molecules through the contin-
uous phase will grow large domains at the expense of smaller ones. The second
mechanism is coalescence induced by Brownian motion or convection.
The concept of spinodal decomposition can be extended to multicomponent
systems and allow the prediction of various morphologies, e.g., core-shell par-
ticles, dual-continuous phases, dual-dispersed phase, and more complex struc-
tures. Spinodal decomposition simulations, in contrast to cellular automata, do
not require adjustable constants to predict non-equilibrium morphologies in

binary and ternary polymer systems.
6. Deposition and Dissolution in Porous Media

One of the important transformation processes in the porous media is the
deposition of solid material in pores and the subsequent clogging of the fluid
flow through the porous medium. Typical examples are filtration and clogging
of well walls in oil fields due to the particles present in the injected fluid. An-
other example is the deposition of clusters of active catalyst species in porous
supports.
Salles et al. (1993) modeled the flowing fluid through the porous medium
coupled with the deposition of solid particles resulting in the modification of the
geometry of the pore space and of the flow through the medium. The velocity
field is determined by solving the Stokes equations (26) in the current state of
the pore space. The deposition rates are calculated by means of random walks,
i.e., the trajectories of individual particles are governed by the convection of the
fluid and by the Brownian diffusion. Each time a particle hits a solid wall, it is
assumed to stick there, and the amount of such particles is recorded for every
elementary area of the solid–liquid interface, i.e., for every face of elementary
voxel lying on this interface. Whenever the cumulative quantity of matter de-
posited on the faces of a given voxel reaches a prescribed threshold, this voxel is
converted to solid.
Control of the deposition of Pd films in porous alumina under supercritical
CO2 conditions was investigated by Gummalla et al. (2004). Their model cap-
tures transport of reactants through the network of capillary pores, homoge-
neous reaction of reagents producing an intermediate species, nucleation, and
growth of the film as a moving boundary problem. The nucleation is treated
stochastically at the finest level, whereas transport and reaction at coarser levels
are treated deterministically.
In the petroleum industry, HCl is routinely injected into carbonate forma-
tions in order to improve oil or gas production. It is known that the porous
medium is not etched uniformly by the reactive fluid but that unstable disso-
lution patterns consisting of highly ramified, empty channels are formed. The
channels are commonly called ‘‘wormholes.’’ As soon as a wormhole pattern
develops, all the fluid will flow through it. Any local increase in the flow rate
results in an increase in the local dissolution rate. A piece of porous medium in
which a wormhole pattern has been created can be considered as composed of
two parts: the first part (wormholes) of very large permeability, and the second
part keeping its original permeability.
Daccord et al. (1993a, b) investigated systematically the types of patterns
that develop by the etching of the porous medium, when the flow rate and
the reaction rate (i.e., rate of dissolution) are varied. When the kinetics of the
surface reaction is limited, the porous medium is etched in a uniform way on
the microscopic scale, and there is a reaction front of macroscopic size. Mass-
transport-limited kinetics creates empty channels (wormholes), which by-
pass most of the porous medium. For low flow rates, all the dissolution takes
place at the entrance of the pores of the medium, and compact patterns are
observed.
C. CHEMICALLY REACTIVE SYSTEMS
The heterogeneous reactors with supported porous catalysts are one of the
driving forces of experimental research and simulations of chemically reactive
systems in porous media. It is believed that the combination of theoretical
methods and surface science approaches can shorten the time required for the
development of a new catalyst and optimization of reaction conditions (Keil,
1996). The multiscale picture of heterogeneous catalytic processes has to be
considered, with hydrodynamics and heat transfer playing an important role on
the reactor (macro-)scale, significant mass transport resistances on the catalyst
particle (meso-)scale and with reaction events restricted within the (micro-)scale
on nanometer and sub-nanometer level (Lakatos, 2001; Mann, 1993; Tian et al.,
2004).
When modeling phenomena within porous catalyst particles, one has to de-
scribe a number of simultaneous processes: (i) multicomponent diffusion of
reactants into and out of the pores of the catalyst support, (ii) adsorption of
reactants on and desorption of products from catalytic/support surfaces, and
(iii) catalytic reaction. A fundamental understanding of catalytic reactions, i.e.,
cleavage and formation of chemical bonds, can only be achieved with the aid of
quantum mechanics and statistical physics. An important subproblem is the
description of the porous structure of the support and its optimization with
respect to minimum diffusion resistances leading to a higher catalyst perform-
ance. Another important subproblem is the nanoscale description of the nature
of surfaces, surface phase transitions, and change of the bonds of adsorbed
species.
Reaction-transport processes in heterogeneous reactors involve spatial scales
in the range from 10–9 to 10 m. Here, we are going to survey simulation ap-
proaches employed on the scales of nanometers to millimeters.
1. Monte Carlo and Molecular Dynamics Methods

Owing to the large number of molecules in a typical catalyst particle, the
sampling to a limited region of the configuration space has to be employed in
order to allow computational feasibility. A typical algorithm utilizing the MC
approach is the Metropolis algorithm, which generates a random walk in the
configurational space (e.g., Allen and Tildesley, 1988). Let us suppose that we
have N molecules in a given configuration confined in a cubic box. Then the

following steps are performed iteratively:
(i) a molecule is randomly chosen;

(ii) a trial move is attempted and the molecule is thus displaced and rotated, so
that a new set of coordinates is generated from the old one;
(iii) the change in energy of the system AE due to the trial move is calculated. If
the energy AE has decreased, the move is accepted. If the energy AE has
increased, the move is accepted with the probability exp(AE/kT), where k
is the Boltzmann constant and T is the temperature of the system.
For the computation of dynamic properties in physical systems with a high

degree of freedom, MD is the most widely used approach. MD calculates the
motion of an ensemble of atoms and molecules by integrating Newton’s equa-
tions. From the motion of the ensemble of atoms and molecules, both micro-
scopic and macroscopic information can be extracted, e.g., transport
coefficients, phase diagrams, and structural properties. The physics of the
model is contained in a potential energy function for the system, from which the
force equations for each atom and molecule are derived. The evaluation of the
total force acting on the atom or molecule is the time-consuming step of MD
simulations because it requires the computation of a wide range of expressions
consisting of terms with fractional powers or transcendental functions, and
because the summation of binary force interactions with the neighboring mol-
ecules confined in a spherical cutoff of properly chosen radius has to be carried
out.
Monte Carlo methods allow the simulation of the reaction kinetics on nano-
meter-sized (1–100 nm) metal or metal-oxide crystallites deposited on the sur-
face of a porous support (McLeod and Gladden, 1998). The complexity of such
systems in combination with relatively high reactant pressures used in the in-
dustry hinders the application of results obtained with surface science micros-
copy and spectroscopy methods (Jansen et al., 2001; Wolff et al., 2003). The
complex factors inherent to nanometer chemistry are the reactant supply via the
surface diffusion, interplay of reaction kinetics on different facets of crystallites
due to the inter-facet diffusion, different rates of reactions on different types of
facets, oscillation, and chaos on the nanometer scale (van Neer and Bliek, 1999;
Zhdanov, 2002). MC simulations allow one to investigate the kinetics of chem-
ically induced Ostwald ripening on facets of nanometer catalyst particles. An-
other important effect associated uniquely with small catalyst crystallites on a
support is the spill-over effect, consisting generally of adsorption on the support
followed by diffusion from the support to the catalyst crystallite or vice versa.
The size of the smallest pores in pellets or washcoat varies from 1–2 to X50 nm.
The size of metal crystallites may vary over a very wide range as well. For
example, Pt and Rh particles of car exhaust cleaning catalysts are typically
10–100 nm large after some time of use.
Zhdanov and Kasemo (2000) considered the reaction 2A+B2-2AB occur-

ring via the standard Langmuir–Hinshelwood (LH) mechanism,
Agas Ð Aads (52)
ðB2 Þgas ! 2Bads (53)
Aads þ Bads ! ðABÞgas (54)
This scheme mimics, e.g., CO or H2 oxidation on the noble metal catalysts Pt,
Pd, or Rh, where symbol A stands for CO or H, and B2 for O2. The reaction was
simulated on a 2D lattice of adsorption sites. To compare the rates of diffusion
and reaction, it is useful to employ the Arrhenius form to represent the rate
constants of diffusion jumps of A and B particles to nearest-neighbor vacant
sites and for the reaction between two nearest-neighbor reactants, respectively.
The diffusion of A is usually rapid when compared to the LH step, while the
rate constant for the LH step might be higher, close to, or lower than that for
the diffusion of B2. The MC algorithm used to simulate the A+B2 reaction is as
follows:
1. A site on the catalyst surface is selected at random. If the site is vacant,

adsorption is attempted (item (2)). If the adsorption trial is successful, the
newly adsorbed molecule tries to react (item (3)).
2. A molecule for adsorption is chosen to be A with the probability p and B2
with the probability 1p, where pp1 is the parameter characterizing the
relative impingement rates of A and B2. To realize adsorption of B2, a con-
figuration for B2 is selected at random. An adsorption trial is considered to
be successful provided that all the sites of the selected configuration are
vacant and belong to the catalyst. Adsorption of the chosen molecule A is
realized with the unit probability.
3. For reaction, all the molecules that neighbor the newly adsorbed molecule
are identified and one of them is chosen to react.
The size of the crystallites affects the selectivity of reactions with larger mol-
ecules, e.g., hydrocarbon chains. Small islands of sites inhibit the adsorption of
larger molecules, because all adsorption sites for the selected configuration of
the larger molecule have to be vacant. As a general rule, catalytic reactions are
structure-sensitive on the scale o10 nm (Piccolo et al., 1999).
A lattice-gas model of the NO+CO/Pt(1 0 0) reaction was found by Makeev
and Kevrekidis (2004) to exhibit bistability and kinetic oscillations. In this
simulation, the catalyst surface was represented by a square lattice with periodic
boundary conditions. The model includes both chemical reactions and transi-
tions of adsorbed reactants from the nearest-neighbor sites, which are in the
context of the MC simulation described by transition probabilities. The mac-

roscopic simulations of the reactor or catalyst particle can be coupled with
microscopic simulations over relatively small ‘‘patches’’ of the macroscopic
computational domain, which could thus provide an approximation of reaction
terms in macroscopic equations.
Lombardo and Bell (1991) reviewed stochastic models of the description of
rate processes on the catalyst surface, such as adsorption, diffusion, desorption,
and surface reaction, which make it possible to account for surface structure of
crystallites, spatial inhomogeneities, and local fluctuations of concentrations.
Comparison of dynamic MC and mean-field (effective) description of the prob-
lem of diffusion and reaction in zeolites has been made by Coppens et al. (1999).
Gracia and Wolf (2004) present results of recent MC simulations of CO ox-
idation on Pt-supported catalysts.
2. Reaction in Fractal Porous Media

Fractals are self-similar objects, e.g., Koch curve, Menger sponge, or Devil’s
staircase. The self-similarity of fractal objects is exact at every spatial scale of
their construction (e.g., Avnir, 1989). Mathematically constructed fractal po-
rous media, e.g., the Devil’s staircase, can approximate the structures of metallic
catalysts, which are considered to be disordered compact aggregates composed
of imperfect crystallites with broken faces, steps, and kinks (Mougin et al.,
1996).
In the fractal porous medium, the diffusion is anomalous because the mol-
ecules are considerably hindered in their movements, cf. e.g., Andrade et al.,
1997. For example, Knudsen diffusion depends on the size of the molecule and
on the adsorption fractal dimension of the catalyst surface. One way to study
the anomalous diffusion is the random walk approach (Coppens and Malek,
2003). The mean square displacement of the random walker hR2 i is not pro-
portional to the diffusion time t, but rather scales in an anomalous way:
lim hR2 i t2=Dw (55)

t!1
where the dimension of the walk Dw is generally greater than 2. The anomalous
scaling law Eq. (55) implies that the classical Brownian or Gaussian diffusion
law is no longer valid in fractal media because the fractal substrate induces
large-scale correlations on the molecular movement. Consequently, Fick’s law is
also invalid in its classical form.
Coppens and Froment (1995a, b) employed a fractal pore model of supported
catalyst and derived expressions for the pore tortuosity and accessible pore
surface area. In the domain of mass transport limitation, the fractal catalyst is
more active than a catalyst of smooth uniform pores having similar average
properties. Because the Knudsen diffusivity increases with molecular size and
decreases with molecular mass, the gas diffusivities of individual species in
multicomponent mixtures will mostly be closer to each other than predicted by

the classical expression. It is generally impossible to construct an equivalent,
smooth cylindrical pore that leads to the same concentration and flux profiles as
those obtained for a fractal pore (Coppens and Froment, 1995a, b). Even when
the concentration or pressure profiles in a cylindrical pore are similar to those in
the fractal pores, the amounts reacted or produced are usually predicted er-
roneously by classical Euclidean models. For example, in the fractal pore, up to
7.1 times more vinyl acetate is predicted to be produced on a Pd/Al2O3 catalyst
than in the smooth pore (Coppens and Froment, 1994).
The selectivity and deactivation processes in pore fractals such as the Sier-
pinski gasket were simulated by Gavrilov and Sheintuch (1997) and Sheintuch
(1999). Their studies investigated, e.g., the effect of the fractal pore structure on
the selectivity of a system that incorporates two parallel reactions. Geometrical
factors, which influence dynamic processes in a porous fractal solid media, were
also investigated by Garza-López and Kozak (1999).
3. Reaction-transport Models in Reconstructed Porous Media

After the spatially 2D or 3D model of the porous catalyst support and the
distribution of catalyst are generated, the multicomponent diffusion, adsorp-
tion, and chemical reaction within this porous structure can be modeled.
Rieckmann and Keil (1997) introduced a model of a 3D network of inter-
connected cylindrical pores with predefined distribution of pore radii and con-
nectivity and with a volume fraction of pores equal to the porosity. The pore
size distribution can be estimated from experimental characteristics obtained,
e.g., from nitrogen sorption or mercury porosimetry measurements. Local he-
terogeneities, e.g., spatial variation in the mean pore size, or the non-uniform
distribution of catalytic active centers may be taken into account in pore-
network models. In each individual pore of a cylindrical or general shape, the
spatially 1D reaction-transport model is formulated, and the continuity equa-
tions are formulated at the nodes (i.e., connections of cylindrical capillaries) of
the pore space. The transport in each individual pore is governed by the Max-
well–Stefan multicomponent diffusion and convection model. Any common
type of reaction kinetics taking place at the pore wall can be implemented.
A similar model has been applied to the modeling of porous media with
condensation in the pores. Capillary condensation in the pores of the catalyst in
hydroprocessing reactors operated close to the dew point leads to a decrease of
conversion at the particle center owing to the loss of surface area available for
vapor-phase reaction, and to the liquid-filled pores that contribute less to the
flux of reactants (Wood et al., 2002b). Significant changes in catalyst perform-
ance thus occur when reactions are accompanied by capillary condensation. A
pore-network model incorporates reaction–diffusion processes and the pore
filling by capillary condensation. The multicomponent bulk and Knudsen dif-
fusion of vapors in each pore is represented by the Maxwell–Stefan model.
Moreover, the convection of the vapors in the pores was considered and the
condensation in the pores was described by the multicomponent Kelvin equa-
tion (Shapiro and Stenby, 1977).
Two levels of intraparticle mass transport resistance are considered in various
grain or multigrain models, e.g., diffusion resistance in the macro- and micro-
pores in the particle of the highly porous char during its combustion (Cai and
Zygourakis, 2004) and monomer transport resistance in macro-pores and pol-
ymer micro-grains in the growing polyolefin particle (Debling and Ray, 1995;
Kosek et al., 2001a).
The techniques of meso-scopic modeling were also applied in the development
of the reaction–diffusion model for the hydration and setting of cement
(Tzschichholz et al., 1996). At the initial stage, the cement particles or powder
(e.g., tricalcium silicate, Ca3SiO5) are mixed with water. Rapidly after mixing,
the dissolution and the reaction of the cement particles starts; the principal
reaction products are Ca2+, OH–, and H2 SiO2 4 ions, which diffuse into water.
However, for a given temperature the ion concentrations are bounded by finite
solubility products above which solid phases start to precipitate from the so-
lution. There are two precipitation reactions: (i) of calcium hydro-silicate (the
so-called ‘‘cement gel’’), and (ii) of calcium hydroxide or ‘‘Portlandite.’’ While
the growth of the cement gel is the basis for the whole cement binding process,
the growth of Portlandite mainly happens in order to compensate for the ac-
cumulation of Ca2+ and OH– ions in the solution. Tzschichholz et al. (1996)
proposed a simple stochastic cellular automation model for cement-hydration
setting to simulate the spatial and temporal evolution of precipitated micro-
structures starting from initial random configurations of anhydrous cement
particles. The model considers the spatial distribution of solid and liquid phases.
The three main chemical reactions (dissolution, precipitation of cement gel, and
Portlandite) are coupled with the transport of ions owing to the diffusion.
IV. Applications
A. MULTI-SCALE RECONSTRUCTION OF A CATALYST PELLET
Experimentally determined effective transport properties of porous bodies,

e.g., effective diffusivity and permeability, can be compared with the respective
effective transport properties of reconstructed porous media. Such a compar-
ison was found to be satisfactory in the case of sandstones or other materials
with relatively narrow pore size distribution (Békri et al., 1995; Liang et al.,
2000b; Yeong and Torquato, 1998b). Critical verification studies of effective
transport properties estimated by the concept of reconstructed porous media
for porous catalysts with a broad pore size distribution and similar materials
are scarce (Mourzenko et al., 2001). Let us employ the sample of the porous
alumina with bimodal pore size distribution for such a comparative study; cf.
the intrusion mercury porosimetry data of the sample in Fig. 14.
The effective diffusivity and permeability of the porous alumina sample G1
were measured in the Graham diffusion cell and in the permeation cell (Salejova
et al., 2004), cf. Table 1. The porosity e is separated into macro-porosity emacro
corresponding to large pores and nano-porosity enano corresponding to small
pores, and ¼ macro þ nano . The boundary in the classification between macro-
and nano-pores is somewhat arbitrary selected as the inflection point on the
integral mercury porosimetry curve in Fig. 14.
The pellets of porous alumina sample G1 were fractured by the application of
the mechanical force, and the bare granular surface of the fracture was exam-
ined by SEM; cf. Fig. 15 displaying the granular morphology on the level of
hundred micrometers and showing the sub-micrometer morphology. The grain
size distribution was evaluated from several SEM images (with the scaling bar
100 mm) as a necessary input for the reconstruction of the spatially 3D porous
0.40
46 nm Sample G1
0.35 0.8
0.30
0.6
Vexp /(cm3/g)
0.25
dV/d log(r)
0.20 2330 nm
0.4
0.15
0.10 0.2
0.05
0.00 0.0
0.001 0.01 0.1 1 10 100
r / µm
FIG. 14. Integral and differential intrusion mercury porosimetry data for sample G1 (from Salej-
ova et al., 2004).
TABLE 1
TEXTURAL ANALYSIS AND EFFECTIVE TRANSPORT PROPERTIES OF SAMPLE G1
Property Value
Porosity e 0.580
Macro-porosity emacro 0.235
Nano-porosity enano 0.345
Graham cell results, c ¼ =t 0.199
Permeation cell results, hric 277 nm
Permeation cell results, hr2 ic 421,000 nm2
Permeation cell results ðhricÞ2 , c ¼ ðhric2 Þ=ðhr2 icÞ 0.182
reconstructed
SEM image, granular structure Grain size distribution
60 small granular structure
50 grains large grains
40 Sample G1
30 200 µm
20
Frequency
1µm 10
0
0 10 20 30 40 50 60 macro = 0.235
Equivalent Grain Diameter /µ m
SEM image, nano-sizedpores
Statistical characteristics,
for example:
• Autocorrelation fucntion.
• Lineal path distribution. 1µm
• Distribution of coverage reconstructed
1µm
radii. nano-porous
structure
nano = 0.345
FIG. 15. Reconstruction of multiscale porous media, sample G1 (from Salejova et al., 2004).
structure. The equivalent grain diameter is the diameter of the circle with the
same area as that of the grain identified in the image.
The reconstruction algorithm of the granular media places spheres with di-
ameters corresponding to the grain size distribution (from Fig. 15) into a rec-
tangular box, where two pairs of opposite walls are subjected to periodic
boundary conditions. The aspect ratio of the height to the width of the rec-
tangular box is 10:1. Because the required target value of porosity of granular
media is macro ¼ 0:235, the individual grains are allowed to partially overlap.
The newly generated sphere is inserted at the lowest position z2 in the z-direc-
tion, satisfying the equation
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðr1 þ r2 Þð1 poverlap Þ ¼ ðx1 x2 Þ2 þ ðy1 y2 Þ2 þ ðz1 z2 Þ2 (56)
where ðx1 ; y1 ; z1 Þ is the coordinate of the center of already placed sphere,

ðx2 ; y2 ; z2 Þ the coordinate of the center of the newly inserted sphere, and r1 and r2
the sphere radii of the placed and newly inserted sphere, respectively. The re-
constructed sample is thus the list of ðx; y; z; rÞ coordinates of sphere centers and
radii. Once the tall rectangular box is filled with spherical grains, its lower half is
removed to avoid the effect of regular configuration of spherical grains at the
bottom. The cube is selected from the remaining upper part of the rectangular
box and is discretized into 101 101 101 or 201 201 201 voxels.
The pore structure of the investigated sample G1 on the sub-micron level
(with the porosity nano ¼ 0:345) is approximated by the Gaussian-correlated
random field; cf. Fig. 15. The correlation length Ly in Eq. (18) was selected as
the modal value of the mercury porosimetry, Ly ¼ 46 nm; cf. Fig. 14.
The ratio of effective and bulk diffusivity Deff/D is calculated in two steps.
First, the effective diffusivity of the nano-porous material with the porosity
nano ¼ 0:345 is calculated by the algorithm introduced in Section II.E.1. The
effective diffusivity of the nano-porous cube with the size 1 mm and porosity
nano ¼ 0:345 displayed in Fig. 15 is cnano ¼ Deff =D ¼ ð0:112 0:004Þ.
Then, the effective diffusivity of the macro-porous granular material is eval-
uated. The transport is again governed by the Fick’s equation (23), but diffusion
also takes place in the ‘‘solid phase,’’ which formally represents the nano-porous
material, cf. Fig. 15. The diffusion coefficient in the ‘‘solid phase’’ is
Dsolid ¼ cnano D ¼ 0:112D, where D is the bulk diffusivity. The concentration
field in the macro-porous media is the solution of Eq. (23) with a diffusivity
(
1 in macro-porous phase
D¼ (57)
0:112ð¼ cnano Þ in solid phase
The boundary condition of zero accumulation on the interface between macro-

pores and ‘‘solid phase’’ is imposed. The effective diffusivity of the porous
sample G1 with bimodal pore size distribution is summarized in Fig. 16, where
the sample macro-porosity emacro is varied on the horizontal axis. This effective
diffusivity is compared with a situation where the diffusion transport in nano-
pores is omitted. The contribution of the transport through the nano-porous
solid phase to the total diffusion flux is significant. The calculated effective
0.35 sample G1
0.30 = Deff/D for diffusion

in both macro- and
0.25
nano-pores
= Deff/ D
0.20 experimental
value
0.15
0.10 =Deff/D for diffusion

only in macro-pores
0.05
0.00
0.10 0.15 0.20 0.25 0.30 0.35 0.40
macro-porosity macro
FIG. 16. Effective diffusivity of the examined porous sample G1 calculated for: (i) Fick’s diffusion
both in macro- and nano-pores, and (ii) Fick’s diffusion in macro-pores only. The experimental
value of c ¼ Deff =D ¼ 0:199 was determined in Grahams diffusion cell. Only grains larger than
10 mm were used in the reconstruction of macro-porous media (from Salejova et al., 2004).
diffusivity is approximately about 10% higher than that determined experi-

mentally in the Graham diffusion cell (Salejova et al., 2004).
The pressure-gradient-driven permeation through nano-pores is considered to
be negligible owing to their small characteristic diameter. Therefore, the Darcy’s
permeability k0 was calculated only in the macro-porous reconstructed media.
The predicted Darcy’s permeability k0 is 3 times larger than the experimental
value measured in the permeation cell. This discrepancy is caused by the dif-
ficulty of determining the fraction of small grains (o10 mm) in the real sample,
the assumption of perfectly spherical grains, and by neglecting the effect of the
clogged pores (Salejova et al., 2004).
B. RECONSTRUCTION OF CLOSED-CELL POLYMER FOAM STRUCTURE
This work was motivated by the need to quantify the pentane vapor diffusion
through closed-cell polystyrene foam (Salejova et al., 2005). Once the polysty-
rene foam is formed, the pentane as the blowing agent has to diffuse out from
the cellular microstructure and is replaced by air. Freshly foamed polystyrene is
not dimensionally stable and it shrinks as it relaxes on the timescale of days to
weeks. Environmental concerns call for the reduced consumption of pentane as
the blowing agent.
Three-dimensional foam structures can be generated by Voronoi tessellation
of sites located on random or regular (e.g., cubic or hexagonal) lattices (Bowyer,
1981). The Voronoi polyhedron is the region of space around a site such that
each point of this region is closer to the site than to any other point of the
system. A simulated annealing algorithm in connection with Voronoi tessella-
tion can be used to produce a replica of the foam with cell-size distribution
corresponding to that of an experimentally processed image of the foam struc-
ture. The foam structure can then be reduced to an equivalent network of
conductors on which the steady-state diffusion problem is solved and the ef-
fective diffusivity determined.
The 3D structure of a foam sample can be obtained by X-ray micro-tomo-
graphy (Elmoutaouakkil et al., 2002) or it can be reconstructed from a series of
2D images (Garcia-Gonzales et al., 1999). The former technique has been used
recently for the structural characterization of PVC foams by means of cell
volume distribution, cell coordination number, wall thickness distribution, and
total porosity (Elmoutaouakkil et al., 2003). It is also possible to generate foam
structures computationally as 3D Voronoi tessellations (Lee, 1999). This was
done, e.g., by Kraynik et al. (2003), who studied the structure of foams gen-
erated by random close packing of monodisperse spheres. Finally, foam struc-
tures can be generated by computer simulation of the foaming process. Models
of the key steps of bubble nucleation and coarsening (Hilgenfeldt et al.,
2001a,b), bubble growth (Koopmans et al., 2000), and foam expansion (Joshi
et al., 1998; Körner et al., 2002) have been implemented.
The Voronoi tessellation has been constructed by the following algorithm

(Lee, 1999), and its purpose is to generate the 3D structure of high-porosity
foams in a computational unit cell consisting of N x N y N z voxels. The
algorithm starts from Np sites distributed in the unit cell either randomly or at
random perturbations of regular lattice positions for two types of lattices—a
cubic lattice and a hexagonal lattice. Then, a distance map (‘‘watershed’’) is
generated by finding the Euclidean distance from each voxel in the unit cell to
the nearest site. A Voronoi region around site j is defined as a set Vj of points i
such that their Euclidean distance from site j is less than from any other site
k 6¼ j, i.e.,
V j ¼ fi : jjxj xi jjojjxk xi jj 8kajg (58)
The Voronoi regions correspond to cells in a computer-generated foam struc-

ture. Examples of cellular structures generated by Voronoi tessellation are
shown in Fig. 17.
The volume of cells is calculated as the number of voxels i belonging to each
Voronoi region Vj, times the voxel volume h3, where h is voxel size. The symbol
Hexagonal
lattice
Random
lattice
FIG. 17. Computer generated structure of hexagonal and random closed-cell foam obtained by
Voronoi tessellation, shown as voxel representation of phase function (left), and network diagram
where nodes correspond to cells and bonds to cell walls (from Salejova et al., 2005).
Vj will hereafter also be used to denote the volume of the Voronoi region Vj. A
wall dividing two adjacent Voronoi regions is a set of points having an equal
distance to the two Voronoi sites. In the discretized voxel space, however, we
define a wall Ajk as a set of voxels, which belong to Voronoi region Vj and have
at least one out of six nearest neighbors in region Vk. The area of the wall
cannot be calculated simply as the sum of points in Ajk times h2 because a
significant discretization error would be introduced. The area per voxel must be
corrected by taking into account the orientation of the wall (the voxel will
account for an area equal to h2 only if the wall is perpendicular to one of the
principal coordinate axes). Let n be the number of nearest neighbors of a voxel
in a wall Ajk that belong to the adjacent Voronoi region Vk. The maximum
number of such nearest neighbors is 3. p The
ffiffiffi contribution ofpeach
ffiffiffi voxel to the
total area is weighted by h2 if n ¼ 1, h2 2 if n ¼ 2, and h2 3=2 if n ¼ 3. The
weights are derived from the PLIC (piecewise linear interface construction)
scheme often used in simulations of free interface flows by the VOF method
(Rider and Kothe, 1998). The symbol Ajk will hereafter also be used to denote
the area of the wall Ajk.
Once the Voronoi tessellation is completed and the areas of the walls dividing
adjacent cells in the foam structure are known, the equivalent conductor net-
work can be constructed. The network is a Delaunay triangulation dual to the
Voronoi diagram (examples of networks are shown in Fig. 17). Its vortices are
the Np sites used for the construction of the Voronoi diagram and its edges
(bonds) link sites that are adjacent in the Voronoi diagram. The structure of the
network can be encoded by the connectivity matrix R, which is an Np Np
symmetric matrix whose elements are rjk ¼ rkj ¼ Ds Ajk =d w if sites j and k are
adjacent, and rjk ¼ rkj ¼ 0 otherwise (this includes diagonal elements). The
conductivity of the bond
Ajk
rjk ¼ Ds (59)
dw
follows from steady-state diffusion across a planar wall of area Ajk, thickness
dw, and diffusion coefficient Ds, assuming Fick’s law. The wall thickness dw is
assumed uniform for all walls within the foam. For high-porosity foams, the
mean wall thickness dw and foam porosity e can be related by the expression
a 1 Atot
¼ 1 dw ¼ 1 dw (60)
2 2 V tot
where a ¼ Atot =V tot is the internal surface area of the foam per unit volume (the
factor 12 must be included because each wall contributes twice to the internal
surfaceParea). On the computational unit cell, V tot ¼ N x N y N z h3 and
N p PN p
Atot ¼ j¼1 k¼1 Ajk .
C. POLYMER PARTICLE MORPHOGENESIS
The growth of porous polyolefin particles from supported catalysts can be

divided into two main stages (Estenoz and Chiovetta, 2001). In the first stage, the
porous-catalyst-support particle with diameter E20 mm is gradually destroyed
into fragments of 10–50 nm size owing to the production of polymer in the
confined pore space. Porous silica or MgCl2 often serves as a catalyst carrier. The
outcome of the first stage is the set of fragments embedded in the continuous but
porous matrix of polymer phase (Kakugo et al., 1989). At the beginning of the
second stage of particle growth, most of the catalyst carrier has already been
fragmented, and the growth and morphogenesis of porous polyolefin particles
then continues owing to the polymerization at active catalyst sites immobilized
typically at the surface of fragments of catalyst carrier (Debling and Ray, 2001).
The typical final size of porous polyolefin particles is 0.5–2 mm.
Let us first consider the catalyst/polyolefin particle in the early stage of its
evolution. The particle consists of the solid catalyst carrier with catalyst sites
immobilized on its surface, polymer phase, and pores. The first-principle-based
meso-scopic model of particle evolution has to be capable of describing the
formation of polymer at catalyst sites, transport of monomer(s) and other re-
actants/diluents through the polymer and pore space, and deformation of the
polymer and catalyst carrier (including its fragmentation). Similar discrete el-
ement modeling techniques have been applied previously to different problems
(Heyes et al., 2004; Mikami et al., 1998; Tsuji et al., 1993).
Let us discretize the catalyst carrier, the polymer phase, and the void phase into
a number of small, spherical micro-elements of several types with elastic or visco-
elastic interactions acting among individual micro-elements, and with transport
processes taking place along the connections of neighboring micro-elements; cf.
Fig. 18. The catalyst carrier is thus represented by microelements C, but micro-
elements with catalyst sites immobilized on their surface are denoted by C. All
catalyst sites are assumed to have constant activity for demonstration purposes in
this chapter. One growing polymer micro-element P is attached to each ‘‘cat-
alyst’’ micro-element C. When the radius of the growing polymer micro-element
P reaches a prescribed value, it becomes the non-growing polymer micro-
element P, and a new microelement of type P is formed at catalyst center C.
The pores are discretized into void-phase micro-elements V in order to describe
the transport of reactants/diluents in pores. For the sake of simplicity, we con-
sider only the transport and reaction of monomer and not of other reactants.
The ith micro-element is characterized by the position of its center xi, velocity
vi, radius ri, monomer concentration ci, and by its type ti. The translational
movement of each micro-element is governed by kinematic equation and
Newton’s equation of momentum
dxi
¼ vi ; ti ¼ C; C
; P; P
; V (61)
dt
Pore fracturing
catalyst site
growing -element of polymer

force
transport of monomer
polymer phase
µ-elements of
catalyst support
Pore clogginng
FIG. 18. Model of the fragmentation of catalyst support based on force interactions and transport
among micro-elements of five types: (i) catalyst support, (ii) catalyst support with active catalyst site,
(iii) growing polymer micro-element, (iv) non-growing polymer microelement, and (v) void-space
transport micro-element (from Grof et al., 2005c).
dðmi vi Þ X binary X ternary

¼ F ij þ F ijk ; ti ¼ C; C
; P; P
dt j j;k
vi ¼ 0; ti ¼ V ð62Þ
where mi ¼ 43pri r3i is the mass of the ith micro-element, and the density of
each micro-element ri is considered to be constant in time. Summations of
forces F on the right-hand side of Eq. (62) are carried out over all binary
and ternary interactions between the ith micro-element and its connected
neighbors. The positions of pore-phase micro-elements V are fixed during the
simulation.
The balance of the monomer in the ith micro-element is
d X
ðci V i Þ ¼ N ij S i ; ti ¼ P; P
; V
dt j
ci ¼ 0; ti ¼ C; C
ð63Þ
where V i ¼ 43pr3i is the volume of the ith micro-element, Nij the molar flux of
monomer from the jth connected micro-element, and Si the rate of monomer
consumption by the reaction. No monomer is assumed to be present inside the
catalyst-support micro-elements C and C.
The rate of the growth of the ith micro-element of type P depends on the
monomer concentration in this micro-element
dmi dri
¼ 4pri r2i ¼ kmcat;j ci ; ti ¼ P
dt dt
dri
¼ 0; ti ¼ C; C
; P; V ð64Þ
dt
where mcat,j is the mass of catalyst in the micro-element C
j to which the ith
micro-element P
i is attached. Let AY be the polymer yield rate at reactor bulk
temperature in kgpol/gcath. Then the rate constant k ¼ AY /(3.6ceq), where ceq ¼
Hcbulk is the concentration of monomer in micro-elements P, which is in equi-
librium with the bulk concentration of monomer in the reactor cbulk; and H is
the equilibrium constant. The rate of monomer consumption in Eq. (63) is
evaluated as
(
ðkmcat;j ci Þ=M M ; ti ¼ P
Si ¼ (65)
0 ti ¼ P; V
where MM is the molecular weight of the monomer.

