Bouke Tuinstra

Development of a

Generic Engineering Model

using Computational Fluid Dynamics

Bouke Tuinstra

Development of a Generic Engineering Model for Packed Bed Reactors using Computational Fluid Dynamics

for Packed Bed Reactors

Development of a generic engineering model

for packed bed reactors

using Computational Fluid Dynamics

Proefschrift
ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,

op gezag van de Rector Magnicus, prof.dr.ir. J.T. Fokkema,

voorzitter van het College voor Promoties,

in het openbaar te verdedigen op dinsdag 23 december 2008 om 15.00 uur

door

Bouke Fokkes TUINSTRA

scheikundig ingenieur

geboren te Cleveland Ohio, Verenigde Staten van Amerika

Dit proefschrift is goedgekeurd door de promotor:

Prof.ir. C.M. van den Bleek

Samenstelling promotiecommissie:
Rector Magnicus, voorzitter

Prof.ir. C.M. van den Bleek, Technische Universiteit Delft, promotor

Prof.dr. R.F. Mudde, Technische Universiteit Delft

Prof.dr.ir. G.B. Marin, Universiteit Gent

Prof.dr.ir. J.C. Schouten, Technische Universiteit Eindhoven

Prof.dr.ir. W.P.M. van Swaaij, Universiteit Twente

Dr.ir. H.P.A. Calis, Shell Gas & Power

Dr.ir. A.N.R. Bos, Shell Global Solutions

Het in dit proefschrift beschreven onderzoek

is gedeeltelijk nancieel ondersteund door:

Koninklijke DSM n.v.

Kiwa Onderzoek & Advies b.v.

Stork Comprimo b.v.

c 2008 by B.F. Tuinstra

Copyright 
All Rights Reserved

ISBN 978-90-9023662-9

Summary
Most industrial chemical processes are possible only through the use of a cat
alyst, i.e., a compound that speeds up the reaction but that is not consumed.
In most cases, heterogeneous catalysis is used, here the catalyst is a solid ma
terial while the reactants are in the liquid or gaseous state. An example of
such a reactor is the one that is used in automotive exhaust systems. There
the catalyst is xed on the surface of a honeycomb-like ceramic structure.
For larger scale chemical reactors, structured packings like this are relatively
expensive. Therefore, in most chemical reactors random packings of porous
ceramic catalyst particles are used instead.
One of the key factors in the operation of these packed bed reactors is the
exchange of reactants and products between the catalyst and the process uid.
In many cases, heat has to be supplied to or removed from the reactor to get
a complete and swift conversion of reactants into desired products. In these
cases also the transfer of heat inside the bed and between the reactor wall and
the bed is crucial. All these transport processes are in turn inuenced by the
macroscopic ow prole in the catalyst bed. Maldistribution of ow occurs
due to the inuence of the wall on the bed packing and, for low pressure drop
reactors, due to the geometry of the ow channel upstream and downstream
the catalyst bed.
Nowadays, computer models are generally used to design packed bed reac
tors. There are various approaches to the modelling of packed bed chemical
reactors, diering in level of detail and complexity as well as computation time.
These models can be thought to be ranging from the engineering approach
on one end of the spectrum to the scientic approach on the other.
In the engineering approach, the goal is to estimate the performance of a
packed bed with sucient accuracy to allow the design, rating and optimisation
of the reactor. The models are usually based on t relations for empirical data.
Typically, the result is a one-dimensional model that gives a relation between
the inlet and the outlet conditions. The advantage of this approach is that the
models dont need much computation time or resources. The disadvantages

are that the models are specic for a given reactor and are limited to those

geometries, for which empirical relations are available.

In the scientic approach, microscopic rst-principle descriptions are used
as much as possible. For a packed bed reactor, the ow around the particles
and transport phenomena on the scale of a particle could be calculated us
ing Computional Fluid Dynamics (CFD) techniques. This approach leads to
generic and detailed models that can be used to better understand the micro
scopic processes. However, they are usually not very suitable to use as a design
tool because of the high level of expertise needed to build and use such a model
and because of forbiddingly high computational eort to calculate reactors of
commercial size.
The central hypothesis of this work is that a practical alternative approach
to the modelling of packed bed reactors can be found that is a combination of
the engineering and scientic approaches and that combines the advantages of
both. This tool consists of a library of routines that can be linked to a CFD
code and that allows the engineer to model a packed bed using engineering
rules within the CFD domain. For this, models needed to be developed in
several areas and scales.
The important processes in a packed bed reactor are reaction, convection
and dispersion. Reaction takes place on the scale of a single particle. For single
reaction systems, simple particle geometry and/or simple kinetics, approxima
tion methods are available that are accurate enough and computationally in
expensive. However, in the most important practical cases, multiple reactions
play a role. For reaction networks in complex catalyst geometry, no approxi
mation method was available. Therefore, a method was developed to extend
the so-called Aris number method to reaction networks with general kinetics
in porous catalyst particles with general geometry.
Convection and dispersion take place on the scale of the bed. Therefore
they are determined by the structure of the bed. Near walls and solid inserts,
the structure of the bed is dierent from the bulk of the bed; therefore also
convection and dispersion are dierent near a wall and far from a wall. In a
CFD model, we need to be able to describe the ow near the wall correctly,
based on local values of the bed parameters. However, engineering models for
convection and dispersion in packed beds that are available from literature are
based on global or average bed parameters. Therefore, new models needed to
be developed.
As a basis, a model was developed to describe the bed structure, especially
near a wall, in terms of local parameters: the porosity, tortuosity and specic
outer particle surface area. To obtain these parameters, a computer program

ii

was written that generates simulated random packings of uniform spherical

particles. General rules for the bed structure parameters were extracted from
a large number of these computer simulated random packings.
A relation was developed that gives ow resistance in the catalyst bed
as a function of the local values of the bed structure parameters only. This
model was based on a physical description of the structure of the ow channels
in a packed bed. It was found that, for standard values of the bed and ow
variables, this model gives identical results as the well-known engineering rules
for pressure drop in packed beds. This is remarkable, because our model was
not tted to experimental packed bed pressure drop data. In addition, our
model can be used to describe the ow near walls.
The dispersion (mixing) behaviour of the bed was also described in terms
of local bed structure parameters (porosity, tortuosity, outer particle surface
area). As heat transfer with a packed bed reactor takes place through a wall,
the wall region is important for many non-isothermal reactions. Available
engineering methods use tted parameters (for instance, wall heat transfer
coecients) or proles that are not consistent with the bed structure. Our
model was derived directly from the bed structure and was found to describe
the literature experimental data for standard cylindrical beds just as well as the
literature models, even though no tting was used. In contrast with literature
models, no assumptions regarding the bed geometry were made in our model,
so it can in principle also be used for non-cylindrical packed beds.
In the nal step, all models were combined with a freely available open
source CFD code. The data generated by the simulations is relatively detailed,
giving pressure, velocity, temperature and concentrations at every point in the
catalyst bed. Therefore, it is hard to nd suitable experimental data in liter
ature for a full validation. A limited validation was done using experimental
literature data for a laboratory scale ethylene oxidation reactor. Good cor
respondence was found between the concentration and temperature proles
predicted by the model and the literature data.
To demonstrate the capabilities of the CFD code, two model reactions were
simulated: the catalytic reduction of NO with ammonia that is used in ue gas
treatment, and the catalytic oxidation of sulphur dioxide that is part of the
sulphuric acid production process. Dierent reactor designs were simulated
and compared, showing that the CFD model is a useful tool to simulate and
optimise such systems.
In general, application of CFD techniques in chemical reactor design is
limited by the high cost of the software and the high level of expertise needed
to use it. The packed bed simulation code that was developed in this work is
distributed as part of the Comow CFD code. This package allows engineers

iii

with a basic knowledge of uid dynamics to simulate packed bed reactors at

a detailed level without the need for in-depth expertise the specics of CFD
software. It is distributed as open source software and therefore it is freely
available for the reactor engineering community.
We conclude that it has been shown that a CFD tool has been developed
which can be used as a design tool for packed bed chemical reactors with
complex reactions, non-standard geometry and high interaction of ow and
heat transfer. The model gives results that are more detailed than the engi
neering relations, allowing more detailed optimisations and alternative reactor
designs. Therefore, we conclude that the central hypothesis of this work can
be accepted.

iv

Contents
1 Introduction
1.1 Randomly Packed Beds . . . . . .
1.2 Modelling approaches . . . . . . .
1.2.1 The engineering approach
1.2.2 The scientic approach . .
1.3 Hypothesis . . . . . . . . . . . . .
1.4 Method . . . . . . . . . . . . . .
1.4.1 Engine . . . . . . . . . . .
1.4.2 Packed bed model . . . . .
1.4.3 Sub-hypotheses . . . . . .
1.5 Structure . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

2 Chemical reaction and diusion in a single catalyst particle

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Theoretical background . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Diusion . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Kinetics . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 Balance equations . . . . . . . . . . . . . . . . . . . . .
2.3 Reduction of the number of dierential equations . . . . . . .
2.3.1 Single reaction . . . . . . . . . . . . . . . . . . . . . .
2.3.2 Multiple reactions . . . . . . . . . . . . . . . . . . . . .
2.4 Approximating the eectiveness factor . . . . . . . . . . . . .
2.4.1 Single reaction systems . . . . . . . . . . . . . . . . . .
2.4.2 Reaction networks . . . . . . . . . . . . . . . . . . . .
2.5 Example: the Sohio process . . . . . . . . . . . . . . . . . . .
2.5.1 Process description . . . . . . . . . . . . . . . . . . . .
2.5.2 Numerical solution . . . . . . . . . . . . . . . . . . . .
2.5.3 Approximation . . . . . . . . . . . . . . . . . . . . . .
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

11

11

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

15

18

21

21

23

25

27

27

29

31

31

34

44

44

48

51

65

3 The
3.1
3.2
3.3

wall eect
Introduction . . . . . . . . . . . . . . . . . . .
A short review of existing literature . . . . . .
Bed packing simulation models . . . . . . . .
3.3.1 Potential energy minimisation model .
3.3.2 Ball placement model . . . . . . . . . .
3.3.3 Modelling approach . . . . . . . . . . .
3.4 Bed packing simulation results . . . . . . . . .
3.4.1 Mean bed porosity and thickness eect
3.4.2 Sphere centre density . . . . . . . . . .
3.4.3 Porosity prole . . . . . . . . . . . . .
3.4.4 Particle surface area proles . . . . . .
3.4.5 Tortuosity proles . . . . . . . . . . .
3.5 Conclusions . . . . . . . . . . . . . . . . . . .