The mathematical model of catalyst/polymer particle evolution consists of the
set of differential-algebraic equations (61)–(64). The constitutive equations de-
scribing the force interactions, transport of monomer, phase equilibria at the
interface between polymer and pore phase as well as the rules for connectivity of
micro-elements have to be specified (Grof and Kosek, 2005; Grof et al., 2005a).
1. Force Interactions between Micro-Elements

Magnitudes of binary and ternary force interactions required in Eq. (62) are
calculated by simple elastic or visco-elastic constitutive equations and then
projected into force vectors. Elastic model is the direct implementation of the
Hook’s law with the stress between micro-elements A and B dependent linearly
on the strain eAB:
F AB =AAB ¼ E t eAB (66)
where Et is the elastic modulus dependent on the type of connected microel-

ements, AAB the contact area, and FAB the binary interaction force. The positive
value of FAB results in the attraction and the negative one in the repulsion of
micro-elements. The strain eAB is defined as
eAB ¼ ðjuj u0 Þ=u0 (67)
where u0 ¼ rA þ rB is the equilibrium distance and u ¼ xB xA the vector con-

necting micro-elements A and B.
Ternary interactions are considered in the case of the deformation of the

triplet of the connected micro-elements A–V–B. If the translation of microel-
ements leads to the change of the angle a ¼ ffAVB, then the force with the
magnitude Fa acts on micro-elements as the resistance against this change. The
magnitude of the force Fa is calculated in the limit of elastic interactions as
F a =AAB ¼ G t ðcos a cos a0 Þ (68)
where Gt is the bending modulus specific to the type of connected microele-

ments, and a0 the original value of the angle ffAVB at the start of the simulation
(or at the time when the triplet A–V–B was formed). The contact area AAB is,
for the sake of simplicity, selected for both the binary and ternary interactions
as AAB ¼ 12pðr2A þ r2B Þ, although for the ternary interaction, the micro-elements
A and B do not have to be in the direct contact.
The visco-elastic interactions of connected micro-elements of semicrystalline
polyolefins are approximated by the Maxwell model, which formally consists of
the spring with modulus Et and the viscous dashpot characterized by the vis-
cosity Zt connected in series,
ds deAB E t s
¼ Et t (69)
dt dt Z
where s ¼ F AB =AAB is the stress. The Maxwell model describes particularly well
the relaxation of stress in a visco-elastic material. When the connection of micro-
elements A and B is subjected to constant strain (deAB =dt ¼ 0), then the stress in
the connection decays exponentially, i.e., s ¼ s0 expðE t t=Zt Þ ¼ s0 expðt=ttR Þ,
where s0 is the initial stress at t ¼ 0 and rtR ¼ Zt =E t the relaxation time. The
Maxwell model also allows for permanent deformation.
The magnitudes of force interactions FAB and Fa defined by constitutive
Equations (66), (68), and (69) have to be projected into vectors required in Eq.
(62); cf. Grof et al. (2005a).
Force interactions among micro-elements were calculated in our simulations
according to the following rules:
F AB ¼ 0 and F a ¼ 0, if at least one of the connected micro-elements is of the

type V (i.e., void phase).
The Maxwell model Eq. (69) is employed to calculate the magnitude of the
binary force interaction if both micro-elements are of the type P or P.
Elastic model given by Eq. (66) with different moduli for the attractive and for
the repulsive forces is used to calculate the magnitude of the binary inter-
action FAB in the remaining cases, i.e., in the cases involving at least one
catalyst micro-element C or C.
All ternary interactions involving micro-elements of type C, C, P, and P are
considered to be elastic and are calculated according to Eq. (68).
2. Connectivity of Micro-Elements
For the purpose of calculation of force interactions, the computational al-
gorithm keeps the list of connected pairs of neighboring micro-elements and the
list of connected triplets of micro-elements. These lists are updated, i.e., new
connections are created or existing connections are removed at each step of
dynamic simulation. Let us describe the rules for the creation and deletion of
bonds of micro-elements.
Micro-elements A and B become connected when they touch each other, i.e.,
when their distance u ¼ juj becomes
jujpu0 ¼ rA þ rB (70)
Two elements become disconnected if the strain between them exceeds max-
imum value etmax
juj u0
eAB ¼ 4etmax (71)
u0
The value of the maximum strain etmax depends on the type of micro-elements A
and B. Thus the superscript ‘‘t’’ in the previous equation can be (a) ‘‘PP,’’ to
denote the connection between two polymer micro-elements P or P; (b) ‘‘CC,’’
to denote the connection between two catalyst micro-elements C or C; and,
finally, (c) ‘‘PC,’’ to denote the connection between polymer and catalyst micro-
elements.
The triplet A–V–B is initiated when both connections A–V and B–V are
formed. For the sake of simplicity and in order to reduce the number of triplets
that have to be processed, the connections A–V and B–V have to be of the same
type, i.e., of the type ‘‘PP,’’ ‘‘CC,’’ or ‘‘PC.’’ The triplet is deleted if (i) one of the
connections A–V or B–V is disconnected, or (ii) the deviation between the actual
angle a and the initial angle a0 exceeds the limiting value atmax , i.e. ja a0 j4atmax
(and in this case also, one of the binary connections A–V and B–V is discon-
nected).
Once the catalyst breaks, it remains broken even if the fragmented parts come
into contact later during the simulation. On the other hand, the polymer micro-
elements P or P are assumed to be ‘‘sticky,’’ and connections between them
created during the simulation are assumed to be as strong as the connections at
the beginning of the simulation.
3. Diffusion Transport of the Monomer

The molar flux of monomer Nij from the jth into the ith micro-element is
described as
Aij
N ij ¼ Dtij ðcj ci Þ (72)
dij
where the area is Aij ¼ 12pðr2i þ r2j Þ and dij ¼ jxi xj j is the distance between the
micro-elements. Diffusion coefficients in the void phase DV and in the polymer
phase DP are different; hence, the diffusivity Dtij depends on the type of con-
nected micro-element
8
>
> 0 ti ¼ C; C
or tj ¼ C; C
>
>
< DP ti ¼ P; P
or tj ¼ P; P
Dtij ¼ (73)
>
> DV ti ¼ V and tj ¼ V
>
>
: DPV ; ðti ¼ V and tj ¼ P; P
Þ or vice versa
No monomer transport through the solid catalyst support is considered. The

diffusivity DPV in Eq. (73) is only a formal notation because the flux between the
void phase V and the polymer P (or P) micro-elements is calculated from the
assumption of no accumulation and sorption equilibrium at the phase interface.
The void-phase micro-elements V are used to model the transport of mon-
omer through the regions where neither polymer nor catalyst micro-elements are
present. The micro-elements V do not change their position or radius during the
dynamic simulation, but they are deleted when the catalyst or the polymer
micro-elements move over them and are inserted as needed to cover any newly
formed void phase. The transport connections between the micro-elements are
different from the connections with the force interactions described in Sections
IV.C.1 and IV.C.2, and are also regularly updated during the dynamic simu-
lation.
4. Evolution of Particle Morphology

The simultaneous evolution of the catalyst support, the polymer phase, and
the concentration field of the monomer is displayed in the sequence of images in
Fig. 19. The catalyst carrier is only partially covered by the catalyst. The ev-
olution of micro-elements is driven by the force interactions and by the growth
of polymer micro-elements P, and is coupled with the transport of monomer
through the dynamically evolving void and polymer phases. The produced pol-
ymer phase contains a number of cavities because the simulation is spatially 2D;
hence, the coordination number of polymer microelements is lower than in 3D
simulations. The initial and the final configurations of the simple catalyst/pol-
ymer particle is shown in Fig. 20. The discrete element modeling of catalyst
particle fragmentation allows one to address a number of practical problems of
catalytic polymerization of olefins, e.g., the dependence of the incorporation of
a co-monomer on the pore space characteristics of catalyst support, the de-
pendence of fragmentation and particle growth on the impregnation of the
porous support by the catalyst system, the discrimination between the ‘‘shrink-
ing core’’ and the ‘‘gradual bi-section’’ fragmentation modes of catalyst sup-
port, etc. (Kittilsen and McKenna, 2001; Knoke et al., 2003; Pater et al., 2003).
Concentration field of monomer
polymer
catalyst
Catalyst and polymer micro-elements

fragmentation separated
catalyst fragments
FIG. 19. Fragmentation of a simple catalyst carrier during early stages of catalytic polymerization
of olefins. The upper row of pictures represents the concentration field of monomer (red ¼ high,
blue ¼ low concentration), the lower row of pictures represents the evolution of particle morphology
(from Grof et al., 2005c).
initial
configuration
polymer
catalyst final configuration
FIG. 20. Detail of the initial and the final configuration of the growing catalyst/polymer particle
from Fig. 19. The polymer phase in the final configuration contains large number of voids because of
the smaller number of connections in spatially 2D simulation when compared to spatially 3D case
(from Grof et al., 2005c).
Alternatively, we can consider the model of the growing particle consisting

only of a single type of the growing micro-element (Grof et al., 2005a); cf. Fig.
21. In this case, the dynamic simulation starts from a hypothetical initial con-
dition just after the fragmentation of the catalyst carrier, and thus describes the
second stage of particle growth. The predictive capabilities of this model were
demonstrated in important problems of the formation of fine particles and the
particles with macro-cavities in dependence on reaction conditions. We have
Ma
xw
ell
mo
de
l
el
od
m
tic
as
El
Growth of
micro-elements
Ternary interactions Formation of cracks...
...and macro-cavities
FIG. 21. Morphogenesis model of porous polyolefin particle based on the elastic and visco-elastic
interactions among micro-elements (from Grof et al., 2005a).
also demonstrated the non-perfect replication of particle shape during its

growth even in the case of an initially uniform distribution of catalyst activity in
the particle (Grof et al., 2005b). The importance of the distribution of catalyst in
individual micro-elements just after the fragmentation was stressed, and its
effect on the final morphology of polyolefin particle was demonstrated. Thus, a
mapping from the distribution of catalyst in fragments as a result of the first
stage of particle evolution into the final morphology of polyolefin particle was
demonstrated, and the importance of the first ‘‘fragmentation’’ stage of particle
evolution on its final morphology observed experimentally has been confirmed.
The effect of the relaxation time tR in the Maxwell model Eq. (69) on the
predicted morphology of polyolefin particle is illustrated in Fig. 22. In this
figure, the simulation started from a uniform distribution of micro-elements,
and the diffusion resistance in the growing particle was considered.
D. GRANULATION AND DISSOLUTION
Many chemical products, such as pigments, fertilizers, detergents, pharma-

ceuticals, or food and beverage products, are in the form of granules, i.e.,
agglomerates of primary particles held together by a binder. The microstructure
of the agglomerate, i.e., the distribution of individual components, binder, and
porosity within the granule, has a significant influence on its end-use properties,
such as rate of dissolution. The granule design problem consists of finding a
Tough / rigid
polymer
R = ∞ Soft / ductile
(elastic) polymer
R = 1s R = 0.01s
R = 10s R = 0.1s
Compact
Fines particle
FIG. 22. Evolution of particle morphology in dependence on the relaxation time tR of the Max-
well model. Lower (higher) polymerization temperature corresponds to longer (shorter) relaxation
time tR. The initial condition of the simulation was a perfectly regular circle of uniformly distributed
micro-elements. The growth of micro-elements in the particle center is slower because of monomer
diffusion limitation (from Grof et al., 2005a).
microstructure that leads to a specified dissolution rate for a given formulation

(i.e., chemical composition of the granule), and establishing processing routes
and conditions that will lead to the formation of that microstructure.
The problem of granule microstructure formation in the specific case of low-
shear, wet-granulation processes (e.g., fluid bed, drum, or pan granulation) has
been considered by Stepanek and Ansari (2005). They have described a method
for constructing ‘‘virtual granules’’ by explicit simulation of primary particle
packing, binder spreading, and binder solidification. A single virtual granule is
built in a unit cell by sequential deposition of solid primary particles and liquid-
binder droplets, as illustrated in Fig. 23. A range of granule microstructures can
then be generated by varying the primary particle properties (size distribution
and morphology), the binder/solids ratio, the binder spreading rate (through
viscosity and dynamic contact angle), and the binder solidification rate. Ste-
panek and Ansari (2005) have systematically investigated the effect of these
parameters on granule microstructure, and generated structure maps relating
granule porosity to the characteristic times of the deposition, spreading, and
solidification processes.
Once the virtual granules have been created, it is of interest to relate their
microstructure to the dissolution rate and release profile of an active compo-
nent. Recently, Stepanek (2004) carried out detailed simulations of virtual
granules. The problem of porous media dissolution is encountered not only
during the application of pharmaceutical or detergent granules and tablets but
also in industrial processes such as leaching (Erlebacher et al., 2001) or extrac-
tion (Adrover et al., 2004), as well as in other engineering applications such as
t' = 0 t' = 16.8
t' = 28.3 t' = 41.0
t' = 53.2 t' = 76.2
FIG. 23. Sequential deposition of primary solid particles and liquid binder droplets leading to the
formation of a virtual granule (from Stepanek and Ansari, 2005).
oil recovery (Békri et al., 1995). Dissolution can be formally regarded as a

morphology transformation operation, i.e., an operation affecting the phase
functions fi(r) defined in Section II.A. The morphology transformation is due to
the dissolution of soluble phases at the solid–liquid interface. The dissolution
can be purely physical or accompanied by a heterogeneous or homogeneous
chemical reaction, coupled with diffusion-only or convection–diffusion trans-
port of the dissolved species. (Combustion can be regarded as a gas-phase
analogy of dissolution and will not be discussed here separately.)
FIG. 24. Velocity (left) and concentration (right) profiles during the simulation of granule dis-
solution (from Stepanek, 2004).
The simulation of the dissolution of a multicomponent porous medium (in-

cluding virtual granules) can be summarized into the following steps:
1. Start from an initial structure—defined by phase functions fi(r, t ¼ 0).

2. Solve a convection–diffusion problem in the fluid phase surrounding (gran-
ules) or penetrating (porous media) the object of interest.
3. Update phase functions according to local dissolution rates.
4. Check the connectivity of the solid skeleton—remove disconnected solid
phase if present.
5. Repeat from Step 2 until the end time is reached.
This methodology has been used by Békri et al. (1995) for the simulation of
porous media dissolution, and recently by Stepanek (2004) for the studying of
the dissolution and breakup of agglomerates; cf. Fig. 24. He has constructed
structure-dissolution maps for two-component porous agglomerates and ob-
tained correlations for the dissolution half-time t50 as function of granule com-
position (binder/solids ratio, porosity), ingredient properties (solubility and
diffusivity of binder and primary solids), and hydrodynamic regime (dissolution
under stagnant and flow conditions).
E. SIMULATION OF CO OXIDATION ON RECONSTRUCTED CATALYTIC WASHCOAT
The effective-scale models have been most often used in the description of
transport and reaction processes within the porous structure of catalysts. Such
models are based on the introduction of an effective diffusion coefficient De,
that is used in the analogy to the Fick’s law for the description of diffusion
fluxes in the porous medium (Satterfield, 1980; Schmidt, 1998). The classical
approach to studies of diffusion and reaction in porous catalysts uses the mod-
els of cylindrical pores. Recent re-examination of the pore diffusion problem
was made by Balakotaiah and Gupta (2000). Discussion of recent approaches to
mesoscopic modeling of diffusion–reaction through a microporous medium is
contained in the recent paper by Chatterjee et al. (2004). Microkinetic modeling
based on mean field theory has been reviewed by Dumesic et al. (1992). Post-
treatments of zeolite-supported catalysts in order to modify their activity and
stability were discussed by Abdullah et al. (2004) and illustrated on TEM im-
ages of actual micropore structures.
The catalytic washcoat of the monolith converters of automotive exhaust gas
usually consists of partly sintered g-Al2O3 particles of different sizes that to-
gether form a (macro-)porous layer. The g-Al2O3 particles themselves are also
porous, with pore sizes belonging to the range of meso-pores (diameter
10 nm). Crystallites of noble metals (typically Pt) are dispersed on the g-Al2O3
support as active catalytic components. Other components, e.g., CeO2-ZrO2, are
added to the washcoat for the stabilization of the porous structure; they may
also act as active catalytic centers (Koci et al., 2004a). Catalytic reactions take
place on the Pt sites located in the meso-pores of g-Al2O3 simultaneously with
the transport of gaseous reaction components. The transport inside the meso-
porous g-Al2O3 particles is characterized by the effective diffusivity based on
Knudsen diffusion, while the transport in the macro-pores can be described by
volume diffusion (Koci et al., 2004b).
SEM and TEM images give detailed information about the porous structure
of a supported heterogeneous catalyst (pore size distribution, typical sizes of the
particles, etc.). The information from SEM and TEM images can be used in the
reconstruction of porous catalytic medium. In the digitally reconstructed cat-
alyst, transport (diffusion, permeation), adsorption, reaction, and combined
reaction–diffusion processes can be simulated (Stepanek et al., 2001a). Para-
metric studies can be performed, and the resulting dependencies can serve as a
feedback for the catalyst development.
The macro-porosity emacro and the correlation function corresponding to the
macro-pore size distribution of the washcoat were evaluated from the SEM
images of a typical three-way catalytic monolith, cf. Fig. 25. The reconstructed
medium is represented by a 3D matrix and exhibits the same porosity and
correlation function (distribution of macro-pores) as the original porous cat-
alyst. It contains the information about the phase at each discretization point—
either gas (macro-pore) or solid (meso-porous Pt/g-Al2O3 particle). In the first
approximation, no difference is made between g-Al2O3 and CeO2 support, and
the catalytic sites of only one type (Pt) are considered with uniform distribution.
In the following example, we use a simple microkinetic model of CO oxi-
dation on Pt together with the reconstructed porous catalyst to follow the
evolution of local concentration profiles within the porous structure. The re-
action–diffusion problem of the CO oxidation on the Pt/g-Al2O3 porous catalyst
FIG. 25. Typical SEM image of porous Pt/Rh/CeO2/g-Al2O3 catalytic washcoat (left) and the
corresponding binarized image (only the macro-pores visible; right). In the SEM image, the black
color corresponds to the macro-pores, the grey color corresponds to the meso-porous g-Al2O3 with
dispersed Pt, and the white color corresponds to CeO2. Length of the section edge is 10 mm. The
average macro-porosity evaluated from more SEM images is approx. macro ¼ 0:15 (from Koci et al.,
2004).
can be described by the following partial differential equations (spatially 3D

distributed problem):
@ck X
¼ Deff 2
k r ck þ nkj tj ; k ¼ CO; O2 ; CO
; O
(74)
@t j
Equation (74) were solved within the section of the reconstructed porous cat-
alyst, represented by the 3D matrix. Here x, y, and z are the spatial coordinates
in the porous catalyst, ck denotes the molar concentration of the kth compo-
nent, Deff
k is the effective diffusivity of the kth component, nkj is the stoichio-
metric coefficient of the component k in the jth reaction step, and rj is the
reaction rate of the jth reaction step.
The characteristic length of the washcoat section to be simulated is 10 mm;
thus, we may consider constant temperature profile on this scale. Since the
volume diffusion in the macro-pores is much faster than the Knudsen diffusion
in the meso-pores of the g-Al2O3 particles, we may further assume that the CO
and O2 gas concentrations in the macro-pores are constant within the simulated
washcoat section. For the surface-deposited components CO and O, a zero
diffusivity is used, i.e., Deff
k ¼ 0. For gaseous CO and O2, the effective diffusivi-
ties are based on the Knudsen diffusivity in the meso-pores (with diameter
10 nm) of the g-Al2O3.
The microkinetics with an explicit consideration of the surface-deposited
species has been employed,
CO þ
2CO
; r1 ¼ kf1 C Pt yCO y
kb1 LPt yCO
O2 þ 2
! 2O
; r2 ¼ k2 C Pt Y O2 y
CO
þ O
! CO2 þ 2
; r3 ¼ k3 C Pt yCO
yO
FIG. 26. Profiles of CO concentration yCO and reaction rate rCO in the reconstructed
Pt/g-Al2O3 catalyst (cross-section), T ¼ 3001C, Lxy ¼ 540 nm, macro ¼ 0:15, C Pt ¼ 50 mol=m3 ,
Deff ¼ 6 108 m2 =s. Free space corresponds to the macro-pores, solid grey corresponds to meso-
porous g–Al2O3 with dispersed Pt. Length of the section edge is 10 mm (from Koci et al., 2004).
Here * denotes the catalytic Pt-site, CPt the concentration of the catalytic sites,
yk the molar fractions of the gaseous components (CO, O2), and yk the relative
surface concentrations of the deposited components (CO*, O*). The values of
kinetic parameters kj are taken from Nibbelke et al. (1998), where the complete
microkinetic scheme for CO oxidation on Pt/Rh/CeO2/g-Al2O3 catalyst is given.
In Fig. 26 we can observe that steep concentration gradients may occur in the
system. There are even regions with zero CO concentration inside the sintered
g-Al2O3 particles, and the reaction takes place only on the surface of the macro-
pores. The Pt distributed deeper is not utilized in the reaction.
The results of the parametric studies (e.g., the influence of noble metal dis-
tribution and correlation length) provide a better understanding of the reaction-
transport effects in porous, supported heterogeneous catalysts (Bhattacharya
et al., 2004). In the combination with semi-deterministic methods of the recon-
struction (simulation of the catalyst preparation process), the results can be used
for the optimization of the washcoat structure.
V. Outlooks
A. BIOLOGICAL SYSTEMS
Plant and animal tissue has a complex microstructure (e.g., Khaled and Vafai,
2003), consisting of cells, extra-cellular matrix (ECM), and porosity. Tissue and
other biological systems are usually structured over several length-scales (Chu
and Lee, 2004). Convective transport in a tissue is controlled by the vascular
network (Kumar et al., 2004). When describing transport at the length-scale of

the network, the tissue surrounding the capillaries is typically treated as an
effective medium with volume-averaged properties (Lasseux et al., 2004), which
depend on the microstructure—phase volume fraction of ECM, etc. The meth-
ods presented in this chapter are applicable with little or no modification to the
modeling of transport in complex biological structures and cellular assemblies
(Picioreanu et al., 2004; Wood et al., 2002a).
One of the aims of the emerging discipline of tissue engineering (Mantalaris,
2001; Mantalaris et al., 2001) is ex vivo growth of tissues from cells, including
stem cells (Rippon and Bishop, 2004). A typical tissue growth protocol would
consist of seeding the cells into a biodegradable porous scaffold, which is then
immersed into a cultivation medium containing the required nutrients and sig-
nals. The cells would then proliferate in the scaffold and produce their own
extracellular matrix that eventually replaces the degradable scaffold as struc-
tural support. The resulting tissue must have a specific microstructure in order
to posses the same functionality as natural tissue. The microstructure depends
on the interplay of initial scaffold morphology, spatial distribution of seed cells,
diffusion of nutrients and metabolites in and out of the scaffold, and the cell
proliferation and scaffold degradation rates. The methods presented in this
chapter are directly applicable to the modeling of morphology transformation
processes occurring during tissue growth.
B. MATERIALS DESIGN
The field of synthesis and fabrication of porous materials with controlled pore
size and morphology has progressed enormously in recent years. Techniques
such as colloidal templating (Linssen et al., 2003; Velev and Lenhoff, 2000; Xia
et al., 2000) or directed self-assembly (Lin et al., 2001; Pilleni, 2001; Texter and
Tirrell, 2001; Xu and Goedel, 2002) have made it possible to fabricate porous
solids with specific microstructures at a wide range of length-scales (Soten and
Ozin, 1999). Moreover, the ‘‘bottom–up’’ materials synthesis techniques
(Shimomura and Sawadaishi, 2001) can be combined with ‘‘top–bottom’’ mi-
crofabrication methods to form truly multi-scale hybrid structures such as
microchannel networks internally coated with templated porous solid (Xu and
Platzer, 2001; Yang et al., 2001).
Owing in part to the fabrication techniques mentioned above, it is legitimate
to pose the problem of rational design of materials with tailor-made micro-
structure. As part of the material microstructure design process, the structure-
property mapping problem has to be solved—i.e., the problem of calculating a
set of end-use properties (e.g., effective transport properties) for a material
microstructure described by a given set of integral characteristics (porosity,
specific surface area, etc.). The second important element of the material
microstructure design process is a solution of the process-structure mapping
problem—i.e., the problem of determining the microstructure that results from

a series of morphology-transformation operations (particle packing, sintering,
leaching, etc.). A materials synthesis recipe can then be derived from the se-
quence of morphology-transformation operations.
The modeling methods reviewed in this chapter (i.e., evaluation of integral
characteristics, calculation of effective transport properties, and diagenesis-
based models of microstructure formation) are directly relevant to the solution
of both structure-property and process-structure problems. While the ‘‘for-
ward’’ problems have a unique solution, the inverse problems usually do not. A
challenge for material microstructure design is to identify suitable optimization
algorithms for the solution of the inverse problems (e.g., Torquato et al., 2002).
Stochastic optimization algorithms applied at one length-scale (e.g., McLeod
and Gladden, 2000) can be combined with deterministic algorithms at a larger
length-scale (e.g., Gheorghiu and Coppens, 2004) to optimize multi-scale struc-
tures such as heterogeneous catalysts. Similarly, discrete and continuous algo-
rithms may need to be combined in situations where one microstructure
parameter, such as phase volume fraction, is inherently continuous, while some
other parameter, such as the order in which microstructure transformation op-
erations are applied, is inherently discrete. Analogous problems are of course
well known from standard process systems engineering, and inspiration will
undoubtedly be found in that field for the solution of material microstructure
design problems.
VI. Conclusions
The methodology for digital representation, characterization and reconstruc-

tion of porous and multi-phase media has been presented in this chapter. Com-
putational techniques for the determination of effective transport properties
(conductivity, diffusivity, permeability, and tensile properties) have been re-
viewed, as well as methods for the solution of transient reaction, transport, and
morphology-transformation processes in those media. The application of the
techniques has been demonstrated on several examples: polymer particle growth
and morphogenesis, granulation and dissolution, and heterogeneous catalysis.
Finally, two emerging application areas of the computational methods have
been reviewed, namely, tissue engineering and structured materials design.
We hope to have demonstrated that computer simulation of transport and
transformation processes on digitally reconstructed multi-phase media can be
beneficial to practical chemical engineering applications. We believe that as
chemical engineering becomes more ‘‘product-oriented,’’ the need to model
phenomena that control material microstructure formation will gain in impor-
tance. We hope that this chapter will provide a useful starting point for those
who wish to familiarize themselves with the relevant computational techniques.
ACKNOWLEDGMENTS
We thank our colleagues and students Zdeněk Grof, Martin Kohout, Petr
Koc̆ı́ and Gabriela Salejová for their work on preparation of illustrative pic-
tures. JK and MM gratefully acknowledge the support of the Grant Agency of
the Czech Republic (project 104/02/0325) and of the Ministry of Education
(project MSM 6046137306).
REFERENCES
Abdullah, A. Z., Bakar, M. Z. A., and Bhatia, S. J. Chem. Technol. and Biotechnol. 79, 761–768
(2004).
Adler, P. M., ‘‘Porous Media: Geometry and Transports’’. Butterworth-Heinemann, Boston (1992).
Adler, P. M. Curr. Top. Phys. Fluids 1, 277–306 (1994).
Adler, P. M., and Thovert, J. F. Appl. Mech. Rev. 51(9), 537–585 (1998).
Adrover, A., Velardo, A., Giona, M., Cerbelli, S., Pagnanelli, F., and Toro, L. Chem. Eng. J. 99,
89–104 (2004).
Allen, M. P., and Tildesley, D. J., ‘‘Computer Simulation of Liquids’’. Clarendon Press, New York
(1988).
Andrade, J. S., Street, D. A., Shibusa, Y., Havlin, S., and Stanley, H. E. Phys. Rev. E 55, 772–777
(1997).
Arenas-Alatorre, J., Avalos-Borja, M., and Diaz, G. Appl. Surf. Sci. 189, 7–17 (2002).
Argento, C., and Bouvard, D. Int. J. Heat Mass Trans. 39, 1343–1350 (1996).
Aris, R., ‘‘The Mathematical Theory of Diffusion and Reaction in Permeable Catalysts’’. Clarendon
Press, Oxford (1975).
Arns, C. H., Knackstedt, M. A., Pinczewski, V. W., and Mecke, K. R. Phys. Rev. E 63, 031112
(2001).
Arns, C. H., Knackstedt, M. A., and Mecke, K. R. Coll. Surf. A: Physicochem. and Eng. Aspects 241,
351–372 (2004).
Atwood, R. C., and Lee, P. D. Acta Mater. 51, 5447–5466 (2003).
Avnir, D., ‘‘Fractal Approach to Heterogeneous Chemistry: Surfaces Colloids and Polymers’’.
Wiley, New York (1989).
Balakotaiah, V., and Gupta, N. Chem. Eng. Sci. 55, 3505–3514 (2000).
Barrie, P. J. Ann. Rep. NMR Spectrosc. 41, 265–316 (2000).
Békri, S., Thovert, J. -F., and Adler, P. M. Chem. Eng. Sci. 50, 2765–2791 (1995).
Bell, A. T. Science 299, 1688–1691 (2003).
Bhatia, S. K. Chem. Eng. Sci. 41, 1311–1324 (1986).
Bhattacharya, M., Harold, M. P., and Balakotaiah, V. Chem. Eng. Sci. 59, 3737–3766 (2004).
Blacher, S., Léonard, A., Heinrichs, B., Tcherkassova, N., Ferauche, F., Crine, M., Marchot, P.,
Loukine, E., and Pirard, J. -P. Coll. Surf. A: Physicochem. Eng. Aspects 241, 201–206 (2004).
Blunt, M. J., Jackson, M. D., Piri, M., and Valvatne, P. H. Adv. Water Resour. 25, 1069–1089
(2002).
Bowyer, A. Comput. J. 24(2), 162–166 (1981).
Cai, Y., and Zygourakis, K. Ind. Eng. Chem. Res. 42, 2746–2755 (2004).
Capek, P., Hejtmánek, V., and Šolcová, O. Chem. Eng. J. 81, 281–285 (2001).
Chatterjee, A., Snyder, M. A., and Vlachos, D. G. Chem. Eng. Sci. 59, 5559–5567 (2004).
Chu, C. P., and Lee, D. J. Chem. Eng. Sci. 59, 1875–1883 (2004).
Coppens, M. -O., and Froment, G. F. Chem. Eng. Sci. 49, 4897–4907 (1994).
Coppens, M. -O., and Froment, G. F. Chem. Eng. Sci. 50, 1013–1026 (1995a).
Coppens, M. -O., and Froment, G. F. Chem. Eng. Sci. 50, 1027–1039 (1995b).
Coppens, M. -O., Bell, A. T., and Chakraborty, A. K. Chem. Eng. Sci. 54, 3455–3463 (1999).
Coppens, M. -O., and Malek, K. Chem. Eng. Sci. 58, 4787–4795 (2003).
Daccord, G., Lenormand, R., and Liétard, O. Chem. Eng. Sci. 48(1), 169–178 (1993a).
Daccord, G., Liétard, O., and Lenormand, R. Chem. Eng. Sci. 48(1), 179–186 (1993b).
Davies, G. M., and Seaton, N. A. AIChE J. 46, 1753–1768 (2000).
Debling, J. A., and Ray, W. H. Ind. Eng. Chem. Res. 34, 3466–3480 (1995).
Debling, J. A., and Ray, W. H. J. Appl. Polym. Sci. 81, 3085–3106 (2001).
Dodds, P. S., and Rothman, D. H. Annu. Rev. Earth Planet. Sci. 28, 571–610 (2000).
Dominquez, A., Bories, S., and Prat, M. Int. J. Multiphas. Flow 26, 1951–1979 (2000).
Dumesic, J. A., Rudd, D. F., Aparicio, L. M., Rekoske, J. E., and Tervino, A. A., ‘‘The Micro-
kinetics of Heterogeneous Catalysis’’. American Chemical Society, Washington (1992).
Elmoutaouakkil, A., Fuchs, G., Bergounhon, P., Péres, R., and Peyrin, F. J. Phys. D: Appl. Phys.
36, A37–A43 (2003).
Elmoutaouakkil, A., Salvo, L., Maire, E., and Peix, G. Adv. Eng. Mater. 4, 803–807 (2002).
El-Nafaty, U. A., and Mann, R. Chem. Eng. Sci. 56, 856–872 (2001).
Erlebacher, J., Aziz, M. J., Karma, A., Dimitrov, N., and Sieradzki, K. Nature 410, 450–453 (2001).
Estenoz, D. A., and Chiovetta, M. G. J. Appl. Polym. Sci. 81, 285–311 (2001).
Farber, L., Tardos, G., and Michaels, J. N. Powder Technol. 132, 57–63 (2003).
Feres, R., and Yablonsky, G. Chem. Eng. Sci. 59, 1541–1556 (2004).
Galani, A. N., Kainourgiakis, M. E., Kikkinides, E. S., Steriotis, Th., Stubos, A. K., and Papa-
ioannou, A. Coll. Surf. A: Physicochem. Eng. Aspects 241, 273–279 (2004).
Garcia-Gonzales, R., Monnereau, C., Thovert, J. -F., Adler, P. M., and Vignes-Adler, M. Coll. Surf.
A: Physicochem. Eng. Aspects 151, 497–503 (1999).
Garza-López, R. A., and Kozak, J. J. J. Phys. Chem. B 103, 9200–9204 (1999).
Gavrilov, C., and Sheintuch, M. AIChE J. 43, 1691–1699 (1997).
Gheorghiu, S., and Coppens, M. -O. AIChE J. 50, 812–820 (2004).
Giona, M., Adrover, A., and Giona, A. R. Chem. Eng. Sci. 50, 1001–1011 (1995).
Gracia, F. J., and Wolf, E. E. Chem. Eng. Sci. 59, 4723–4729 (2004).
Gregg, S. J., and Sing, K. S. W., ‘‘Adsorption, Surface Area and Porosity’’. Academic Press, London
(1982).
Grof, Z., Kosek, J., Marek, M., and Adler, P. M. AIChE J. 49(4), 1002–1013 (2003).
Grof, Z., and Kosek, J., 2005. Dynamics of fragmentation of catalyst carriers in catalytic polym-
erization of olefins. In preparation (2005).
Grof, Z., Kosek, J., and Marek, M. AIChE J. 51(7), 2048–2067 (2005a).
Grof, Z., Kosek, J., and Marek, M. Ind. Eng. Chem. Res. 44, 2389–2404 (2005b).
Groot, R. D., and Warren, P. B. J. Chem. Phys. 11, 4423–4435 (1997).
Gummalla, M., Tsapatsis, M., Watkins, J. J., and Vlachos, D. G. AIChE J. 50, 684–695 (2004).
Gupper, A., Wilhelm, P., Schmied, M., Kazarian, S. G., Chan, K. L. A., and ReuXner, J. Appl.
Spectrosc. 56, 1515–1523 (2002).
Hazlett, R. D. Math. Geol. 29(6), 801 (1997).
Heyes, D. M., Baxter, J., Tüzün, U., and Qin, R. S. Philos. Trans. Roy. Soc. London A 362,
1853–1865 (2004).
Hilgenfeldt, S., Kraynik, A. M., Koehler, S. A., and Stone, H. A. Phys. Rev. Lett. 86, 2685–2688
(2001a).
Hilgenfeldt, S., Koehler, S. A., and Stone, H. A. Phys. Rev. Lett. 86, 4704–4707 (2001b).
Hoebink, J. H. B. J., and Marin, G. B. Modeling of Monolithic Reactors for Automotive Exhaust
Gas Treatment, in A. Cybulski and J. A. Moulijn, ‘‘Structured Catalysts and Reactors’’.
Marcel Dekker, NY (1998).
Hosier, I. L., Alamo, R. G., and Lin, J. S. Polymer 45, 3441–3455 (2004).
Ingram, G. D., Cameron, I. T., and Hangos, K. M. Chem. Eng. Sci. 59, 2171–2187 (2004).
Ichikawa, S., Akita, T., Okumura, M., Kohyama, M., and Tanaka, K. JEOL News 38, 6–9 (2003).
Jackson, R., ‘‘Transport in Porous Catalysts’’. Elsevier, Amsterdam (1977).
Jansen, W. P. A., Harmsen, J. M. A., Denier, A. W., Hoebink, J. H. B. J., Schouten, J. C., and
Brongersma, H. H. J. Catal. 204, 420–427 (2001).
Janssen, A. H., Koster, A. J., and de Jong, K. P. Angew. Chem. Int. Edit. 40, 1102–1104 (2001).
Joshi, K., Lee, J. G., Shafi, M. A., and Flumerfelt, R. W. J. Appl. Polym. Sci. 67, 1353–1368 (1998).
Kainourgiakis, M. E., Steriotis, Th., Kikkinides, E. S., Romanos, G., and Stubos, A. K. Coll. Surf.
A: Physicochem. Eng. Aspects 206, 321–334 (2002).
Kakugo, M., Sadatoshi, H., Yokoyama, M., and Kojima, K. Macromolecules 22, 547–551 (1989).
Kanamori, K., Nakanishi, K., Hirao, K., and Jinnai, H. Coll. Surf. A: Physicochem. Eng. Aspects
241, 215–224 (2004).
Keil, F. J. Chem. Eng. Sci. 51, 1543–1568 (1996).
Khaled, A. -R. A., and Vafai, K. Int. J. Heat Mass Trans. 46, 4989–5003 (2003).
Kim, J. C., Auh, K. H., and Martin, D. M. Model. Simul. Mater. Sci. Eng. 8, 159–168 (2000).
Kittilsen, P., and McKenna, T. F. J. Appl. Polym. Sci. 82, 1047–1060 (2001).
Knoke, S., Korber, F., Fink, G., and Tesche, B. Macromol. Chem. Phys. 204(4), 607–617 (2003).
Koci, P., Kubicek, M., and Marek, M. Chem. Eng. Res. Dev. 82, 284 (2004a).
Koci, P., Stepanek, F., Kubicek, K., and Marek, M. Modeling of CO Oxidation in Digitally Re-
constructed Porous Pt/gAl2O3 Catalyst. ‘‘Proceedings of CHISA 2004’’, 22–25 August 2004,
Prague, Czech Republic (2004b).
Kohout, M., Collier, A. P., and Stepanek, F. Int. J. Heat Mass Trans. 47, 5565–5574 (2004).
Kohout, M., Collier, A.P., and Stepanek, F. Powder Technol., in press (2005).
Kohout, M., Collier, A.P., and Stepanek, F. Vacuum Contact Drying: Multi-Scale Modeling and
Experiments, in ‘‘European Symposium on Computer Aided Process Engineering – 14’’ (A.
Barbosa-Povoa, and H. Matos Eds.), pp. 1075–1080. Elsevier Science, Amsterdam (2004).
Koopmans, R. J., den Doelder, J. C. F., and Paquet, A. N. Adv. Mater. 12, 1873–1880 (2000).
Körner, C., Thies, M., and Singer, R. F. Adv. Eng. Mater. 4, 765–769 (2002).
Kosek, J., Grof, Z., Novak, A., Stepanek, F., and Marek, M. Chem. Eng. Sci. 56, 3951–3977
(2001a).
Kosek, J., Stepanek, F., Novak, A., Grof, Z., and Marek, M. Multi-Scale Modeling of Growing
Polymer Particles in Heterogeneous Catalytic Reactors, in ‘‘European Symposium on Com-
puter Aided Process Engineering – 11’’ (R. Gani, S. B. Jorgensen Eds.), pp. 177–182. Elsevier
Science, Amsterdam (2001b).
Kothe, D. B., Rider, W. J., Mosso, S. J., Brock, J. S., and Hochstein, J. I. Volume tracking of
interfaces having surface tension in two and three dimensions. Technical Report AIAA 96-
0859 (1996).
Kraynik, A. M., Reinelt, D. A., and van Swol, F. Phys. Rev. E 67, 031403 (2003).
Krishna, R. Chem. Eng. Sci. 48, 845–861 (1993).
Krishna, R., and Wesseling, J. A. Chem. Eng. Sci. 52(6), 861–911 (1997).
Kumar, R., Stepanek, F., and Mantalaris, A. Food Bioprod. Process. 82, 105–116 (2004).
Lakatos, B. G. Chem. Eng. Sci. 56, 659–666 (2001).
Lasseux, D., Ahmadi, A., Cleis, X., and Garnier, J. Chem. Eng. Sci. 59, 1949–1964 (2004).
Lee, D.T., Computational Geometry II – Proximity, in ‘‘Algorithm & Theory of Computation
Handbook, Chapter 20.2’’ (M. J. Atallah, Ed.), CRC Press, Boca Raton, USA (1999).
Lee, P. D., Chirazi, A., Atwood, R. D., and Wang, W. Mater. Sci. Eng. A 365, 57–65 (2004).
Li, X., and Yortsos, Y. C. Chem. Eng. Sci. 50, 1247–1271 (1995).
Liang, Z., Ioannidis, M. A., and Chatzis, I. J. Colloid Interf. Sci. 221, 13–24 (2000a).
Liang, Z., Ioannidis, M. A., and Chatzis, I. Chem. Eng. Sci. 55, 5247–5262 (2000b).
Liger-Belair, G., Vignes-Adler, M., Voisin, C., Robillard, B., and Jeandet, P. Langmuir 18,
1294–1301 (2002).
Lin, C. L., and Miller, J. D. Chem. Eng. J. 80, 221–231 (2000).
Lin, X. M., Jaeger, H. M., Sorensen, C. M., and Klabunde, K. J. J. Phys. Chem. B 105, 3353–3357
(2001).
Linssen, T., Cassiers, K., Cool, P., and Vansant, E. F. Adv. Colloid Interf. Sci. 103, 121–147 (2003).
Lombardo, S. J., and Bell, A. T. Surf. Sci. Rep. 13(1–2), 1–72 (1991).
Losic, N., Thovert, J. -F., and Adler, P. M. J. Colloid Interf. Sci. 186, 420–433 (1997).
Ma, M. C., and Sonka, M. Computer Vision and Image Understanding 64(3), 420 (1996).
Makeev, A. G., and Kevrekidis, I. G. Chem. Eng. Sci. 59, 1733–1743 (2004).
Mann, R. Chem. Eng. Res. Des. 71, 551–562 (1993).
Mantalaris, A. Chem. Eng. 720, 32–34 (2001).
Mantalaris, A., Panoskaltsis, N., Sakai, Y., Bourne, P., Chang, C., Messing, E. M., and Wu, J. H.
D. J. Pathol. 193, 361–366 (2001).
Manwart, C., Torquato, S., and Hilfer, R. Phys. Rev. E 62, 893–899 (2000).
Martys, N. S., and Garboczi, E. J. Phys. Rev. B 46, 6080–6090 (1992).
Mason, E. A., and Malinauskas, A. P., ‘‘Gas Transport in Porous Media: The Dusty-Gas Model’’.
Elsevier, Amsterdam (1983).
McLeod, A. S., and Gladden, L. F. J. Catal. 173, 43–52 (1998).
McLeod, A. S., and Gladden, L. F. J. Chem. Inf. Comp. Sci. 40, 981–987 (2000).
Mikami, T., Kamiya, H., and Horio, M. Chem. Eng. Sci. 53(10), 1927–1940 (1998).
Mourzenko, V. V., Thovert, J. -F., and Adler, P. M. Eur. Phys. J. B 19, 75–85 (2001).
Mougin, P., Pons, M., and Villermaux, J. Chem. Eng. J. 64, 63–68 (1996).
Nam, J. H., and Kaviany, M. Int. J. Heat Mass Trans. 46, 4595–4611 (2003).
Nauman, E. B., and He, D. Q. Chem. Eng. Sci. 56, 1999–2018 (2001).
Nibbelke, R. H., Nievergeld, A. J., Hoebink, J. H., and Marin, G. B. Appl. Catal. B: Environ. 19, 245
(1998).
Pater, J. T. M., Weickert, G., and van Swaaij, W. P. M. J. Appl. Polym. Sci. 87, 1421–1435 (2003).
Perré, P. Meshpore: A Software Able to Apply Image-Based Meshing Techniques to Anisotropic
and Heterogeneous Porous Media. ‘‘Drying 2004 – Proceedings of the 14th International
Drying Symposium’’, pp. 664–671, Sao Paulo, Brazil, 22–25 August 2004.
Piccolo, I., Becker, C., and Henry, C. R. Eur. Phys. J. D 9, 415–419 (1999).
Picioreanu, C., Kreft, J. -U., and van Loosdrecht, M. C. M. Appl. Environ. Microbiol. 70, 3024–3040
(2004).
Pilleni, M. P. J. Phys. Chem. B 105, 3358–3371 (2001).
Poulet, J., Manzoni, D., Hage-Chehade, F., Jacquin, C. J., Boutéca, M. J., Thovert, J. -F., and
Adler, P. M. J. Mech. Phys. Solids 44(10), 1587–1620 (1996).
Quintard, M., Kaviany, M., and Whitaker, S. Adv. Water Resour. 20, 77–94 (1997).
Ramaswamy, S., Gupta, M., Goel, A., Aaltosalmi, U., Kataja, M., Koponen, A., and Ramarao, B.
V. Coll. Surf. A: Physicochem. Eng. Aspects 241, 232–333 (2004).
Rider, W. J., and Kothe, D. B. J. Comput. Phys. 141, 112–152 (1998).
Rieckmann, C., and Keil, F. J. Ind. Eng. Chem. Res. 36, 3275–3281 (1997).
Rigby, S. P., and Daut, S. Adv. Colloid Interf. Sci. 98, 87–119 (2002).
Rippon, H. J., and Bishop, A. E. Cell Proliferat. 37, 23–34 (2004).
Roberts, A. P., and Knackstedt, M. A. Phys. Rev. E 54, 2313–2328 (1996).
Roberts, A. P., and Torquato, S. Phys. Rev. E 59, 4953–4963 (1999).
Rozman, M. G., and Utz, M. Phys. Rev. E 63, 066701 (2001).
Sahimi, M. Rev. Mod. Phys. 65, 1393–1534 (1993).
Sahimi, M., Gavalas, G. R., and Tsotsis, T. T. Chem. Eng. Sci. 45, 1443–1502 (1990).
Sahimi, M., Heidarinasabb, A., and Dabirb, B. Chem. Eng. Sci. 59, 4289–4301 (2004).
Salejova, G., Kosek, J., Nevoral, V., Solcova, O., and Schneider, P. Effective Transport Properties
of Reconstructed Porous Catalyst Carriers. ‘‘Proceedings of CHISA 2004’’, 22–25 August
2004, Prague, Czech Republic (2004).
Salejova, G., Kosek J., and Stepanek F. Effective diffusivity of random and regular cellular struc-
tures. In preparation (2005).
Salles, J., Thovert, J. -F., and Adler, P. M. Chem. Eng. Sci. 48, 2839–2858 (1993).
Sapoval, B., Andrade, J. S., and Filochea, M. Chem. Eng. Sci. 56, 5011–5023 (2001).
Satterfield, Ch. N., ‘‘Heterogeneous Catalysis in Practice’’. McGraw-Hill, New York (1980).
Scardovelli, R., and Zaleski, S. Annu. Rev. Fluid Mech. 31, 567–603 (1999).
Schmidt, L. D., ‘‘The Engineering of Chemical Reactions’’. Oxford University Press, Oxford (1998).
Schneider, P., and Gelbin, D. Chem. Eng. Sci. 40, 1093–1099 (1984).
Sethian, J. A., ‘‘Level Set Methods’’. Cambridge University Press, New York (1996).
Shapiro, A. A., and Stenby, E. H. Fluid Phase Equilibr. 134, 87–101 (1977).
Sheintuch, M. Ind. Eng. Chem. Res. 38, 3261–3269 (1999).
Shimomura, M., and Sawadaishi, T. Curr. Opin. Colloid Interf. Sci. 6, 11–16 (2001).
Silverstein, D. L., and Fort, T. Langmuir 16, 839–844 (2000).
Soten, I., and Ozin, G. A. Curr. Opin. Colloid Interf. Sci. 4, 325–337 (1999).
Stepanek, F. Chem. Eng. Res. Desig. 82(A11), 1458–1466 (2004).
Stepanek, F., and Ansari, M. A. Chem. Eng. Sci. 60(14), 4019–4029 (2005).
Stepanek, F., Hanika, J., Marek, M., and Adler, P. M. Catal. Today 66, 249–254 (2001a).
Stepanek, F., Marek, M., and Adler, P. M. Chem. Eng. Sci. 56, 467–474 (2001b).
Stepanek, F., Kubicek, M., Marek, M., and Adler, P. M., Computer-Aided Screening of Adsorbents
and Porous Catalyst Carriers, in ‘‘European Symposium on Computer Aided Process En-
gineering – 10’’ (S. Pierucci Ed.), pp. 667–672. Elsevier Science, Amsterdam (2000).
Stepanek, F., Marek, M., and Adler, P. M. AIChE J. 45, 1901–1912 (1999).
Texter, J., and Tirrell, M. AIChE J. 47, 1706–1710 (2001).
Thompson, K. E. AIChE J. 48, 1369–1389 (2002).
Thomson, K. T., and Gubbins, K. E. Langmuir 16, 5761–5773 (2000).
Thornton, C., Ciomocos, M. T., and Adams, M. J. Powder Technol. 105, 74–82 (1999).
Thornton, C., Ciomocos, M. T., and Adams, M. J. Powder Technol. 140, 258–267 (2004).
Thovert, J. -F., Salles, J., and Adler, P. M. J. Microsc. 170, 65–79 (1993).
Thovert, J. -F., Yosefian, F., Spanne, P., Jacquin, G., and Adler, P. M. Phys. Rev. E 63, 061307
(2001).
Tian, Z., Xu, Y., and Lin, L. Chem. Eng. Sci. 59, 1745–1753 (2004).
Torquato, S. Int. J. Solids Struct. 37, 411–422 (2000).
Torquato, S., ‘‘Random Heterogeneous Materials: Microstructure and Macroscopic Properties’’.
Sprinder-Verlag, New York (2002).
Torquato, S., Hyun, S., and Donev, A. Phys. Rev. Lett. 89, 266601 (2002).
Tsuji, Y., Kawaguchi, T., and Tanaka, T. Powder Technol. 77, 79–87 (1993).
Tzschichholz, F., Herrmann, H. J., and Zanni, H. Phys. Rev. E 53, 2629–2637 (1996).
van Neer, F., and Bliek, A. Chem. Eng. Sci. 54(20), 4483–4499 (1999).
Velev, O. D., and Lenhoff, A. M. Curr. Opin. Colloid Interf. Sci. 5, 56–63 (2000).
Weyland, M., Midgley, P. A., and Thomas, J. M. J. Phys. Chem. B 105, 7882–7886 (2001).
Whitaker, S., ‘‘The Method of Volume Averaging’’. Kluwer Academic Publishers, Dordrecht (1999).
Wolff, J., Papathanasion, A. G., Rotermund, H. H., Ertl, G., Li, X., and Kevrekidis, I. G. J. Catal.
216(1–2), 246–256 (2003).
Wood, B. D., Quintard, M., and Whitaker, S. Biotechnol. Bioeng. 77, 495–516 (2002a).
Wood, J., Gladden, L. F., and Keil, F. J. Chem. Eng. Sci. 57, 3047–3059 (2002b).
Xia, Y., Gates, B., Tin, Y., and Lu, Y. Adv. Mater. 12, 693–713 (2000).
Xu, H., and Goedel, W. A. Langmuir 18, 2363–2367 (2002).
Xu, Y., and Platzer, B. Chem. Eng. Technol. 24, 773–783 (2001).
Yang, P., Rizvi, A. H., Messer, B., Chmelka, B. F., Whitesides, G. M., and Stucky, G. D. Adv.
Mater. 13, 427–431 (2001).
Yao, J., Zhang, Y., Wang, C. -H., Matsusaka, S., and Masuda, H. Ind. Eng. Chem. Res. 43,
7181–7199 (2004).
Yeong, C. L. Y., and Torquato, S. Phys. Rev. E 57, 495–506 (1998a).
Yeong, C. L. Y., and Torquato, S. Phys. Rev. E 58(1), 224–233 (1998b).
Zalc, J. M., Reyes, S. C., and Iglesia, E. Chem. Eng. Sci. 58, 4605–4617 (2003).
Zalc, J. M., Reyes, S. C., and Iglesia, E. Chem. Eng. Sci. 59, 2947–2960 (2004).
Zhdanov, V. P., and Kasemo, B. Surf. Sci. Rep. 39, 25–104 (2000).
Zhdanov, V. P. Surf. Sci. Rep. 45(7–8), 231–326 (2002).
SPATIALLY AVERAGED MULTI-SCALE MODELS FOR
CHEMICAL REACTORS
Saikat Chakraborty and Vemuri Balakotaiah
Department of Chemical Engineering, University of Houston, Houston, Texas 77204-4004,

USA
I. Introduction 206
A. A Brief History of Chemical Reactor Models 208
B. Multi-scale Nature of Homogeneous and Catalytic
Reactors 211
C. Different Approaches to Multi-scale Averaging or
Dimension Reduction 214
II. Spatial Averaging of Convection–diffusion–reaction Mod-
els using the L–S Method 217
III. Spatially Averaged Models for Describing Dispersion
Effects in Tubes and Packed Beds 221
A. A Hyperbolic Averaged Model for Describing Disper-
sion Effects in Tubes/Capillaries 222
B. Multi-mode Hyperbolic Averaged Models for Describ-
ing Dispersion Effects in Chromatographs 233
C. A Hyperbolic Model for Describing Dispersion Effects
in Monoliths with Diffusion into the Solid Phase 238
IV. Spatially Averaged Multi-mode (Multi-scale) Models for
Homogeneous Reactors 239
A. Isothermal Tubular Reactors 239
B. Loop and Recycle Reactors 248
C. Tank Reactors (CSTRs) 250
D. Non-isothermal Reactor Models 252
E. Multiple Reactions 259
F. Examples Illustrating Use of Multi-mode Homoge-
neous Reactor Models 260
V. Spatially Averaged Multi-mode (Multi-scale) Models for
Catalytic Reactors 273
A. Wall-catalyzed Reactions 273
B. Coupled Homogeneous and Wall-catalyzed Reactions 277
C. Non-isothermal Reactor Models 278
D. Examples Illustrating Use of Multi-mode Catalytic
Reactor Models 279
VI. Accuracy, Convergence and Region of Validity of Multi-
mode/Multi-scale Averaged Models 283
A. Accuracy 283
E-mail: bala@uh.edu
205
DOI 10.1016/S0065-2377(05)30004-4
206 SAIKAT CHAKRABORTY AND VEMURI BALAKOTAIAH
B. Convergence 284
C. Regularization of the Local Equation 288
VII. Summary, Conclusions, and Recommendations for Future
Work 293
Acknowledgments 294
References 296
Abstract
The classical homogeneous and catalytic reactor models such as

CSTR, PFR, and two-phase models, make a priori assumptions on the
length and time scales of diffusion, convection, and reaction and apply
the conservation laws at the macroscopic level. This intuitive averaging
leads to incorrect and often non-physical description of the small-scale
mixing/diffusion effects, accumulation and reaction terms in the
averaged form of the conservation laws. We consider the full
convection–diffusion–reaction equations with homogeneous and cataly-
tic reaction and spatially average them rigorously over small length/time
scales using the Liapunov–Schmidt method. This results in low-
dimensional averaged models defined in terms of multiple concentration
and temperature modes (variables), each of which is representative of a
physical scale of the system. These multi-mode models are markedly
different from the classical reactor models and are hyperbolic for most
cases of practical interest. While their solution requires a numerical effort
that is comparable to that of the classical reactor models, they retain all
the parameters and qualitative features of the full governing equations
(within their region of convergence) and overcome the shortcomings of
the classical models (such as dependence of effective transport
coefficients on kinetics, inability to describe micromixing effects, infinite
speed of propagation of signals even in convection dominated systems,
etc). Examples are presented to illustrate and compare the predictions of
the new averaged multi-mode models with those of the classical models.
I. Introduction
Modeling and analysis of chemical reactors is at the core of the chemical

engineering discipline. It is one activity that is unique to our discipline and
distinguishes it from other branches of engineering. One fundamental difference
between the modeling of chemical reactors and non-reacting systems is that the
SPATIALLY AVERAGED MULTI-SCALE MODELS 207
length and time scales associated with reacting flows can vary from the
molecular scale to the macro or process scale. Depending on the level of detail
included at various length and time scales, mathematical models that describe
chemical reactors can vary in complexity as well as in the number of physico-
chemical parameters. In addition, due to the strong coupling between the
transport and reaction processes and the dependence of the kinetic and
transport rates on the state variables, the model equations are highly nonlinear
and are known to exhibit a variety of complex spatio-temporal patterns. For
most cases of practical interest, even with the present day computational power,
it is impractical to solve such detailed models and explore the different types of
solutions that may exist in the multi-dimensional parameter spaces. Even in
cases where detailed solutions can be obtained, the numerical results have to be
coarse-grained to determine quantities (such as the average conversion or speed
of a thermal front) that are of interest to the designer. Accurate low-dimensional
models in terms of measurable variables (such as cup-mixing concentrations or
temperatures) are desired for the purpose of design, control, and optimization
of chemical processes.
The most common procedure in chemical reaction engineering to develop
low-dimensional models of reactors is to make certain a priori assumptions on
the length and time scales of reaction, diffusion and convection and apply the
conservation laws at the meso or macroscales only. For example, the most
famous and widely used chemical reactor model, namely that of the continuous-
flow stirred-tank reactor (CSTR), is obtained by applying the species and energy
balances at the macro (reactor) scale and consists of ordinary differential
equations. This model ignores the physics at small length scales (and hence is
independent of the transport properties such as the viscosity of the fluid or the
diffusivities of the various species). The shortcomings of this model (such as its
inability to predict micromixing effects on conversion and yield/selectivity of an
intermediate product for the case of fast reactions) are well known. The same
can be said of many other classical reactor models such as the plug-flow tubular
reactor model, the axial dispersion model with Danckwerts boundary
conditions, the two-phase catalytic reactor model, psuedohomogeneous models
of multi-phase reactors, etc.
At the other extreme, it may be argued that the traditional low-dimensional
models of reactors (such as the CSTR, PFR, etc.) should be abandoned in favor
of the detailed models of these systems and numerical solution of the full
convection–diffusion–reaction (CDR) equations using computational fluid
dynamics (CFD). While this approach is certainly feasible (at least for single-
phase systems) due to the recent availability of computational power and tools,
it may be computationally prohibitive, especially for multi-phase systems with
complex chemistry. It is also not practical when design, control and
optimization of the reactor or the process is of main interest. The two main
drawbacks/criticisms of this approach are: (i) It leads to discrete models of very
high dimension that are difficult to incorporate into design and control schemes.
In addition, for highly nonlinear cases, the mesh size needed to avoid spurious
solutions may be so small that this approach is not feasible. (ii) The CFD
approach also uses averaged models (e.g., k–e model for turbulent flows) with
closure schemes that are not always justified and contain adjustable constants.
The main goal of this chapter is to demonstrate an intermediate approach and
a systematic method of obtaining low-dimensional models of chemical reactors
by spatial averaging of the full CDR equations. Our method of spatial
averaging is rigorous and is based on the Liapunov–Schmidt (L–S) technique of
classical bifurcation theory. Intuitively speaking, the L–S method of averaging
is equivalent to Taylor expansion of a more detailed model in terms of one or
more small parameters representing the ratio of length or time scales present in
the detailed model. In such an expansion, the lowest (zeroth) order term (which
ignores the physics at the small scales) is the simplified model (such as the ideal
PFR or CSTR model), while the higher order corrections modify it by including
the small but significant physical phenomena (such as local velocity gradients,
molecular diffusion, finite rates of adsorption, reaction, etc.) present at various
length and time scales.
Before we illustrate the new approach, we give a brief history of some
important reactor models. We also outline the multi-scale nature of both
homogeneous and catalytic reactors and present a brief review of prior efforts at
describing the scale coupling using low-dimensional models. In Section II, we
outline briefly the L–S procedure for spatial averaging of CDR equation. We
illustrate the spatial-averaging procedure in Section III by considering some
examples of non-reacting systems describing dispersion in tubes, packed beds
and monoliths. In Sections IV and V we present low-dimensional averaged
models for homogeneous, catalytic and coupled homogeneous and catalytic
reactors, respectively. Examples illustrating the use of these models and a
comparison of their predictions with the classical reactor models is also given in
these sections. In Section VI, we discuss briefly the convergence and the region
of validity of the low-dimensional models. Finally, in the last section, we outline
some possible extensions of our approach to multi-phase reacting systems.
A. A BRIEF HISTORY OF CHEMICAL REACTOR MODELS
The most widely used homogeneous reactor models are the three classical
ideal reactor models, namely the plug-flow reactor (PFR) model, the CSTR
model, and the batch reactor (BR) model. While the BR model and the PFR
model (which are identical for constant density systems with time replaced by
space time or dimensionless distance along the tube) have existed since the late
eighteenth century, a conceptual leap came in the form of the CSTR model
through the work of Bodenstein and Wolgast (1908). Unlike the PFR model,
which assumes no gradients in the radial direction and no mixing in the axial
direction, the CSTR model assumes complete mixing at all scales. For constant
density systems, the three classical homogeneous reactor models for the case of
a single reaction of the form
X
M
nj A j ¼ 0
j¼1
(where M is the number of species, nj 40 for products and nj o0 for reactants)

are described by the following equations:
dhC j i
PFR : hux i ¼ nj RðhCiÞ with hC j i ¼ C j;in at x ¼ 0 (1)
dx
dhC j i
BR : ¼ nj RðhCiÞ with hC j i ¼ C j;in at t0 ¼ 0 (2)
dt0
hC j i C j;in
CSTR : ¼ nj RðhCiÞ ðj ¼ 1; 2; . . . ; MÞ (3)
tC
where hC j i is the spatially (or cross-sectional) averaged reactant concentration
of species Aj, Cj,in is the mean inlet (or initial) concentration of the reactant,
RðhCiÞ the rate of the homogeneous reaction, x the distance coordinate along
the length of the PFR, hux i the mean fluid velocity in the reactor in the axial
direction, t0 the time, and tC is the total residence time of the reactor. (Remark:
No distinction is made between spatially averaged concentration, hCi, and cup-
mixing concentration, Cm, in these ideal one-mode reactor models.)
Irving Langmuir (1908) first replaced the assumption of no axial mixing of the
PFR model with finite axial mixing and the accompanying Dirichlet boundary
condition ðhC j i ¼ C j;in at x ¼ 0Þ by a flux-type boundary condition
dhC j i
Dm;j ¼ hux i½hC j i C j;in at x ¼ 0 (4)
dx
where Dm,j is the molecular diffusivity of the species Aj. The above boundary
condition was rediscovered several times later: first by Förster and Geib (1934),
which was quoted and applied by Damköhler (1937); and then finally by
Danckwerts (1953); and it has since then been known as ‘‘Danckwerts’
boundary condition’’. In his paper, Langmuir dealt with both the limiting cases
of ‘‘mixing nearly complete’’ and ‘‘only slight mixing’’.
Thirty years later, Gerhard Damköhler (1937) in his historic paper,
summarized various reactor models and formulated the two-dimensional
CDR model for tubular reactors in complete generality, allowing for finite
mixing both in the radial and axial directions. In this paper, Damköhler used
the flux-type boundary condition at the inlet and also replaced the assumption
of plug flow with parabolic velocity profile, which is typical of laminar flow in
tubes.
Förster and Geib (1934) first introduced the concept of residence time
distribution (RTD) to study the problem of longitudinal dispersion in tubes.
They also obtained the RTD curves for the axial dispersion model with
Danckwerts boundary condition using the Laplace transform theory. Twenty
years later, Danckwerts (1953), in his much celebrated paper, devised a
generalized treatment of RTD and introduced the concepts of ‘‘hold-back’’ and
‘‘segregation’’. Following this, it was Zwietering (1959) who quantified the
degrees of mixing with the ideas of ‘‘complete segregation’’ and ‘‘maximum
mixedness’’ and brought forth the concept of micromixing, i.e. mixing caused by
local diffusion, local velocity gradients and reaction at the small scales.
The study of mixing effects on chemical reactions has been an active area of
research since the pioneering papers of Danckwerts (1958) and Zwietering
(1959). The topic has become a part of classical Chemical Reaction Engineering
and has been discussed in textbooks (Froment and Bischoff, 1990; Levenspiel,
1999; Westerterp et al., 1984) and review articles (Villermaux, 1991).
Historically, this study has progressed in two parallel branches, based on the
Lagrangian and Eulerian frameworks of description, respectively.
The mechanistic models based on the Lagrangian description were initiated
by the studies of Danckwerts (1958) and Zwietering (1959), using the RTD
theory and the ideas of ‘‘complete segregation’’ and ‘‘maximum mixedness’’.
The two- and the three-environment models (Miyawaki et al., 1975; Ng and
Rippin, 1965) that followed, described micromixing as an exchange between the
environments of ‘‘complete segregation’’ and ‘‘maximum mixedness’’. Other
mechanistic models that have been formulated in the last 40 years include the
coalescence–redispersion model (Curl, 1963; Harada, 1962), the interaction by
exchange with mean (IEM) model (Villermaux and Devillon, 1972), models that
assume diffusion with chemical reaction in deforming fluid lamellae (Ottino
et al., 1979), and the engulfment–deformation–diffusion (EDD) models
(Baldyga and Bourne, 1984). Vatistas and Marconi (1992) extended the IEM
model to address the issue of non-isothermal micromixing in exothermic
reactions, Although many of these models were low-dimensional (and therefore
are inexpensive to compute) and could relate the global to the local interactions
quite effectively, they are phenomenological in nature and are derived based on
a top-down approach involving simplifying assumptions on the different time
and length scales of the system.
The Eulerian (bottom-up) approach is to start with the convective–diffusion
equation and through Reynolds averaging, obtain time-smoothed transport
equations that describe micromixing effectively. Several schemes have been
proposed to close the two terms in the time-smoothed equations, namely, scalar
turbulent flux in reactive mixing, and the mean reaction rate (Bourne and Toor,
1977; Brodkey and Lewalle, 1985; Dutta and Tarbell, 1989; Fox, 1992; Li and
Toor, 1986). However, numerical solution of the three-dimensional transport
equations for reacting flows using CFD codes are prohibitive in terms of
the numerical effort required, especially for the case of multiple reactions with
fast/non-isothermal kinetics. As a result, in spite of the simplifying assumptions

present, the century-old ideal classical reactor models [Eqs. (1)–(3)] are still the
most popular choices among chemical engineering practitioners. The classical
ideal reactor models, which are easy-to-solve ordinary differential or algebraic
equations with no adjustable parameter, are particularly preferred to the full
CDR models (which are partial differential equations in two or more
independent variables) in case of multiple reactions with complex kinetics.
In contrast to the homogeneous reactor models, catalytic and multi-phase
reactor models (using the interphase transfer coefficient concept) appeared only
in the late 1950s and early 1960s. By this time, the concept of heat and mass
transfer coefficients was widely accepted and used in the design and analysis of
process equipment (mostly non-reacting systems such as absorbers, heat
exchangers and separation columns). These transfer coefficient models were
extended to reacting systems by just adding the reaction terms to appropriate
balance equations for the phase. Wicke (1961) and Liu and Amundson (1962)
developed and analyzed two-phase models of packed-bed catalytic reactors.
Trambouze and Piret (1960) and Schmitz and Amundson (1963) developed and
analyzed models of two-phase systems with heat and mass exchange and
reaction in one of the phases. [Prior to these studies, mostly pseudohomoge-
neous models were used to describe packed-bed catalytic reactors. Also, while
the concepts of effectiveness and enhancement factors (Hatta, 1932; Thiele,
1939; Zeldovich, 1939) were used much earlier in the 1930s and 1940s, the use of
these in the design and analysis of multi-phase reactors occurred much later, i.e.
in the late 1950s and early 1960s.] During the 1970s through 1990s, the literature
on the modeling of homogeneous, catalytic, and multi-phase reactors has
proliferated. More recently, CFD has also been applied in the modeling of
bubble column and other single- and multi-phase reactors (Ranade, 2002).
B. MULTI-SCALE NATURE OF HOMOGENEOUS AND CATALYTIC REACTORS
As stated earlier, chemical reactors are characterized by multiple length (or

time) scales and these disparate scales are typically characterized by three
representative ones, namely, micro (molecular), meso (catalyst particle or tube
diameter) and macro (reactor or process) scales. Figures 1–4 show examples of
simple to complex chemical reactors, in each of which scale separation exists.
Figure 1 shows the schematic of a tubular reactor, of radius a and length L,
where a ¼ a=L is the aspect ratio. Clearly, if a 1, or a 1, a physical length
scale separation exists in the reactor. This length scale separation could also be
interpreted in terms of time scales. For example, a 1 implies that the time
scale for radial diffusion is much smaller than that of either convection and axial
diffusion, and concentration gradients in the transverse direction are small
compared to that in the axial direction.
a a
ξ α=
x L
FIG. 1. Schematic diagram of a tubular reactor, showing the length scales in transverse and axial
directions.
macroscale
mesoscale/cell
(exchange to other cells)
∂ijc
y' micro/continuum
z' scale
x' ∂ei (Reactor outlet)
uin
i
ith cell
i
∂bi
∂iin
(shaded area
(Reactor inlet) ∂jic with no mass flux)
(exchange from other cells)
FIG. 2. Schematic diagram of a CSTR illustrating the macro-, meso-, and microscales and the
scale-coupling (Bhattacharya et al., 2004).
The issue of multiple time scales is also present in the case of homogeneous
tank reactors. As in Fig. 2, a tank reactor consists of several circulating flow
loops which exchange material with each other and within which micromixing
occurs at the continuum scale. Therefore, there are three physical length scales
present in a tank reactor. The size of the reactor is the macroscale. The meso
length scale could be anywhere from the size of a circulation loop to the size of
an eddy (or cell). The continuum scale is the microscale. The time scales
A B
L(m) 1 10−3 10−6 10−9

t(s) 10-103 1 10−5 10−7
FIG. 3. Scale separation in a packed-bed reactor: L and t are the corresponding length and time
scales in meters and seconds, respectively.
Precious Metals
(Pt, Pd, Rh)
Channel Diameter : 0.5 – 2 mm Pore Diameter: 10–50 Å

Monolith: Radius 10 cm, Washcoat
Ceramic or Washcoat Thickness: 10 – 50 µm
Length 30–50 cm
Metallic support Support Thickness: 100 – 180 µm
FIG. 4. Schematic diagram illustrating the different length scales in a catalytic monolith reactor.
associated with these three length scales are residence time of the reactor,
circulation/macromixing time, and micromixing time, respectively.
More complex reactors, like packed-bed reactors or catalytic monoliths,
consist of many physically separated scales, with complex nonlinear interactions
between the processes occurring at these scales. Figure 3 illustrates scale
separation in a packed-bed reactor. The four length (and time) scales present in
the system are the reactor, catalyst particle, pore scale, and molecular scale. The
typical orders of magnitude of these four length scales are as follows: reactor,
1 m; catalyst particle, 102 m (1 cm); macropore scale, 1 mm (106 m); micro-
pore/molecular scale, 10 Å (109 m). The corresponding time scales also vary
widely. While the residence time in the reactor varies between 1 and 1000 s, the
intraparticle diffusion time is of the order of 0.1 s and is 105 s inside the
pores. The time scale associated with molecular phenomenon like adsorption is
typically less than a microsecond and could be as small as a nanosecond.
Figure 4 is a picture of another reactor with widely varying scales—a catalytic
monolith. Like the tubular reactor, the monolith itself has two intrinsic length
scales, the radius and the length, which are typically 10–20 cm and between 30
and 50 cm, respectively. The monolith cross-section has a honeycomb structure
consisting of 100–1000 channels, with washcoat and catalyst support. Typical

channel width or hydraulic diameter is 0.5–2 mm, typical washcoat thickness is
10–50 mm, and support thickness is 100–180 mm, while the mean diameter of the
pores present in the catalyst support is 10–100 Å. Thus, it could be seen that the
reaction, which occurs inside the pores of the catalyst support, requires
transport of reactants and products across length scales that vary from a few
angstroms to meters.
In all of the above cases, a strong non-linear coupling exists between reaction
and transport at micro- and mesoscales, and the reactor performance at the
macroscale. As a result, the physics at small scales influences the reactor and
hence the process performance significantly. As stated in the introduction, such
small-scale effects could be quantified by numerically solving the full CDR
equation from the macro down to the microscale. However, the solution of the
CDR equation from the reactor (macro) scale down to the local diffusional
(micro) scale using CFD is prohibitive in terms of numerical effort, and
impractical for the purpose of reactor control and optimization. Our focus here
is how to obtain accurate low-dimensional models of these multi-scale systems
in terms of average (and measurable) variables.
C. DIFFERENT APPROACHES TO MULTI-SCALE AVERAGING OR DIMENSION

REDUCTION
As stated in the introduction, dimension reduction of the governing partial

differential equations describing reactors is necessary for the purpose of design,
control, and optimization of chemical processes, and is typically achieved by
three different approaches, as illustrated in Fig. 5.
The first approach is the discretization of the convection and the diffusion
operators of the PDEs, which gives rise to a large (or very large) system of
effective low-dimensional models. The order of these low-dimensional models
depend on the minimum mesh size (or discretization interval) required to avoid
spurious solutions. For example, the minimum number of mesh points (Nxyz)
necessary to perform a direct numerical simulation (DNS) of convective–diffu-
sion equation for non-reacting turbulent flow is given by (Baldyga and Bourne,
1999)
N xyz Re9=4 Sc3=2 (5)
where Re and Sc are the Reynolds and turbulent Schmidt numbers, respectively.
(For Re ¼ 104 and Sc ¼ 103, this leads to a Nxyz of 3 1013, which is quite
large.) For the case of laminar reacting flows in a tube of radius a and length L,
the minimum number of mesh points required to solve the steady-state one-
dimensional CDR problem has been shown (Dommeti and Balakotaiah, 2000)
Bottom-up Approach Averaging Approach Top-down Approach
Physical System Simplifications/