4 Flow resistance in randomly packed beds

4.1 Introduction . . . . . . . . . . . . . . . . .
4.2 A short overview of existing literature . . .
4.3 Flow resistance in innite packed beds . .
4.4 Flow resistance: the wall eect . . . . . . .
4.5 Velocity prole . . . . . . . . . . . . . . .
4.6 Conclusions . . . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

5 Dispersion in randomly packed beds

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . .
5.2 Dispersion in packed beds . . . . . . . . . . . . . . .
5.3 A short overview of existing literature . . . . . . . . .
5.3.1 Experimental methods . . . . . . . . . . . . .
5.3.2 The standard model . . . . . . . . . . . . . .
5.3.3 The wall heat conduction model . . . . . . . .
5.3.4 Other models . . . . . . . . . . . . . . . . . .
5.4 Dispersion modelling . . . . . . . . . . . . . . . . . .
5.4.1 Correlations for Pe at stagnant conditions . .
5.4.2 Correlations for Pe at high Reynolds numbers
5.4.3 Dispersion of momentum . . . . . . . . . . . .
5.4.4 Dispersion model . . . . . . . . . . . . . . . .
5.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . .

vi

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

71

74

76

87

87

88

92

92

92

96

100

102

106

113

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

119

. 121

. 122

. 123

. 136

. 136

. 139

.
.
.
.
.
.
.
.
.
.
.
.
.
.

147
. 149
. 150
. 156
. 156
. 157
. 158
. 161
. 162
. 163
. 170
. 176
. 178
. 178
. 181

6 Modelling of packed bed reactors using CFD

191

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

6.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

6.2.1 Basic ow and transport equations . . . . . . . . . . . . 195

6.2.2 Porosity, specic particle outer surface area and tortuosity197

6.2.3 A note on discretisation . . . . . . . . . . . . . . . . . . 199

6.2.4 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 200

6.2.5 Flow resistance . . . . . . . . . . . . . . . . . . . . . . . 213

6.2.6 Dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . 214

6.2.7 Pellet scale models . . . . . . . . . . . . . . . . . . . . . 215

6.3 Model verication and validation . . . . . . . . . . . . . . . . . 221

6.3.1 Flow in a packed tube . . . . . . . . . . . . . . . . . . . 221

6.3.2 Heat transfer in a steam heated packed tube . . . . . . . 226

6.3.3 Catalytic oxidation of ethane . . . . . . . . . . . . . . . 239

6.4 Demonstration . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

6.4.1 Reduction of NO in a packed tube . . . . . . . . . . . . . 243

6.4.2 Reduction of NO in a radial ow reactor . . . . . . . . . 247

6.4.3 Non-isothermal chemical reaction in a packed tube: ox

idation of SO2 . . . . . . . . . . . . . . . . . . . . . . . . 252

6.4.4 Oxidation of SO2 : alternative reactor design . . . . . . . 257

6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266

7 Conclusions

273

Samenvatting

281

vii

Chapter 1
Introduction
1.1

Randomly Packed Beds

Most industrial chemical processes are possible only through the use of a cat
alyst, i.e., a compound that participates in the reaction but that is not con
sumed. One of the key factors in designing industrial reactors for catalytic
processes is that the contact between the process uid and the catalyst needs
to be as intimate as possible during the reaction, while it should be easy to sep
arate the catalyst and the product stream after the reaction. In homogeneous
catalysis, the catalyst is part of the reaction mixture so contact is excellent
but recovery of catalyst after the reactor requires a separation step. In het
erogeneous catalysis, the catalyst is a solid material with active sites on its
surface. Therefore, separation of catalyst and process uid is relatively easy,
but transport of mass and heat between the uid and the catalyst is critical.
Although many reactor concepts for heterogeneous catalysis can be conceived,
the most common type found in the process industry is the randomly packed
bed reactor.
In randomly packed bed reactors, the solid catalyst usually consists of
porous particles of an inert (e.g., ceramic) carrier material that is doped or
coated with catalytically active material. These particles are placed in the re
actor vessel and kept in place by means of, for instance, sieve plates, through
which the process uid enters and leaves the reactor. Randomly packed bed
reactors are widely used throughout the chemical process industry, for example
in reactors for steam reforming and subsequent reaction steps like water-shift
and partial oxidation in the production of hydrogen from hydrocarbons, oxi
dation reactors in the sulphuric acid production, ethane oxidation, methanol
synthesis and in many oil rening processes like hydrocracking and hydrodesul
phurisation. Apart from these examples of chemical reactors, randomly packed

1. Introduction

beds are also widely used in physical process equipment like mixers, absorbers

and scrubbers.
Although many of the topics apply to randomly packed beds in general, the
focus in this work is on the packed bed reactor. Such a reactor typically con
tains a large number of catalyst particles. Reactants are transported through
the bed by the owing uid and are then transported to the active surface by
diusion. After the reaction has taken place, the products are transported in
the reverse order. The ow pattern in a packed bed is complex: the maze-like
structure of the ow paths in between the particles causes the ow to contin
uously change direction, split up into separate side streams and merge with
other side streams. This ow pattern causes a relatively high pressure loss in a
packed bed, but also provides the high degree of mixing that is often desired.
The wall is an important factor in the operation of a packed bed reac
tor. The reactor can in principle have any shape, but cylindrical vessels are
the most common. In many cases, especially where heat has to be removed
from or added to the reactor, the catalyst material is contained in a number
of smaller tubes. The space between these tubes is lled with a heating or
cooling medium and heat is transferred through the tube wall. Inside the bed,
heat is transported by conduction and convection through the uid, but also
by radiation and by conduction through the solid particles. The porosity of
the bed, i.e., the fraction of open space in between the particles, plays an im
portant role in these processes. Far from any wall or other solid object, the
porosity of a packed bed is about independent of position (if averaged over
a volume that is small compared to the reactor but large compared to the
particles). When we get near a wall, the structure of the bed changes be
cause the particles cannot penetrate the wall. This change in structure results
in a porosity prole perpendicular to the wall. Due to this prole the local
ow resistance and transport parameters change compared with the far-from
the-wall values, causing ow maldistribution and changing concentration and
temperature proles near the wall.
The combination of the practical importance and relative complexity of
these phenomena has spurred research from the 1930s onward. Through the
years, an increasing detail in both measurements and models can be seen.

1.2

Modelling approaches

There are various approaches to the modelling of packed bed chemical reactors,
diering in level of detail and complexity as well as computation time, ranging
from the engineering approach on one end of the spectrum to the scientic
approach on the other.

1.2 Modelling approaches

1.2.1

The engineering approach

In the engineering approach, the goal is to estimate the performance of a packed

bed with sucient accuracy to allow the design, rating and optimisation of the
reactor. To be of practical relevance, the engineering models need to be sim
ple. Therefore, a selection has to be made of the processes and phenomena
that have the largest impact on the most relevant output parameters for a
specic system. This selection depends on the specic features of the process
and therefore requires expertise and intimate knowledge of the system. As the
model becomes more accurate, taking into account more physical phenomena,
the mathematical complexity also increases. As a result, such models often
turn out to be not very simple. The parameters used in an engineering model
are often obtained from measurements on an existing reactor. Therefore, they
are specic for a given design. A disadvantage of this approach is that for each
dierent combination of catalyst, reactor, reactants and process conditions a
new model with a dierent choice of sub-models, needs to be developed. Math
ematically, the model is typically a set of 1-dimensional dierential equations,
but depending on the complexity of the process and on the type of output data
required (spatial or temporal variations), the solution of multi-dimensional or
time-dependent partial dierential equations may be necessary. As output, the
engineering model gives predictions for macroscopic variables that can often
be measured directly in the actual reactor. Comparison between experimental
data and model predictions leads either to acceptance or to adaptation of the
assumptions on which the model is based and thus to insight. In addition,
the parameters of the model are often obtained by tting the model results
to experimental data. These parameters then contain not only measurement
errors but also the eects of all phenomena that were not included (correctly)
in the model: therefore these parameters are coupled specically to the model.
This is shown schematically in gure 1.1.