Assumptions
Detailed Model Dimension

Reduction Low Dimensional
Model
Solution
Solution
Coarse Graining Solution
Macro Description/ Properties
FIG. 5. Different approaches to dimension reduction.
to be
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
Pe3 2Da 2Da2
Nx ¼ q1 þ (6)
24q Pe Pe2
where
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
4Da
q¼ 1þ (7)
Pe
Da is the Damköhler number, Pe the Peclet number based on the length of the
reactor and is given by
L
Pe ¼ ReSc
2a
For very fast reactions i.e. Da 1, criterion equation (6) gives the minimum
number of mesh points required (to avoid spurious solutions) for a three-
dimensional scalar CDR equation as
3=2
f3
N xyz ¼
24
(where f2 ¼ Pe Da), which results in ODEs of very high order, especially for the
case of diffusion-limited reactions, i.e. when f2 is large. For the case of non-
isothermal kinetics, f2 ¼ f20 exp[B], where f20 is the value based on the inlet/
reference temperature and B is the Zeldovich number (dimensionless adiabatic
temperature rise). Therefore, for a typical value of B of 20 (which is very
common for partial oxidation and combustion reactions) and f20 ¼ 1, N xyz ¼
3 1017 and the problem is rendered numerically unsolvable. Dynamical
systems of such high order cannot be easily incorporated in the existing control
strategies. In such cases, low-dimensional models are a natural alternative.
The second approach is the one that chemical engineers have followed
historically to develop low-dimensional models of reactors by making certain a
priori assumptions on the length and time scales of reaction, diffusion, and
convection and applying the conservation equations only at the macroscopic
level. As stated earlier, the CSTR model is an example illustrating this ‘‘top-
down approach’’ (Fig. 5). The assumptions made in developing such low-
dimensional models are usually not justified since it involves comparison of the
solutions obtained with more detailed (fundamental) models, which are not
available. Because of the a priori assumptions, these models cannot explain
many experimentally observed features that arise due to the coupling between
transport processes and chemical reaction at the local scales. When the
predictions of such ad-hoc models did not match with experimental results, the
low-dimensional models were modified by expanding the degrees of freedom
using concepts such as RTD, non-ideal flow and mixing, and introducing
empirical constants such as effective dispersion coefficients (Danckwerts, 1958;
Levenspiel, 1999; Zwietering, 1959). The shortcomings of this approach (such as
the dependence of the effective dispersion coefficients on the kinetic parameters
and inconsistencies such as infinite propagation speed of signals even in
convection-dominated systems) are well known (Balakotaiah and Chang, 2003).
An alternative method of obtaining low-dimensional models is the bottom-up
approach based on the averaging of the detailed models derived from the
fundamental laws. Several different empirical as well as rigorous averaging
techniques (with different terminology such as homogenization, dimension
reduction, adiabatic elimination, multi-scale method, slaving principle, etc.) are
used in different fields for obtaining low-dimensional models. For dynamical
systems with scale separation, the Center Manifold theorem (Carr, 1981) has
been used extensively in recent years to eliminate the slave (or fast decaying)
modes and obtain low-dimensional models described by a few ordinary
differential equations (Balakotaiah and Chang, 1995; Balakotaiah and
Dommeti, 1999; Mercer and Roberts, 1990). While this is a powerful technique,
a major limitation of this technique is that it can only describe the asymptotic
behavior of a physical system close to a fixed point (such as a trivial solution).
This chapter presents spatially averaged chemical reactor models based on the
L–S technique of classical bifurcation theory. This method is best suited for
spatial averaging near a zero eigenvalue (corresponding to the vanishing of a
small parameter representing the ratio of length or time scales in the system).
For the case of CDR problems, local equilibrium exists in the limit when local
diffusion is very fast as compared to convection and reaction [Remark: The
convection time scale varies as L=hux i, where L is the macro length scale and
hux i is the average velocity in the axial direction. The reaction time scale varies
as CR/R(CR), where CR is a reference concentration and R(CR) the reaction
rate. The local diffusion time scale varies as ‘2 =Dm , where ‘ is the meso or micro
length scale and Dm the molecular (or effective) diffusivity. When ‘ ! 0, local
diffusion becomes dominant as the other scales are independent of ‘.
Conversely, by choosing the appropriate length scale ‘, the local diffusion can
be made to be the dominant process and, hence, the spatial degrees of freedom
associated with this length scale can be eliminated]. Dimension reduction of
such systems using L–S technique consists in orthogonal projection of the CDR
operator. The resultant low-dimensional models are described by multiple
concentration and temperature variables, unlike the traditional low-dimensional
models which are described by a single concentration and a single temperature
variable. Each of these variables is representative of a physical scale of a system
and is called a ‘‘mode’’, and the averaged models are called ‘‘multi-mode
models’’. Moreover, spatial averaging by the L–S method retains all the
parameters present in the full CDR equation. Therefore, within their region of
validity, these low-dimensional models retain the complex spatio-temporal
behaviors (multiple solutions, oscillations, micromixing effects, etc.) that are
exhibited by the detailed model that are often missed by the traditional low-
dimensional models.
II. Spatial Averaging of Convection–diffusion–reaction Models

using the L–S Method
The L–S method is a well-known technique for eliminating the degrees of

freedom near a zero eigenvalue. It has been used for analyzing the bifurcation
behavior of a nonlinear operator near a zero eigenvalue (Balakotaiah et al.,
1984; Golubitsky and Schaeffer, 1984). Recently, this technique has been used
for spatial averaging of CDR models provided a scale separation exists
(Balakotaiah, 2004; Balakotaiah and Chakraborty, 2003; Balakotaiah and
Chang, 2003; Chakraborty and Balakotaiah, 2002a, b, 2003, 2004). Here, we
illustrate the main steps of this averaging procedure by considering a single
partial differential equation of the CDR type and show how the local spatial
degrees of freedom (present in the diffusion operator) may be eliminated by
averaging. Application of the procedure for more general cases may be found in
the cited references.
We consider a nonlinear partial differential equation of the form
F ðc; pÞ r2 c pf ðx; y; z; t; c; p; p Þ ¼ 0 (8)
where cðx; y; z; tÞ is a concentration variable dependent on the local coordinates

ðx; yÞ as well as other independent variables z and t, r2 is the diffusion
(Laplacian) operator in the local coordinates x and y in a region O subject to
either zero flux or periodic boundary conditions on the boundary @O. The
parameter p is assumed to be small and is the ratio of local diffusion time to
(global scale) convection time. f is a nonlinear operator that accounts for large
scale mixing (diffusion) and convection effects as well as the source/sink terms
due to reaction, adsorption, etc. It is assumed that f has a Taylor series
expansion in terms of p and the parameters p* appearing in f are of order unity.
We note that the diffusion operator with Neumann (or periodic) boundary
conditions is symmetric and has a simple zero eigenvalue with a constant
eigenfunction. Equivalently, the eigenvalue problem
Lc r2 c ¼ mc in O (9)
rc
n ¼ 0 on @O (10)
is self-adjoint ðL ¼ adjoint operator ¼ LÞ with a zero eigenvalue m0 ¼ 0 and a

constant eigenfunction c0 ðx; yÞ ¼ 1. Moreover, for jX1, mj 40 and for all jX0,
the eigenfunctions cj ðx; yÞ can be chosen to satisfy the orthogonality condition
Z Z (
1 0; iaj
hci ; cj i ¼ ci cj dx dy ¼ dij ¼ (11)
AO O 1; i ¼ j
As shown below, these properties can be used to eliminate the spatial degrees of
freedom in Eq. (8).
Averaging Eq. (8) over the domain O and use of divergence theorem gives
hf ðx; y; z; t; c; p; p Þ; c0 i ¼ 0 (12)
(Here, the angle brackets stand for inner product or average over the domain
O.) For obvious reasons, we shall refer to Eq. (12) as the averaged model. To
write it in a more useful form, we express c as
cðx; y; z; tÞ ¼ hciðz; tÞc0 þ c0 ðx; y; z; tÞ (13)
where
Z Z
1
hci ¼ cðx; y; z; tÞ dx dy ¼ hc; c0 i (14)
AO O
is the spatially averaged concentration (over the domain O). It follows from Eq.
(13) that
hc0 i ¼ hc0 ; c0 i ¼ 0 (15)
For obvious reasons, c0 ðx; y; z; tÞ will be referred to as the local variation. The
L–S procedure uses orthogonal complementary spaces in the domain to split c
as given by Eq. (13). Similarly, in the codomain, Eq. (8) is satisfied iff
EF ðhcic0 þ c0 Þ ¼ 0 (16)
ðI EÞF ðhcic0 þ c0 Þ ¼ 0 (17)

where E is the projection operator onto range L. The complementary projection
(onto ker L) is given by
ðI EÞF ¼ hF ; c0 ic0 (18)
Then, Eq. (17) is identical to Eq. (12), which may be written as
hf ðx; y; z; t; hcic0 þ c0 ; p; p Þ; c0 i ¼ 0 (19)
Simplification of Eq. (16) gives
Lc0 ¼ pf ðx; y; z; t; hcic0 þ c0 ; p; p Þ phf ðx; y; z; t; hcic0 þ c0 ; p; p Þ; c0 ic0 (20)
We refer to this as the local equation. Since L : range L ! range L is invertible,
it follows from the implicit function theorem that the local equation [Eq. (20)]
with the constraint given by Eq. (15) can be solved uniquely for c0 in terms of
hci. Substitution of this in Eq. (19) gives the reduced or averaged model.
The local equation may be solved perturbatively for c0 . Writing
X
1
c0 ¼ pi c i (21)
i¼1
we get
Lc1 ¼ f ðhcic0 ; 0; p Þ hf ðhcic0 ; 0; p Þ; c0 ic0 ; hc1 i ¼ 0 (22)
Lc2 ¼ Dc f ðhcic0 ; 0; p Þ
c1 þ Dp f ðhcic0 ; 0; p Þ
hDc f ðhcic0 ; 0; p Þ
c1 Dp f ðhcic0 ; 0; p Þ; c0 ic0 ¼ 0; hc2 i ¼ 0
etc. Taylor series expansion of Eq. (19) gives
hf ðhcic0 ; 0; p Þ; c0 i þ hDc f ðhcic0 ; 0; p Þ

c0 ; c0 i þ phDp f ðhcic0 ; 0; p Þ; c0 i
þ 2!1 hD2cc f ðhcic0 ; 0; p Þ
ðc0 ; c0 Þ; c0 i þ phD2cp f ðhcic0 ; 0; p Þ
c0 ; c0 i
þ 2!1 p2 hD2pp f ðhcic0 ; 0; p Þ; c0 i þ

¼ 0 ð23Þ
[For simplicity of notation, we have written f ðx; y; z; t; hcic0 ; 0; p Þ as

f ðhcic0 ; 0; p Þ.] Thus, the averaged model to order p2 is given by
hf ðhcic0 ; 0; p Þ; c0 i þ phDc f ðhcic0 ; 0; p Þ

c1 ; c0 i þ phDp f ðhcic0 ; 0; p Þ; c0 i
2
þ p2! hD2cc f ðhcic0 ; 0; p
ðc1 ; c1 Þ; c0 i þ p2 hD2cp f ðhcic0 ; 0; p Þ
c1 ; c0 i
þ p2 hDc f ðhcic0 ; 0; p Þ
c2 ; c0 i þ 2!1 p2 hD2pp f ðhcic0 ; 0; p Þ; c0 i þ

¼ 0 ð24Þ
Here, Dc f , Dp f , D2cc f , D2cp f are the Fréchet derivatives of the nonlinear

operator f and D2cc f
ðci ; cj Þ, is a symmetric bilinear form.
The following observations may be made from the structure of the averaged
model given by Eq. (24): (i) The (zeroth-order) first term is the averaged model
to the lowest order and can be obtained by setting c0 ¼ 0 and p ¼ 0 in Eq. (19).
(ii) The second and third terms represent the order-p corrections. The second
term arises due to elimination of local spatial degrees of freedom. (In physical
terms, this is the combined effect of the interaction of local diffusion, velocity
gradients, and convection/reaction.) The third term is due to order-p effect that
is already present in the function f of the original model. (iii) If the Taylor
expansion of f in powers of p has terms up to order pq (qX0), then the averaged
model has to be derived to order pq so that all the physical phenomena present
at different scales in the original detailed model are also represented in the
averaged model. If this is not the case, then some of the physical phenomena
represented in the original model are not important and can be ignored. (iv)
When Eq. (24) is truncated at order pq (qX1), the truncation error arises from
two sources, the first being the truncation of the Taylor series of the averaged
equation (19), the second being the truncation error of the perturbation
expansion (21) of the local equation. As we show in the following sections, the
first truncation error may be zero in some practical cases (e.g. linear kinetics,
wall reaction case, or solutal dispersion problems in which f is linear in c) and
the averaged equation may be closed exactly, i.e. higher order Fréchet
derivatives are zero and the Taylor expansion of f terminates at some finite
order (usually after the linear and quadratic terms in most applications). In such
cases, the only error is due to the truncation of the solution of the local
equation.
It should be noted that while in principle, it is possible to obtain a single
averaged equation in terms of hci by elimination of c0 (to some desired order in
p), this equation may not be explicit in hci. In addition, this may not be useful or
what is desired in applications. For example, in chemical reactors, it is not the
spatially averaged concentration ðhciÞ that is measured experimentally but the
so-called ‘‘cup-mixing’’ or velocity-weighted concentration defined by
cm ¼ hcðx; y; z; tÞgðx; yÞ; c0 i (25)
where gðx; yÞ (with hgi ¼ hg; c0 i ¼ 1) is the local velocity profile, The relation-
ship between cm and hci may be obtained from Eq. (13) as
cm ðz; tÞ ¼ hciðz; tÞ þ hgðx; yÞc0 ; c0 i ¼ hciðz; tÞ þ phgðx; yÞc1 ; c0 i þ Oðp2 Þ (26)
Now, the averaged model to order p is defined in terms of cm and hci by the
global equation (19) and the local equation (26). The local equation can be
extended to any desired order in p by using higher order approximations of c0 .
This form of the reduced model, expressed in terms of two concentration
variables, will be referred to as the ‘‘two-mode model’’, and is convenient for
physical interpretation of various limiting cases as well as to extend the range of
validity of the local equation and hence the averaged model by a procedure
called regularization.
The above averaging procedure described for Neumann boundary conditions
may be extended to general flux boundary conditions of the form
rc
n þ pDa rw ðcÞ ¼ 0 on @O (27)
where Da is a (reactor scale) Damköhler number (which is of order unity) and rw

(c) is some nonlinear function describing reaction or adsorption at the wall.
When p ! 0, the local gradients vanish and the leading-order operator remains
the same. Thus, the splitting given by Eq. (13) is again valid. However, for this
case, the averaged equation as well as the local equation may contain extra
terms that appear due to the inhomogeneous boundary condition at the wall.
In addition, since the wall reaction term is evaluated at a different concen-
tration (cw or cs) than cm or hci, the resulting model may contain three
concentration variables or three modes. These multi-mode models couple the
local and global equations in terms of physically meaningful and measurable
concentrations.
This spatial averaging theory could also be extended to the non-isothermal
reacting flows to perform spatial averaging of coupled species and energy
balances. This has been discussed in detail by Chakraborty and Balakotaiah
(2004). It should also be pointed out that the above averaging theory is
applicable to discrete interacting units, which is called L–S averaging in finite
dimensions. In this case, Eq. (8) is modified to
Ac þ pfðc; p; p Þ ¼ 0 (28)
where c is a concentration vector ðc 2 Rn Þ and A is an n n interaction or

exchange matrix. Due to mass conservation, A is singular (with columns
summing to zero) and is symmetric (or can be made symmetric with a proper
choice of the inner product in Rn ). For such cases, Eqs. (19) and (20) are valid
(with L replaced by A) and the averaged model is an algebraic equation for the
average concentration hci.
III. Spatially Averaged Models for Describing Dispersion Effects

in Tubes and Packed Beds
In this section, we illustrate the spatial averaging procedure by considering

several simple examples of non-reacting systems that describe dispersion effects
in tubes, packed beds, and monoliths.
A. A HYPERBOLIC AVERAGED MODEL FOR DESCRIBING DISPERSION EFFECTS IN

TUBES/CAPILLARIES
As our first application, we consider the classical Taylor–Aris problem (Aris,

1956; Taylor, 1953) that illustrates dispersion due to transverse velocity
gradients and molecular diffusion in laminar flow tubular reactors. In the
traditional reaction engineering literature, dispersion effects are described by the
axial dispersion model with Danckwerts boundary conditions (Froment and
Bischoff, 1990; Levenspiel, 1999; Wen and Fan, 1975). Here, we show that the
inconsistencies associated with the traditional parabolic form of the dispersion
model can be removed by expressing the averaged model in a hyperbolic form.
We also analyze the hyperbolic model and show that it has a much larger range
of validity than the standard parabolic model.
The dispersion of a non-reactive solute in a circular tube of constant cross-
section in which the flow is laminar is described by the convective–diffusion
equation

@C x02 @C Dm @ 0 @C @2 C
þ 2hu x i 1 ¼ x þ Dm ,
@t0 a2 @x x0 @x0 @x0 @x2
0ox0 oa; x40; t0 40 ð29Þ
with the following boundary and initial conditions:
@C
¼0 at x0 ¼ 0; a (30)
@x0
IC : Cðx; x0 ; 0Þ ¼ f ðx; x0 Þ (31)
BC : Cð0; x0 ; t0 Þ ¼ gðx0 ; t0 Þ (32)
Here, hux i is the average velocity in the pipe, a the radius, and Dm the molecular
diffusivity of the species. [Remark: In writing the boundary condition at x ¼ 0,
we have assumed that L=a 1. If this is not the case, it can be modified to
include the diffusive flux. Also, when L/a is of order unity, we need boundary
conditions at the inlet (x ¼ 0) as well as exit (x ¼ L).] Defining dimensionless
variables
x hux it0 x0 a2 hux i ahux i

z¼ ; t¼ ; x¼ ; p¼ ; Per ¼ (33)
L L a LDm Dm
we can write Eqs. (29) and (30) as

1 @ @C @C 2 @C p @2 C @C
LC x ¼p þ 2ð1 x Þ ; ¼ 0 at x ¼ 0; 1
x @x @x @t @z Pe2r @z2 @x
(34)
Here, p is the local (transverse) Peclet number, which is the ratio of transverse
diffusion time to the convection time. Per is the radial Peclet number (ratio of
transverse diffusion time to a convection time based on pipe radius). We assume
that p 1 while Per is of order unity. (Remark: The parameter Pe2r =p ¼
hux iL=Dm is also known as the axial Peclet number. Also note that for any finite
Per or tube diameter, the axial Peclet number tends to infinity as p tends to
zero.) When such scale separation exists, we can average the governing equation
over the transverse length scale using the L–S technique and obtain averaged
model in terms of axial length and time scales.
We note that the transverse operator L is symmetric with respect to the inner
product
Z 1
ðv; wÞ ¼ 2xvðxÞwðxÞ dx (35)
0
It has a zero eigenvalue with normalized eigenfunction of unity. We define the

mixing-cup (velocity weighted) and spatial average concentrations by
Z 1
Cm ¼ 4xð1 x2 ÞCðx; z; tÞ dx (36)
0
Z 1
hCi ¼ 2xCðx; z; tÞ dx (37)
0
Transverse averaging of Eq. (34) gives
@hCi @C m p @2 hCi
þ ¼0 (38)
@t @z Pe2r @z2
We note that when p ¼ 0, hCi ¼ C m and substitution of this into Eq. (38) gives
the leading (or zeroth) order evolution equation for the averaged concentration:
@hCi @hCi
þ ¼0 (39)
@t @z
(Remark: This zeroth-order model is the ideal plug-flow model.) To obtain the
averaged equation to order p, we write
Cðx; z; tÞ ¼ hCiðz; tÞ þ C 0 ðx; z; tÞ; C 0 2 ker L (40)

and solve for the slave variable C 0 ðx; z; tÞ in terms of hCiðz; tÞ using the
procedure outlined above. To leading order, we have

@hCi 1 x2 x4
C 0 ðx; z; tÞ ¼ p þ þ Oðp2 Þ (41)
@z 12 4 8
Substitution of the Eq. (40) and transverse averaging (after multiplying by the
velocity profile) gives the local equation relating Cm and hCi:
p @hCi p @C m
C m hCi ¼ þ Oðp2 Þ ¼ þ Oðp2 Þ (42)
48 @z 48 @z
This local equation (when written in dimensional form) defines a characteristic

transfer time between the slowly evolving mode Cm and the slave mode
C m hCi. Thus, the averaged model to order p is given by
@hCi @C m p @2 hCi
þ ¼0 (43)
@t @z Pe2r @z2
p @C m
C m hCi ¼ (44)
48 @z
For obvious reasons, we shall refer to Eq. (43) as the global equation while Eq.
(44) that couples the local and global scales as the local equation. Although we
can leave the reduced model in this two-mode form, in this special case, we can
combine the two equations to obtain a single equation either for Cm or hCi. In
this specific example, the model is linear in the concentration and hence both Cm
and hCi satisfy the same equation (as they are linearly related). Since the cup-
mixing concentration (which is often measured in experiments) is more relevant
in applications, we write the reduced model in terms of Cm:
@C m @C m p @2 C m p @2 C m
þ þ þ Oðp2 Þ ¼ 0 (45)
@t @z 48 @z@t Pe2r @z2
We note that this averaged model is hyperbolic. The third term in Eq. (45)
represents the Taylor dispersion term (due to velocity gradients and transverse
molecular diffusion) while the last term is the Aris-correction term (representing
the influence of axial molecular diffusion). In dimensional form, the reduced
model may be written as
@C m @C m @2 C m @2 C m
þ hu x i þ hu x itD Dm ¼ 0; t0 tD ; x ‘ D (46)
@t0 @x @x@t0 @x2
where the local diffusion or mixing time is defined by

a2
tD ¼ (47)
48Dm
The corresponding local length scale is given by ‘D ¼ hux itD while the diffusivity
may be written as Deff ¼ hux i2 tD . When axial molecular diffusion is neglected
(or equivalently, in the limit of Per ! 1) Eq. (46) simplifies to
@C m @C m @2 C m
0
þ hux i þ hux itD ¼0 (48)
@t @x @x@t0
In his famous paper, Taylor (1953) used the leading-order approximation
@C m @C m
¼ hux i (49)
@t0 @x
to express the mixed derivative term as a dispersion term and simplified Eq. (48) to
@C m @C m @2 C m
þ hux i ¼ Deff ; Deff ¼ hux i2 tD (50)
@t 0 @x @x2
In the literature, Deff is also known as the Taylor dispersion coefficient.
However, the approximation used by Taylor transforms a hyperbolic equation
into a parabolic equation. In the chemical engineering literature, this
approximation is made worse by the further requirement of an artificial
boundary condition at the exit of the tube. During the last 50 years, the
parabolic model with Danckwerts’ boundary conditions is used extensively to
describe dispersion effects in chemical reactors. We show here that the
hyperbolic form of the model is more accurate, retains the proper physics,
can describe dispersion effects more accurately than the parabolic model and is
valid in a much larger domain of the physical parameter space. [Remark: A
good analogy between the parabolic and hyperbolic models is the approxima-
tion of the function ex for small positive x by f p ðxÞ ¼ 1 x and
f h ðxÞ ¼ 1=ð1 þ xÞ. Both approximations have the same accuracy for x ! 0
but the first approximation breaks down qualitatively for x41 while the second
approximation is valid qualitatively for all positive x. The second (Pade)
approximation is a regularized version of the first function. This regularization
is closely connected with how we write the local equation. Although the local
equation is an infinite series in powers of p, we can truncate it (often at the first
term) and rewrite it so that it is qualitatively valid for all positive values of p.
This regularization is discussed in Section VI.]
When the Taylor approximation is used in Eq. (46) the averaged model is
again parabolic but now the effective dispersion coefficient is given by
a2 hux i2
Deff ¼ hux i2 tD þ Dm ¼ þ Dm (51)
48Dm
This result was first derived by Aris (1956) using the method of moments. While
the resulting model now includes both the effects (axial molecular diffusion and
dispersion caused by transerverse velocity gradients and molecular diffusion) it
has the same deficiency as the Taylor model, i.e. converting a hyperbolic model
into a parabolic equation.
We note that when Dm hux i2 tD , or equivalently, the radial Peclet number
Per 6:93, axial diffusion can be neglected. The local Peclet number p, which is
equal to Per times the aspect ratio (a/L), can be small even when Per 6:93,
provided the aspect ratio is sufficiently small. Thus, the conditions Per 6:93
and p 1 are usually satisfied in most tubular reactors. In such cases, it is more
appropriate to use the leading-order approximation to modify the small axial
dispersion term to a mixed derivative term and write the averaged model as
@C m @C m @2 C m Dm a2 Dm
þ hu x i þ hu x i ^D
t ¼ 0; t^D ¼ tD þ ¼ þ (52)
@t0 @x @x@t0 hux i2 48Dm hux i2
Now, the averaged hyperbolic model, Eq. (52), defines a characteristic initial-
value problem (Cauchy problem). To complete the model, we need to specify
Cm only along the characteristic curves x0 ¼ 0 and t0 ¼ 0. Thus, the initial and
boundary conditions for the averaged model are obtained by taking the mixing-
cup averages of Eqs. (31) and (32):
Z 1
0
C m ðx; t ¼ 0Þ ¼ 4xð1 x2 Þf ðx; RxÞ dx f m ðxÞ (53)
0
Z 1
C m ðx ¼ 0; t0 Þ ¼ 4xð1 x2 ÞgðRx; t0 Þ dx gm ðt0 Þ (54)
0
When the assumption Per 6:93 is not valid, it is better to leave the averaged
model in the more general hyperbolic form given by Eqs. (45) or (46) with
boundary and initial conditions given by Eqs. (53)–(54). The important point to
be made is the wave forms of the averaged model [either Eqs. (46) or (52)] have
much larger domain of validity than the parabolic form as shown below.
(Remark: The hyperbolic model is also much easier to solve numerically
compared to the parabolic model.)
The L–S method can be used to derive the averaged model to higher orders in
p but we will not be pursue it here (see Section VI). In fact, since our averaged
model at order p is also regularized, higher order approximations are not
necessary to see the qualitative behavior for all positive values of p.
Before closing this section, we compare the hyperbolic model derived here
with the wave model of Westerterp et al. (1995). For the classical Taylor
problem (without axial dispersion), the model of Westerterp et al. may be
written in the present notation as
@hCi @hCi @J
þ þ ¼0 (55)
@t @z @z
p @J p @J p @hCi
Jþ þ ¼ (56)
15 @t 12 @z 48 @z
where J is the dimensionless flux (the flux normalized by hux iC R , CR being a
reference concentration). The following observations may be made about this
model: (i) The second and third terms on the l.h.s. of Eq. (56) are Oðp2 Þ terms.
Hence, this model is parabolic at order p and is not different from the classical
Taylor model (at this order). In contrast, the model derived by the L–S method
is hyperbolic at order p. (ii) The model is not expressed in terms of measurable
quantities such as the cup-mixing concentration. (iii) The steady-state version of
the model predicts that there is dispersion even in the case of constant (and time
invariant) inlet conditions. The dispersion coefficient predicted by the steady-
state version of this model is quite different from the transient value. This
dispersion is spurious and arises due to the choice of the concentration variable
hCi since integration of the full PDE model (without the axial dispersion term)
over the tube cross-section shows that the cup-mixing concentration remains
constant with axial position. (iv) The model predicts that there are two waves
moving in the axial direction, whereas in the physical problem there is only one
wave.
1. Comparison of Solutions of Parabolic and Hyperbolic Models

We now present the solution of the hyperbolic model defined by Eqs. (52) and
(53)–(54) and compare the solution to that of the classical parabolic model with
Danckwerts boundary conditions. We use the axial length and convective time
scales to non-dimensionalize the variables and write the hyperbolic model in the
following form:
@C m @C m @2 C m
þ þZ ¼ 0; t Z; z Z (57)
@t @z @z@t
C m ðz; t ¼ 0Þ ¼ f ðzÞ (58)
C m ðz ¼ 0; tÞ ¼ gðtÞ (59)
where Z is now the effective local Peclet number (or the dimensionless local
mixing time) defined by

hux it^D a2 hux i Dm 1 1
Z¼ ¼ þ ¼p þ (60)
L 48LDm hux iL 48 Pe2r
and is the ratio of effective transverse diffusion time to the convection time.
The exit concentration C m ðz ¼ 1; tÞ for the case of a unit impulse (Delta

function) input (f ðzÞ ¼ 0, gðtÞ ¼ dðtÞ) is known as the dispersion (or RTD)
curve. For the hyperbolic model, this can be found either by Laplace
transformation or from the general solution of the model [see Balakotaiah
and Chang (2003), for a general analytical solution of Eqs. (57)–(59)]. It is easily
seen that the Laplace transform of the dispersion curve is given by

b ¼ Exp s
EðsÞ (61)
1 þ sZ
while the dispersion curve is given by

pffiffi
ð1 þ tÞ 1 2 t
E h ðtÞ ¼ C m ð1; tÞ ¼ Exp dðtÞ þ pffiffi I 1 (62)
Z Z t Z
We note that the second central moment (or the dimensionless variance) of the
dispersion curve is given by
s2 ¼ 2Z (63)
Thus, for Z ! 0, the variance approaches zero and the behavior approaches
that of plug flow. For Z small and t 1, Eq. (62) may be simplified to
pffiffi
1 ð1 tÞ2
E h ðtÞ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Exp (64)
4pZt3=2 Z
Hence, the dispersion curve has a peak at t ¼ 1 and is slightly asymmetrical (but
the deviation from the Gaussian curve is small). As Z increases to 0.5, it can be
shown by analyzing Eq. (62) that the peak moves to t ¼ 0 and for any Z40:5,
the peak remains at t ¼ 0. For any Z40:5, the dispersion curve has a slow
decaying (long) tail and the variance can exceed unity. Figure 6 shows the
dispersion curves for different values of Z. [Note: For simplicity, the delta
function of magnitude Expð1=ZÞ at t ¼ 0 is not shown in the figure.]
We now compare the solution of the hyperbolic model with that of the
parabolic model used widely in the literature to describe dispersion in tubular
reactors. The parabolic model with Danckwerts boundary conditions (in
dimensionless form) is given by
@C @C 1 @2 C
þ ¼ ; 0ozo1; t40 (65)
@t @z Pe @z2
1 @C
C ¼ gðtÞ at z ¼ 0 (66)
Pe @z
1.4
1.2
η = 0.05
1
0.8 0.1
E(t)
0.6
0.
0.3
0.4 0.
0.4
0.5
0.2
0
0 1 2 3 4 5 6
t
0.5 η = 0.5
0.4
1
E(t) 0.3
0.2
2
0.1
5
0
0 1 2 3 4 5 6
t
FIG. 6. Dispersion curves predicted by the hyperbolic model [Eq. (57)] for various values of the
effective local Peclet number, Z.
@C
¼ 0 at z ¼ 1 (67)
@z
Cðz; t ¼ 0Þ ¼ f ðzÞ (68)
where Pe is the axial Peclet number. For the parabolic model (with f ðzÞ ¼ 0 and
gðtÞ ¼ dðtÞ), the dispersion curve is given by
pffiffiffiffiffi pffiffiffiffiffi
2 X1
Pe Pe2 þ 4ln ln sin ln ðPe2 þ 4ln Þ
E p ðtÞ ¼ Cðz ¼ 1; tÞ ¼ Exp t
Pe n¼1 2 4Pe ðPe2 þ 4Pe þ 4ln Þ
(69)
where
pffiffiffiffiffi
pffiffiffiffiffi ln Pe
cot ln ¼ pffiffiffiffiffi ; n ¼ 1; 2; . . . (70)
Pe 4 ln
The dimensionless variance can be found more easily from the Laplace
transform and is given by
2 2
s2 ¼ ð1 ePe Þ (71)
Pe Pe2
We note that the variance for the parabolic model is always bounded between
zero (for Pe ! 1 or plug flow) and unity (Pe ! 0 or ideal CSTR behavior).
Thus, the parabolic model can only describe dispersion behavior that lies
between these two extremes. In contrast, the hyperbolic model can describe the
same behavior when Z varies between 0 and 0.5 as well as the bypassing,
stagnant region or solute retaining behavior (with long tails as in segregated
laminar flow) when Z40:5. For very small Z, the dispersion curves predicted by
the two models are very close to each other but the hyperbolic model predicts an
asymmetric curve with a slightly higher peak than the parabolic model. In
addition, the parabolic model predicts upstream diffusion and infinite
propagation speed. Both these non-physical phenomena are not present in the
hyperbolic model which retains the qualitative behavior of the full model for all
values of Z. Thus, we conclude that the hyperbolic model describes dispersion
effects better than the parabolic model and is valid over a wider range of the
physical parameter space.
We also present solutions of the full hyperbolic model
@C m @C m @2 C m @2 C m 48
þ þZ lZ ¼ 0; l¼ (72)
@t @z @z@t @z2 Pe2r
for the case of a unit impulse input. The Laplace transform of the dispersion
curve is now given by
( pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi)
b 1 þ sZ s2 Z2 þ 2sZð1 þ 2lÞ þ 1
EðsÞ ¼ Exp (73)
2Zl
from which we can obtain the second central moment (variance) as
s2 ¼ 2^Z; Z^ ¼ Zð1 þ lÞ (74)
This shows again that when l is small (or equivalently, Per 1), we can
combine the small axial dispersion term with the mixed derivative term and
simplify the general hyperbolic model [Eq. (72)] to the simpler model [Eq. (57)].
However, for l values of order unity or larger, this cannot be justified. The
inverse transform of Eq. (73) can be found by integrating around the branch
points but we will not pursue it here. Instead, we show in Figs. 7 and 8 the
numerically determined dispersion curves for Z ¼ 0:1; 1 and various values of l.
As can be expected, the qualitative behavior of the full hyperbolic model [Eq.
(72)] is similar to that of the simpler case of l ¼ 0. Only for lX1, the peak value
changes and shifts to lower times.
1.5
2
λ = 48/Per
η = 0.1
λ = 10
0.1
E(t)
0.5
0.5
0
0 1 2 3 4 5 6
t
FIG. 7. Dispersion curves predicted by the full hyperbolic model [Eq. (72)] for Z ¼ 0:1.
2. Hyperbolic Model for Describing Dispersion Effects in Turbulent flow in a

Tube
The hyperbolic model given by Eq. (48) is also applicable for describing
dispersion in turbulent flow in a tube. In this case, we replace the local diffusion/
mixing time tD by the effective local (turbulent) mixing time tmix, which can be
calculated using the Reynolds-averaged convective–diffusion model, and is
given by Eq. (164) and b1 could be obtained by using an universal velocity
profile (as shown in the following section) as given in Eq. (165). For the case of
turbulent flow with L=a 1, axial diffusion effects can be ignored and the
hyperbolic model becomes even more appropriate compared to the laminar flow
case.
3. Hyperbolic Model for Describing Dispersion Effects in Packed Beds

The traditional parabolic model with Danckwerts’ boundary conditions is
also used in the literature to describe dispersion effects in packed beds (and
porous media). However, unlike the case of capillaries and straight tubes, the
flow field in packed beds is more complex and is three-dimensional. However,
for many cases of interest, the average velocity in the transverse directions is
zero. In such cases, dispersion in the flow direction can be described by the
4
λ = 48/Per2
3.5 η =1
3
λ = 10
2.5
E(t)
2
5
1.5
1
1
0.1
0.5
0
0 0.5 1 1.5 2 2.5 3
t
FIG. 8. Dispersion curves predicted by the full hyperbolic model [Eq. (72)] for Z ¼ 1.
hyperbolic model. For the case of isotropic beds with no mean flow in the
transverse directions, the hyperbolic model may be expressed as
@C m @C m @2 C m
0
þ hux i þ hux itD ¼ DT r2T C m (75)
@t @x @xqt0
where e is the bed porosity (assumed to be constant), hux i the velocity in the
mean flow direction, and r2T is the Laplacian operator in transverse coordinates
(radial and azimuthal). Here, the local mixing time (tD) and the transverse
dispersion coefficient (DT) may be expressed as
DT ¼ Dm þ aT hux id (76)
Dm d
tD ¼ 2
þ aX (77)
hux i hu xi
where d is the particle diameter and the numerical coefficients aX and aT depend
on the structure of the bed. For the case of a randomly packed bed of spherical
particles, typical values are aT ¼ 0:1 and aX ¼ 0:5. The molecular contribution
to dispersion, represented by the first term in the above expressions, can be
neglected except at very small values of the local Peclet number ðhux id=Dm Þ.
We do not pursue the solution of this new packed-bed dispersion model but
note that it can be used to explain Hiby’s (1962) experimental result, for which
the traditional parabolic model cannot provide a good explanation.
B. MULTI-MODE HYPERBOLIC AVERAGED MODELS FOR DESCRIBING DISPERSION

EFFECTS IN CHROMATOGRAPHS
We now extend the averaging method to derive hyperbolic models to describe

dispersion effects in chromatographs. We consider the case of a single solute
being adsorbed on the wall of a tube in which the flow is laminar. Assuming
Langmuir adsorption and neglecting axial molecular diffusion (Per 1), the
governing partial differential equations (assuming azimuthal symmetry) may be
written as

@C A x02 @C A Dm @ 0 @C A
þ 2hux i 1 2 ¼ 0 x ; 0ox0 oa; x40 (78)
@t0 a @x x @x0 @x0
@C A 0 @C A 0
Dm ðx ¼ a; z; tÞ ¼ ka C Aw C s kd C As ; ðx ¼ 0; z; tÞ ¼ 0 (79)
@x0 @x0
@C As
¼ ka C Aw C s kd C As (80)
@t0
C Aw ¼ C A ðx0 ¼ a; z; tÞ; C s þ C As ¼ C 0 (81)

with appropriate inlet and initial conditions. Here, ka and kd are the adsorption
and desorption rate constants and C Aw is the solute concentration at the wall
and the other symbols have their usual meaning. Scaling the solute
concentration using some reference inlet concentration (C A0 ), adsorbed
concentration by the total concentration of sites C 0 ðy ¼ C As =C 0 Þ, time, radial,
and axial coordinates as in the Taylor problem (using convection time, tube
radius and length, respectively) and defining dimensionless parameters
2 C0 ka C A0 a2 ka C A0
G¼ ; K¼ ; Dal ¼ (82)
a C A0 kd Dm
the dimensionless model equations may be written as