1.2.2

The scientic approach

In the scientic approach, the purpose of the model is to increase the under
standing of the dierent phenomena that take place in a given process. Ideally,
the model starts from rst principles and general parameters like physical prop
erties that are not specic for the process or geometry. A general schematic of
this process is given in gure 1.2. The model describes the physical phenom
ena on a microscopic level. For example, detailed 2 or 3-dimensional Com
putational Fluid Dynamic (CFD) models are built to describe the motion of
particles in a uidised bed or the interaction between liquid ow and gas bub
bles in a bubble column. In such simulations, the particles or bubbles are

1. Introduction

specific
parameters

parameter
fitting

simple
model

assumptions

macroscopic
prediction

compare

experiments

model
adaptation

insight

Figure 1.1: The typical structure of the engineering approach

modelled as separate entities, where additional microscopic models describe

the interaction between two particles and between particles and the uid. For
packed bed reactors, models like these have been proposed to calculate pressure
drop, heat and mass transfer parameters (Derckx and Dixon, 1996).
In order to describe the inuence of a particle on the ow, the ow equations
need to be solved on a length scale smaller than the particle size. Consequen
tially, these models are computationally intensive. In practice only a limited
number of particles typically much less than a thousand can be taken
into account. For instance, to calculate ow and heat transfer in a packed
bed of only 10 particles already requires a computational grid of more than
a million of cells. The volume of an actual chemical reactor is often many
cubic meters, and the particle size is usually in the order of one millimetre.
Therefore, a typical reactor contains billions of particles. It is clear that it
is not practical to compute a complete reactor using these models. Instead,
these models should be regarded as research tools for a better understanding of
the microscopic processes. A better understanding leads to better macroscopic
approximating relations that can be used in improved engineering models.

1.2 Modelling approaches

A comparison of model results with experimental data (often from detailed

lab-scale experiments) is required to validate the choice of microscopic models,

but in contrast with the engineering approach, in the scientic approach insight
in the behaviour of reactors on a macroscopic scale is gained mainly from
analysis of the results of the model.

general
parameters

insight

complex
model

first
principles

microscopic
results

macroscopic
relations

model
adaptation

compare

experiments

Figure 1.2: The typical structure of the scientic approach

1. Introduction

1.3

Hypothesis

The central hypothesis of the present work is that a practical alternative ap

proach to the modelling of packed bed reactors can be found that is a com
bination of the engineering and scientic approaches and that combines the
advantages of both.
In this alternative approach, the engineering rules are adapted and applied
on a local scale. A simplied scientic CFD solver is used as the compu
tational engine (see gure 1.3). To be of practical use, the approach needs
to give results at a higher accuracy than the engineering method, but with
a considerably lower computational eort and requiring a lower level of CFD
expertise than the scientic approach. In addition, it should be accessible and
easy to use for the reactor engineer. The model should be suciently generic
to be applied to the majority of reactor designs without major modications.

engineering
approach

error

els ve
od pro
bm m
su t / i
ap

ad

adap
this
work

engi
ne
t /sim
plify

scientific
approach
complexity / computation time

Figure 1.3: Graphical representation of the engineering approach and scientic

approach on both ends of the error computational eort spectrum and the
approach developed in this work

1.4 Method

1.4

Method

To be able to accept or reject the central hypothesis, we will develop a tool that
consists of a packed bed model, implemented in an existing CFD code. The
CFD code calculates the balances of mass, momentum, enthalpy and chemical
species, while the packed bed model supplies values for a.o., friction force, dis
persion coecients and source terms, as is shown in gure 1.4. The packed bed
model consists of four sub-models: the particle model, that describes reaction
and diusion within a porous catalyst particle; the packing simulation that
gives the porosity, tortuosity and particle outer surface area at each location
in the bed; the friction model and the dispersion model.

Packed bed model

CFD
friction model

packing simulation

dispersion model

mass balance

momentum balance

enthalpy balance

species balance

S(T)
particle model
S(c)

Figure 1.4: Schematic representation of the CFD and packed bed models;
major interactions and variables. The variables shown are the pressure p, the
uid velocity vector v, the temperature T and the concentrations c. The linking
coecients are the friction force vector F , the dispersion coecients D and
the temperature and component sources S(T ) and S(c). The bed packing is
characterised by the porosity , tortuosity and specic particle outer surface
area a at every point in the bed.

1. Introduction

For use in day to day engineering practice, the tool should have the follow
ing characteristics:
The tool should be generic, so that many dierent reactor systems can
be described without the need to modify the code
The input of the model should be in terms of parameters known to the
reactor engineer, avoiding parameters that are familiar only to the CFD
expert
The tool should run on a standard desktop or laptop PC
The calculation time should be low enough to simulate several cases in
a single day for a typical calculation
The model should be able to handle single and multiple reactions, com
plex kinetics and complex particle shapes
The model should also be valid for packed beds with a low bed-to-particle
diameter ratio
The tool should be freely available

1.4.1

Engine

The engine for the randomly packed bed reactor model will be a simplied
version of the scientic CFD codes. To be useful for use in reactor design,
the CFD package should also have a preprocessor to set-up a simulation and
a postprocessor for displaying and analysing the results. The CFD package
selected in this work is Comow. In contrast with many larger CFD systems
that require a ow dynamics expert to be operated, Comow was developed as
a tool for the process engineer. It is simple, in the sense that it supports only
a subset of the general CFD functionality: 2-dimensional mostly hexagonal
grids and only very simple turbulence modelling. On the other hand, there are
comparatively extensive models for physical properties and ow restrictions.
Comow is a pre- and postprocessor for an external CFD solver, and it extends
the solver with a library of routines to support the additional models, e.g., tube
bundles and perforated plates (Roelofs, 1998).
In the past, Phoenics (http://www.cham.co.uk) has been used as a solver
for Comow, as well as a custom solver based on the (now historic) champion
code (see e.g., Pun and Spalding, 1976). For the present work, it was decided
to create a new version of Comow based on a more modern CFD solver. The
requirements for the solver, that it should be lightweight and that the source

1.4 Method

code should be available, led to the selection of Dolfyn (www.dolfyn.net). This

is an open source, 3-D unstructured, single phase CFD code written in Fortran
90 and based on the numerical method described by Ferziger and Peric (2002).
Since an open source CFD code is used as engine, and the source code of the
packed bed model will also be released as open source, the result of this eort
will be a tool that is freely available for every reactor engineer.
Dolfyn is a relatively simple CFD code that does not include many of the
facilities needed to simulate packed bed reactors. Therefore, the code will
be extended to implement variable physical properties of ideal gas mixtures,
custom source terms (allowing the specication of the drag force F and the
reaction source terms S) and anisotropic dispersion (D in gure 1.4).

1.4.2

Packed bed model

In a CFD code, the ow domain is divided into a large number of computational

cells. The balance equations for mass, momentum, enthalpy and concentra
tions are linearised, discretised and then solved for each cell. As the resulting
algebraic equations for each cell depend on all of the cells neighbours, a large
(sparse) system of equations is formed. Because of the non-linear nature of
the original equations the solution of this system requires an iterative pro
cedure. After each iteration, new estimates for the state variables (pressure,
temperature, velocity, temperature and composition) are made for each cell.
The packed bed model interacts with the CFD code through transport
parameters (e.g., dispersion) and source terms (e.g., heat source and ow re
sistance). It provides the values of the parameters that are specic for a packed
bed, based on the local values of constants (e.g., porosity) and state variables
(e.g., temperature) that are calculated by the CFD code. These values are
then used in the CFD code for the next iteration.
The local values of the transport parameters and the reaction rates are
evaluated once for each cell for each iteration. Hence they are calculated very
often during a simulation. Each sub-model calculation should therefore be
computationally inexpensive.
The transport processes in a packed bed are diusion (dispersion, conduc
tion), convection and radiation. These processes have a limited characteristic
length over which information can be passed, typically in the order of a few
particle diameters. Therefore, theoretically, the data needed to calculate the
transport parameter in a computational cell should depend only on the local
values of the ow variables, i.e., the values for the cell itself and its direct
neighbours. For instance, the radial dispersion in a cell should be given in
terms of the velocity in that cell (and possibly that of a neighbouring cell); it
cannot depend directly on the velocity at the inlet or at the centre of the bed.

1. Introduction

As another example, the diameter of the bed should not show up in any of
the equations used to calculate the transport parameters. Many engineering
models, however, use this kind of global information. Therefore, they cannot
be used in a CFD model without modication.
In summary, we can conclude that to be able to calculate transport param
eters locally in a CFD code, the submodels should have the following charac
teristics:
the calculation should be computationally inexpensive
the parameters should not depend directly on the geometry of the bed
the parameters should not depend directly on global information
The important processes in a packed bed reactor are reaction, convection
and dispersion. Reaction takes place on the scale of a single particle. For
single reaction systems, simple particle geometry and/or simple kinetics, ap
proximation methods are available that are accurate enough and computa
tionally inexpensive. However, in the most important practical cases, multiple
reactions play a role. For reaction networks in complex catalyst geometry, no
approximation method is available. Therefore, this needs to be developed.
Convection and dispersion take place on the scale of the bed. Therefore
they are determined by the structure of the bed. Near walls and solid inserts,
the structure of the bed is dierent from the bulk of the bed; therefore also
convection and dispersion are dierent near a wall and far from a wall. In a
CFD model, we need to be able to describe the ow near the wall correctly,
based on local vales of the bed parameters. Engineering models for convection
and dispersion in packed beds that are available from literature are based on
global or average bed parameters and therefore need to be adapted.
We need to describe the bed structure, especially near a wall, in terms of
local parameters like the local porosity, tortuosity and the local specic outer
particle surface area. These parameters are hard to determine experimentally,
but easy to calculate for a simulated packing. A computer program can be used
to generate simulated random packings and general rules for the bed structure
parameters can be extracted from these.
Once the bed structure is known, a relation needs to be developed that
describes ow in the catalyst bed (and especially ow resistance) as a function
of the local values of the structure parameters (porosity, tortuosity, outer par
ticle surface area) only. This model should not only give the correct pressure
drop but also correctly describe the velocity proles near a wall.
The dispersion (mixing) behaviour of the bed should also be described in
terms of local bed structure parameters. As heat transfer with a packed bed

10

1.5 Structure

reactor takes place through a wall, the wall region is important for many nonisothermal reactions. Available engineering methods use tted parameters (for
instance, wall heat transfer coecients) or proles that are not consistent with
the bed structure.

1.4.3

Sub-hypotheses

In order to full the main hypothesis, existing engineering models need to be

re-evaluated or adapted and new models need to be developed. We therefore
formulate a set of sub-hypotheses that are conditions for the main hypothesis:
1. It is possible to develop a method to calculate chemical reaction with dif
fusion limitation inside a catalyst particle that requires a limited amount
of computation time and yet is accurate enough to use in a reactor model.
An approximation method that handles complex reaction networks, gen
eral kinetics and general catalyst particle shape can be developed.
2. The structure of a packed bed near a wall can be characterised such
that the main parameters that inuence transport processes in the bed
(porosity, particle outer surface area and tortuosity) can be described as
a function of the location in the bed. These functions can be derived
from computer simulations of contained random packed beds.
3. The existing engineering rules to calculate ow resistance of packed beds
can be adapted and rewritten in terms of the main parameters (porosity,
particle outer surface area and tortuosity) so they can be used on a local
scale in a CFD-based packed bed model. These models should correctly
predict the pressure drop as well as velocity proles near the wall, without
using global t parameters.
4. Dispersion models can be developed that describe heat and mass trans
port in a packed bed based on the main bed parameters in such a way
that they can be used locally in a CFD-based packed bed model. These
models should be consistent with the bed structure and cannot use su
percial t parameters

1.5

Structure

Each of the following chapters will address one of the sub-hypotheses given in
the previous section. We start at the smallest length scale: the processes inside
a catalyst particle. In chapter 2, the methods used to calculate the reaction
rate in a porous catalyst particle, taking into account diusion limitation are

11

1. Introduction

evaluated. A method is needed that is valid for general particle shape, general
kinetics and also for reaction networks. For simple particle geometry and/or
for simple reaction kinetics, analytic, numeric or approximative solutions are
available in the form of the Aris number method (Wijngaarden et.al. 1998).
However, if multiple reactions take place in a catalyst particle with a complex
geometry, the approximation methods cannot be used. In that case, the eec
tive reaction rate in the particle can only be found by integration of a set of
partial dierential equations. The computational eort for this is higher than
we can aord in a CFD packed bed model. To remove this limitation, we will
extend the Aris number method so that it can be used for reaction networks.
Chapter 3 focuses on the packing of the bed, especially near a solid wall.
It is well known that near a solid surface like the container wall, the particles
in a packed bed are no longer distributed in space in a random way. Due to
the fact that the wall cannot be penetrated by the particles, a more or less
regular pattern occurs near it. This pattern is most clear and well-dened for
packings of smooth, identical spheres. For these, the porosity has a value of 1
at the wall, drops to a minimum value of about 0.2 at one radius from the wall,
increases again and approaches the mean bed value (usually about 0.4) in a few
oscillations, at a distance of ve particle diameters from the wall. The prole
is less well-dened for non-spherical particles or packings of particles with a
broad particle size distribution. This so-called wall eect has been included in
packed bed models in the past through an exponential approximation of the
porosity prole. Although this may give a reasonable description of the poros
ity proles for packings of irregular particles, it does not describe the minimum
in the porosity prole found for regular particles. In addition, it is not solely
the porosity that is of interest. The specic particle outer surface area and the
tortuosity of the ow paths also determine the transfer of momentum, mass
and heat between the uid and the particles. Both these parameters are in
uenced by the presence of the wall. Therefore a procedure will be developed
to determine the distribution of particles near the wall for perfectly spherical,
mono-disperse particles through computer simulation of bed packings. The
porosity, tortuosity and particle outer surface area proles can be determined
from the simulated particle distribution.
In chapter 4 the porosity, specic outer particle surface and tortuosity
proles from chapter 3 are used to calculate the ow distribution and pressure
drop in a packed bed. The basic equation used to calculate ow resistance in
a packed bed is the well-known Ergun equation. However, in its most common
form, this equation does not contain the specic surface area or the tortuosity
explicitly. It is clear that these two parameters have a large inuence on the
ow resistance. Therefore, a new equation will be derived from a parallel

12

1.5 Structure

twisted tubes model, using common engineering expressions for the pressure
loss in tubes, at bends, expansions and contractions. The shape of the tubes
is a function of the distance to the wall and related to the bed structure. With
this equation, the radial velocity proles in a packed bed can be calculated.
These will be compared to measured packed bed velocity proles that are
available in the open literature.
Chapter 5 builds on the results from chapter 3 and 4 to model the transfer
of mass and heat in a randomly packed bed. The eective heat conduction is a
combination of molecular or stagnant terms (conduction through the uid and
solid phase) and the mixing that is caused by the random ow path through
the bed. Both contributions depend on the porosity and tortuosity proles,
as well as on the velocity prole in the bed. A new dispersion model will be
developed that takes these proles into account.
Each of the sub-models will be veried and/or validated using literature
measurement data for packed tubes. The use of the models for non-tubular
geometries cannot be validated by lack of experimental data. However, pro
vided the correct particle density distribution or porosity, tortuosity and par
ticle outer surface areas are supplied, there is no fundamental reason why the
model would not apply to dierent geometries.
Finally, in chapter 6 the sub-models developed in the previous chapters will
be combined with the CFD code to address the main hypothesis.
Just like the individual sub-models, the resulting combined model will be
validated using literature data measured in packed tube reactors. Unfortu
nately, the number of useful cases in literature even for cylindrical geometry
is limited. For other geometries, no literature measurement data is available.
Therefore, the application of the model for non-tubular geometries will be
demonstrated instead of validated.
The data generated by the CFD simulations is relatively detailed, giving
pressure, velocity, temperature and concentrations at every point of the grid.
Therefore, it is quite a task to duplicate this resolution in experiments and
measurements, and validation of the model on a detailed level in a complex
simulation is not possible at this time. This applies to this work in particular,
but also more in general as the possibilities for detailed modelling improve
rapidly. To make similar improvements in measurement techniques for chemi
cal reactors is outside the scope of this study and will be left as a challenge to
future experimentalists.

13

1. Introduction

Literature

Derckx, O.R., Dixon, A.G. (1996), Determination of the xed bed wall heat
transfer coecient using computational uid dynamics, Num. Heat Transfer
part A: Applications, 29(8), 777-794
Ferziger, J.H., Peric, M, Computational methods for uid dynamics, 3rd Ed.,
Springer Verlag, 2002
Pun, W.M., D.B. Spalding, A General Computer Program for Twodimensional Elliptic Flows, HTS 76/2, CHAM Ltd. (1976), as described on
http://www.simuserve.com/cfd-shop/hts76-2.htm
Roelofs, H.J. (1998), CFD-programma Comow is van alle markten thuis,
npt procestechnologie, 5 (september-oktober), 27-29
Wijngaarden, R.J., A. Kronberg, K.R. Westerterp (1998), Industrial
catalysis: optimizing catalysts and processes, Wiley-VCH, Weinheim

14

Chapter 2
Chemical reaction and diusion
in a single catalyst particle
Summary
The goal of this chapter is to develop a method to calculate chemical reac
tion with diusion limitation inside a catalyst particle that requires a limited
amount of computation time and yet is accurate enough to use in a CFD reac
tor model. This method should be able to handle complex reaction networks,
general kinetics and general catalyst particle shape.
The reaction rate usually depends in a non-linear way on several reactant
(or product) concentrations and on the temperature. Therefore, the mass and
heat balances lead to a set of non-linear dierential equations, one for each
component and one for the temperature. For a single reaction, it is easily
found that these equations are not independent. It suces to solve only the
dierential equation for a key component.
For reaction networks, the number of degrees of freedom is equal to the
number of independent reactions, so the set of dierential equations can also
be reduced. However, the selection of key components is not always straight
forward and not easy to do in a generic way. It is shown here that the number
of equations in the set of dierential equations can be reduced eciently to the
minimum number by adding a virtual key component to each reaction, with
a pseudo concentration zj for each reaction j. The resulting dierential equa
tions, one for each virtual key component, can be solved simultaneously. The
real concentrations and the temperature can be calculated from the pseudo
concentrations through algebraic equations.
For single reactions, the Aris number approximation can be used to calcu
late the eectiveness factor for any particle shape and kinetics to a claimed