1 @ @C @C @C
LC x ¼p þ 2ð1 x2 Þ (83)
x @x @x @t @z
@C
¼0 at x ¼ 0 (84)
@x

1 @C G y
ðx ¼ 1; z; tÞ ¼ C w ð1 yÞ ; C w ¼ Cðx ¼ 1; z; tÞ (85)
Dal @x 2 K

G @y y
¼ C w ð1 yÞ (86)
Dal @t K
with initial and inlet conditions

Cðx; z; t ¼ 0Þ ¼ C 0 ðx; zÞ (87)
yðz; t ¼ 0Þ ¼ y0 ðzÞ (88)
Cðx; z ¼ 0; tÞ ¼ C in ðx; tÞ (89)
Here, Cðx; z; tÞ is the scaled solute concentration in the fluid phase, Cw the solute
concentration at the wall, y the normalized adsorbed concentration ð0pyp1Þ,
K the adsorption equilibrium constant, p the transverse Peclet number, G
represents the adsorption capacity (ratio of adsorption sites per unit tube
volume to the reference solute concentration), and Dal is the local Damköhler
number (ratio of transverse diffusion time to the characteristic adsorption time).
We shall assume that p 1 while G and Dal are order-one parameters. (In
physical terms, this implies that transverse molecular diffusion and adsorption
processes are much faster compared to the convection.)
Transverse averaging of the above model using the procedure outlined in
Section II gives the following averaged model to order p:
@hCi @C m @y
þ þG ¼0 (90)
@t @z @t

@y y
p ¼ Dal C w ð1 yÞ (91)
@t K
p @hCi p @hCi
C w hCi ¼ þ (92)
8 @t 6 @z
p @hCi p @hCi
C m hCi ¼ (93)
24 @t 16 @z
hCiðz; t ¼ 0Þ ¼ hC 0 ðx; zÞi; yðz; t ¼ 0Þ ¼ y0 ðzÞ; C m ðz ¼ 0; tÞ ¼ C m;in ðtÞ (94)
The averaged model is a four-mode model, the four modes (or variables) being
the fluid-phase cup-mixing concentration (Cm), the fluid-phase average
concentration (hCi), the solute concentration at the wall (Cw), and the adsorbed
concentration (y). [Note: Since the initial and boundary conditions for the
averaged model are obtained in the same manner as in the Taylor problem by
taking transverse averages of Eqs. (87) and (89), we do not consider them any
further.] We now consider various limiting cases of this model.
For the case of p ¼ 0 (which corresponds to adsorption, desorption, and
transverse diffusion time scales going to zero), we have
KhCi
C m ¼ C w ¼ hCi; y¼ (95)
1 þ KhCi
and the above model reduces to the widely used zeroth-order hyperbolic model
(with no dispersion) (Rhee et al., 1986):
@hCi @hCi @y
þ þG ¼0 (96)
@t @z @t
The first non-trivial case we consider is that of linear adsorption and desorption.
For this case, we have y 1 and the model becomes linear. For small p, we can
use the leading-order approximation to further simplify the model by
eliminating the variables Cw and hCi and write it in terms of Cm:
@C m 1 @C m @2 C m
þ þ pL ¼0 (97)
@t ð1 þ gÞ @z @z@t
1 1 þ 6g þ 11g2 1 g2
L¼ þ (98)
48 ð1 þ gÞ2 GDal ð1 þ gÞ2
2 ka C 0
g ¼ GK ¼ (99)
a kd
(Note: Since the model is linear for the special case considered, the same
equation is also satisfied by the other three variables.) The following
observations may be made from Eq. (98) that expresses the dimensionless
dispersion coefficient L: (i) The first term describes dispersion effects due to
velocity gradients when adsorption equilibrium exists at the interface. We note
that this expression was first derived by Golay (1958) for capillary
chromatography with a retentive layer. (ii) The second term corresponds to
dispersion effects due to finite rate of adsorption (since this term vanishes if we
assume that adsorption and desorption are very fast so that equilibrium exists at
the interface). (iii) The effective dispersion coefficient reduces to the Taylor limit
when the adsorption rate constant or the adsorption capacity is zero. (iv) As is
well known (Rhee et al., 1986), the effective solute velocity is reduced by a factor
(1 þ g). (v) For the case of irreversible adsorption (g ! 1 and Dal ! 1), the
dispersion coefficient is equal to 11 times the Taylor value. It is also equal to the
reciprocal of the asymptotic Sherwood number for mass transfer in a circular
channel with constant wall flux boundary condition. (vi) When the local
Damköhler number is small (adsorption is slow or capacity is low), L can be
large, leading to long tails in the dispersion curve.
The second limiting case we consider is that of Langmuir adsorption with
equilibrium at the interface. This assumption is equivalent to assuming Dal !
1 and replacing Eq. (91) by
KC w
y¼ (100)
1 þ KC w
Combining Eqs. (92) and (93), we get
p @hCi 11p @hCi

Cw ¼ Cm þ þ
6 @t 48 @z
p @C m 11p @C m
¼ Cm þ þ þ Oðp2 Þ ð101Þ
6 @t 48 @z
Thus, we can obtain the following single evolution equation for the cup-mixing
concentration:
@C m @C m p @2 C m p @2 C m
þ þ 2
þ
@t @z 24 @t 16 @z@t
0 1
pK @C m 11pK @C m
@B KC m þ þ
þG @ 6 @t 48 @z C A¼0 ð102Þ
@t pK @C m 11pK @C m
1 þ KC m þ þ
6 @t 48 @t
This equation reduces to the zeroth-order hyperbolic model [Eq. (95)] for p ¼ 0.
It also reduces to the hyperbolic model treated in the previous section for either
G ¼ 0 or K ¼ 0 (after combining the third and fourth terms using the leading-
order approximation). For finite p, G, and K, this rigorously derived averaged
model is quite different from the intuitively written models in the literature.
The last case we consider is that of a flat velocity profile with linear
adsorption and desorption. In this case, the dimensionless model equations are
given by

1 @ @C @C @C
x ¼p þ (103)
x @x @x @t @z
@C
¼0 at x ¼ 0 (104)
@x

1 @C G y
ðx ¼ 1; z; tÞ ¼ C w ð1 yÞ ; C w ¼ Cðx ¼ 1; z; tÞ (105)
Dal @x 2 K

@y y
p ¼ Dal C w ð1 yÞ (106)
@t K
Now, because of flat velocity profile, the distinction between cup-mixing and
averaged concentrations disappears and the averaged model to order p is a
three-mode model given by
@hCi @hCi @y
þ þG ¼0 (107)
@t @z @t

@y y
p ¼ Dal C w ð1 yÞ (108)
@t K

p @hCi @hCi
C w hCi ¼ þ (109)
8 @t @z
For the case of linear adsorption, we can combine these equations into a single
hyperbolic equation:
@hCi 1 @hCi @2 hCi

þ þ pL ¼0 (110)
@t ð1 þ gÞ @z @z@t
1 g2 1 g2
L¼ þ (111)
8 ð1 þ gÞ2 GDal ð1 þ gÞ2
Comparing Eqs. (97) and (110), we see that the adsorption induced dispersion is
independent of the velocity profile. We also note that for the case of flat velocity
profile, there is no dispersion when g ¼ 0. When equilibrium is assumed at the
wall, we can eliminate y and Cw and write the averaged equation as
0 1
pK @hCi pK @hCi
@hCi @hCi @B KhCi þ þ
þ þG @ 8 @t 8 @z C A¼0 (112)
@t @z @t pK @hCi pK @hCi
1 þ KhCi þ þ
8 @t 8 @z
As expected, for p ¼ 0, this model reduces again to the zeroth-order hyperbolic
model, Eq. (95), but for any finite p, it does not simplify to any of the standard
models in the literature.
The order p terms that appear in Eqs. (102) and (112) modify the leading-
order hyperbolic behavior by introducing dispersion (which is always present in
real systems due to velocity gradients and finite rates of adsorption). As is well
known in the literature, the leading (zeroth)-order hyperbolic models may have
discontinuous solution profiles (Rhee et al., 1986). As stated in the introduction,
in the literature, these models are often modified by adding a dispersion term
and transforming them to a parabolic form. The above analysis shows that the
parabolic form of the reduced model is not only a poor approximation but also
cannot be justified based on physical grounds (and rigorous derivation). The
L–S method of averaging modifies the zeroth-order hyperbolic models by
adding order p corrections and gives the averaged models in multi-mode
(hyperbolic) form. As shown above, in some special cases (e.g. linear adsorption
or equilibrium at the wall) it is possible to transform them to a single-mode form
which is a hyperbolic equation containing the dispersion effects. When this
simplification is not possible, it is best to leave the averaged models in multi-
mode form, which is more convenient for numerical solution.
C. A HYPERBOLIC MODEL FOR DESCRIBING DISPERSION EFFECTS IN MONOLITHS

WITH DIFFUSION INTO THE SOLID PHASE
We can extend the hyperbolic model to cases in which the solute diffuses in
more than one phase. A common case is that of a monolith channel in which the
flow is laminar and the walls are coated with a washcoat layer into which the
solute can diffuse (Fig. 4). The complete model for a non-reacting solute here is
described by the convection–diffusion equation for the fluid phase coupled with
the unsteady-state diffusion equation in the solid phase with continuity of
concentration and flux at the fluid–solid interface. Transverse averaging of such
a model gives the following hyperbolic model for the cup-mixing concentration
in the fluid phase:
@C fm @C fm @2 C fm @2 C fm
þ þ L1 ð; mÞp L2 ð; mÞp ¼0 (113)
@t @z @z@t @z2
Here, t and z have been scaled with respect to the convection time and length of
monolith, respectively. The transverse or local Peclet number p and the radial
Peclet number Per are defined as in the case of the Taylor problem, e is the
volume fraction of the fluid phase and

1 1 Des
L2 ð; mÞ ¼ 2 þ ; m¼ (114)
Per m Dm
(Des is the effective diffusivity of the solute in the solid phase while Dm is the
molecular diffusivity in the fluid phase.) The function L1 ð; mÞ depends on the
shape of the channel and the washcoat surrounding it. For the case of a circular
channel with a uniform washcoat thickness, it can be shown that
62 16 þ 11 m
L1 ð; mÞ ¼ þ ð4 2 2 2 ln Þ (115)
48 8
We note that when ¼ 1, Eq. (113) reduces to the Taylor problem (with axial
diffusion included). As e decreases or m increases, the magnitude of dispersion
increases leading to long tails in the dispersion curve. Once again, the hyperbolic
model describes this solute retaining behavior much better than the traditional
parabolic models.
IV. Spatially Averaged Multi-mode (Multi-scale) Models for

Homogeneous Reactors
In this section, we present spatially averaged multi-scale models for different

types of homogeneous reactors. We consider a single homogeneous reaction
involving M species, which is given by
X
M
nj A j ¼ 0 (116)
j¼1
where, nj is the stoichiometric coefficient of species Aj, and we use the usual
convention of nj 40 if Aj is a product and nj o0 if Aj is a reactant.
A. ISOTHERMAL TUBULAR REACTORS
1. Laminar Flows
For a constant density system, the scalar concentration C j ðx0 ; j; x; t0 Þ of
species j in a tubular reactor of uniform cross-section O with unidirectional
laminar flow (Fig. 1) obeys the CDR equation
@C j @C j
0
þ ux ðx0 Þ nj RðC 1 ; C 2 ; . . . ; C M Þ
@t @x

1 @ 0 @C j 1 @2 C j @2 C j
¼ Dm;j 0 0 x þ 02 ð117Þ
x @x @x0 x @j2 @x2
where ux ðx0 Þ is the fully developed velocity field, Dm;j the molecular diffusivity of
species j, RðC j Þ the intrinsic reaction rate, and x,j, and x0 are the axial,
azimuthal, and radial coordinates, respectively. In the transverse direction, Eq.
(117) is subject to the no-flux boundary condition at the wall,
rC j
n ¼ 0 (118)
where n is the unit normal to the boundary @O and r is the gradient operator in
O, while the inlet condition is of Danckwerts’ type,
@C j
Dm;j ¼ ux ðx0 Þ½C j ðx0 ; j; x; t0 Þ C j;in ðx0 ; j; t0 Þ at x ¼ 0 (119)
@x
and the exit condition is
@C j
¼0 at x ¼ L (120)
@x
Using a (radius of the pipe) and L (length of the pipe) as the characteristic
lengths in the transverse and axial directions, respectively, CR as the reference
concentration, and Dm,R as the reference molecular diffusivity, we obtain four
time scales in the system associated with convection ðtC Þ, local/transverse
diffusion (tD), axial diffusion (tZ), and reaction (tR),
V L a2 L2 CR
tC ¼ ¼ ; tD ¼ ; tZ ¼ ; tR ¼ (121)
qin hux i Dm;R Dm;R RðC R Þ
where V and qin are the volume of the reactor and the volumetric flow rate of the
reactants, respectively. The ratios of these time scales give rise to the
dimensionless parameters: p (transverse or local Peclet number), Per (radial
Peclet number), Da (reactor scale Damköhler number), and f2 (local
Damköhler number), given by
a2 hux i tD hux ia hux iL tZ Pe2r

p¼ ¼ ; Per ¼ ; Pe ¼ ¼ ¼
LDm;R tC Dm;R De;R tC p
Dm;R LRðC R Þ tC a2 RðC R Þ tD
kj ¼ ; Da ¼ ¼ ; f2 ¼ pDa ¼ ¼ ð122Þ
Dm;j huz iC R tR Dm;R C R tR
We use the following dimensionless variables:
t0 x0 x ux C RðC 1 ; C 2 ; . . . ; C M Þ
t¼ ; x¼ ; z¼ ; u¼ ; cj ¼ ; rðcÞ ¼
tC a L hux i CR RðC R Þ
to rewrite Eq. (117) in dimensionless form as

1 @ @cj 1 @2 c j @cj p @2 c j @cj
x þ 2 ¼ pkj þ uðxÞ n j Da rðcÞ (123)
x @x @x x @j2 @t kj Pe2r @z2 @z
D pf ðc; p; p Þ (124)
As in the previous problems, we assume that p 1 while Da, kj, and Per are
order-one parameters. The boundary and initial conditions on the model are
given by
p @cj
¼ uðxÞ½cj cj;in ðx; j; tÞ at z ¼ 0 (125)
kj Pe2r @z
@cj
¼0 at z ¼ 1 (126)
@z
@cj
¼0 at x ¼ 0; 1 (127)
@x
cj ðx; j; z; tÞ ¼ cj ðx; j þ 2p; z; tÞ (128)
cj ðx; j; z; t ¼ 0Þ ¼ cj;0 ðx; j; zÞ (129)
Transverse averaging of the above model using the procedure outlined in the
previous section for the case of laminar flow in a tube, i.e. uðxÞ ¼ 2ð1 x2 Þ, with
azimuthally symmetric feeding gives the following two-mode model for the jth
species (j ¼ 1; 2; . . . ; M), involving the spatially averaged concentration hcj i and
the mixing-cup concentration cj;m to order p:
@hcj i @cj;m p @2 hcj i

þ nj Da rðhciÞ þ Oðp2 Þ ¼ 0 (130)
@t @z kj Pe2r @z2
@cj;m
hcj i cj;m ¼ b1 pkj þ Oðp2 Þ (131)
@z
with boundary and initial conditions given by
p @hcj i
¼ cj;m cj;m;in ðtÞ; at z ¼ 0 (132)
kj Pe2r @z
@hcj i
¼ 0; at z ¼ 1 (133)
@z
hcj i ¼ hcj;0 iðzÞ; at t ¼ 0 (134)
where hcj i is the transverse-averaged concentration in the reactor and cj;m is the
mixing-cup concentration, which for the tubular geometry are given by
Z Z
1 x¼1 j¼2p
hcj i ¼ 2xcj ðx; j; zÞ dj dx (135)
2p x¼0 j¼0
R x¼1 R j¼2p
x¼0 j¼0 cj uðxÞ2x dj dx
cj;m ¼ R1 (136)
2p 0 uðxÞ2x dx
1
respectively. The numerical coefficient b1 ¼ 48 for fully developed laminar flow
2 1
in a pipe. (Note: b1 ¼ 105 for plane Poiseuille flow while it is 30 for Couette
flows.) It may be noted that the low-dimensional model defined by Eqs.

(130)–(134) has the same number of parameters as the full CDR equations. For
p ¼ 0, the model reduces to the zeroth-order plug-flow model, while for any
p40, the plug-flow model is modified by including the effects of local velocity
gradients and diffusion (micromixing) as well as reactor scale mixing. In this
reduced model, reactor scale mixing (macromixing) is captured by the third
term in Eq. (130) while the local mixing (micromixing) is captured through the
local equation [Eq. (131)].
We now consider some important limiting cases of this model. First, we note
that when axial molecular diffusion is negligible, Eqs. (130)–(131) reduce to
hyperbolic model:
@cj;m @cj;m @2 cj;m

þ þ b1 pkj nj Da rðhciÞ ¼ 0 (137)
@t @z @z@t
@cj;m
hcj i cj;m ¼ b1 pkj (138)
@z
cj;m ¼ cj;m;in ðtÞ at z ¼ 0; hcj i ¼ hcj;0 iðzÞ; at t ¼ 0 (139)
Under steady-state conditions (and assuming that the inlet condition is

independent of time), this two-mode model can be further simplified to
dcj;m
¼ nj Da rðhciÞ with cj;m jz¼0 ¼ cj;m;in (140)
dz
dcj;m
hcj i cj;m ¼ b1 pkj ¼ b1 pkj nj Da rðhciÞ (141)
dz
which in dimensional form may be written as
dC j;m
hux i ¼ nj RðhCiÞ with C j;m jx¼0 ¼ C j;m;in (142)
dx
hC j i C j;m ¼ tmix;j nj RðhCiÞ (143)
where
a2 tmix
tmix;j ¼ b1 ¼ (144)
Dm;j kj
is the local mixing time of the jth species and tmix is the characteristic local
mixing time of the system defined in terms of the reference species. The
dimensionless form of the characteristic local mixing time of the system (Z), is
obtained by non-dimensionalizing tmix w.r.t. the residence time in the reactor
and is given by
tmix a2 =Dm;R
Z¼ ¼ b1 ¼ b1 p (145)
tC tC
The solutions of the steady-state two-model model given by Eq. (140)–(141)
should be compared to the parabolic axial dispersion model with Danckwerts
boundary conditions (Danckwerts, 1953; Wehner and Wilhelm, 1956):
1 d2 hcj i dhcj i
þ nj Da rðhciÞ ¼ 0; 0ozo1 (146)
kj Pe dz2 dz
1 dhcj i dhcj i
¼ hcj i hcj;in i at z ¼ 0; ¼ 0 at z ¼ 1 (147)
kj Pe dz dz
The Danckwerts model lumps the combined effect of axial diffusion, velocity
gradients, and transverse molecular diffusion into an effective axial dispersion
coefficient (axial Peclet number, Pe). However, when L=a 1 (or Per 1), the
two-mode model represents dispersion using the local mixing time and two
concentration variables. As in the Taylor problem, the exit concentration
predicted by the Danckwerts model is always bounded between the two limiting
cases of plug flow (Pe ¼ 1) and CSTR (Pe ¼ 0). This is not the case for the two
models defined by Eqs. (140)–(141). While for p ! 0 (or Z ! 0) the solution
approaches the plug-flow limit, as p (or Z) increases the conversions can be
below those obtained in a CSTR. This is the so-called mixing limited asymptote
which is similar to the mass transfer controlled case for the case of catalytic
reactions. The single-mode Danckwerts model cannot describe (even qualita-
tively) this mixing (or more precisely micromixing) limited regime.
This feature is illustrated in Fig. 9, which compares the solution of the ideal
PFR, ideal CSTR, and the Danckwerts model [Eqs. (146)–(147)] with the two-
mode convection model [Eqs. (140)–(141)] for the case of steady-state and single
first-order homogeneous reaction of the form A ! B. The solution of the
steady-state Danckwerts model is given by
4q expðPe=2Þ
X ¼1 (148)
Peq Peq
ð1 þ qÞ2 exp ð1 þ qÞ2 exp
2 2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where q ¼ 1 þ 4Da=Pe and X is the conversion of A, while the solution of the
steady-state two-mode convection model is given by

Da
X ¼ 1 exp (149)
1 þ ZDa
and the variances ðs2 Þ of the models are given by Eqs. (71) and (63),
1
PFR Model Ideal CSTR Model
Danckwerts Model (Pe = 10) η=5
η = 0.09
0.8
η = 0.75
η=1
0.6
0.4
Two-mode convection model
0.2
0
0.1 1 10 100
Da
FIG. 9. Comparison of the solutions of ideal PFR, CSTR, and Danckwerts models with that of
two-mode convection model for an isothermal first-order homogeneous reaction.
respectively. We note that for small values of s2 , both models predict
1 X ¼ expðDaÞ½1 þ s2 Da2 þ OðZ2 Þ (150)
Needless to mention that the Danckwerts model [Eqs. (146)–(147)] reduces to

the ideal PFR and the ideal CSTR models in the limits of Pe ! 1 and Pe ! 0,
respectively. As seen in Fig. 9, the ideal PFR model and the ideal CSTR model
bound the solution of the Danckwerts model [Eqs. (146)–(147)] for the case of
any finite Pe number. As shown in Fig. 9 (and from Eq. (150)), for PeX10, the
two-mode model predicts very close to the Danckwerts model when the
variances of the two models are matched using Eqs. (71) and (63). However, the
two-mode models can predict conversions in regions in the parameter space
which are inaccessible to the Danckwerts models. These are the micromixing
limited conversions which are even below conversions predicted by the ideal
CSTR model. As illustrated in Fig. 9, these solutions of the two-mode model

attain mixing limited asymptotes for large values of Da.
The second important limiting case of the general model [Eqs. (130)–(134)] we
consider is that of the steady-state case in which macro (axial) and micro
(transverse) mixing effects are of the same order of magnitude. Assuming that
the inlet concentrations are independent of time, the general model simplifies to
dcj;m p d2 hcj i
nj Da rðhciÞ þ Oðp2 Þ ¼ 0 (151)
dz kj Pe2r dz2
dcj;m dhcj i
hcj i cj;m ¼ b1 pkj þ Oðp2 Þ ¼ b1 pkj þ Oðp2 Þ (152)
dz dz
with boundary conditions given by
p dhcj i
¼ cj;m cj;m;in ; at z ¼ 0 (153)
kj Pe2r dz
dhcj i
¼ 0; at z ¼ 1 (154)
dz
Eqs. (151)–(153) may be combined and written as
2
dhcj i p d hcj i
b1 pkj þ nj Da rðhciÞ þ Oðp2 Þ ¼ 0 (155)
dz kj Pe2r dz2

p dhcj i
b1 pkj þ 2
¼ hcj i cj;m;in ; at z ¼ 0 (156)
kj Per dz
Thus, we recover the Danckwerts model only if no distinction is made between

the cup-mixing and spatial average concentrations (with this assumption, the
effective axial dispersion coefficient is given by the Taylor–Aris theory). This
derivation also shows that the concept of an effective axial dispersion coefficient
and lumping the macro- and micromixing effects into one parameter is valid
only at steady-state, constant inlet conditions and when the deviation from plug
flow is small. [Remark: Even with all these constraints, the error in the model
because of the assumption hcj i ¼ cj;m is of the same order of magnitude as the
dispersion effect!]
The last limiting case of the general model we consider is that of a tubular
reactor in which macromixing effect is dominant compared to the micromixing
effect. (This could be the case when the tube diameter or length are of the same
order of magnitude or when there is physical backmixing). Here, the general
model given by Eqs. (130)–(134) is valid again but we replace p=Pe2r by 1/Pe,
where Pe is the axial Peclet number, which is of order unity. Now, we have a
two-parameter two-mode mixing model where the magnitude of Pe is an

indication of the macromixing while that of p gives the micromixing effect. For
example, for the case of steady state, the model may be written as
1 d2 hcj i dcj;m
þ nj Da rðhciÞ ¼ 0; 0ozo1 (157)
kj Pe dz2 dz
dcj;m
hcj i cj;m ¼ b1 pkj (158)
dz
1 dhcj i dhcj i
¼ cj;m cj;m;in at z ¼ 0; ¼ 0 at z ¼ 1 (159)
kj Pe dz dz
For p ¼ 1=Pe ¼ 0, this model reduces to the ideal PFR model, for p ¼ 0, it
reduces to the Danckwerts model, for 1/Pe ¼ 0, it reduces to the two-mode
plug-flow model while for Pe ¼ 0, it reduces to the two-mode CSTR model
(discussed below). Thus, we have shown that the hyperbolic two-mode model
given by Eqs. (130)–(134) has a much larger region of validity than the
traditional homogeneous tubular reactor models.
2. Turbulent Flows
The extension of the two-mode axial dispersion model to the case of fully
developed turbulent flow in a pipe could be achieved by starting with the time-
smoothed (Reynolds-averaged) CDR equation, given by Eq. (117), where the
reaction rate term R(C) in Eq. (117) is replaced by the Reynolds-averaged
reaction rate term Rav(C), and the molecular diffusivity Dm;j is replaced by the
effective diffusivity De;j in turbulent flows given by
De;j ¼ Dm;j þ DT (160)
where DT is the turbulent diffusivity, which could be obtained on the basis of

turbulent shear stress and expressed in terms of Reynolds number using one of
the correlations presented by Taylor (1954), Sittel et al. (1968), or Wen and Fan
(1975).
The Reynolds-averaged reaction rate Rav(c) could be evaluated by using
simple closure models of Bourne and Toor (1977), Brodkey and Lewalle (1985),
Li and Toor (1986), Dutta and Tarbell (1989), etc. It should also be pointed out
that the spatial averaging presented here is independent of the methodology
using which Rav(c) and DT are evaluated, or in other words, spatial averaging
follows time averaging.
We use an universal velocity distribution obtained by Churchill (2001) to
approximate the fully developed velocity profile uðxÞ across the turbulent core,
given by
rffiffiffiffi " rffiffiffiffi# !
ft ft 15 2 10 3
uðxÞ ¼ 5:5 þ 2:5 ln ð1 xÞRe þ ð1 xÞ ð1 xÞ (161)
2 8 4 3
where ft is the Fanning friction factor and Re is the Reynolds number.

Using the velocity profile given by Eq. (161) to solve the local equations, we
obtain the leading-order fluctuations cj;1 as
" #
pffiffiffiffi @hcj i 1 2:684x 0:895x2 þ 1:79x3 3:355x4 þ 1:431x5
cj;1 ¼ 0:1647 f t
@z þ0:536oð2x2 1Þ þ 2:68ðx2 1Þ ln½1 x
(162)
where
rffiffiffiffi! sffiffiffiffi
ft 2
o ¼ 5:5 þ 2:5 ln Re (163)
8 ft
The important result is that the two-mode models for a turbulent flow tubular
reactor are the same as those for laminar flow tubular reactors. The two-mode
axial dispersion model for turbulent flow tubular reactors is again given by Eqs.
(130)–(134), while the two-mode convection model for the same is given by Eqs.
(137)–(139), where the reaction rate term rðhciÞ is replaced by the Reynolds-
averaged reaction rate term rav ðhciÞ. The local mixing time tmix,j for turbulent
flows is given by
a2
tmix;j ¼ b1 (164)
Dm;j þ DT
where b1 is given by
b1 ¼ 0:1f t ðo 3:45Þ
"rffiffiffiffi# sffiffiffiffi!
ft 2
¼ 0:1f t 2:05 þ 2:5 ln Re ð165Þ
8 ft
We note that in turbulent flows, typically Dm;j DT ðj ¼ 1; . . . ; MÞ, a result of

which, the local mixing time tmix is practically independent of the molecular
diffusivity or the molecular Schmidt number, i.e. tmix;j tmix;T ðj ¼ 1; . . . ; MÞ,
where
a2
tmix;T ¼ b1 (166)
DT
B. LOOP AND RECYCLE REACTORS
In this section, we present the two-mode models for loop and recycle reactors.
In a loop reactor (Fig. 10) of loop length L, a flow rate of qin with an average
velocity of huin i enters and leaves the reactor at points x ¼ 0 and x ¼ l,
respectively (where x is the length coordinate along the loop), and the total flow
rate in the loop is Q+qin between points x ¼ 0 and x ¼ l, and is Q between
points x ¼ l and x ¼ L, due to a recycle rate of Q. The recycle ratio L is the
ratio of the volume of fluid returned to the reactor entrance per unit time to the
volume of fluid leaving the system per unit time, and is given by L ¼ Q=qin .
Neglecting axial diffusion (or assuming L=a 1), the three-dimensional
CDR equation of a laminar flow loop reactor is given by

@C j 0 @C j 1 @ 0 @C j 1 @2 C j
þ u in ðx Þ n j RðC 1 ; C 2 ; . . . ; C M Þ ¼ D m;j x þ
@t0 @x x0 @x0 @x0 x02 @j2
(167)
with boundary and initial conditions being given by
C j;in þ LC j ðx0 ; j; x ¼ LÞ
C j ðx0 ; j; x ¼ 0Þ ¼ (168)
1þL
C j ðx0 ; j; x ¼ l þ Þ ¼ C j ðx0 ; j; x ¼ l Þ (169)
@C j
¼ 0 at x0 ¼ 1 (170)
@x0
Q+q
q x=0
x=L
q
x=l
FIG. 10. Schematic diagram of a loop reactor with a single (premixed) feed and product streams.
C j ðx0 ; j; xÞ ¼ C j ðx0 ; j þ 2p; xÞ (171)
C j ðt0 ¼ 0Þ ¼ C j0 (172)
Using the spatial averaging procedure illustrated in Section II, we average Eqs.
(167)–(171) in the transverse direction to obtain the two-mode model for a loop
reactor, which is given by
8
> 1 C j;m hC j i
>
> ; 0pxol
@C j;m < 1 þ L tmix;j
huin i ¼ (173)
@x >
> 1 C j;m hC j i
>
: L ; lpxpL
tmix;j
@hC j i C j;m hC j i
¼ þ nj RðhCiÞ; 0pxoL (174)
@t0 tmix;j
(for j ¼ 1; 2; . . . ; M), with the boundary and initial conditions being

C j;m;in; þ LC j;m ðx ¼ LÞ
C j;m ðx ¼ 0Þ ¼ (175)
1þL
hC j iðx ¼ l Þ ¼ hC j iðx ¼ l þ Þ (176)
hC j iðt0 ¼ 0Þ ¼ C j0 (177)
For the special case when no reaction occurs between x ¼ l and x ¼ L, i.e.
C j;m ðx ¼ lÞ ¼ C j;m ðx ¼ LÞ, the loop reactor reduces to a recycle reactor of length
l, the two-mode model for which is given by

@C j;m 1 C j;m hC j i
huin i ¼ (178)
@x 1þL tmix;j
@hC j i C j;m hC j i
¼ þ nj RðhCiÞ; 0pxol (179)
@t0 tmix;j
with the boundary condition
C j;m;in þ LC j;m ðx ¼ lÞ
C j;m ðx ¼ 0Þ ¼ (180)
1þL
and initial condition being given by Eq. (177).
The two-mode loop and recycle reactor models are two-parameter two-mode
models. Here, the two parameters are the recycle ratio L, and the dimensionless
local mixing time tmix =tc , which describe macro- and micromixing effects in the
system, respectively.
C. TANK REACTORS (CSTRS)
1. Two-mode Model for Premixed Feed

It is well known that as the recycle ratio L of a recycle reactor is increased, the
behavior shifts from a PFR at L ¼ 0 (no macromixing) to a CSTR at L ¼ 1
(perfect macromixing). We use this limit to obtain the two-mode model for a
perfectly macromixed CSTR, by integrating Eq. (178) along the length of the
reactor x and simplifying the resulting equation for L 1. This gives the two-
mode model for a CSTR as
dhC j i C j;m hC j i
¼ þ nj RðhCiÞ; with hC j iðt0 ¼ 0Þ ¼ C j;0 (181)
dt0 tmix;j
C j;m hC j i C j;m;in C j;m

¼ (182)
tmix;j tC
where tC ð¼ V =qin Þ is the total residence time in the reactor and tmix is the
characteristic local mixing time. Equations (181) and (182) constitute a two-
mode one-parameter model for a perfectly macromixed CSTR where feed enters
the tank as a single (premixed) stream. Micromixing effects are captured
through the local mixing time tmix, and in the limit of complete micromixing (i.e.
tmix ! 0), it reduces to the ideal one-mode zero-parameter CSTR model.
2. Two-mode Two-mixing Time Model for a General CSTR

In real tanks, both micro- and macromixing effects are important, and are
influenced by several factors including the type and speed of impellers, number,
and position of baffles, and the manner of feed distribution. Macromixing
effects in tanks have often been modeled by using compartment models in the
mixing literature (Baldyga and Bourne, 1999). Recently, Bhattacharya et al.
(2004) used the L–S technique to develop a two-mode two-mixing time CSTR
model that in addition to capturing micromixing, accounts for macromixing
effects resulting from unmixed feed. Here, we summarize this model for the
isothermal case.
Consider a single-phase homogeneous stirred-tank reactor with a time-
invariant velocity field Uðx0 ; y0 ; z0 Þ a single reaction of the form A-B. (This
approach can be extended to the case of time-dependent velocity fields. If the
flow in the tank is turbulent, then the velocity field is the solution of the
Reynolds averaged Navier–Stokes equations). The tank is divided into a three-
dimensional network of n spatially fixed volumetric elements, or n-interacting
cells. Theoretically speaking, n can be arbitrarily large but the size of any cell is
such that the length scale associated with it is at least one order of magnitude
larger than the continuum scale. Here, the continuum scale is the microscale, the
length scale of any cell is the mesoscale, and the reactor scale is the macroscale.
In practice, the division of the reactor into n cells depends intricately on various
design parameters including stirrer and baffle positions, feed distributions,
number of circulation zones, etc. Inside each cell, the reactant is transported by
diffusion and convection and is consumed by reaction, while interacting with
the other cells through the cell boundaries. Depending on the location, the
boundary of cell i may be divided into the following types: (i) @Oin i is the
boundary through which the reacting fluid enters the tank through cell i with
in in e
flow rate qini ðC i ; T i Þ, (ii) @Oi is the boundary through which the reacting fluid
leaves the reactor through cell i with flow rate qei ðC ei ; T ei Þ, (iii) @Ocij , j ¼ 1; 2; . . . ; n
ðjaiÞ is the boundary through which cell i interacts with cell j with circulation
flow rate qcij ðC ci;m ; T ci;m Þ, (iv) @Ocji , j ¼ 1; 2; . . . ; n ðjaiÞ is the boundary through
which cell j interacts with cell i with circulation flow rate qcji ðC cj;m ; T cj;m Þ, and (v)
@Obi is the boundary with no mass flux (as shown in Fig. 2), where the quantities
in the bracketed terms represent the concentration and temperature of the
corresponding streams, respectively. Here, cci;m ðT ci;m Þ and ccj;m ðT cj;m Þ are the mean
cup-mixing concentrations (temperatures) of the circulating stream leaving cell i
and j, respectively. [Remark: For non-isothermal CSTR with external or internal
cooling (or heating), cell i may have boundary @Oyi , through which it exchanges
heat with cooling (or heating) fluid ðT i;c Þ.] In the first step of the reduction
process, the mass and energy balance (CDR) equations in each cell, which
describes mixing at the local scale and are infinite-dimensional in nature, are
reduced by L–S technique as described in Section II. This is followed by a finite-
dimensional reduction on n cells, resulting in the final reduced model for the
whole tank containing both micro- as well as macromixing effect (see
Bhattacharya et al., 2004, for details). The final reduced model that describes
both macro- and micromixing effects in an isothermal CSTR is given by
dhCi 1
þ RðhCiÞ ¼ ðC in CmÞ (183)
dt0 tc m
1
C m hCi ¼ ðtmix;2 C in
m tmix;1 C m Þ (184)
tc
where tmix,1 is the overall mixing time of the tank, which depends on the local
variables (such as local velocity gradients, local diffusion length, diffusivity) as
well as reactor scale variables (such as baffle position, stirrer type, circulation
time or stirrer speed, feed pipe locations, etc.), while tmix,2 captures the effect of
non-uniform feeding. For the case of all feed streams premixed and entering as a
single stream, i.e. uniform feeding:
tmix;2 ¼ tmix;1 ¼ tmix (185)
and Eqs. (183) and (184) reduce to Eqs. (181)–(182). For the general case, both
tmix,1 and tmix,2 have contributions from two terms and can be written as
tmix;1 ¼ td n1 þ tE n2 ¼ tmicro;1 þ tmacro;1 (186)

|{z} |{z}
micromixing macromixing
tmix;2 ¼ td n3 þ tE n4 ¼ tmicro;2 þ tmacro;2 (187)

|{z} |{z}
micromixing macromixing
where td ¼ l 2d =De is the local diffusion time, which is a measure of local scale
mixing or micromixing present within the tank and tE is the exchange or
macromixing time of the whole tank. (Remark: tE is inversely proportional to
the impeller rpm.) The coefficients ni, i ¼ 1; 2; 3; 4, are functions of reactor
geometry (including number, design and positions of baffles and stirrers, feed
positions, etc.) as well as feed distributions, where the coefficients n3 and n4,
associated with tmix,2, capture the effect of non-uniform or distributed feeding
through their additional dependency on feed compositions. For turbulent flow
within the tank, the local diffusion length ld depends on kinematic viscosity as
well as the power input per unit mass and can be as small as Kolmogorov length
scale, while ld for laminar flow is much bigger and can be as large as reactor
length scale. On the other hand, the exchange time, which depends on both
circulation or exchange flow rate and reactor volume, is in general inversely
proportional to the impeller speed. The local diffusion time enters in the
expressions for mixing times from the reduction of local CDR equation from
micro to meso scale, while the exchange time appears through the second step of
reduction from meso scale to reactor scale. The contribution of these two scales
of mixing appears in the overall mixing time as a combined function of reactor
geometry, feed distribution and composition, etc. and cannot be separated from
each other. It should be pointed that the coefficients (ni) can be positive as well
as negative.
D. NON-ISOTHERMAL REACTOR MODELS
In this section, we present the low-dimensional multi-mode models for

different types of non-isothermal homogeneous reactors. Here, we skip the
details of the averaging process and summarize the results.
1. Tubular Reactors
In obtaining the low-dimensional models for non-isothermal tubular reactors,
we start with the full energy balance equations in conjunction with the species
balance Eqs. (117)–(119). For the case of a non-isothermal homogeneous
reactor the energy balance equations are given by fluid- and solid-state thermal
balances which are coupled through the boundary conditions. The fluid
temperature T f ðx0 ; j; x; t0 Þ obeys the following energy balance equation