15

2. Chemical reaction and diusion in a single catalyst particle

accuracy of 5-10 % (Wijngaarden and Westerterp, 1994). The method is writ

ten here in a slightly more general way in terms of the virtual key component
concentration. In this way, a single set of equations can be used to estimate
the eectiveness factor for a single, isothermal or non-isothermal reaction with
any number of components in any catalyst particle geometry.
For reaction networks, the particle geometry is important. In 1-dimensional
particles, the balances lead to a set of ordinary dierential equations. Although
not trivial, it is feasible to solve these with an acceptable computational eort.
For particles with a more complex geometry, the mass and heat balances lead
to a coupled set of partial dierential equations. The numerical solution of this
set involves a relatively large computational eort. It is not feasible to solve
this system for each cell and iteration in a CFD computation of a complete
chemical reactor. Therefore, an approximation method is developed in this
work to estimate the eectiveness factors for reaction networks with general
kinetics in general particle geometry.
Our approach is based on the Aris number method. An estimate for the
eectiveness factor for each reaction is calculated for the fast reaction regime
and for the slow reaction regime; the actual value is then found by interpolation
between the two regimes.
For the fast reaction regime, the reaction takes place in a thin shell of the
particle. The proles for the pseudo concentrations become independent of the
particle size. Each strictly increases from zero at the particle edge to approach
an equilibrium value at a certain penetration depth j . The equations reduce
to a set of ordinary dierential equations. This set is integrated to nd the
zeroth Aris number for each reaction.
For the slow reaction regime, the eectiveness factor for the components will
be close to one. The kinetics are linearised around this point. The deviation
of the eectiveness factor for a component i from one will depend on the
reactions in which component i is involved, but also on other reactions that
produce or consume components that appear in the rate equation of each of
these reactions. It is shown that the rst Aris number An1,i for component i
can be approximated by:

 2 
nr
nr


Vp
j 
2
An1,i lim(1 i )
(2.1)
i,j
k,s

1
i,s Ap
z
k
0
z
k
j=1
k=1
Note the cross terms on the right hand side that give the dependence of
reaction j on the value of the pseudo concentration of all other reactions k.
To obtain an estimate for the eectiveness factor we need to interpolate
between the fast and slow regimes. For single reaction systems, Wijngaarden

16

and Westerterp (1994) propose an implicit interpolation formula to obtain the

eectiveness factor from the values of An0 and An1 . This approach can not be
used for reaction networks. Instead, an explicit method is presented, based on
the penetration depth j of reaction j.
The approximation method is demonstrated for the ammoxidation of pro
pene in the Sohio process. The kinetics for this process are known from litera
ture. The reaction network consists of six reactions. Each reaction is assumed
to be rst order in its hydrocarbon reactant concentration and includes an
Arrhenius type temperature dependence. The eectiveness factors are calcu
lated using the approximation method as well as by direct integration of the
reduced equations for a range of slab-shaped catalyst particle and for a ring
shaped catalyst particle. The approximation method correctly describes the
trend of the eectiveness factor curves as a function of particle size. The curves
take the correct values in the limits of low and high reaction rate, and show
the correct trend in between. Eectiveness factors higher than one and lower
than zero are correctly predicted and in the right order of scale. It is shown
that the error made in the approximation is in the order of 5 % under most
conditions and for the primary reactions, but may be as large as 50 % for the
secondary reactions in conditions where the eectiveness factor is much larger
than one. The large errors are probably caused by the fact that the cross terms
in the calculation of An1 are taken at bulk conditions. The eect of the cross
terms is only large if the reaction rate itself is very low compared to the other
reaction rates. Consequently, the larger errors do not have a large eect on
the mass balances.
To judge the eect of the errors in a practical calculation, a very simple
model of an ideal tube reactor packed with ring-shaped particles is used to
calculate the conversion and selectivity towards the desired product and the
by-products for the Sohio process. The maximum dierence between the selec
tivity curves based on the approximation and those based on the full numerical
solution is less than 3 percent point, which is satisfactory. It is concluded that
for the systems that have been modelled, the approximation method can be
used with sucient accuracy to calculate the eectiveness factor. In principle,
the method is valid for any particle shape and for general kinetics. Only rst
order kinetics with Arrhenius type temperature dependence were tested in this
work. Application of the approximation method for other particle geometries
and other reaction systems is left for future studies.

17

2. Chemical reaction and diusion in a single catalyst particle

2.1

Introduction

In packed bed reactors, usually porous particles are used as carrier for the
catalyst. These particles have a large internal surface area (hundreds of square
meters per gram). On this surface, catalytically active sites, for instance nickel,
platinum or cobalt atoms, are dispersed. The catalyst particles can have many
shapes; spherical catalyst particles are sometimes used but cylindrical particles
are more common as the are more easily produced. More complex extruded
shapes like rings, trilobes and quadrulobes are often chosen to optimise the
pressure drop and conversion for a specic process.
For a heterogeneously catalysed reaction, the following steps can be distin
guished (see gure 2.1):
1. diusion of the reactant from the bulk of the uid through the hydrody
namic boundary layer to the outer edge of the particle
2. diusion of the reactant through the pores into the particle to reach an
active catalytic site
3. reaction on the active site to form products with a certain reaction rate
4. diusion of products through the pores to the outer edge of the particle
5. diusion of products through the hydrodynamic boundary layer into the
bulk of the uid
In this chapter we will only consider the processes inside the particle itself;
transport through the boundary layer will be discussed elsewhere in this work.
Inside a porous catalyst particle, there is a balance between transport of
reactants and products by diusion and conversion by the chemical reaction.
It is clear that if the particle is large (or if the reaction rate is high compared
to the diusion rate), the reactant concentration will decrease strongly inside
the particle. In this case, the reaction will mainly take place in a thin layer
on the outside of the particle and the central part of the particle will not be
used. This is a waste of reactor space and catalyst material. If the particles
are small (or the reaction slow compared to diusion), the concentration in
the particle will be (almost) the same as in the bulk gas and all catalytic
sites of the catalyst are used to an equal degree. However, small particles
give rise to a high pressure drop over the bed which is usually not desirable.
Therefore, packed bed reactors are often designed in a way where the particle
size is a trade-o between catalyst usage and pressure drop. In other words,
most packed bed reactors are operated in a range where the reaction can be
considered neither fast nor small. In these cases, the eective reaction rate

18

2.1 Introduction

boundary layer
catalyst particle

CA,b

1
3

CB,s
CB,b

CA,s

2
5

Figure 2.1: Schematic representation of the steps involved in chemical reaction

in a porous catalyst particle for the reaction A B and typical concentration
proles.

depends not only on the bulk composition and temperature but also on the
proles inside the catalyst particle.
To calculate the conversion of reactants to products in a packed bed reac
tor, we need to be able to predict the performance of a single catalyst particle
as a function of the bulk conditions. Sometimes, only one reaction takes place
between reactants and only the overall conversion needs to be calculated. How
ever, often there are side reactions leading to dierent products or the desired
product can react with a reactant to form an undesired by-product. In those
cases, the selectivity of the reactor is an even more important factor than the
conversion, because it determines the economy with which reactants are used
and the amount of (possibly hazardous) waste that is produced.
Computational Fluid Dynamics (CFD) can be used as a tool in the design
of more ecient reactors using heterogeneous catalysis. In a CFD model, the
reactor is divided in to a large number (typically thousands to millions) of
small cells. The ow equations (mass, momentum and energy balances) are

19

2. Chemical reaction and diusion in a single catalyst particle

solved for each of these cells. To do this, we need to be able to predict the
performance of particles at the local conditions in each cell. Ideally, we should
be able to use simple (spherical, cylindrical, slab-shaped) and complex particle
shapes (trilobes and other extrudates), for single reaction systems, for parallel
and consecutive reactions and for even more complex reaction networks. Also,
the computation time available to calculate the eective reaction rate in the
particles is small because the calculation needs to be carried out many times
(once for each cell for each iteration). The goal of this chapter is to give e
cient methods to calculate internal diusion limited reaction in porous catalyst
particles that can be used in such a CFD code.
We will rst briey discuss discuss the basic diusion-reaction equations
for reaction networks. Although the balance equations lead to one (partial)
dierential equation for each active component in the system, it is known that
in fact the number of independent dierential equations is in fact equal to the
number of (independent) reactions which is always lower or equal. We will
introduce a generic way to nd the independent dierential equations by the
introduction of a virtual key component for each reaction. For one-dimensional
particles (spheres, long cylinders and large slabs), the balance equations lead
to a set of coupled ordinary dierential equations. This set of equations may
be solved numerically with an acceptable computational eort.
For more complex particle shapes, a system of coupled partial dierential
equations needs to be solved to calculate the conversion in a single particle.
Although this is possible for a single particle, the purpose of this work is to
calculate a complete packed bed reactor in a Computational Fluid Dynamics
code. It is certainly not desirable to solve a set of partial dierential equations
for each computational cell and for each iteration in a CFD model.
Even though CFD codes may be able to compute a solution to a high nu
merical accuracy, the errors made in modelling and determination of parame
ters is usually in the order of several percents. Therefore. there is no need to
spend a lot of eort to compute the reaction rate to a much higher accuracy
than that. Therefore, approximation methods to estimate the conversion in a
catalyst particle are acceptable for our purposes.
To estimate the conversion in a catalyst particle for simple cases, in Thiele
(1939) introduced the notion of the eectiveness factor of a catalyst particle,
which is dened as the conversion in the particle as a fraction of the conversion
that would be found if no mass transfer limitation would be present. The value
of this eectiveness factor can be calculated analytically for single reactions
with some specic forms of the kinetic equations (for instance n-th order reac
tions) in one-dimensional particles. Wijngaarden and Westerterp (1994) have
introduced the so-called Aris number method that allows the estimation of the

20

2.2 Theoretical background

eectivity for single reactions with arbitrary single reaction kinetics and arbi
trary shapes. For multiple reactions, no approximation methods are available
so the only option is to use numerical methods. Therefore, the situation in the
existing literature can be summarised as is shown in table 2.1.
Table 2.1: Options available for the calculation of reaction and diusion in
catalyst particles for dierent particle shapes and kinetics

particle shape
kinetics
single simple

simple (1-D)

complex

analytical solution

single complex

Aris method

numerical (PDE)
Aris method

network simple

numerical
(ODEs)

numerical (PDEs)
?

network complex

As the solution of partial dierential equations is too computationally in

tensive to be performed within a CFD calculation, there are currently no op
tions for network kinetics in a complex particle shape. To be able to calculate
this class of problems in a CFD packed bed simulation, a method will be
developed to use the Aris number approach for reaction networks.
The methods will be demonstrated using literature kinetic data for the
Sohio process for the ammoxidation of propene. Note that, although the re
mainder of this work is mainly focused on packed bed reactors, the results
of this chapter are equally applicable to other reactor types where catalyst
particles are used like uidised bed reactors.