@T f 0 @T f
rf C pf þ u x ðx Þ þ ðDH R ÞRðC 1 ; C 2 ; . . . ; C M ; T f Þ
@t0 @x

1 @ 0 @T f 1 @2 T f @ 2 T f
¼ kf 0 0 x þ 02 þ ð188Þ
x @x @x0 x @j2 @x2
while the temperature T S ðx0 ; j; x; t0 Þ in the solid wall (of thickness al) of the
channel is given by the equation

@T S @ @T S
rf C pf 0 ¼ r?
ðkS r? T S Þ þ kS (189)
@t @x @x
Equation (188) is subject to the boundary and initial conditions
@T f
¼0 at x0 ¼ 0 (190)
@x0
T f ðx0 ¼ a; j; x; t0 Þ ¼ T S ðx0 ¼ a; j; x; t0 Þ (191)
T f ðx0 ; j; x; t0 Þ ¼ T f ðx0 ; j þ 2p; x; t0 Þ (192)
kf @T f @T f
¼ ux ðx0 Þ½T f ðx0 ; xÞ T in at x ¼ 0; ¼ 0 at x ¼ L (193)
rf C pf @x @x
T f ðx0 ; j; x; t0 ¼ 0Þ ¼ T f0 (194)
respectively, while the boundary and initial conditions for Eq. (189) describing
the solid temperature are
@T S @T f @T S
kS ¼ kf 0 at x0 ¼ a; kS ¼ hðT S T C Þ at x0 ¼ að1 þ lÞ
@x0 @x @x0
(195)
T S ðx0 ; j; x; t0 Þ ¼ T S ðx0 ; j þ 2p; x; t0 Þ (196)

@T S
¼ 0 at x ¼ 0; L (197)
@x
T S ðx0 ; j; x; t0 ¼ 0Þ ¼ T S0 (198)
respectively, where TC is the coolant temperature and al is the thickness of the

solid wall. In addition to the dimensionless parameters defined in Eq. (122), we
define some new parameters to non-dimensionalize the energy balance
equations. These dimensionless parameters are Lef (fluid Lewis number), St
(Stanton number), Bi (Biot number), B (Zeldovich number), g (dimensionless
activation energy), a (coolant capacity), etc., which are given by
kf 2hL
Lef ¼ ; St ¼ ,
De;R rf C pf rf C pf að1 þ lÞhux i
2hC R St
a¼ ¼
rf C pf að1 þ lÞRðC R ; T in Þ Da

2ah p E lf kf rS C pS
Bi ¼ ¼ St; g¼ ; ¼ (199)
kf ð1 þ lÞ Lef RT in lS kS rf C pf
kf rS C pS hux iLrf C pf p gðDH R ÞC R

k¼ ; ssf ¼ ; PeS ¼ ¼ ; B¼
kS rf C pf kS kLef Pe2r rf C pf T in
The dimensionless temperature variables for the fluid and solid phases (yf and
yS) and the dimensionless reaction rate ðrðc; yf ÞÞ are given by
T f T in T S T in RðC 1 ; C 2 ; . . . ; C M ; T f Þ
yf ¼ g ; yS ¼ g ; rðc; yf Þ ¼ (200)
T in T in RðC R ; T in Þ
Using the above dimensionless parameters and variables, Eqs. (188)–(198) are
written in dimensionless form and spatially averaged over transverse dimensions
to obtain the low-dimensional model for non-isothermal homogeneous tubular
reactors, which is given to order p by Eqs. (130)–(134) with rðhciÞ being replaced
by rðhci; hyf iÞ and
0 2 1
p @ hyf i 1 @2 hyS i
@hyf i @hyS i @yfm B 2 þ C
þ ssf ð1 Þ þ ¼ @ Per Lef @z2 k @z2 A
@t @t @z
þBDa rðhci; hyf iÞ StðhyS i yC Þ
(201)

yfm hyf i @yfm 1 @hyS i
¼ þ b2 ssf ð1 Þ þ StðhyS i yC Þ (202)
ZH @z @t

hyS i hyf i @yfm b4 1 @hyS i
¼ b3 ssf ð1 Þ þ StðhyS i yC Þ (203)
ZH @z b1 @t
with the boundary and initial conditions being given by
p @hyf i @yfm
¼ yfm yfm;in ; at z ¼ 0; ¼ 0 at z ¼ 1,
Lef Pe2r @z @z
@hyS i
¼ 0; at z ¼ 0; 1 ð204Þ
@z
hyf i ¼ hyf0 i; and hyS i ¼ hyS0 i at t ¼ 0 (205)
where, e is the volume fraction of the fluid phase in the system, hyf i and hyS i are
the transverse averaged temperatures of the fluid- and solid phases, respectively,
and yfm is the mixing-cup temperature of the fluid phase, which are given by
1
¼ (206)
ð1 þ lÞ2
R x¼1 R j¼2p
x¼0 j¼0 xyf ðx; j; z; tÞ dj dx
hyf i ¼ R x¼1 R j¼2p (207)
x¼0 j¼0 x dj dx
R x¼1 R j¼2p
x¼0 j¼0 xuðxÞyf dj dx
yfm ¼ R x¼1 R j¼2p ¼ hyf i þ hu0 y0f i (208)
x¼0 j¼0 xuðxÞ dj dx
R x¼1þl R j¼2p
x¼1 j¼0 xyS ðx; j; z; tÞ dj dx
hyS i ¼ R x¼1þl R j¼2p (209)
x¼1 j¼0 x dj dx
In Eqs. (202)–(203), ZH is the characteristic dimensionless local thermal mixing

time, which is given by
tmix;H a2 =lf p
ZH ¼ ¼ b1 ¼ b1 (210)
tC tC Lef
The coefficients b1, b2, and b3 depend on the flow profile and the local shear
rates of the system, and b4 depends on the reactor geometry and for a tubular
geometry,
b4 ¼ 18 (211)
For the case of laminar flow in a tube,

1
b1 ¼ 48 (212)
b2 ¼ 2 (213)
b3 ¼ 2 (214)
while for the case of fully-developed turbulent flows in tubes,
b1 ¼ 0:1f t ðo 3:45Þ (215)
1:41
b2 ¼ pffiffiffiffi (216)
f t ðo 3:45Þ

0:89 1:55
b3 ¼ pffiffiffiffi 1 (217)
ft o 3:45
where o is given by Eq. (163).

In the above model [Eqs. (201)–(203)], thermal micromixing is captured
through the dimensionless local thermal mixing times of the system, ZH, as an
exchange of energy between the two temperature modes yf;m and hyf i by the
local equation (202). The other local equation [Eq. (203)] captures the energy
transfer between the fluid- and solid phases as an exchange between the average
fluid ðhyf iÞ and solid ðhyS iÞ temperatures. It should also be pointed out that our
low-dimensional model retains all the parameters ðp; Per ; Lef ; B; Da; St; k; ssf Þ
present in the full CDR model, and in the limit of complete mass and thermal
micromixing (Z ! 0, ZH ! 0), our model reduces to the plug-flow model for
tubular reactors with wall cooling.
Low-dimensional models for loop, recycle, and tank reactors could similarly
be derived starting from the coupled mass and thermal balances. Here, we
present the reduced models and refer to a previous publication (Chakraborty
and Balakotaiah, 2004) for the derivation of these models.
2. Loop and Recycle Reactors

a. Loop reactors. The low-dimensional model for non-isothermal homoge-
neous loop reactor (as shown in Fig. 10) is given by
@hcj i cj;m hcj i

¼ þ nj Da rðhci; hyf iÞ; 0ozo1 (218)
@t Zkj
8
> 1 cj;m hcj i l
>
>
< ; 0ozol;^ where l^ ¼
@cj;m 1 þ L Zk j L
¼ ðj ¼ 1; 2; . . . ; MÞ
@z >
> 1 cj;m hcj i ^
>
: L ; lozo1
Zkj
(219)

@hyf i b yfm hyf i hyS i hyf i
¼z b3 þ 4 þ ð1 þ b2 Þ þ BDa rðhci; hyf iÞ,
@t b1 ZH ZH
0ozo1 ð220Þ
8
> z b4 yfm hyf i hyS i hyf i
>
>
< þ b2 ; 0ozol^
@yfm 1 þ L b1 ZH ZH
¼ (221)
@z >
> z b4 yfm hyf i hyS i hyf i ^
>
: L b þ b 2 ; lozo1
1 ZH ZH
for the fluid-phase mass and thermal balances, respectively, with boundary and
initial conditions being given by
cj;m;in þ Lcj;m jz¼1
cj;m jz¼0 ¼ ; hcj ijz¼l^ ¼ hcj ijz¼l^þ (222)
1þL
hcj i ¼ cj;0 at t ¼ 0 (223)
yfm;in þ Lyfm jz¼1

yfm jz¼0 ¼ ; hyf ijz¼l^ ¼ hyf ijz¼l^þ (224)
1þL
hyf i ¼ yf0 at t ¼ 0 (225)

The solid-phase thermal balance is given by

@hyS i yfm hyf i hyS i hyf i
ssf ð1 Þ ¼ z b3 þ StðhyS i yC Þ (226)
@t ZH ZH
for 0pzp1, with the initial condition being given by hyS i ¼ yS0 at t ¼ 0, and
where z ¼ b1 =ðb4 b1 b2 b3 Þ.
b. Recycle reactors. For the special case when no reaction occurs between
z ¼ l^ and z ¼ 1, i.e. cj;m jz¼l^ ¼ cj;m jz¼1 , the loop reactor reduces to a recycle
reactor (0ozo1), the low-dimensional model for which is given by
@hcj i cj;m hcj i
¼ þ nj Da rðhci; hyf iÞ (227)
@t Zkj

@cj;m 1 cj;m hcj i
¼ (228)
@z 1þL Zkj

@hyf i b4 yfm hyf i hyS i hyf i
¼z b3 þ þ ð1 þ b2 Þ þ BDa rðhci; hyf iÞ (229)
@t b1 ZH ZH

@yfm z b4 yfm hyf i hyS i hyf i
¼ þ b2 (230)
@z 1 þ L b1 ZH ZH
cj;m;in þ Lcj;m jz¼1
cj;m jz¼0 ¼ (231)
1þL
yfm;in þ Lyfm jz¼1

yfm jz¼0 ¼ (232)
1þL
The solid-phase balance [Eq. (226)] with accompanying boundary and initial
conditions remain unaltered.
3. Tank Reactors
The low-dimensional model for a non-isothermal homogeneous tank reactor
with premixed feed is given by
dhcj i
þ ðcj;m cj;m;in Þ ¼ nj Da rðhci; hyf iÞ (233)
dt
hci cj;m
cj;m cj;m;in ¼ (234)
Zkj
as the fluid-phase species balance equation,

dhyf i b yfm hyf i hyS i hyf i
¼ z b3 þ 4 þ ð1 þ b2 Þ þ BDa rðhci; hyf iÞ (235)
dt b1 ZH ZH

b4 yfm hyf i hyS i hyf i
hyf i ¼ z þ b2 (236)
b1 ZH ZH
as the fluid-phase thermal balance equation, and

dhyS i yfm hyf i hyS i hyf i
ssf ð1 Þ ¼ z b3 þ StðhyS i yC Þ (237)
dt ZH ZH
as the solid-phase thermal balance, where z ¼ b1 =ðb4 b1 b2 b3 Þ, and the initial
conditions are given by hci ¼ c0 , hyf i ¼ yf0 , and hyS i ¼ yS0 at t ¼ 0. The bj’s and
z are the five independent constants that could be obtained for each tank
depending on the local flow profile and the shape of the tank.
For the special case of a simple reaction A ! B, the low-dimensional model
for a CSTR with premixed feed consists of three differential equations and two
algebraic equations. When the mass and thermal micromixing effects are
ignored (Z ¼ ZH ¼ 0), cm ¼ hci, hyf i ¼ yfm ¼ hyS i, and we get the classical
pseudohomogeneous CSTR model
dhci
¼ ðcin hciÞ Da rðhci; hyf iÞ (238)
dt
dhyf i St
Le ¼ hyf;in i hyf i þ BDa rðhci; hyf iÞ ðhyf i yC Þ (239)
dt
where
1
Le ¼ 1 þ ssf (240)

E. MULTIPLE REACTIONS
Here, we extend the low-dimensional models derived for the case of a single
reaction to the case of multiple homogeneous reactions represented by
X
M
nij Aj ¼ 0; i ¼ 1; 2; . . . ; N H (241)
j¼1
involving M species in NH homogeneous reactions occurring in a constant

density system, in which the species obey the laws of binary diffusion. We use
the usual convention of nij 40 if Aj is a product and nij o0 if Aj is a reactant.
Here, we use the same notations as in the single reaction case, with the exception
of Dai (Damköhler number), Bi (Zeldovich number), and ri (reaction rate) of the
ith reaction, which are defined as
LRi ðC R ; T in Þ g ðDH R Þi C R Ri ðC 1 ; C 2 ; . . . ; C M ; T f Þ
Dai ¼ ; Bi ¼ i ; ri ðhci; yf Þ ¼
hux iC R rf C pf T in Ri ðC R ; T in Þ
where Ri ðC 1 ; C 2 ; . . . ; C M ; T f Þ is the intrinsic rate of the ith reaction, and gi and

ðDH R Þi are the dimensionless activation energy and the heat of reaction of the
ith reaction, respectively.
The low-dimensional models obtained for the case of multiple reactions are
the same as that obtained for the case of the single reactions in the previous
examples, with the term nPj rðhci; hyf iÞ in the species balance equation for the jth
species being replaced by N H
i¼1 nij ri ðhci; hyf iÞ,P
and the term BDa rðhci; hyf iÞ in the
energy balance equation being replaced by N H
i¼1 Bi Dai ri ðhci; hyf iÞ. For example,
the steady state averaged model for a tubular reactor with negligible axial
dispersion ðPer 1Þ in which multiple reactions [given by Eq. (241)] occur, is
given by the following global equations:
dcj;m XNH
¼ nij Dai ri ðhci; hyf iÞ (242)
dz i¼1
dyfm XNH
St
¼ Bi Dai ri ðhci; hyf iÞ ðhyS i yC Þ (243)
dz i¼1

while the local equations are given by
hcj i cj;m dcj;m

¼ (244)
Zkj dz
yfm hyf i dyfm StðhyS i yC Þ

¼ þ b2 (245)
ZH dz
hyS i hyf i dyfm b4 StðhyS i yC Þ

¼ b3 (246)
ZH dz b1
with cj;m jz¼0 ¼ cj;m;in , and yfm jz¼0 ¼ yfm;in . The low-dimensional models for
other types of homogeneous reactors for the multiple reaction case are similarly
obtained.
F. EXAMPLES ILLUSTRATING USE OF MULTI-MODE HOMOGENEOUS REACTOR

MODELS
In this section, we present examples to illustrate the usefulness of multi-mode

homogeneous reactor models in predicting micromixing effects on yield and
selectivity, reactor runaway, etc.
1. Single Bimolecular Reaction in a Tubular Reactor: Mixing Effects on

Conversion
k
Here, we consider the case of a bimolecular reaction of the type A þ B !
Products, with RðhC A i; hC B iÞ ¼ khC A ihC B i, for the case of premixed feed as well
as unmixed feed.
a. Premixed feed. Our two-mode model equations are given by
dC A;m Da
¼ hC A ihC B i (247)
dz C A;in
dC B;m Da
¼ hC A ihC B i (248)
dz C A;in
Da
C A;m hC A i ¼ ZA hC A ihC B i (249)
C A;in
Da
C B;m hC B i ¼ ZB hC A ihC B i (250)
C A;in
with the boundary conditions C A;m ¼ C B;m ¼ C A;in at z ¼ 0 (i.e. for the case of
uniform and stoichiometric feeding of reactants). In the above Eqs. (247)–(250),
ZA and ZB are the dimensionless mixing times of A and B, respectively, given by
ZA ¼ tA;mix =tC , ZB ¼ tB;mix =tC , where tC ð¼ L=hux iÞ is the total residence time in
the reactor, tA;mix ¼ b1 a2 =DAm and tB;mix ¼ b1 a2 =DBm , and the Damköhler
number Da ¼ kC A;in tC .
We first analyze the case where the mixing times of A and B are equal, i.e.
ZA ¼ ZB ¼ Z. Figure 11 illustrates how the conversion X ð¼ 1 C A;m ðz ¼
0Þ=C A;m;in Þ varies with the Damköhler number Da for different values of the
dimensionless mixing time Z, for the case of stoichiometric feeding of reactants.
In order to examine the effects of differences in mixing times on conversion,
we solve Eqs. (247)–(250) for a wide range of values of k, where k ¼ ZA =ZB ð¼
DBm =DAm Þ is the ratio of the mixing times of A and B. The results, which have
been plotted in Fig. 12, enable us to capture solely the effects of differences in
transport properties (i.e. species diffusivities) on product formation. It may be
noted from Fig. 12, that conversion attains a maximum at k ¼ 1 because for the
case of premixed feed, the two species A and B are interchangeable and the
curves are therefore symmetric about k ¼ 1.
b. Unmixed feed. In this example, we show how the two-mode models, unlike
the traditional tubular reactor models, can capture the effects of non-uniform
reactant feeding on reactor performance.
100
η =0
(PFR)
80
η = 0.1
Conversion, X (%)
60
η =1
40 k
A+B >P
CA,in= CB,in
Da = k CA,in C η =2
20 η = tmix / C
0
0.01 0.1 1 10
Da
FIG. 11. Plot of conversion vs. Da for a bimolecular second-order reaction in a tubular reactor,
for different values of the dimensionless mixing time Z.
In order to capture the effects of non-uniform reactant feeding at the reactor

entrance, we rederive our two-mode models by introducing a delta function
source in the species balance equation for the jth species [Eq. (123)]

1 @ 0 @C j @C j @C j
Dm;j x þ nj RðCÞ ¼ þ ux ðx0 Þ ux ðx0 ÞC j;in ðx0 ÞdðxÞ (251)
x0 @x0 @x0 @t @x
@C j
¼0 at x0 ¼ 0; a (252)
@x0
C j ðx0 ; xÞ ¼ 0 at x ¼ 0 (253)
Transverse averaging of Eqs. (251)–(253) leads to the following global evolution
and local equations, respectively:
@hC j i @C j;m
þ hux i ¼ RðhCiÞ þ C j;m;in dðxÞ (254)
@t @x
100
87.5
η =0.1, D a=10
A
Conver sion X ( % )
75
η A=0.1, D a=100
k
A+B >P
62.5 C =C
A,in B,in
Da = k C τ
A,in C
η =t / τ
A A,mix C
κ = ηA / ηB
50
1 2 3 4 5
=D /D
Bm Am
FIG. 12. Influence of the difference in local mixing times of species A and B on conversion in a
second-order bimolecular reaction in a tubular reactor.

@C j;m
hC j i C j;m ¼ tmix;j hux i w1;j dðxÞ (255)
@x
where
C j;m;in ¼ hux ðx0 ÞC j;in ðx0 Þi=hux i (256)
w1;j ¼ hux ðx0 Þf j ðx0 Þi=ðb1 hux iÞ (257)
Z Z *Z Z +
x0 k x0 k
1 1
f j ðx0 Þ ¼ zuðzÞJ j ðzÞ dz dk zuðzÞJ j ðzÞ dz dk (258)
0 k 0 0 k 0
J j ðx0 Þ ¼ C j;m;in C j;in ðx0 Þ

In order to illustrate how this formulation can capture mixing effects resulting
from non-uniform reactant feeding at the reactor inlet, we apply the steady-state
version of our model [Eqs. (254)–(258)] to the simple case of bimolecular
k
second-order reaction of the type A þ B ! Products, with RðhC A i; hC B iÞ ¼
khC A ihC B i, where the reactants are fed as
0 2
0 x
C A;in ðx Þ ¼ 2 hC in i
a
"
0 2 #
0 x
C B;in ðx Þ ¼ 2 1 hC in i
a
such that hC A;in ðxÞi ¼ hC B;in ðxÞi ¼ hC in i but C A;m;in aC B;m;in . Here
C A;m;in ¼ 23hC in i; C B;m;in ¼ 43hC in i; w1;A ¼ 60
1 1
hC in i, and w1;B ¼ 60 hC in i.
Figure 13 shows how the steady-state exit conversion X ½¼ 1 C B;m ðz ¼
1Þ=C B;m;in varies with the Damköhler number Da for different values of the
dimensionless mixing time Zð¼ tmix =tÞ. The figure shows how non-uniform
feeding could significantly reduce the conversion as compared to premixed feed
for the case of a bimolecular second-order reaction (e.g. by a factor of 2 for the
case of Z ¼ 0:1), when mixing limitations are present in the system.
100
k
A+B P Premixed Feed
80 Da = kCA,inC
η = tmix/C = 0.1
CA,in(ξ)=2ξ2 <Cin>
CB,in (ξ) = 2(1−ξξ2) <Cin>
Conversion, X (%)
60
Non-premixed Feed
40
20
0
0.1 1 10
Da
FIG. 13. Comparison of conversion for a bimolecular second-order reaction in a homogeneous

tubular reactor for premixed and unmixed reactant feeding.
2. Single Non-isothermal Reaction in a Tubular Reactor: Mixing Effects on

Multiplicity Features
It is well known that a tubular reactor model with no macromixing (i.e.
Per 1) and perfect micromixing (Z ¼ 0) exhibits no multiple solutions, even in
the presence of autocatalytic (e.g. non-isothermal) kinetics. However, even in
the presence of small micromixing limitations (i.e. Z40), the reaction–diffusion
problem at the local scale starts generating multiple solutions (if the kinetics is
autocatalytic), leading to multiplicity in the solution of the full CDR equation at
the global scale. While this feature could be captured by the full CDR
equations, it is completely missed by the traditional low-dimensional models,
such as the plug-flow model.
The multi-mode model for a tubular reactor, even in its simplest form (steady
state, Per 1), is an index–infinity differential algebraic system. The local
equation of the multi-mode model, which captures the reaction–diffusion
phenomena at the local scale, is algebraic in nature, and produces multiple
solutions in the presence of autocatalysis, which, in turn, generates multiplicity
in the solution of the global evolution equation. We illustrate this feature of the
multi-mode models by considering the example of an adiabatic ða ¼ 0Þ tubular
reactor under steady-state operation. We consider the simple case of a non-
isothermal first order reaction
k
A ! Products
with Arrhenius kinetics, where the rate of consumption of A, rðhci, and hyf i) is
given by

hyf i
rðhci; hyf iÞ ¼ hci exp (259)
1 þ hyf i=g
and the multi-mode model is given by

dcm hyf i
¼ hci exp (260)
dDa 1 þ hyf i=g

hyf i
cm hci ¼ Daloc hci exp (261)
1 þ hyf i=g

dyfm hyf i
¼ Bhci exp (262)
dDa 1 þ hyf i=g

B hyf i
yfm hyf i ¼ Daloc hci exp (263)
Lef 1 þ hyf i=g
with the initial conditions cm ¼ 1 and yfm ¼ 0 at Da ¼ 0, where Daloc is the local
Damköhler number, given by
tmix
Daloc ¼ ¼ ktmix ¼ ZDa
tR
and B is the Zeldovich number (dimensionless adiabatic temperature rise).
Equations (260)–(263) form a set of differential-algebraic equations which has
a unique solution when the two algebraic equations [(261) and (263)] themselves
have unique solution of hci (and hyf i) for any fixed cm (and yfm). Equivalently,
the above system has multiple solutions only when Eqs. (261) and (263)
evaluated at the reactor exit conditions begin to have multiple solutions. For
Lef X1 (typical fluid Lewis numbers vary between 1 and 100), and for g ! 1,
the hysteresis variety for the above set of equations is given by
@F @2 F
F¼ ¼ ¼0 (264)
@hyf i @hyf i2
where
Daloc ðB yfm Þ expðhyf iÞ

F ¼ hyf i yfm (265)
Lef 1 þ Daloc expðhyf iÞ
Solving Eq. (264), the hysteresis locus is obtained in parametric form as
ð4 sÞ exp½s 2 s
yfm;exit ¼ ; 1oso2 (266)
1 þ 1 Le1 f exp½s 2
hyf iexit ¼ yfm;exit þ 2 (267)
B ¼ yfm;exit þ 4Lef (268)
Daloc ¼ ZDa ¼ exp½2 yfm;exit (269)
Z 2þyfm;exit
exp½y ð3 þ yfm;exit yÞ exp½2 yfm;exit
Da ¼ dy (270)
sþyfm;exit ½4Lef þ yfm;exit y
For very small values of the micromixing time, Z and for Lef ¼ 1 (turbulent
flows), Eqs. (266)–(270) may be simplified to
yfm;exit ¼ B 4 (271)
hyf iexit ¼ B 2 (272)

0.18 14
Region of Multiple Solutions
0.16 12
0.14 10
Da
0.12 8 B
0.1 6
k
A Products
Lef = 1
0.08 Da = kC 4
0.06 2
0.0001 0.001 0.01 0.1 1
η (= tmix/C)
FIG. 14. Hysteresis loci in (Da, Z) and (B, Z) planes for a first-order non-isothermal reaction in an
adiabatic tubular reactor.
1
B Da 1 þ (273)
B
Z ðB 1Þ exp½2 B (274)
Figure 14 presents the hysteresis loci [given by Eqs. (266)–(270)] in the (Da, Z)
and the (B,Z) planes, for the case of Lef ¼ 1. For any finite value of the
micromixing time, Z, the hysteresis locus gives the minimum value of B required
to produce multiplicity. It could be observed from Fig. 14 that as Z increases,
the critical value of B required to produce multiple solutions, decreases, finally
reaching the asymptotic limit of 4. It could also be noticed that for even for very
small values of Z, the critical value of B is reasonably low (e.g. for gas-phase
reactions Z 0:05, Bcritical ¼ 6:74; for liquid-phase reactions Z 1, Bcritical ¼ 4),
indicating that in practice, multiple solutions resulting from micromixing
limitations and autocatalytic kinetics are present in all tubular reactors. Thus,
the above results clearly contradict the traditional belief that an adiabatic
laminar flow reactor (with no axial dispersion) has only one possible steady
state, even in the presence of autocatalytic kinetics.
Numerical computations and experiments have revealed that a small change

in the operating variables leads to a very large change in the behavior of
adiabatic tubular reactors. This phenomenon has been referred to as
‘‘parametric sensitivity’’ in the prior literature (Bilous and Amundson, 1955;
Morbidelli and Varma, 1982). The reason for ‘‘parametric sensitivity’’ could be
attributed to the presence of multiple steady states in tubular reactors with
mixing limitations (at the local scale). As shown above, if the local mixing
limitations exceed Zcrit [given by Eq. (274)], the local equation can have multiple
solutions and the reactor might ignite locally. Under such conditions (i.e.
ZXZcrit ), when one of the operating variables is changed slightly, the local
temperature jumps from the stable extinguished steady-state branch to the
stable ignited steady-state branch, resulting in local hot spots. Thermal
micromixing limitations prevent these local hot spots from being extinguished
immediately, and they are carried downstream, where they are eventually
stabilized due to transverse diffusion and depletion of the reactant.
The most important observation that follows from the above analysis of the
multi-mode model is that in almost all practical cases, tubular reactor
instabilities arise due to mixing/diffusional limitations at the small scales and
spread over the reactor. In contrast, pseudohomogeneous models predict
(erroneously) that reactor instabilities (ignition, multiple solutions, etc.) arise
due to macromixing limitations at the reactor scale.
3. Single Bimolecular Reaction in a CSTR: Micromixing Effects on Conversion

Second-order reactions provide the simplest example of nonlinear kinetics,
where micromixing limitations have significant effects on reactant conversion.
We use the two-mode model to determine the same for a typical bimolecular
second-order reaction of the type
k
A þ B ! P with rate ¼ kC A C B
occurring in a CSTR. We consider the case of species A and B being fed in

stoichiometric amounts, but use two different feeding strategies, namely
uniform (premixed) and distributed feeding. In uniform feeding, the reactants
are mixed completely before entering the tank and fed as a single stream, while
in distributed feeding species A and B are fed separately as two different feed
streams of different concentrations (but in stoichiometric amount). Defining
qin in in in
A ðcA Þ and qB ðcB Þ to be the flow rates (concentrations) of entering feed streams
containing pure A and B, respectively, with qin in in in
A cA ¼ qB cB , the mean inlet
concentration becomes cm ¼ qA cA =ðqA þ qB Þ ¼ qB cB =ðqA þ qin
in in in in in in in in
B Þ. Therefore, by
changing qin in
A : qB from 1:2 to 1:19 [curve (a)–(d) in Fig. 15] with corresponding
change of cin in in
A and cB keeping cm constant, the feed stream containing B becomes
more diluted, while A feed stream becomes more concentrated. This
corresponds to more mixing limitations within tank and changes mixing times
80 (a)
premixed feed
(b)
distributed feed
60
Conversion (%)
(c)
40
20
(d)
0
10−2 10−1 100 101 102 103
Da
FIG. 15. Variation of conversion with Damköhler number for a bimolecular second-order
reaction for uniform and distributed feeding in a CSTR.
TABLE I
MIXING TIMES FOR FIG. 15
tA
mix;1 =tc tA
mix;2 =tc tBmix;1 =tc tBmix;2 =tc
Premixed 0.1 0.1 0.1 0.1

Unmixed (a) 0.1028 0.1417 0.1028 0.0833
Unmixed (b) 0.1090 0.2350 0.1090 0.0775
Unmixed (c) 0.1160 0.4800 0.1160 0.0756
Unmixed (d) 0.1203 0.9775 0.1203 0.0751
as shown in Table I (cases a, b, c, and d corresponding to qin in 1 1 1 1

A =qB ¼ 2; 4; 9, and 19,
respectively). The mixing times are calculated by considering a two-zone tank
(see Bhattacharya et al. (2004), for details). The results are shown in Fig. 15,
where the conversion is plotted as a function of reactor Damköhler number
(Da). As expected intuitively, the unmixed feed can significantly affect the
reactor conversion compared to the premixed feed, as the difference between the
concentrations (mixing times) of the two entering streams increases.
4. Competitive– Consecutive Reaction in a CSTR: Micromixing Effects on

Selectivity
Next we consider the competitive–consecutive reaction between species A and
B of the type
k1
A þ B ! R (275)
k2
B þ R ! S (276)
which is encountered very often in many multi-step homogeneous reactions

including diazo coupling between 1-napthanol and diazotized sulfanilic acid,
where k1 =k2 1. The latter has been studied extensively, both theoretically
(Angst et al., 1982a, b) and experimentally (Angst et al., 1982a, b; Bourne et al.,
1981; Zlokarnik, 2002), where it has been shown that the selectivity or yield of
the products (R or S) is extremely sensitive to mixing and can differ significantly
depending on the absolute value of qin in in in
A =qB , where qA and qB are the flow rates of
in in in
entering streams containing pure A and B, respectively with qin A C A ¼ qB C B (as
considered in the previous example). The selectivity is in general expressed as
the fraction of B converting to S and can be defined as X S ¼ 2C S =ð2C s þ C R Þ.
In the literature, two opposing facts of XS increasing and decreasing with
increasing qin in
A =qB have been reported under different reaction conditions. In
order to validate these experimental observations, Bhattacharya et al. (2004)
changed the ratio of the flow rates qin in
A =qB keeping total number of moles of A
in in in
and B entering the CSTR constant ðqin A C A ¼ qB C B Þ. Following the same
procedure as illustrated in previous example (two-zone tank), the mixing times
for species A, B, R, and S are calculated with the assumption of td being same
for all species. Then, the exit cup-mixing concentrations are calculated for k1 ¼
100k2 and plotted X S ¼ 2C mS =ð2C mS þ C mR Þ as a function of Da ¼ k2 C in m tc in
Fig. 16. In Fig. 16(A), the results are calculated for a case where micromixing
time is of same order as macromixing time, whereas in Fig. 16(B), the results are
calculated for a case where micromixing effect is negligible compared to
macromixing. In both the figures, curves (a), (b), (c), and (d) correspond to
qin in 1 1 1 1
A =qB ¼ 1; 4; 9, and 19, respectively. As can be observed, the yield of S decreases
with increasing qA =qin
in
B for micromixing dominated situations, while it increases
with increasing qin in
A =qB for macromixing dominated reactor. Another point to be
mentioned is that for the first case all mixing times are positive, while the second
mixing time of concentrated species is negative for macromixing dominated
case.
5. Prediction of Micromixing Effects on Polymer MWD in Tank Reactors

It is well known that in polymerization reactions, mixing affects monomer
conversion, copolymer distribution, and molecular weight distribution (MWD)
(Villermaux, 1991). In linear polymerization systems, imperfect mixing is found
to broaden the MWD, while in nonlinear polymerization with significant
branching, depending on reaction conditions, imperfect mixing can broaden or
narrow the MWD (Zhang and Ray, 1997). Here, we verify the first of the two
above-mentioned observations by examining the case of an anionic polymeriza-
tion using the two-mode model for a CSTR. Anionic polymerization, often used
industrially to produce polymers of narrow MWD, is typically characterized by
1
(d)
0.8
(c)
0.6
Xs
0.4
(b)
0.2
(a)
0
10−3 10−2 10−1 100 101 102
Da
(A)
0.15
(c)
(a)
Xs 0.1 (d)
(b)
0.05
0
10−3 10−2 10−1 100 101 102
Da
(B)
FIG. 16. Variations of selectivity (XS) with Damköhler number (Da) for qin in 1 1 1 1
A =qB ¼ 1; 4; 9 and 19
corresponding to curves (a), (b), (c), and (d), respectively. (A) tmicro ¼ tmacro ¼ 0:1tc and (B)
tmicro ¼ 0, tmacro ¼ 0:1tc ( Bhattacharya et al., 2004).
the lack of a termination step. The kinetics is
kI
I þ M ! P1
kI
Pj þ M ! Pjþ1
where the rate of the initiation (assuming a constant initiator concentration) is

given by
RI ¼ k I M
0.5
0.4
MWD
η=
0.3
η=
5
η=
2
η=
1
0.2
0
0.1
0
0 500 1000 1500 2000 2500 3000
n
FIG. 17. Effects of mixing on polymer properties: variation of MWD with chain length n, for
different values of the dimensionless mixing time, Z.
while the rate of propagation is given by
RP ¼ kP MPj
where kI and kP are the initiation and propagation rate constants, res-
pectively.
Application of the two-mode models for a CSTR to the above kinetics results
in a set of nonlinear algebraic equations, which when solved gives the MWD
and the polydispersity index (PDI). For the case of premixed feed, Fig. 17 shows
the variation of the MWD with n, where MWD is defined as
n2 Pn
MWD ¼ P1 (277)
n¼1 nPn
and n is the chain length of a polymer chain Pn. The MWDs shown in Fig. 17
correspond to the parameter values of Da ¼ kI tC ¼ 103 and kP =kI ¼ 2 105 ,
where Da is the Damköhler number and tC is the residence time of the tank.
While the Z ¼ 0 case shows that the MWD for a perfectly mixed CSTR is fairly
narrow, a significant broadening of the MWD is observed as the dimensionless
mixing time of the system, Zð¼ tmix =tC Þ, is increased.
V. Spatially Averaged Multi-mode (Multi-scale) Models for

Catalytic Reactors
A. WALL-CATALYZED REACTIONS
We consider the case of a single heterogeneous wall-catalyzed reaction

involving M species given by
X
M
mj Aj ¼ 0
j¼1
where mj is the stoichiometric coefficient of species j, with mj 40 if Aj is a product

and mj o0 if Aj is a reactant. The governing CDR equation for the jth species
(j ¼ 1; 2; . . . ; M) in the above reaction occurring in a tubular reactor with fully
developed laminar flow is given in dimensionless form by