2.2

Theoretical background

2.2.1

Diusion

It is assumed that the diusion ux of component i inside a porous particle is

proportional to the concentration gradient:
i = De,i

ci
x

(2.2)

21

2. Chemical reaction and diusion in a single catalyst particle

The proportionality constant is the eective diusion coecient De,i for

component i in the gas mixture in the particle pores. This eective diusion
coecient diers from the bulk diusion coecient of component i in the gas
mixture due to three eects:
part of the space is occupied by solids and therefore not accessible to gas
molecules
the pores in the particle are often not straight. This increases the length
of the path a molecule has to travel to penetrate a certain distance into
the particle
the size of the pores is often in the same range as the free path length
of the gas molecules, so collisions of the molecules with pore walls (i.e.,
Knudsen diusion) needs to be taken into account
The shape of the channels is characterised by the pore tortuosity c , which
is dened as the length of a channel divided by the distance between the start
and end points. Commonly, the eective diusion coecient is expressed as
De,i = Dt,i

p
c

(2.3)

where Dt,i is the diusion coecient of component i in a straight pore and p

the particle porosity. The tortuosity is usually determined experimentally, and
typically values in the order of 3 or higher are found (e.g., Fogler, 1986). This
is remarkable, as a channel needs to be quite labyrinth-like to get such a high
value for the tortuosity. For a porous conglomerate of small solid particles (as
the porous materials usually are), one would sooner expect values between 1
and 2.
On closer inspection, it can be concluded that equation (2.3) is not correct.
The tortuosity of the channels has two eects: the distance along the length of a
channel that needs to be travelled is longer, but also the concentration gradient
needs to be taken along the channel axis and not along the coordinate axis.
Therefore, the eective diusion coecient decreases not by the tortuosity but
by its square
De,i = Dt,i

p
2
c,act

(2.4)

As the general practice is to use equation (2.3) instead of equation (2.4), we

will do so here, too. However, it should be kept in mind that the tortuosity
used in that equation is actually the square of the actual, geometric tortuosity
c = c,2 act

22

(2.5)

2.2 Theoretical background

This of course also explains the high values reported for c .

The error made in the estimation of the eective diusion coecients is
relatively large. Binary bulk diusion coecients are commonly estimated
with an error of about 5 %. In addition, most porous particles will contain
a distribution of pore sizes, and not all pores may be accessible to all gas
molecules. The shape of the pores and the characteristics of the maze structure
with interconnected or blocked-o pores are hard to model and the tortuosity
factor is often a rough guess. The resulting error in the eective diusion
coecient may well be in the order of 10% or more. Therefore, replacing the
diusion ux (equation 2.2) by the Maxwell-Stefan equations to account for
non-equimolar diusion (drift ux eects) will in most cases not drastically
improve the accuracy, but it will drastically increase computation time.

2.2.2

Kinetics

In this work, we will try to write equations that are independent on the choice
of the kinetic or reaction rate equations . However, there will be some re
strictions on the kinetics that can be used, so we will discuss a simple kinetic
model here.
In theory a gas phase reaction can be thought to be the result of a collision
between reactant molecules. The probability that a molecule of component A
and a molecule of component B collide is proportional to the concentration
of A and the concentration of B. The energy of the collision should be large
enough to overcome the energy barrier; the fraction of collisions that lead
to a reaction therefore increases with temperature with an Arrhenius factor.
Apart from that, the fraction of successful collision also depends on other
conditions like the orientation of the molecules when they collide, so there is
also a temperature independent factor. Therefore, for a reaction
A A + B B C C
the reaction rate should in theory be


EA A B
cA c B
A = k0 exp
RT

(2.6)

(2.7)

All reactions are in principle reversible; for the above example, if C mole
cules of C collide with sucient energy, they could react to reproduce the
original number of molecules A and B. The rate at which this happens of
course depends on the value of C : the larger it is the lower the chance that so

23

2. Chemical reaction and diusion in a single catalyst particle

many product molecules collide at (more or less) the same time. The reaction

should be written as:

A A + B B C C
with reaction rate




EA,b C
EA,f A B
A = kf,0 exp
cA cB kb,0 exp
cC
RT
RT

(2.8)

(2.9)

Starting from pure A and B, the reaction will continue until there is thermo
dynamic equilibrium; there is a relation between the equilibrium constant K
and the kinetics:


nc

kb,0
EA,f EA,b
i
K
ci =
exp
(2.10)
k
RT
f,0
i=1
For a heterogeneous reaction, the situation is somewhat more complicated.
For the reaction to occur, at least some of the reactants need to be adsorbed
on the catalytically active site on the surface. The absorption of a reactant is
in itself an equilibrium step in series with the actual reaction. The product(s)
formed are often also adsorbed on the surface and therefore need to desorb
to make the active sites available for another reaction. If one considers these
steps, the rate equation in the form of a Langmuir-Hinshelwood equation is
found, for instance,
i =

C
kb c C
kf cAA cBB

1 + kf cAA cBB
1 + kb cCC

(2.11)

In reality, there are many complicating factors and it is in general impos

sible except in a few cases to calculate the reaction rate based on theoretic
considerations. Therefore, the kinetic equations are determined by tting to
experimental data. the kinetic equations found often lump together many
reaction details like intermediate steps. One possibility for the form of the
equation that is often encountered is power law kinetics:


EA
n m p
(2.12)
A = k0 CA cB cC exp
RT
which is similar to the theoretical gas phase reaction but the exponents n and
m are no longer connected with the stoichiometric coecients and could take
any value; usually they are not integer numbers and sometimes they are even

24

2.2 Theoretical background

negative. The validity of the equations is limited to the range of conditions

that were included in the experiment.

For many kinetic equations, the reaction rate will be zero when the concen
trations of the (main) reactants are zero. For equilibrium reactions, some nite
reactant concentration will be found at equilibrium. For power law kinetics
with reaction orders of zero or lower, no equilibrium state exists; the reaction
rate is constant or even increases with decreasing reactant concentration. In
this case, the kinetic equation (which is usually determined experimentally at
higher reactant concentrations) can be modied to tend to zero for lower reac
tant concentrations, without modifying its behaviour at higher concentrations,
e.g., by multiplying with



ci

 =  1 exp
(2.13)
ci,low
where ci,low is a low value of the concentration of component i for the specic
operating conditions.
The concentration of reactants that are present in a large excess, will only
change by a small factor as the reaction takes place. Therefore, the observed
reaction rate is almost independent of these concentrations, and the coecient
is almost zero. Usually the reaction rate equation is simplied by neglecting
these factors. However, it may be that these concentrations are depleted some
where in a catalyst particle; as the reaction rate has been made independent
of this concentration the reaction will continue and unphysical negative con
centrations can occur in the model. To prevent this, the modication shown
above can also be applied to these reactants.

2.2.3

Balance equations

The microscopic mass, concentration and and heat balances inside a catalyst
particle can be written as a set of partial dierential equations that describe
the diusion and reaction processes
ci

i,x +
+
i,y + i,z = i
t
x
y
z

(2.14)

where i gives the reaction rate for component i, which may be a non-linear
function of the temperature and gas composition in the particle. Under the
assumption that molecular transport inside the particle is by diusion only (no
convective transport) and that the eective diusion coecient is constant:
 2

ci
ci 2 ci 2 ci
+ De,i
+
+ 2 = i
(2.15)
y 2
z
t
x2

25

2. Chemical reaction and diusion in a single catalyst particle

We will assume stationary behaviour for the particles

 2

ci 2 ci 2 ci
+
+
= i
De,i
y 2 z 2
x2

(2.16)

The way in which this set of equations can be solved depends on the geometry
of the catalyst particles and the complexity of the reaction scheme. When
diusion inside a particle takes place in one (coordinate) direction only, the
set of partial dierential equations reduces to a set of ordinary dierential
equations. This is the case for spherical particles and for cylindrical and slabshaped particles that are either innitely long (and wide) or have blocked-o
faces (that can be treated as symmetry planes). For brevity, these particles
are referred to as one-dimensional particles; in fact the particles are perfectly
three-dimensional; it is the balance equations that are one-dimensional.
For one-dimensional catalyst particles, the partial dierential equation 2.16
reduces to an ordinary dierential equation:
 2

d ci
m dci
+
= i
(2.17)
De,i
1 x dx
dx2
Here m is the power with which the surface area of the particle perpendicular
to the transport ux changes as a function of the coordinate x. For innite
slabs, cylinders and spheres, m has the value 0, 1 and 2, respectively. The
boundary conditions for equation (2.17) are that the concentration is assumed
to be known at the outer edge of the particle and that the concentration is
bounded inside the particle:
x=0

x=1

ci = ci,s
dci

=0
dx

(particle edge)

(particle center)