1 @ @cj 1 @2 c j @cj p @2 c j @cj
x þ 2 ¼ p kj þ uðxÞ 9pf ðcj ; p; p Þ
x @x @x x @j2 @t kj Pe2r @z2 @z
(278)
with initial and boundary conditions given by

@cj
Das
¼p kj mj rw ðc1;s ; c2;s ; . . . ; cM;s Þ (279)
@x
x¼1 2
cj ðx; j; z; tÞ ¼ cj ðx; j þ 2p; z; tÞ (280)
p @cj @cj
¼ uðxÞ½cj cj;in at z ¼ 0; ¼ 0 at z ¼ 1 (281)
kj Pe2r @z @z
cj ðx; j; z; t ¼ 0Þ ¼ cj0 (282)
In the above equations, cj,s is the surface/wall concentration of species j, rw (c1,s,

c2,s,y,cM,s) is the dimensionless intrinsic rate of surface reaction, Das is the
reactor scale Damköhler number, which are given by
cj;s ¼ cj jx¼1 ¼ hcj i þ c0j jx¼1
Rw ðC 1;s ; C 2;s ; . . . ; C M;s Þ

rw ðc1;s ; c2;s ; . . . ; cM;s Þ ¼
Rw ðC R Þ
2LRw ðC R Þ
Das ¼
hux iaC R
while the other symbols retain their usual meanings. As in the single
homogeneous reaction case, we represent rw(c1,s,c2,s,y,cM,s) by rw(cs) and
Rw(C1,s,C2,s,y,CM,s) by Rw(Cs) for the sake of simplicity.
It may be noted that the above set of equations has a zero eigenvalue and a
corresponding constant eigenfunction for p ¼ 0, making spatial averaging by
L–S technique possible. Using the averaging theory outlined in Section II, the
spatially averaged low-dimensional model for heterogeneous wall-catalyzed
reactors to O(p) is obtained as

þ mj Das rw ðcs Þ ¼ 0 (283)
@t @z kj Pe2r @z2
cj;m hcj i @cj;m

¼ b2 mj Das rw ðcs Þ (284)
Z kj @z
cj;s hcj i @cj;m b4

¼ b3 þ mj Das rw ðcs Þ (285)
Z kj @z b1
where b4 is given by Eq. (211) for tubular geometry and b1 2b3 are given as by
Eqs. (212)–(214). The boundary and initial conditions for this averaged model
are same as those in the case of homogeneous tubular reactor [Eqs. (132)–(134)].
It may be noticed that unlike in the isothermal homogeneous tubular reactor
models, which were ‘‘two-mode models’’, the catalytic reactor models are
‘‘three-mode models’’, the three modes being the spatially averaged ðhcj iÞ,
mixing-cup (cj,m), and surface or wall (cj,s) concentrations. The cup-mixing and
wall concentrations are necessary to describe the mass transfer between the bulk
and the wall [Eq. (285)] while the two-modes cj,m and hcj i describe micromixing
that occurs in the fluid phase [Eq. (284)] due to transverse velocity gradients and
transverse molecular diffusion. Traditional two-phase models of catalytic
reactors that use only the wall and the cup-mixing concentrations ignore this
term which can be important in transient operation of the reactor.
It should be noted that it is possible to eliminate hci from Eq. (283) and write
the model in two-mode form using cm and cs. For the simple case of single
reaction A ! B, the two-mode model for a wall-catalyzed reactor is given by

p @rw @cm @cm p @2 c m p @2 c m
1 Das ðcs Þ þ þ 2 2 þ Das rw ðcs Þ ¼ 0 (286)
24 @c @t @z 48 @z@t Per @z
p @cm p
cs cm ¼ Das rw ðcs Þ (287)
16 @z 6
where the symbol c now represents the concentration of the reactant A. In this
form, the applicability of the model is limited to the parameter range in which
the term in square bracket does not vanish.
1. Limiting Cases
We now consider various limiting cases using the example of a single reaction
A ! B.
The first limiting case we consider is that of steady-state limit with negligible
axial dispersion ðPer 1Þ. For this case, Eqs. (286) and (287) reduce to the two-
mode form given by
cs cm dcm
11
¼ ¼ Das rw ðcs Þ; with cm jz¼0 ¼ cm;in (288)
48p
dz
In this form, the two-mode model is identical to the classical steady-state two-
phase model of a tubular catalytic reactor with negligible axial dispersion. There
is also a striking structural similarity between the two-mode models for
homogeneous reactions and two-phase models for catalytic reactions in the
practical limit of Per 1. This could be seen more clearly when Eqs. (137) and
(138) are rewritten as
dcm hci cm
¼ ¼ DarðhciÞ with cm jz¼0 ¼ cm;in (289)
dz b1 p
and Eq. (288) is written as
dcm cs cm
¼ ¼ Das rw ðhciÞ with cm jz¼0 ¼ cm;in (290)
dz b5 p
1
where b1 ¼ 48 and b5 ¼ 11
48 for laminar flow tubular reactors, and
ZTP ¼ b5 p (291)
is the dimensionless two-phase transfer time. The reciprocals of b1 and b5 are
the asymptotic Sherwood numbers or dimensionless mass transfer coefficients
for exchange between the two modes. There is an exact one to one
correspondence between the two-mode models of homogeneous reactions and
the two-phase models of catalytic reactions. For example, just as the reaction
rate in the two-phase model is not evaluated at the mixing-cup concentration cm
but at the wall concentration cs, similarly the reaction rate term in the two-mode
homogeneous reactor model is evaluated at the spatially averaged concentration
hci. Also analogous to the dimensionless two phase transfer time, ZTP ð¼ b5 pÞ, in
the two-phase model is the dimensionless local mixing time, Zð¼ b1 pÞ, in the
two-mode model.
The second limiting case we consider is that of linear kinetics with negligible
axial diffusion ðPer 1Þ. For this case, the averaged model can be written in
terms of the mixing-cup concentration as

" # " # " #
2 1 1 þ f2s =24 @cm f2s =16 @cm p 1 þ f2s =24 @2 cm
þ þ 1þ þ
3 3 1 þ f2s =6 @t 1 þ f2s =6 @z 48 1 þ f2s =6 @z@t
Das
þ cm ¼ 0 ð292Þ
1 þ f2s =6
where
f2s ¼ pDas (293)
is the local Damköhler number. For the case of f2s ! 0 (slow wall reaction), Eq.
(292) reduces to
@cm @cm p @2 c m
þ þ þ Das cm ¼ 0 (294)
@t @z 48 @z@t
while for the case of f2s ! 1 (infinitely fast wall reaction or the mass-transfer-
controlled limit), it may be written as
@cm 11 @cm p @2 c m 8
þ þ þ cm ¼ 0 (295)
@t 6 @z 144 @z@t p
Comparing this with the slow reaction case, we note that the effective velocity
has increased (by a factor 1.83), the dispersion coefficient is reduced by a factor
3 while the apparent reactor scale Damköhler number changed from Das to 8/p.
The last limiting case we consider is the practical case of long tubes where the
axial dispersion term may be neglected. For this case, it is more convenient to
write the three-mode model given by Eqs. (283)–(285) in the following form:
0 1 0 10 1
p @hci
24 @t 11 8 3 hci
B C B CB C
B p @cm C¼@ 8 6 2 A@ cm A (296)
@ 24 @z A
Das rw ðcs Þ 3 2 1 cs
with initial and boundary conditions
hciðz; t ¼ 0Þ ¼ c0 ðzÞ; cm ðz; t ¼ 0Þ ¼ cm0 ðzÞ; cm ðz ¼ 0; tÞ ¼ cm;in ðtÞ (297)
This model reduces to the two-phase model given by Eq. (288) under steady-
state conditions. However, for the general case of time-varying inlet conditions
this model retains all the qualitative features of the full partial differential
equation model and while the traditional two-phase model which does not
distinguish between cm and hci ignores the dispersion effect in the fluid phase.
B. COUPLED HOMOGENEOUS AND WALL-CATALYZED REACTIONS
We now consider the case of coupled homogeneous and wall-catalyzed

reactions involving M species occurring in a tubular reactor. The homogeneous
reaction is of the form
X
M
nj A j ¼ 0
j¼1
while the wall-catalyzed reaction is of the form
X
M
mj Aj ¼ 0
j¼1
where nj and mj are the stoichiometric coefficient of species j in the homogeneous

and the wall-catalyzed reactions, respectively, with nj 40 if Aj is a product in the
homogeneous reaction and mj 40 if Aj is a product in the wall-catalyzed
reaction. If species Aj does not participate in the homogeneous (wall-catalyzed)
reaction, then nj ðmj Þ is zero.
The governing CDR equation for the jth species is given by Eq. (123), with
initial and boundary conditions being given by Eqs. (279)–(282). As in the case
of wall-catalyzed reactions, the local/transverse diffusion operator has a zero
eigenvalue and a corresponding constant eigenfunction, thus enabling spatial
averaging by L–S technique. In this case, the low-dimensional model is
described by three modes, namely, the spatially averaged concentration hcj i, the
cup-mixing concentration cj,m, and the wall (or surface) concentration cj,s, and is
given for the jth species ðj ¼ 1; 2; . . . ; MÞ by

þ ¼ nj Da rðhciÞ þ mj Das rw ðcs Þ (298)
@t @z kj Pe2r @z2
along with Eqs. (284)–(285), with Eqs. (132)–(134) as boundary and initial
conditions. The coefficients b1 b4 are given by Eqs. (211)–(217).
It is interesting to note the above equations that mixing in the fluid phase
(which is described by the exchange between cj,m and hcj i), and transfer between
phases (which is described by the exchange between cj,m and cj,s) are both
influenced not only by the rate of the homogeneous reaction and the local
mixing time but also by the rate of the wall reaction and the two-phase transfer
time. These rigorously derived low-dimensional models thus illustrate that for
the case of coupled homogeneous–heterogeneous reactions, the system cannot
be described by a single transfer/exchange time, as has been traditionally done.
C. NON-ISOTHERMAL REACTOR MODELS
1. Wall-catalyzed Reactions
The low-dimensional model for non-isothermal wall-catalyzed reaction in a
tubular reactor is given by Eqs. (283)–(285) and
0 2 1
p @ hyf i 1 @2 hys i
@hyf i @hys i @yfm B þ
þ ssf ð1 Þ þ ¼ @ Pe2r Lef @z2 k @z2 C A
@t @t @z
þBs Das rw ðhci; hyf iÞ Stðhys i yC Þ
(299)
" #
yfm hyf i @yfm 1 ssf ð1 Þ@hy
@t
si
¼ þ b2 (300)
ZH @z Bs Das rw ðhci; hyf iÞ þ Stðhys i yC Þ
" #
hys i hyf i @yfm b4 1 ssf ð1 Þ@hy
@t
si
¼ b3 (301)
ZH @z b1 Bs Das rw ðhci; hyf iÞ þ Stðhys i yC Þ
with boundary and initial conditions being given by Eqs. (132)–(134), (204), and
(205), where Bs is the Zeldovich number for the wall-catalyzed reaction and all
other symbols retain their usual meanings.
2. Coupled Homogeneous and Wall-catalyzed Reactions

The low-dimensional model for coupled non-isothermal homogeneous and
wall-catalyzed reactions in a tubular reactor is given by Eqs. (284), (285), (298),
(300), and (301) and
0 2 1
p @ hyf i 1 @2 hys i
þ
@hyf i @hys i @yfm B 2
B Pe Lef @z2 k @z2 C
C
þ ssf ð1 Þ þ ¼B C
@t @t @z @ þ½BDa rðhci; hyf iÞ þ Bs Das rw ðcS ; ys Þ A
Stðhys i yC Þ
(302)
with boundary and initial conditions being given by Eqs. (132)–(134), (204), and
(205).
D. EXAMPLES ILLUSTRATING USE OF MULTI-MODE CATALYTIC REACTOR MODELS
1. Wall-catalyzed Reactions
We apply the low-dimensional convection model [Eq. (288)] to the simple case
of a isothermal bimolecular wall-catalyzed reaction occurring in a tubular
reactor. For the reaction
ks
A þ B ! P; with rate ¼ ks C A C B
where ks is the second-order surface rate constant, and for the case in which the
molecular diffusivities of A and B are assumed to be equal, the balance
equations in dimensionless form are given by
dcA;m dcB;m
¼ Das hcA ihcB i ¼ with cA;m jz¼0 ¼ cA;in ; cB;m jz¼0 ¼ cB;in (303)
dz dz
cA;m hcA i ¼ ZTP Das hcA ihcB i ¼ cB;m hcB i (304)
where
2ks C A;in tC
Das ¼ ; ZTP ¼ b5 p (305)
a
Figure 18 shows the variation of conversion
cA;m jz¼1
X ¼1 (306)
cA;in
with reactor-scale Damköhler number Das for different values of dimensionless

two-phase transfer time ZTP , for the case of stoichiometric feeding
(cA;in ¼ C B;in ). The case of ZTP ¼ 0 corresponds to the case of an ideal PFR,
where conversion X is given by X ¼ Das =ð1 þ Das Þ. In the limit of mass transfer
control, i.e. ZTP 40, Das ! 1 (but ZTP Das is finite), cA;s ! 0, cB;s ! 0, and
conversion is given by

1
X 1 ¼ 1 exp (307)
ZTP
Figure 18 illustrates these asymptotic limits.
2. Coupled Homogeneous and Wall-catalyzed Reactions

In this example, we examine the effects of mixing and mass transfer
limitations on the yields of competitive–consecutive reactions of the type
k1
A þ B ! R ðhomogeneousÞ (308)
100 ηTP = 0 (ideal PFR)
ks
A+B→P
80 CA,in= CB,in
ηTP = 0.1
Das = 2 ks CA,in C / a
ηTP = tTP/C
Conversion, X (%)
60
ηTP = 1
40
ηTP = 2
20
0
0.01 0.1 1 10 100
Das
FIG. 18. Variation of conversion (X) with the Damköhler number, Das, for a bimolecular second-
order wall-catalyzed reaction occurring in a tubular reactor.
k2S
B þ R ! S ðcatalyticÞ (309)
occurring in an isothermal tubular reactor in which the first reaction [Eq. (308)]
is a homogeneous one, which occurs in the bulk fluid phase, and the second
reaction [Eq. (309)] is catalytic. As in the homogeneous case, if the first reaction
is infinitely fast as compared to the second one (i.e. k1 =k2 s ! 1), and A and B
are fed in stoichiometric ratio, under perfectly micromixed conditions, B is
completely consumed in the homogeneous reaction and the catalytic reaction
does not occur. However, if micromixing limitations are present in the fluid
phase, the homogeneous reaction attains a mixing limited asymptote resulting in
a local excess of B, which can take part in the catalytic reaction. We use the
multi-mode steady convection model for coupled homogeneous–heterogeneous
reactors to quantify the yield of S (YS) under such conditions, where
2cS;m jz¼1
YS ¼ (310)
2cS;m jz¼1 þ cR;m jz¼1
The multi-mode model for the above reaction scheme in an isothermal tank
reactor is given by
dcA;m
¼ DahcA ihcB i (311)
dz
dcB;m
¼ ½DahcA ihcB i þ Das cB;s cR;s (312)
dz
dcR;m
¼ DahcA ihcB i Das cB;s cR;s (313)
dz
cA;m hcA i
¼ DahcA ihcB i (314)
Z
cB;m hcB i
¼ DahcA ihcB i þ ð1 þ b2 ÞDas cB;s cR;s (315)
Z
cC;m hcR i
¼ DahcA ihcB i þ ð1 þ b2 ÞDas cB;s cR;s (316)
Z
cA;s hcA i
¼ b3 DahcA ihcB i (317)
Z

cB;s hcB i b
¼ b3 DahcA ihcB i b3 þ 4 Das cB;s cR;s (318)
Z b1

cR;s hcR i b4
¼ b3 DahcA ihcB i þ b3 þ Das cB;s cR;s (319)
Z b1
where Dað¼ k1 C A;in tC Þ is the Damköhler number of the homogeneous reaction,

Das ð¼ 2k2s C A;in tC =aÞ is the Damköhler number of the catalytic reaction, and
Zð¼ tmix =tC Þ is the dimensionless local mass micromixing time, which is assumed
to same and equal for all species. For the case of a laminar flow reactor
1
considered here, b1 ¼ 48 , b2 ¼ 2, b3 ¼ 2, b4 ¼ 18. In this example, we consider the
case, where the first reaction is infinitely fast as compared to the second one, i.e.
Da ¼ k1 C A;in tC ! 1, and attains a mixing limited asymptote within a very
short residence time in the reactor. Using the formula for mixing limited
conversion in a tubular reactor (Chakraborty and Balakotaiah, 2002a), the local
excess of B that remains after the first reaction attains its mixing-limited
asymptote is obtained as

c1B;m 1
¼ exp (320)
cin Z
the amount of R formed from the first reaction is given by

c1R;m 1
¼ 1 exp (321)
cin Z
The catalytic reaction is simply a bimolecular reaction between B and R, with

boundary conditions given by cB;m jz¼0þ ¼ c1B;m ; cR;m jz¼0þ ¼ c1R;m . The yield of S
increases monotonically as the Damköhler number of the catalytic reaction,
Das, increases, and finally attains an asymptotic value when the catalytic
reaction reaches its mass transfer limited asymptote. This feature is illustrated in
Fig. 19, where the variation of YS with Das is shown. It is interesting to note
from Fig. 19, that the value of the mass transfer limited asymptote depends on
the micromixing limitation of the homogeneous reaction. Larger is the
micromixing limitation (Z) of the homogeneous reaction, more is the local
12
k
1
A + B ----->R (Homogeneous reaction)
k
η=1. 4427
2S
10 B + R ---->S (Catalytic reaction)
C =C =C
A,in B,in in
Da = 2 k C τ /a η=1
S 2S in C
η= t /τ
mix C
8 η=0. 8
Yield of S, YS (%)
η=0.5
4
2
η=0. 3
0
0.1 1 100
Damkohler Number of Catalytic Reaction, Das
FIG. 19. Influence of micromixing and mass transfer limitations on the yield of competitive-
consecutive reactions (of which one reaction is homogeneous and the other is wall-catalyzed) in a
tubular reactor.
excess of (unreacted) B that can participate in the catalytic reaction. However,

as could be seen from Fig. 19, the maximum yield of S(YS) is obtained when
c1B;m ¼ c1R;m , i.e. when Z ¼ 1=ln 0:5 ¼ 1:4427. If Z41:4427, R becomes the
limiting reactant in the catalytic reaction (instead of B), leading to decrease in
the yield of S.
VI. Accuracy, Convergence and Region of Validity of Multi-mode/

Multi-scale Averaged Models
In this section, we consider briefly the accuracy and convergence aspects of

the multi-mode models derived by the L–S method. We also illustrate the
regularization procedure used for the local equation(s) to increase the region of
convergence of the multi-mode models.
A. ACCURACY
It follows from the procedure explained in Section II that the global equation
is a Taylor series expansion of the nonlinear operator around some base point,
while the local equation is a perturbation series in p. Thus, the accuracy and
convergence properties of the averaged equation depend on the parameter p, the
specific nonlinear operator f(c), and the initial and boundary conditions. For
example, the global equation converges iff the Taylor series expansion of f(c)
around the base point c ¼ hci converges. Similarly, if the perturbation
expansion for c0 does not converge, then the local equation and hence the
averaged model does not exist.
The accuracy of the averaged model truncated at order pq ðqX0Þ thus depends
on the truncation of the Taylor series as well as on the truncation of the
perturbation expansion used in the local equation. The first error may be
determined from the order pqþ1 term in Eq. (23) and may be zero in many
practical cases [e.g. linear or second-order kinetics, wall reaction case, or
thermal and solutal dispersion problems in which f and rw(c) are linear in c] and
the averaged equation may be closed exactly, i.e. higher order Fréchet
derivatives are zero and the Taylor expansion given by Eq. (23) terminates at
some finite order (usually after the linear and quadratic terms in most
applications). In such cases, the only error is the second error due to the
perturbation expansion of the local equation. This error e for the local Eq. (20)
truncated at O(pq) may be expressed as
X
1
ðz; t; p; p Þ ¼ piþ1 E iþ1 ðz; t; p Þ (322)
i¼q
where the function E iþ1 ðz; t; p Þ depends on the specific nature of the nonlinear
operators f and rw(c). The magnitude of the truncation error of the local
equation depends on the convergence properties of the series given by Eq. (322),
which could only be evaluated on a case to case basis.
It should also be noted that this ‘‘model accuracy’’ is distinct from the
‘‘solution accuracy’’ (or error), which is defined by
^ ðz; t; pÞ ¼ khciðz; t; pÞ hcie ðz; t; pÞk (323)
where hciðz; t; pÞ is the solution of the truncated (at order pq) averaged model,
while hcie ðz; t; pÞ is the exact solution of the CDR equation and jj
jj is a norm in
the appropriate Banach space. For example, for steady-state problems, the
quantity of practical interest is the exit concentration. In this case, we simply
take
^ðpÞ ¼ jhciðz ¼ 1; pÞ hcie ðz ¼ 1; pÞj (324)
In general, it is not possible to obtain exact expressions for ^ as it requires

knowledge of the exact solution of hcie as well as the solution of the truncated
model. However, an estimate of the order of magnitude of the error may be
obtained by simply expanding it in a Taylor series around p ¼ 0. For example,
for p ! 0, we have
^ ðz; t; pÞ ¼ pqþ1 E 1 ðz; t; pÞ (325)
where the function E1 (with E 1 ðz; t; 0Þa0) depends on the specific nonlinear
operator f. Thus, for the common case of q ¼ 1 and p 1; ^ ¼ Oðp2 Þ.
The accuracy of low-dimensional models derived using the L–S method has
been tested for isothermal tubular reactors for specific kinetics by comparing the
solution of the full CDR equation [Eq. (117)] with that of the averaged models
(Chakraborty and Balakotaiah, 2002a). For example, for the case of a single
second order reaction, the two-mode model predicts the exit conversion to three
decimal accuracy when for f2 ð¼ pDaÞp1, and the maximum error is below 6%
for f2 20, where f2 ð¼ pDaÞ is the local Damköhler number of the reaction.
Such accuracy tests have also been performed for competitive–consecutive
reaction schemes and the truncated two-mode models have been found to be
very accurate within their region of convergence (discussed below).
B. CONVERGENCE
We now consider the convergence aspects of the averaged models by

considering some specific cases.
1. Homogeneous Reactions
As our first example, we consider the case of a first-order homogeneous
reaction A ! B in a laminar flow tubular reactor for which the global equation
is linear in c (i.e. rðhciÞ ¼ hci) and is therefore completely closed. To obtain the
range of convergence of the two-mode model, we need to consider only the local
equation. In this specific case, the reduced model equations to all orders of p are
then given by
dcm
¼ Dahci (326)
dz
" #
X
1
i
cm ¼ hci 1 þ bi ðpDaÞ (327)
i¼1
1 1 59
where b1 ¼ 48 , b2 ¼ 11;520 , b3 ¼ 77;41;440, etc. Substituting hci from Eq. (138) to
Eq. (137), we get
dcm cm
¼ Da P1 (328)
dz 1 þ i¼1 bi ðpDaÞi
The averaged model given by Eq. (328) can be expanded in a convergent power
series expansion in f2 ð¼ pDaÞ provided P
the infinite series in the denominator of
Eq. (328) is convergent. The series ð1 þ 1 2i
i¼1 bi f Þ is convergent if ro1, where
r is defined by
bnþ1 2
lim f ¼r (329)
n!1 bn
The first 20 of the coefficients bi have been calculated and from these, the
convergence criterion for the above series is obtained as (Chakraborty and
Balakotaiah, 2002a)
pDao41:17
that is
Daloc o0:858 (330)
where Daloc is the local Damköhler number, given by Daloc ¼ tmix =tR ¼ Z
Da ¼ b1 f2 , where Z is the dimensionless local mixing time, given by
Z ¼ tmix =tC
Criterion (330) specifies the range of validity of the two-mode models for the
case of uniform inlet feeding.
A similar convergence test for the case loop and recycle reactors (with large
recycle ratio) shows that the terms in the local equation when arranged in
ascending order of p change sign alternately (with a finite ratio), thus
significantly increasing the radius of convergence, and guarantees convergence
for most cases of practical interest. As a result, the two-mode models for
homogeneous reactors remain accurate as long as the local Damköhler number,
Daloc does not exceed a critical value, which is of order unity. When this is not
the case (i.e. Daloc 1) the two-mode models do not converge indicating that
the scale separation breaks down. In this case of fast reactions, the reaction
length scale becomes smaller than the meso length scale (e.g. tube diameter) so
that diffusion is no longer the dominant mechanism at the local scale. In such
cases, it is possible to have solutions of the CDR equation (e.g. for autocatalytic
kinetics) with length scales smaller than the meso scale, i.e. patterned solutions
(Balakotaiah and Chakraborty, 2003; Balakotaiah et al., 2002).
2. Catalytic Reactors
For the case of the wall-catalyzed reaction A ! B, the global equation is
closed for any type of kinetic expression rw(c). [Unlike the homogeneous
reaction case, here the reaction rate r is evaluated at the surface concentration
cs, which contains all the fluctuation modes (c0 ) about the mean /cS.] Thus,
convergence of the model depends only on the local equation. For the special
case of linear kinetics, flat velocity, and Per ! 1, the local equation was
determined to be
" #
X
1
2i
cm cs ¼ cs bi fs (331)
i¼1
where b1 ¼ 18, b2 ¼ 96

1 1
, b3 ¼ 1024 1
, b4 ¼ 11;520 , etc. This series in Eq. (331) has
alternating signs and converges absolutely if
f2s o10:67 (332)
We note that the local equation (331) may be rearranged as
f2s cs
cm cs ¼ (333)
Shðf2s Þ
where
!1
X
1
1
Shðf2s Þ ¼ bi f2i2
s ¼ (334)
i¼1 b1 þ b2 f2s þ

is the Sherwood number or dimensionless mass transfer coefficient. It may be

shown that (Gupta and Balakotaiah, 2001) Shðf2s Þ decreases monotonically
from 8 to 5.78 as f2s increases from 0 to N. Thus, the rearranged form of the
local equation (334) converges for all values of f2s .
The special case of linear kinetics with the axial Peclet number, Pe ¼ 0 and
flat velocity may be examined analytically. This case corresponds to the so-
called short monolith model, which is given by

1 d dc
x þ pð1 cÞ ¼ 0; 0oxo1
x dx dx
dc
¼0 at x ¼ 0
dx
dc f2
¼ sc at x ¼ 1; f2s ¼ pDa
dx 2
The exact solution of the model is given by
pffiffiffi
f2s I 1 ð pÞ
cm ¼ 1 pffiffiffi pffiffiffi pffiffiffi pffiffiffi (335)
p½ pI 1 ð pÞ þ ðf2s =2ÞI 0 ð pÞ
pffiffiffi pffiffiffi
pI 1 ð pÞ
cs ¼ pffiffiffi pffiffiffi pffiffiffi (336)
pI 1 ð pÞ þ ðf2s =2ÞI 0 ð pÞ
f2s cs
cm cs ¼ (337)
ShðpÞ
where the exact Sherwood number as a function of p is given by

pffiffiffi
1 I o ð pÞ 1
¼ pffiffiffi pffiffiffi (338)
ShðpÞ 2 pI 1 ð pÞ p
The L–S procedure gives the global and the local equations as
cm;in cm ¼ Da cs (339)

1 p p2 p3
cm cs ¼ f2s cs þ þ
(340)
8 192 3072 46; 080
Comparison of Eqs. (337) and (340) shows that the term in the brackets in Eq.
(340) is just the Taylor series expansion of 1/Sh(p) around p ¼ 0. Thus, the local
equation converges for all values of f2s and p. For this special case, we can
also estimate the error in the solution when the local equation is truncated at
order p as
2
p2 Da
¼ þ
(341)
192 1 þ Da
Thus, for the practical case of po1 (high conversion branch), the averaged
model gives at least two decimal accuracy for the mass transfer-limited case
ðDa ! 1Þ and higher accuracy in the kinetic regime.
C. REGULARIZATION OF THE LOCAL EQUATION
Here, we illustrate a mathematical technique called ‘‘regularization’’, that we

use to increase the region of convergence of the low-dimensional multi-mode
models.
When a function is defined by an infinite power series in terms of a parameter
p, the traditional approach is to truncate the power series, retaining terms up to
pq. However, if the power series fails to converge (i.e. outside the region of
convergence of the local equation), including higher order terms does not save
the truncated series from failure, and the truncated series may lead to non-
physical results in the limit of p ! 1.
Fortunately, there is an remedy for a poorly convergent power series, known
as Padé approximation. Consider a function c(k), given in terms of the power
series, as
c ¼ c0 þ c1 k þ c 2 k 2 þ

(342)
The power series may be only poorly convergent or even non-convergent, in
which case the truncated series becomes a poor approximation to c. Unlike the
power series, which tries to express c in terms of a single polynomial, the Padé
approximation expresses c as a ratio of two polynomials. The procedure to
determine the two polynomials involves converting the power series [Eq. (342)]
into another power series
f ¼ g c ¼ a0 þ a1 k þ a2 k 2 þ

(343)
such that c ¼ f =g, where
g ¼ 1 þ b1 k þ b2 k 2 þ

þ bn k n (344)
It is expected that a suitable choice of g will improve the convergence of the
power series as the singularity of c may be cancelled by the zero of g. In most
cases, the Padé approximation f =g provides with a better approximation than
the corresponding power series truncated at any order q, especially when|k|is
comparable to (or even greater than) the convergence radius of the power series
[Eq. (342)] (Takeshi, 1999).
Here, we use the example of the Taylor dispersion problem discussed Section
III to illustrate the regularization procedure. For simplicity, we illustrate this
only for the case of Per ! 1 (negligible axial diffusion). In this case, the global
equation is given by
@hCi @C m
þ ¼0 (345)
@t @z
while the local equation is given by
C m ¼ hCi þ huðxÞC 0 i (346)
where C0 is determined from

1 @ @C 0
x ¼ p f ðhCi þ C 0 Þ (347)
x @x @x
with
@C @C
f ðCÞ ¼ þ uðxÞ ; uðxÞ ¼ 2ð1 x2 Þ
@t @z
Writing
X
1
C0 ¼ pi c i (348)
i¼1
the functions ci satisfy the boundary and solvability conditions

@ci
¼0 at x ¼ 0; 1 (349)
@x
hci i ¼ 0; iX1 (350)

Following the procedure outlined in Section II, c1 and ci (iX2) are obtained by
solving

1 @ @c1 @hCi
x ¼ ðu 1Þ (351)
x @x @x @z

1 @ @ci @ci1 @ci1
x ¼ þu ; iX2 (352)
x @x @x @t @z
along with the above boundary and solvability conditions [Eqs. (349) and (350)].
Solving Eq. (351), c1 is obtained as
@hCi
c1 ¼ h1 ðxÞ (353)
@z
where

1 x2 x4
h1 ðxÞ ¼ þ (354)
12 4 8
and
1 @hCi
huc1 i ¼ (355)
48 @z
Substituting for c1 and using the leading-order approximation
@hCi @hCi
¼ þ OðpÞ
@t @z
the equation for c2 ðxÞ may be written as

1 @ @c2 @2 hCi
x ¼ ðu 1Þh1 ðxÞ (356)
x @x @x @z2
which when solved along with Eqs. (349) and (350) gives
@2 hCi
c2 ¼ h2 ðxÞ (357)
@z2
where
1
h2 ðxÞ ¼ ½45x8 200x6 þ 300x4 180x2 þ 31 (358)
11; 520
and
1 @2 hCi
huc2 i ¼ (359)
2880 @z2
Therefore, the local equation to O(p2) is
p @hCi p2 @2 hCi
C m hCi ¼ þ
(360)
48 @z 2880 @z2
Similarly, it is easily seen that
@i hCi
ci ¼ hi ðxÞ ; iX1 (361)
@zi
The local equation that relates hCi to Cm could now be written as
C m hCi ¼ huðxÞC 0 i
X1
¼ pi huðxÞci i
i¼1
X
1
@i hCi
¼ pi huðxÞhi i
i¼1
@zi
X
1
¼ mi ki hCi ð362Þ
i¼1
where
mi ¼ huðxÞhi i; iX1
@
k¼p (363)
@z
The parameters mi’s can be calculated using a symbolic manipulation

package. The first 20 of these are listed in Table II. We note that these
coefficients have alternate positive and negative signs (with approximate
periodicity of 2), indicating that the series given by Eq. (362) is convergent.
Using the method outlined by Mercer and Roberts (1990), the radius of
convergence of this series may be estimated to be 13.8. The physical meaning of
this result is that the local equation is meaningful only for wavelengths
TABLE II
COEFFICIENTS IN THE LOCAL EQUATION FOR THE TAYLOR DISPERSION PROBLEM
i mI I mi
1 2.0833 102 11 3.2220 1015

2 3.4722 104 12 7.5088 1017
3 1.5889 105 13 1.0632 1017
4 1.1755 106 14 7.0908 1019
5 4.9257 109 15 1.9357 1020
6 3.6895 109 16 3.8083 1021
7 1.0715 1010 17 5.7558 1023
8 9.3087 1012 18 1.4781 1023
9 7.2057 1013 19 8.1784 1025
10 1.0244 1014 20 3.5907 1026
exceeding the critical value
2pa2 hui a2 hui

Lc ¼ ¼ (364)
13:8Dm 2:2Dm
or equivalently, for p42:2 (where the p value is based on the minimum

wavelength contained in the initial conditions).
Regularization of the local equation may be used to increase the region of
validity of the truncated local equation. In fact, the regularized local equation
may be written such that it gives a physically meaningful solution even in the
limit of large p.
Using Eq. (362), we write
C m ¼ m0 hCi þ m1 khCi þ m2 k2 hCi þ

(365)
where m0 ¼ 1. Following the regularization principle outlined above [Eqs.

(342)–(344)], we express Cm as a ratio of two series:
½m0 þ ðb1 m0 þ m1 Þk

Cm ¼ hCi
1 þ b1 k
that is
ð1 þ b1 kÞC m ¼ ½m0 þ ðb1 m0 þ m1 ÞkhCi (366)
We choose b1 ¼ 2m1 , and Eq. (367) simplifies to
hCi ¼ ð1 m1 kÞC m (367)
and the regularized form of the local equation [Eq. (362)] is obtained from Eq.
(367) as
@C m
C m hCi ¼ m1 p (368)
@z
@C m
¼ b1 p (369)
@z
The regularized form of the two-mode model is now given by
@hCi @C m
þ ¼0 (370)
@t @z
@C m
C m hCi ¼ b1 p (371)
@z
This regularized form [Eqs. (370) and (371)] has a much larger region of validity
than the original low-dimensional model and gives qualitatively correct results
for any p40. As discussed in Section III, we can combine the above two
equations to get a single hyperbolic regularized equation for Cm.
VII. Summary, Conclusions, and Recommendations for Future

Work
The classical ideal reactor models such as the PFR and CSTR, obtained by
applying the conservation laws at the meso or macroscale ignore the small-scale
physics that is very important in determining the behavior of real reactors.
While these models are very simple to analyze and may be easily incorporated in
the design and control schemes, they are not realistic as they do not retain the
qualitative features of the full CDR equations. However, accurate low-
dimensional models that retain most of the qualitative features can be derived
by rigorous averaging of the CDR equations using the L–S method. In this
chapter, we have demonstrated this for the case of well-defined flow fields for
dispersion problems, homogeneous and wall-catalyzed reactors. We have also
illustrated the accuracy, convergence, and application of the low-dimensional
multi-mode/multi-scale models with commonly used examples from chemical
reaction engineering.
Generally speaking, the averaged models exist only when the local diffusion
time is much smaller compared to the convection and characteristic reaction
times ðp 1; Daloc 1Þ, i.e. physical length scale separation also corresponds
to separation of time scales. The accuracy of the averaged model depends on the
order of truncation in the small parameter p as well as the magnitude of the
other parameters p* (e.g. the reactor scale Damköhler number). An averaged
model derived to order pq (qX1) retains all the parameters of the original CDR
equations and the error in the solution is of the order pqþ1 . While
pseudohomogeneous ideal reactor models (corresponding to q ¼ 0) do not
retain the qualitative features of the full CDR equations, the next order
truncation (at q ¼ 1) retains all the important features such as the mass transfer
or mixing limited asymptote and multiplicity of solutions for autocatalytic
kinetics.
It should be noted that for non-isothermal case (and also for isothermal case
with autocatalytic kinetics) the local equation may have multiple solutions.
When this occurs, the averaged model obtained by the L–S method captures the
complete set of solutions of the full CDR equations only within the region of
convergence of the local equation. For example, for the wall-catalyzed non-
isothermal reaction case, we have shown that the averaged two-mode model can
capture only the three azimuthally symmetric solutions of the full CDR
equation. The latter has three symmetric solutions (of which two are stable) as
well as an arbitrarily large number of asymmetrical solutions (Balakotaiah

et al., 2002). These asymmetrical solutions with length scales smaller than the
tube radius exist only in the region of parameter space in which the local
equation derived by L–S technique does not converge. For the case of fast
homogeneous or heterogeneous reaction ðDaloc 1Þ, averaged models do not
exist as the local time scale can go to zero for Daloc ! 1. In this limit, the
spectrum of the local diffusion operator (with source/sink term either in the
equation or boundary condition) becomes continuous and time scale separation
breaks down. Thus, for the fast reaction case, it is not possible for any averaged
model to capture all the solutions of the full CDR equations.
A major limitation of the present work is that it deals only with well-defined
(and mostly unidirectional) flow fields and simple homogeneous and catalytic
reactor models. In addition, it ignores the coupling between the flow field and
the species and energy balances which may be due to physical property
variations or dependence of transport coefficients on state variables. Thus,
a major and useful extension of the present work is to consider two- or
three-dimensional flow fields (through simplified Navier–Stokes or Reynolds
averaged equations), include physical property variations and derive low-
dimensional models for various types of multi-phase reactors such as gas–liquid,
fluid–solid (with diffusion and reaction in the solid phase) and gas–liquid–solid
reactors.
ACKNOWLEDGMENTS
This work was supported by the Robert A. Welch Foundation.
NOMENCLATURE
Roman Letters
a radius of the tube
B Zeldovich number (or dimensionless adiabatic temperature rise)
Bi Biot number
Cj (cj) reactant (dimensionless) concentration of the jth species
Da reactor scale Damköhler number (homogeneous reaction)
Das reactor scale Damköhler number (catalytic reaction)
Daloc local Damköhler number
De,j effective diffusion coefficient of jth species
Dm,j molecular diffusivity of jth species
DT turbulent diffusivity
ft friction factor in turbulent flow
DHR heat of reaction

L length of the reactor
Lef fluid Lewis Number
M number of species
NH number of homogeneous reactions
NW number of heterogeneous wall-catalyzed reactions
p transverse Peclet number
Pe axial Peclet number
Per radial Peclet number
Q volumetric flow rate of recycle
qin volumetric flow rate of reactants
Ri intrinsic rate of ith homogeneous reaction (dimensional)
Rw,i intrinsic rate of ith wall reaction (dimensional)
St Stanton number
Sh Sherwood number
t dimensionless time
tmix mass micromixing time of the reference species
tmix,j mass micromixing time of the jth species
tmix,H thermal micromixing time
T temperature
ux velocity in the axial direction
X conversion
x coordinate along the length of the reactor (dimensional)
z dimensionless coordinate along the length of the reactor
Greek Letters
a cooling parameter
b exchange coefficient
g dimensionless activation energy
e volume fraction of fluid phase in the reactor
Z dimensionless mass micromixing time based on the reference species
(tmix/tC)
Zj dimensionless mass micromixing time of the jth species (tmix,j/tC)
ZH dimensionless thermal micromixing time
y temperature (dimensionless)
L recycle ratio
kj ratio of the diffusivities of the reference species to that of the jth species
mij stoichiometric coefficient of species Aj in the ith wall reaction
nij stoichiometric coefficient of species Aj in the ith homogeneous reaction
x dimensionless radial coordinate
tC total residence time on the reactor
f2 Thiele modulus
f2s surface Damköhler number
j azimuthal coordinate
REFERENCES
Angst, W., Bourne, J. R., and Sharma, R. N. Chem. Eng. Sci. 37(4), 585–590 (1982a).
Angst, W., Bourne, J. R., and Sharma, R. N. Chem. Eng. Sci. 37(8), 1259–1264 (1982b).
Aris, R. Proc. Roy. Soc. Lond. A 235, 67 (1956).
Balakotaiah, V. Korean J. Chem. Eng. 21(2), 318–328 (2004).
Balakotaiah, V., and Chakraborty, S. Chem. Eng. Sci. 58(21), 4769–4786 (2003).
Balakotaiah, V., and Chang, H. -C. Phil. Trans. Roy. Soc. Lond. A 351, 39–75 (1995).
Balakotaiah, V., and Chang, H. -C. SIAM. J. Appl. Math. 63(4), 1231–1258 (2003).
Balakotaiah, V., and Dommeti, S. M. S. Chem. Eng. Sci. 54, 1621–1638 (1999).
Balakotaiah, V., Gupta, N., and West, D. Chem. Eng. Sci. 57, 435 (2002).
Balakotaiah, V., Luss, D., and Keyfitz, B. L. Chem. Eng. Commun. 36, 121 (1985).
Baldyga, J., and Bourne, J. R. Chem. Eng. Sci. 39, 329 (1984).
Baldyga, J., and Bourne, J. R., ‘‘Turbulent Mixing and Chemical Reactions’’. Wiley, New York
(1999).
Bhattacharya, M., Harold, M. P., and Balakotaiah, V. Chem. Eng. Sci. 59, 5587 (2004).
Bilous, O., and Amundson, N. R. AIChE J. 1, 513–521 (1955).
Bodenstein, M., and Wolgast, K. Ztschr. Phys. Chem. 61, 422 (1908).
Bourne, J. R., Kozicki, F., and Rys, P. Chem. Eng. Sci. 36(10), 1643–1648 (1981).
Bourne, J. R., and Toor, H. L. AIChE J. 23, 602 (1977).
Brodkey, R. S., and Lewalle, J. AIChE J. 31, 111 (1985).
Carr, J., ‘‘Applications of Center Manifold Theory’’. Springer, Berlin (1981).
Chakraborty, S., and Balakotaiah, V. Chem. Eng. Sci. 57, 2545 (2002a).
Chakraborty, S., and Balakotaiah, V. AIChE J 48, 2571 (2002b).
Chakraborty, S., and Balakotaiah, V. Chem. Eng. Sci. 58, 1053 (2003).
Chakraborty, S., and Balakotaiah, V. Chem. Eng. Sci. 59(17), 3495–3736 (2004).
Churchill, S. W., Turbulent Flow and Convection: The Prediction of Turbulent Flow and
Convection in a Round Tube, in ‘‘Advances in Heat Transfer’’ Academic Press, New York (J.
P. Hartnett, T. F. Irvine, Y. I. Cho, and G. A. Greene, Eds.), (2001).
Curl, R. L. AIChE J. 9, 175 (1963).
Damköhler, G. Z. Elektrochem. 43, 1 (1937).
Danckwerts, P. V. Chem. Eng. Sci. 2, 1 (1953).
Danckwerts, P. V. Chem. Eng. Sci. 8, 93 (1958).
Dommeti, S. M. S., and Balakotaiah, V. Chem. Eng. Sci. 55, 6169 (2000).
Dutta, A., and Tarbell, J. M. AIChE J. 35, 2013 (1989).
Förster, V. T., and Geib, K. H. Ann. Phys. 5, 250 (1934).
Fox, R. O. Chem. Eng. Sci. 47, 2853 (1992).
Froment, G. F., and Bishchoff, K. B., ‘‘Chemical Reactor Analysis and Design’’. Wiley, New York
(1990).
Golay, M. J. E., Theory of Chromatography in Open and Coated Tubular Columns with Round and
Rectangular Cross-sections, in ‘‘Gas Chromatography’’ (D. H. Desty Ed.), pp. 36–49.
Butterworth, London (1958).
Golubitsky, M., and Schaeffer, D. G. ‘‘Singularities and Groups in Bifurcation Theory’’. Vol. 1.
Springer, Berlin (1984).
Gupta, N., and Balakotaiah, V. Chem. Eng. Sci. 56, 4771 (2001).
Harada, M. Mem. Facul. Eng., Kyoto Univ. 24, 431 (1962).
Hatta, S. Technol. Rep., Tohoku Univ. 10, 119 (1932).
Hiby, J. W. Longitudinal and Transverse Mixing During Single-phase Flow Through Granular
Beds, in ‘‘Proceedings of the Symposium on Interaction between Fluids and Particles’’ pp.
312–320. Institution of Chemical Engineers, London (1962).
Langmuir, I. J. Am. Ceram. Soc. 30, 656 (1908).
Levenspiel, O., ‘‘Chemical Reaction Engineering’’. Wiley, New York (1999).
Li, K. T., and Toor, H. L. AIChE J. 32, 1312 (1986).

Liu, S. -L., and Amundson, N. R. I&EC Fund. 1, 200 (1962).
Mercer, G. N., and Roberts, A. J. SIAM. J. Appl. Math. 50, 1547–1565 (1990).
Miyawaki, O., Tsujikawa, H., and Uraguchi, Y. J. Chem. Eng. Japan 8, 63 (1975).
Morbidelli, M., and Varma, A. AIChE J. 28, 705 (1982).
Ng, D. Y. C., and Rippin, D. W. T. Chem. Eng. Sci. 22, 65 (1965).
Ottino, J. M., Ranz, W. E., and Macosko, C. W. Chem. Eng. Sci. 34, 877 (1979).
Ranade, V. V., ‘‘Computational Flow Modeling for Chemical Reactor Engineering’’. Academic
Press, New York (2002).
Rhee, H. -K., Amundson, N. R., and Aris, R., ‘‘First-order Partial Differential Equations: Volume
1: Theory and Application of Single Equations’’. Prentice-Hall, New York (1986).
Schmitz, R. A., and Amundson, N. R. Chem. Eng. Sci. 18, 415 (1963).
Sittel, C. N., Threadgill, W. D., and Schnelle, K. B. I&EC Fund. 7, 39–43 (1968).
Takeshi, O. Phys. Fluids 11, 3247–3269 (1999).
Taylor, G. I. Proc. Roy. Soc. Lond. A 219, 186 (1953).
Taylor, G. I. Proc. Roy. Soc. Lond. A 223, 446 (1954).
Theile, E. W. Ind. Eng. Chem. 31, 916 (1939).
Trambouze, P. J., and Piret, E. L. AIChE J. 6, 574 (1960).
Vatistas, N., and Marconi, P. F. Chem. Eng. Sci. 47, 1727 (1992).
Villermaux, J. Rev. Chem. Eng. 7(1), 51 (1991).
Villermaux, J., and Devillon, J. C. Représentation de la coalescence et de la redispersion des
domaines de ségrégation dans un fluide per modèle d’interaction phénoménologique,
Proceedings of the Second Industrial Symposium on Chemical Reactors Engineering,
Amsterdam, B1 (1972).
Wehner, J. F., and Wilhelm, R. H. Chem. Eng. Sci. 6, 89 (1956).
Wen, C. Y., and Fan, L. T., ‘‘Models for Flow Systems and Chemical Reactors’’. Marcel Dekker,
New York (1975).
Westerterp, K. R., Dilman, V. V., and Kronberg, A. E. AIChE J. 41, 2013 (1995).
Westerterp, K. R., van Swaaij, W. P. M., and Beenackers, A. A. C. M., ‘‘Chemical Reactor Design
and Operation’’. Wiley, New York (1984).
Wicke, E. Z. Elektrochem. 65, 267 (1961).
Zeldovich, Ya. B. Zhur. Fiz. Khim. 13, 163 (1939).
Zhang, S. X., and Ray, W. H. AIChE J. 43, 1265 (1997).
Zlokarnik, M., ‘‘Scale-Up in Chemical Engineering’’. Wiley, New York (2002).
Zweitering, T. N. Chem. Eng. Sci. 11, 1 (1959).
INDEX
A Effective diffusivity, 153, 155, 167, 176,

178
Archie’s power law, 154 Effective permeability, 160
Axial dispersion model, 222 Effective scale, 159
Effective physical properties, 157
Electron microscopy, 138, 142, 145
B Engineering, 63–65, 67–69, 71, 79, 87–88,
90, 92–93, 96
Equivalent network, 165, 179
Ballistic packing, 152
Biology, 1–2, 6, 10, 38
Bubble growth, 165
F
Foam structure, 179

C
Formation factor, 155
Formation of Gas Cavities, 164
Capillary condensation, 164, 174
Fractal Porous Media, 173
Catalytic washcoat, 138, 193
Catalysis, 103, 105
Chord-length distribution, 145
CO oxidation, 192–193 G
Coarse graining, 2–3, 13–14, 21, 24, 30, 32,
35–37, 55–56 Gaussian correlated random field, 150
Correlation function, 144, 148, 150 Granulation, 190
Correlation length, 145 Granule dissolution, 192
Covering radius, 148–149 Growth, 2–3, 6, 10, 12, 15, 17–21, 25–31,
38, 40, 43, 46, 54–55, 58–60
D H
Darcy’s law, 156, 160, 179 Heterogeneous, 13, 26
Delaunay triangulation, 181 Heterogeneous catalysis, 138, 140, 197
Deposition, 169 Hierarchical, 2, 9–10, 50–51, 53
Discrete element modeling, 182, 187 Hybrid, 2, 6, 9, 12–26, 28–32, 35, 38,
Dissolution, 169 40–44, 50–51, 54
Dusty gas model (DGM), 159 Hyperbolic model, 222
E I
Effective conductivity, 153 Interfacial Flows, 161
299
300 INDEX
K Phase transition of microstructure, 160,

162
Knudsen diffusion, 155, 159, 194 Phase volume fraction, 143–144, 154–155
Poissonian generation of polydisperse
spheres, 147
Pore size distribution, 142, 145
L
Pore-network diagram, 140, 142, 161
Porous catalyst carrier, 182
Level set method, 165, 202
Porous polyolefin particles, 182
Length scale, 138, 139, 142, 143,
195
Liapunov–Schmidt, 208
R
Loop reactor, 248
Low-dimensional models, 208
Reaction engineering, 64–65, 67, 101, 103,
112
M Reconstructed porous media, 145, 151,
164, 174
Macromixing, 242 Reconstruction of porous media,
Magnetic Resonance, 63–64, 125, 131 146, 147
Materials, 1–2, 4, 6, 10, 12, 17, 23–25, 38, Recycle reactor, 249
46, 53
Mean transport-pore model (MTPM), 159 S
Metropolis rule, 147
Micromixing, 242
Simulated annealing, 146, 151, 163,
Microstructure, 140, 160, 166, 195–197 179
Molecular Dynamics Method, 170 Skeletonization, 145, 161
Monte Carlo, 2–3, 10–11, 56, 59 Solidification, 166
Monte Carlo method, 138, 171 Spatial averaging, 208
Morphogenesis of particles, 182, 187 Spinodal Decomposition, 167
Morphological descriptors, 143, 145 Stiffness, 32, 35–36, 43
Multi-mode models, 217 Stirred tank reactors, 256
Multigrid, 2, 8, 13–15, 18, 24–26, 28, Stochastic, 2, 6–11, 14, 16, 18, 23, 26–28,
31–32, 38 32–38, 40–44, 46–49, 51–52, 54–56, 58
Multiscale, 1–9, 11–21, 23–33, 35–41, Structure-property correlations, 153, 157
43–47, 49–57, 59, 61
T
O
Thresholding of correlated random fields,
Ostwald ripening, 168, 171 149
Tubular reactor, 245
P
U
Parabolic model, 222
Phase function, 140, 141, 143, 146, 161 Unit cell, 141, 158
INDEX 301
V W
Virtual granules, 190 Wet granulation, 190

Volume of fluid method (VOF), 163,
181 Y
Voronoi tessellation, 179, 181
Voxel, 141, 147, 163, 166, 169, 177, 180 Young-Laplace equation, 165
CONTENTS OF VOLUMES IN THIS SERIAL
Volume 1
J. W. Westwater, Boiling of Liquids

A. B. Metzner, Non-Newtonian Technology: Fluid Mechanics, Mixing, and Heat Transfer
R. Byron Bird, Theory of Diffusion
J. B. Opfell and B. H. Sage, Turbulence in Thermal and Material Transport
Robert E. Treybal, Mechanically Aided Liquid Extraction
Robert W. Schrage, The Automatic Computer in the Control and Planning of Manufacturing
Operations
Ernest J. Henley and Nathaniel F. Barr, Ionizing Radiation Applied to Chemical Processes and to
Food and Drug Processing
Volume 2
J. W. Westwater, Boiling of Liquids

Ernest F. Johnson, Automatic Process Control
Bernard Manowitz, Treatment and Disposal of Wastes in Nuclear Chemical Technology
George A. Sofer and Harold C. Weingartner, High Vacuum Technology
Theodore Vermeulen, Separation by Adsorption Methods
Sherman S. Weidenbaum, Mixing of Solids
Volume 3
C. S. Grove, Jr., Robert V. Jelinek, and Herbert M. Schoen, Crystallization from Solution
F. Alan Ferguson and Russell C. Phillips, High Temperature Technology
Daniel Hyman, Mixing and Agitation
John Beck, Design of Packed Catalytic Reactors
Douglass J. Wilde, Optimization Methods
Volume 4
J. T. Davies, Mass-Transfer and Inierfacial Phenomena

R. C. Kintner, Drop Phenomena Affecting Liquid Extraction
Octave Levenspiel and Kenneth B. Bischoff, Patterns of Flow in Chemical Process Vessels
Donald S. Scott, Properties of Concurrent Gas–Liquid Flow
D. N. Hanson and G. F. Somerville, A General Program for Computing Multistage Vapor–Liquid
Processes
Volume 5
J. F. Wehner, Flame Processes—Theoretical and Experimental

J. H. Sinfelt, Bifunctional Catalysts
S. G. Bankoff, Heat Conduction or Diffusion with Change of Phase
George D. Fulford, The Flow of Liquids in Thin Films
K. Rietema, Segregation in Liquid–Liquid Dispersions and its Effects on Chemical Reactions
303
304 CONTENTS OF VOLUMES IN THIS SERIAL
Volume 6
S. G. Bankoff, Diffusion-Controlled Bubble Growth

John C. Berg, Andreas Acrivos, and Michel Boudart, Evaporation Convection
H. M. Tsuchiya, A. G. Fredrickson, and R. Aris, Dynamics of Microbial Cell Populations
Samuel Sideman, Direct Contact Heat Transfer between Immiscible Liquids
Howard Brenner, Hydrodynamic Resistance of Particles at Small Reynolds Numbers
Volume 7
Robert S. Brown, Ralph Anderson, and Larry J. Shannon, Ignition and Combustion of Solid Rocket
Propellants
Knud Østergaard, Gas–Liquid–Particle Operations in Chemical Reaction Engineering
J. M. Prausnitz, Thermodynamics of Fluid–Phase Equilibria at High Pressures
Robert V. Macbeth, The Burn-Out Phenomenon in Forced-Convection Boiling
William Resnick and Benjamin Gal-Or, Gas–Liquid Dispersions
Volume 8
C. E. Lapple, Electrostatic Phenomena with Particulates

J. R. Kittrell, Mathematical Modeling of Chemical Reactions
W. P. Ledet and D. M. Himmelblau, Decomposition Procedures foe the Solving of Large Scale
Systems
R. Kumar and N. R. Kuloor, The Formation of Bubbles and Drops
Volume 9
Renato G. Bautista, Hydrometallurgy

Kishan B. Mathur and Norman Epstein, Dynamics of Spouted Beds
W. C. Reynolds, Recent Advances in the Computation of Turbulent Flows
R. E. Peck and D. T. Wasan, Drying of Solid Particles and Sheets
Volume 10
G. E. O’Connor and T. W. F. Russell, Heat Transfer in Tubular Fluid–Fluid Systems

P. C. Kapur, Balling and Granulation
Richard S. H. Mah and Mordechai Shacham, Pipeline Network Design and Synthesis
J. Robert Selman and Charles W. Tobias, Mass-Transfer Measurements by the Limiting-Current
Technique
Volume 11
Jean-Claude Charpentier, Mass-Transfer Rates in Gas–Liquid Absorbers and Reactors

Dee H. Barker and C. R. Mitra, The Indian Chemical Industry—Its Development and Needs
Lawrence L. Tavlarides and Michael Stamatoudis, The Analysis of Interphase Reactions and Mass
Transfer in Liquid–Liquid Dispersions
Terukatsu Miyauchi, Shintaro Furusaki, Shigeharu Morooka, and Yoneichi Ikeda, Transport
Phenomena and Reaction in Fluidized Catalyst Beds
CONTENTS OF VOLUMES IN THIS SERIAL 305
Volume 12
C. D. Prater, J, Wei, V. W. Weekman, Jr., and B. Gross, A Reaction Engineering Case History: Coke
Burning in Thermofor Catalytic Cracking Regenerators
Costel D. Denson, Stripping Operations in Polymer Processing
Robert C. Reid, Rapid Phase Transitions from Liquid to Vapor
John H. Seinfeld, Atmospheric Diffusion Theory
Volume 13
Edward G. Jefferson, Future Opportunities in Chemical Engineering

Eli Ruckenstein, Analysis of Transport Phenomena Using Scaling and Physical Models
Rohit Khanna and John H. Seinfeld, Mathematical Modeling of Packed Bed Reactors: Numerical
Solutions and Control Model Development
Michael P. Ramage, Kenneth R. Graziano, Paul H. Schipper, Frederick J. Krambeck, and Byung C.
Choi, KINPTR (Mobil’s Kinetic Reforming Model): A Review of Mobil’s Industrial Process
Modeling Philosophy
Volume 14
Richard D. Colberg and Manfred Morari, Analysis and Synthesis of Resilient Heat Exchange
Networks
Richard J. Quann, Robert A. Ware, Chi-Wen Hung, and James Wei, Catalytic Hydrometallation of
Petroleum
Kent David, The Safety Matrix: People Applying Technology to Yield Safe Chemical Plants and
Products
Volume 15
Pierre M. Adler, Ali Nadim, and Howard Brenner, Rheological Models of Suspensions
Stanley M. Englund, Opportunities in the Design of Inherently Safer Chemical Plants
H. J. Ploehn and W. B. Russel, Interactions between Colloidal Particles and Soluble Polymers
Volume 16
Perspectives in Chemical Engineering: Research and Education

Clark K. Colton, Editor
Historical Perspective and Overview
L. E. Scriven, On the Emergence and Evolution of Chemical Engineering
Ralph Landau, Academic—industrial Interaction in the Early Development of Chemical Engineering
James Wei, Future Directions of Chemical Engineering
Fluid Mechanics and Transport
L. G. Leal, Challenges and Opportunities in Fluid Mechanics and Transport Phenomena
William B. Russel, Fluid Mechanics and Transport Research in Chemical Engineering
J. R. A. Pearson, Fluid Mechanics and Transport Phenomena
Thermodynamics
Keith E. Gubbins, Thermodynamics
J. M. Prausnitz, Chemical Engineering Thermodynamics: Continuity and Expanding Frontiers

H. Ted Davis, Future Opportunities in Thermodynamics
Kinetics, Catalysis, and Reactor Engineering
Alexis T. Bell, Reflections on the Current Status and Future Directions of Chemical Reaction
Engineering
James R. Katzer and S. S. Wong, Frontiers in Chemical Reaction Engineering
L. Louis Hegedus, Catalyst Design
Environmental Protection and Energy
John H. Seinfeld, Environmental Chemical Engineering
T. W. F. Russell, Energy and Environmental Concerns
Janos M. Beer, Jack B. Howard, John P. Longwell, and Adel F. Sarofim, The Role of Chemical
Engineering in Fuel Manufacture and Use of Fuels
Polymers
Matthew Tirrell, Polymer Science in Chemical Engineering
Richard A. Register and Stuart L. Cooper, Chemical Engineers in Polymer Science: The Need for an
Interdisciplinary Approach
Microelectronic and Optical Material
Larry F. Thompson, Chemical Engineering Research Opportunities in Electronic and Optical
Materials Research
Klavs F. Jensen, Chemical Engineering in the Processing of Electronic and Optical Materials: A
Discussion
Bioengineering
James E. Bailey, Bioprocess Engineering
Arthur E. Humphrey, Some Unsolved Problems of Biotechnology
Channing Robertson, Chemical Engineering: Its Role in the Medical and Health Sciences
Process Engineering
Arthur W. Westerberg, Process Engineering
Manfred Morari, Process Control Theory: Reflections on the Past Decade and Goals for the Next
James M. Douglas, The Paradigm After Next
George Stephanopoulos, Symbolic Computing and Artificial Intelligence in Chemical Engineering: A
New Challenge
The Identity of Our Profession
Morton M. Denn, The Identity of Our Profession
Volume 17
Y. T. Shah, Design Parameters for Mechanically Agitated Reactors

Mooson Kwauk, Particulate Fluidization: An Overview
Volume 18
E. James Davis, Microchemical Engineering: The Physics and Chemistry of the Microparticle
Selim M. Senkan, Detailed Chemical Kinetic Modeling: Chemical Reaction Engineering of the Future
Lorenz T. Biegler, Optimization Strategies for Complex Process Models
Volume 19
Robert Langer, Polymer Systems for Controlled Release of Macromolecules, Immobilized Enzyme
Medical Bioreactors, and Tissue Engineering
J. J. Linderman, P. A. Mahama, K. E. Forsten, and D. A. Lauffenburger, Diffusion and Probability
in Receptor Binding and Signaling
Rakesh K. Jain, Transport Phenomena in Tumors
R. Krishna, A Systems Approach to Multiphase Reactor Selection
David T. Allen, Pollution Prevention: Engineering Design at Macro-, Meso-, and Microscales
John H. Seinfeld, Jean M. Andino, Frank M. Bowman, Hali J. L. Forstner, and Spyros Pandis,
Tropospheric Chemistry
Volume 20
Arthur M. Squires, Origins of the Fast Fluid Bed

Yu Zhiqing, Application Collocation
Youchu Li, Hydrodynamics
Li Jinghai, Modeling
Yu Zhiqing and Jin Yong, Heat and Mass Transfer
Mooson Kwauk, Powder Assessment
Li Hongzhong, Hardware Development
Youchu Li and Xuyi Zhang, Circulating Fluidized Bed Combustion
Chen Junwu, Cao Hanchang, and Liu Taiji, Catalyst Regeneration in Fluid Catalytic Cracking
Volume 21
Christopher J. Nagel, Chonghum Han, and George Stephanopoulos, Modeling Languages:

Declarative and Imperative Descriptions of Chemical Reactions and Processing Systems
Chonghun Han, George Stephanopoulos, and James M. Douglas, Automation in Design: The
Conceptual Synthesis of Chemical Processing Schemes
Michael L. Mavrovouniotis, Symbolic and Quantitative Reasoning: Design of Reaction Pathways
through Recursive Satisfaction of Constraints
Christopher Nagel and George Stephanopoulos, Inductive and Deductive Reasoning: The Case of
Identifying Potential Hazards in Chemical Processes
Keven G. Joback and George Stephanopoulos, Searching Spaces of Discrete Soloutions: The Design
of Molecules Processing Desired Physical Properties
Volume 22
Chonghun Han, Ramachandran Lakshmanan, Bhavik Bakshi, and George Stephanopoulos,

Nonmonotonic Reasoning: The Synthesis of Operating Procedures in Chemical Plants
Pedro M. Saraiva, Inductive and Analogical Learning: Data-Driven Improvement of Process
Operations
Alexandros Koulouris, Bhavik R. Bakshi and George Stephanopoulos, Empirical Learning through
Neural Networks: The Wave-Net Solution
Bhavik R. Bakshi and George Stephanopoulos, Reasoning in Time: Modeling, Analysis, and Pattern
Recognition of Temporal Process Trends
Matthew J. Realff, Intelligence in Numerical Computing: Improving Batch Scheduling Algorithms
through Explanation-Based Learning
Volume 23
Jeffrey J. Siirola, Industrial Applications of Chemical Process Synthesis

Arthur W. Westerberg and Oliver Wahnschafft, The Synthesis of Distillation-Based Separation
Systems
Ignacio E. Grossmann, Mixed-Integer Optimization Techniques for Algorithmic Process Synthesis
Subash Balakrishna and Lorenz T. Biegler, Chemical Reactor Network Targeting and Integration: An
Optimization Approach
Steve Walsh and John Perkins, Operability and Control in Process Synthesis and Design
Volume 24
Raffaella Ocone and Gianni Astarita, Kinetics and Thermodynamics in Multicomponent Mixtures
Arvind Varma, Alexander S. Rogachev, Alexandra S. Mukasyan, and Stephen Hwang, Combustion
Synthesis of Advanced Materials: Principles and Applications
J. A. M. Kuipers and W. P. M. van Swaaij, Computional Fluid Dynamics Applied to Chemical
Reaction Engineering
Ronald E. Schmitt, Howard Klee, Debora M. Sparks, and Mahesh K. Podar, Using Relative Risk
Analysis to Set Priorities for Pollution Prevention at a Petroleum Refinery
Volume 25
J. F. Davis, M. J. Piovoso, K. A. Hoo, and B. R. Bakshi, Process Data Analysis and Interpretation
J. M. Ottino, P. DeRoussel, S., Hansen, and D. V. Khakhar, Mixing and Dispersion of Viscous
Liquids and Powdered Solids
Peter L. Silverston, Li Chengyue, Yuan Wei-Kang, Application of Periodic Operation to Sulfur
Dioxide Oxidation
Volume 26
J. B. Joshi, N. S. Deshpande, M. Dinkar, and D. V. Phanikumar, Hydrodynamic Stability of

Multiphase Reactors
Michael Nikolaou, Model Predictive Controllers: A Critical Synthesis of Theory and Industrial Needs
Volume 27
William R. Moser, Josef Find, Sean C. Emerson, and Ivo M, Krausz, Engineered Synthesis of
Nanostructure Materials and Catalysts
Bruce C. Gates, Supported Nanostructured Catalysts: Metal Complexes and Metal Clusters
Ralph T. Yang, Nanostructured Absorbents
Thomas J. Webster, Nanophase Ceramics: The Future Orthopedic and Dental Implant Material
Yu-Ming Lin, Mildred S. Dresselhaus, and Jackie Y. Ying, Fabrication, Structure, and Transport
Properties of Nanowires
Volume 28
Qiliang Yan and Juan J. DePablo, Hyper-Parallel Tempering Monte Carlo and Its Applications
Pablo G. Debenedetti, Frank H. Stillinger, Thomas M. Truskett, and Catherine P. Lewis, Theory of
Supercooled Liquids and Glasses: Energy Landscape and Statistical Geometry Perspectives
Michael W. Deem, A Statistical Mechanical Approach to Combinatorial Chemistry
Venkat Ganesan and Glenn H. Fredrickson, Fluctuation Effects in Microemulsion Reaction Media
David B. Graves and Cameron F. Abrams, Molecular Dynamics Simulations of Ion–Surface
Interactions with Applications to Plasma Processing
Christian M. Lastoskie and Keith E. Gubbins, Characterization of Porous Materials Using Molecular
Theory and Simulation
Dimitrios Maroudas, Modeling of Radical-Surface Interactions in the Plasma-Enhanced Chemical
Vapor Deposition of Silicon Thin Films
Sanat Kumar, M. Antonio Floriano, and Athanassiors Z. Panagiotopoulos, Nanostructured
Formation and Phase Separation in Surfactant Solutions
Stanley I. Sandler, Amadeu K. Sum, and Shiang-Tai Lin, Some Chemical Engineering Applications of
Quantum Chemical Calculations
Bernhardt L. Trout, Car-Parrinello Methods in Chemical Engineering: Their Scope and potential
R. A. van Santeen and X. Rozanska, Theory of Zeolite Catalysis
Zhen-Gang Wang, Morphology, Fluctuation, Metastability and Kinetics in Ordered Block Copolymers
Volume 29
Michael V. Sefton, The New Biomaterials

Kristi S. Anseth and Kristyn S. Masters, Cell–Material Interactions
Surya K. Mallapragada and Jennifer B. Recknor, Polymeric Biomaterias for Nerve Regeneration
Anthony M. Lowman, Thomas D. Dziubla, Petr Bures, and Nicholas A. Peppas, Structural and
Dynamic Response of Neutral and Intelligent Networks in Biomedical Environments
F. Kurtis Kasper and Antonios G. Mikos, Biomaterials and Gene Therapy
Balaji Narasimhan and Matt J. Kipper, Surface-Erodible Biomaterials for Drug Delivery
Volume 30
Dionisios G. Vlachos, A Review of Multiscale Analysis: Examples from Systems Biology, Materials
Engineering, and Other Fluid Surface Interacting Systems
Lynn F. Gladden, M.D. Mantle and A.J. Sederman, Quantifying Physics and Chemistry at Multiple
Length-Scales using Magnetic Resonance Techniques
Juraj Kosek, František Štěpánek, and Milos̆ Marek, Modelling of Transport and Transformation
Processes in Porous and Multiphase Bodies
Saikat Chakraborty and Vemuri Balakotaiah, Spatially Averaged Multiscale Models for Chemical
Reactors

Advances in Chemical Engineering - Vol 30 Multiscale Analysis

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advances in Chemical Engineering - Vol 30 Multiscale Analysis

Uploaded by

Copyright:

Available Formats

Advances in Chemical Engineering,

Volume 30: Multiscale Analysis

A Review of Multiscale Analysis: Examples from Systems Biology,

Quantifying Physics and Chemistry at Multiple Length-scales using

Modeling of transport and transformation processes in porous and

D. Digital Reconstruction of Multiphase Media . . . . . . . . . . . . . . . . . . . . . . . 145

Spatially Averaged Multi-Scale Models for Chemical Reactors

C. Non-isothermal Reactor Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

VEMURI BALAKOTAIAH, Department of Chemical Engineering, University of

Multiscale simulation is an emerging scientific ﬁeld that spans many

A decadal report recently issued by the National Research Council (NRC),

Engineering (NRC, 2003a), advances 13 ‘‘Grand Challenges’’ for the ﬁeld.

10-2 Mesoscopic CFD

10-12 10-9 10-3 103

II. Deterministic, Continuum Models

Traditionally, modeling in chemical engineering has invoked continuum de-

where substantial mathematical and computational contributions have been

Continuum modeling has often been based on algebraic equations (AEs),

B. SOLVING DETERMINISTIC, CONTINUUM DIFFERENTIAL EQUATION MODELS:

Substantial advances in computational power (Moore’s law) have had a tre-

III. Overview of Discrete, Particle Models

Discrete models treat individual atoms, molecules, or particles and can be

A. HIERARCHY OF STOCHASTIC MODELS FOR WELL-MIXED, CHEMICALLY REACTING

recovered in the limit of an inﬁnite-size system ðN ! 1Þ. The concept of hi-

B. SOLVING MASTER EQUATIONS STOCHASTICALLY: MONTE CARLO METHODS

The introduction of the Metropolis MC algorithm in 1953 (Metropolis et al.,

IV. Classification of Multiscale Simulation Approaches

Multiscale simulation enables coupling of phenomena at various scales from

and concurrent simulation have been employed, depending on the algorithm

Finer scale Coarser scale

MULTIGRID-TYPE HYBRID MULTISCALE MODELS

Fine (discrete, particle)

V. Hybrid Multiscale Simulation

Hybrid multiscale simulation is the most developed branch of multiscale

is ﬁrst covered, followed by a discussion of its application to crystal growth and

A. ONION-TYPE HYBRID MULTISCALE SIMULATIONS AND ALGORITHMS

problems. The solution can be obtained iteratively by passing the steady-state

B. APPLICATION OF ONION-TYPE HYBRID MULTISCALE SIMULATION TO GROWTH OF

Fabrication of polycrystalline ﬁlms is an inherently multiscale problem of

2. Physical and Chemical Fluid Deposition of Thin Films (Fluid– Surface

Boundary condition, e.g., External field,

TEG Flux (mmoles/min)

There is a limited number of experiments from coupled systems to enable

3. Deposition by Molecular Beam Epitaxy (Uncoupled Fluid– Surface Systems)

as continuum points where either partial equilibrium or Robin boundary con-

C. APPLICATIONS OF ONION-TYPE HYBRID MULTISCALE SIMULATION TO OTHER

There have been many hybrid multiscale simulations published recently in

successful multiscale approaches in solid mechanics is the quasi-continuum

D. MULTIGRID-TYPE HYBRID MULTISCALE SIMULATIONS

The above problems exhibit phenomena with well-separated length scales,

There is another very important class of problems where no well-deﬁned sep-

hybrid multiscale simulations for the example of a traveling wave in a catalytic

E. AN EXAMPLE OF MULTIGRID-TYPE HYBRID MULTISCALE SIMULATION FOR

An example of the aforementioned multigrid-type hybrid multiscale simula-

nucleation has a higher probability of occurrence wherever concentrations are

Pðtnuc Þ ¼ 1 exp½tnuc Po (3)

F. CHALLENGES IN HYBRID MULTISCALE SIMULATIONS

The major issue in hybrid multiscale simulation is ‘‘patching’’ of models used

brief, coupling may lack convergence (Raimondeau and Vlachos, 2002b)

where D ¼ Do ebU o is the diffusion coefﬁcient, Do the diffusion coefﬁcient at

j ¼ DfebJc ½rc bcð1 cÞrJ cg (6)

and the constitutive relation is

j ¼ Do f½rc bcð1 cÞrJ c (8)

VI. Coarse Graining of Stochastic Models

M z0 ðtÞ ¼ M 0 ½1 expðt=T 1 Þ (3)