For a system with nc components involved in nr reaction, the system is

written as:
 2

nr

d ci
m dci
+
=

i,j j
(i = 1..nc )
(2.18)
De,i
dx2
1 x dx
j=1
The energy balance leads to an additional dierential equation that is anal
ogous to equation (2.18)

cp

26

 2
 
nc
m dT
r Hj
d T
+
=
j
2
dx
1 x dx
cp
j=1

(2.19)

2.3 Reduction of the number of dierential equations

The rst factor on the left hand side /(cp ) has the same dimensions as
the diusion coecient (m2 /s) in equation (2.18), the factor r Hf /(cp )
is analogous to the stoichiometric constant ij . Note that the factor cp is
included to demonstrate this analogy only and cancels out on both sides of the
equation. The boundary conditions for equation (2.19) are:
x=0

x=1

T = Ts
dT

=0
dx

(particle edge)

(particle center)

2.3 Reduction of the number of dierential

equations
For one-dimensional particles the balance equations lead to a set of coupled,
usually non-linear, second order ordinary dierential equations, where the
number of equations is equal to the number of components nc plus (2.18) plus
one for the temperature (2.19). The concentrations of the components are not
independent of each other since they are coupled through the stoichiometry
of the reactions. Likewise, the temperature is coupled to the concentrations
through the heat of reaction. For simple, single reaction systems, this is often
directly clear from the balance equations. For multi-reaction systems this is
not as obvious. In this section a method is presented to write down the system
of balance equations in an ecient way so that the number of equations that
need to be solved is reduced from the number of dierent chemical compo
nents nc (plus one for the temperature) to the number of stoichiometrically
independent reactions nr . Note that for any reaction system, nc nr .
This procedure is generic in that it can be done without any prior knowledge
about the reaction scheme (e.g., key components) and the reaction kinetics and
therefore it is ideal for application in a exible computer code.

2.3.1 Single reaction

Consider a general (isothermal) chemical reaction where nc species are in
volved:
(a,1 )A1 + (a,2 )A2 + ... + (a,na )Ana
na+1 B1 + na+2 B2 + ... + nc Bnb (2.20)
The reaction takes place in a spherical particle. Now consider a spherical subvolume with radius r inside the particle. In a stationary situation, every a,i
molecules of reactant Ai that enter the sphere are converted to b,j molecules

27

2. Chemical reaction and diusion in a single catalyst particle

of each product Bj . To keep the mass balance, these product molecules must
be transported out of the spherical sub-volume. Therefore, the mole uxes of
all species A and B at the radial position r are related:
j
i
=
j
i

(i = 1..nc ; j = 1..nc )

or, with diusion according to Ficks law:



De,i dci 
De,j dcj 
=
j dx x=r
i dx x=r

(2.21)

(2.22)

Hence all concentration gradients are linked to each other by a constant factor.
Equation (2.22) can be integrated from the edge of the particle (x = 0) to some
position r:
r
r
r
r
De,i dci
De,j dcj
De,i 
De,j 
dx =
dx
ci =
cj 
(2.23)
i dx
j dx
i  0
j
0
0
0
Since the concentrations at the particle edge ci,s are known as the boundary
conditions, there is an algebraic relation between all concentrations at each
position inside the catalyst particle:
De,i
De,j
(ci ci,s ) =
(cj cj,s )
j
i
De,i j
cj =
(ci ci,s ) + cj,s
De,j i

(2.24)
(2.25)

Now we choose a key component i. With the aid of equations (2.24), the con
centrations of the non-key components j can be eliminated from the reaction
rate. Thus, the balance equation can be written as a function of the chosen key
component i only. For the key component, the mole balance equation (2.18)
reduces to:


De,i d2 ci
m dci
+
= (ci )
(2.26)
dx2
1 x dx
i
Depending on the complexity of the rate equation, this equation may be solved
analytically or numerically to yield the concentration prole of component i.
From the solution, the concentration proles of the other components (and the
temperature prole) can be calculated using equations (2.24). Thus, we have
reduced the set (2.18) from nc equations to only one.

28

2.3 Reduction of the number of dierential equations

2.3.2

Multiple reactions

Slab-shaped porous particles

The simplest shape for a one-dimensional particle is a slab; a at plate that
is either innitely wide and long or has side-faces that are closed for diusion
and thermally insulated. For a slab-shaped particle, the cross-sectional area
does not depend on the depth in the particle, so the geometric parameter m
is 0. Thus, the equations (2.18) simplify to:
nr

d2 c i
i,j j
(i = 1..nc )
(2.27)
De,i 2 =
dx
j=1
For each reaction j we dene a variable zj such that
d2 zj
= j
(j = 1..nr )
dx2
with boundary conditions

(2.28)

(particle edge)
x = 0 zj = 0
dzj
x=1
=0
(particle center)
dx
The variable zj can be interpreted as the (scaled) concentration of an imaginary
key component that is produced by reaction j with stoichiometry 1, with
a bulk concentration 0 and that is not involved in any other reaction. On
the right hand side of the equations, a concentration always appears together
with its diusion coecient (see for example equation 2.26). For the pseudo
concentration, we will incorporate the diusion coecient into the variable zj ,
which consequently has units [mol/m3 m2 /s] or [mol/m/s]. Note that the zj
are just a mathematical construction; it is irrelevant whether a key component
actually exists for a reaction j or not.
The reaction rates j are functions of the concentrations ci . If the con
centrations can be rewritten as functions of the variables zj , then the reaction
rates can be rewritten as functions of the zj only and the set of nr equations
(2.28) can in principle be solved. In order to nd this relationship between the
zj and the ci , we substitute equation (2.28) into equation (2.27):
nr
d2 c i 
d2 zj
De,i 2 =
i,j 2
(i = 1..nc )
(2.29)
dx
dx
j=1
Equation (2.29) is essentially a sum of second derivatives to x which can be
integrated between the center of the particle (x = L) and certain x:
x 2
x 2
nr

d ci
d zj
dx =
i.j
dx
(2.30)
De,i
2
2
x=L dx
x=L dx
j=1

29

2. Chemical reaction and diusion in a single catalyst particle

so
x
x
nr

dci 
dzj 
=
i,j
De,i 
dx x=L j=1
dx x=L

(2.31)

and with the boundary conditions at x = L:

r
dzj
dci 
=
i,j
De,i
dx
dx
j=1

(2.32)

This equation simply states that the ux of each chemical species is related
to the ux of the key component of each of the reactions that takes place
through the stoichiometry. Equation (2.32) can be integrated once more, this
time from the edge of the particle (x = 0) to a certain depth x:

De,i
x=0


dci
dx =
i,j
dx
j=1
x

x=0

dzj
dx
dx

(2.33)

which, with the boundary conditions at x = 0 leads to:

De,i (ci ci,s ) =

nr


i,j zj

(2.34)

j=1

or
ci = ci,s +

nr
1 
i,j zj
De,i j=1

(2.35)

The concentration of component i at a location inside the particle is seen to

depend on the boundary concentration and a contribution from each reaction
that takes place. Equation (2.35) again has an equivalent for the temperature:
r
1
r Hj zj
T = Ts +
j=1

(2.36)

The reaction rate equations j are non-linear functions of the concentrations ci

and the temperature T . Using equations (2.35) and (2.36), the reaction rates
can be expressed in the variables zj only. Thus the set of (nc + 1) dierential
equations (2.18, 2.19) with (nc + 1) unknowns can be reduced to the set of nr
dierential equations (2.28) that we used to dene the variables zj .

30

2.4 Approximating the eectiveness factor

Cylindrical and spherical particles

The procedure described above can be generalised to spherical and cylindrical
particles. Hereto, we dene the variable zj by:
m dzj
d2 zj
+
j
2
dx
1 x dx

(2.37)

with boundary conditions

x = 0 zj = 0

dzj

x=1
=0
dx
so that

De,i


 2

nr

d2 ci
m dci
m dzj
d zj
+
=
i,j
+
2
1 x dx
dx
1 x dx
dx2
j=1

(i = 1..nc )
(2.38)

Equations (2.28) and (2.29) for slab-shaped particles are special cases of equa
tion (2.37) and (2.38). It can easily be seen that the solution reached for a
slab-shaped particle (equation 2.35) is also a solution of equation (2.38) and
conforms to the boundary conditions of equation (2.37).

2.4

Approximating the eectiveness factor

Although the set of dierential equations has been reduced to the minimum
size, the solution still requires quite some computational eort. Especially for
more complex particle shapes, for which a set of non-linear partial dierential
equations needs to be solved numerically. This usually involves dening a 2 or
3-dimensional grid of (typically thousands of) nodes to describe the geometry,
discretisation of the equations and an iterative solution procedure. As it is
not desirable to do this for each computational cell and each interation of
a CFD model, we will be looking at approximation methods to reduce the
computational load.

2.4.1

Single reaction systems

Wijngaarden and Westerterp have introduced generalised formula for the ap

proximation of the eectiveness factor for a single reaction with arbitrary ki
netics in a catalyst particle of any shape (Westerterp and Wijngaarden, 1992;

31

2. Chemical reaction and diusion in a single catalyst particle

Wijngaarden and Westerterp, 1994; Wijngaarden et.al., 1998). In this section,

this so-called Aris number approach is rewritten along the same lines, making
use of the imaginary key component z as was introduced above.
Wijngaarden et.al. dene the zeroth and rst Aris number as:
 



1 
2 
An0
An

(2.39)
1
1
2  0
An0 is the square of a Thiele modulus that (by denition) brings together
the eectiveness curves for all geometries for fast reactions. For fast reactions,
only a thin layer of the catalyst particle contributes in the reaction. If this
layer is thin enough, it can be considered as a at plate, i.e., a slab. Hence,
the reaction-diusion process can be described by:
d2 z
= 
dx2

(2.40)

or, after some manipulation:

 2
dz
d
= dz
dx

(2.41)

At the edge of the particle (x = 0), the values of z is known to be 0 (by

denition). Since the reaction takes place only in the outer layer of the particle
with thickness , the composition will approach thermodynamic equilibrium
at some distance X > from the edge. Since every possible composition
is characterised by a single value of z, the equilibrium value zeq exists that
corresponds to the equilibrium composition. This value can be found by setting
the reaction rate to zero:
(zeq ) = 0

(2.42)

The concentration gradients will approach zero at x = X. We can integrate

equation (2.41) between x = 0 and x = X:
 2 
 2 
zeq

dz 
dz 

=2
dz
(2.43)


dx 
dx 
0
x=X

x=0

After some rearrangement:

|0

32

Ap

dz
dx




x=0

Vp |z=0


=

Ap
Vp


z
2 0 eq dz
|z=0

(2.44)

2.4 Approximating the eectiveness factor

The integral on the right-hand side can be calculated (either analytically or

numerically) for well-behaved reaction kinetics. The denition of An0 can be
rewritten as:
 
 2
( |z=0 )2
1 
Vp
=
(2.45)
An0 =
0
Ap
2 0
2 zeq dz
Due to the denition of zj this expression for the zeroth Aris number can
be used for both isothermal and non-isothermal reactions with any number of
components.
The rst Aris number is dened at low reaction rates, where the eective
ness factor diers only slightly from 1.


An1 1 2 1
(2.46)
Note that An0 is the square of the Thiele modulus, so approaches one as An0
approaches zero. We use a Taylor expansion around An0 = 0 to approximate
the eectiveness factor at slow reaction rates:


|1 1 +
An0
(2.47)
An0 An0 0
To calculate the derivative, we linearise the reaction rate around the rate at
bulk composition



 z
(2.48)
 = |z=0 +
z z=0
where the star superscript indicates a value for the linearised kinetics. We can
write the rst-order relation for An0 :

 2
Vp
 

(2.49)
An0 =
z z=0
Ap
Now we can write the denition as:

2





An1 =
An
2
An
An0 An0 0 0
An0 An0 0 0
For small values of An0 , we can neglect the quadratic term, so that



An1 = 2
An0
An0

(2.50)

(2.51)

33

2. Chemical reaction and diusion in a single catalyst particle

The derivative of the eectiveness factor for the linearised reaction rate on the
right hand side depends on the particle shape only and is called the geometry
function :
2

d
dAn0

(2.52)

The geometry function can be calculated for any particle geometry; for a slab
it has the value 2/3; for a cylinder 1 and for a sphere 6/5. In this way, the
rst Aris number can be calculated as the product of a part that depends on
geometry only and a part that depends on reaction kinetics only:

 2
Vp
 
(2.53)
An1 =
z z=0
Ap
The eectiveness factor for fast and slow reactions can be approximated
using the denition of An0 and An1 respectively:
1

An
 0
1 An1

(fast reaction)

(2.54)

(slow reaction)

(2.55)

For practical use, Wijngaarden and Westerterp (1994) recommend the follow
ing (implicit) interpolation formula for the entire An0 range:
1

1 + An1 + (1 )An0

(2.56)

The ux in to the particle for each of the chemical components i can be cal
culated from the reaction eectiveness as:

 
dci 
Vp

(2.57)
i = De,i
=
i |z=0

Ap
dx x=0
for the heat ux:
H

2.4.2

dT
=
=
cp dx

Vp
Ap

r H
|z=0
cp

(2.58)

Reaction networks

If more than one reaction is taking place at the same time, the situation is in
herently more complex than in the single reaction case. For a reaction network,
distinction needs to be made between eectiveness factors for reactions and

34

2.4 Approximating the eectiveness factor

eectiveness factors for components. The eectiveness factor for components

may depend on several reactions, some fast and some slow. The eectiveness
factor for a reaction in a reaction network only depends on the rate of one re
action. The eectiveness factor j for reaction j is identical to the eectiveness
factor for pseudo concentration zj .


Ap
j =

dzj 
dx 

x=0

Vp j |zj =1

(2.59)

Of course, the reaction rate j probably depends on at least one reactant

concentration and that reactant may be involved in other reactions in the
network. The other reactions inuence the concentration proles and therefore
the rate and eectiveness factor of this reaction.
No general approximation method is available for reaction networks. There
fore, we will develop one here. The approach will be the same as for the single
reaction Aris number method: we will calculate both Aris numbers for each
reaction and interpolate between the fast reaction estimate for the eectiveness
factor based on An0 and the slow reaction estimate based on An1 . The inter
actions between the reactions need to be taken into account in the calculation
of the Aris numbers.
For a given process, the reaction kinetics and diusion coecients are given
and cannot be changed. The only parameter that we are in direct control of
is the particle size. It will be useful to think of the slow reaction regime as
the limit for small particles and the fast reaction regime as the limit for large
particles.
The zeroth Aris number
In the fast reaction regime, the concentration proles inside the particle can be
quite complex: in contrast with the single reaction situation where all proles
are strictly increasing or decreasing, for reaction networks maxima and minima
can appear in the concentration proles. This occurs if a component is e.g.,
produced by one reaction and consumed by another. In contrast, the proles
for the pseudo concentrations are all strictly ascending as the zj are involved
in only one reaction by denition. Therefore, we will base the calculation of
An0 on the z-proles.
We cannot use the regular equation (2.45) to calculate An0,j because we
cannot calculate the integral in the denominator as it depends on the values of
all other z. Therefore, the basis for the calculation of An0,j is the intermediate

35

2. Chemical reaction and diusion in a single catalyst particle

result of equation (2.44):





dz
Ap dxj 

x=0,0
|0
Vp j |
zj =0

(2.60)

The term in the numerator gives the gradient of zj at the edge of the particle
when the eectiveness factor approaches zero, e.g., for a very large particle.
If we take a slab-shaped particle with closed sides and increase the thickness
of the slab, the concentration of the reactants in the center will decrease (the
concentration zj will increase) and the conversion in the particle will increase;
therefore the gradient at the edge of the particle must also increase. However,
the eect of increasing the particle size will become less pronounced when the
concentration in the center becomes low (or: close to the equilibrium value).
At a certain point the concentration prole will be located for the largest part
near the edge of the particle and increasing its size only adds more volume
to the inactive central part of the particle. Hence, the total conversion, and
therefore the gradient at the particle edge, will not change if the particle size
is increased any further. In practice, this gradient at the edge quite rapidly
approaches the one shown in the numerator of equation (2.60).
Of course, the size of the particle where this occurs depends on the rate
of the reaction; for fast reactions this will be at a smaller size than for slow
reactions. However, there is a certain minimum particle size that is large for
all reactions. In the center of such a particle, all zj are at their equilibrium
values which can be calculated by setting all component reaction rates to
zero simultaneously (provided the rate equations are well posed in the sense
explained earlier).
To calculate the gradient at innitesimal eectiveness, we need to solve the
dierential equations
d2 zj
= j
dx2

(2.61)

with boundary conditions

x = 0 zj = 0
dzj
x=X
=0
dx
In fact, there is an additional known boundary condition: zj = zj,eq at the
particle center, where the equilibrium value of zj can be calculated by setting
all reaction rates to zero. However, it is probably easier to solve the dierential

36

2.4 Approximating the eectiveness factor

equations to obtain the value of zj,eq instead. Of course, we need an estimate

of X, but this is usually not hard to give and the result is not sensitive to the
actual value provided that X is chosen large enough; this does not need to
inuence the calculation time or accuracy if an adaptive grid method is used.
The solution of the dierential equations and then taking the gradient at
the surface is equivalent to the integral calculated in the original Aris number
approach (equation 2.45); however, for multiple reactions all dierential equa
tions are solved simultaneously so that the inuence of one reaction on the
other is taken into account in a natural way.
Apart from the gradient at the particle surface, the solution of the dier
ential equations give us the equilibrium values zeq,j as well as an estimate of
the thickness of the reaction zone for each reaction.
Once we have the particle edge gradient at zero eectiveness, we can easily
calculate An0 :

2
Vp
An0,j =
Ap

j |z=0

dzj 
dx 

(2.62)

z=0,0

Note that as before, An0 does not depend on the actual particle shape.
The rst Aris number
The rst Aris number is dened as (1 2 ) for eectiveness factors close to
one. In the limit for very small particles, all reactions in the network will be
slow in this sense; all zj will be close to zero so all concentrations will be close
to the bulk uid concentrations. To simplify the expressions, the slow reaction
regime is based on the eectiveness factor for components instead of reactions;
of course once these are known, the numbers for the individual reactions can
also be calculated.
As a starting point we will take the single reaction equation for An1 (2.53):

An1 =

Vp
Ap

2


 
z z=0

(2.63)

In the slow reaction regime the form of the kinetic equation is not important, so
it may be expected that a similar form should also exist for each component in
a multi-reaction system. The derivative in the last factor should of course take
into account the eect of all reactions in which a given component is involved.
Furthermore, there are cross terms because of the inuence of reaction i on the

37

2. Chemical reaction and diusion in a single catalyst particle

rate of reaction j (because reaction i may consume or produce components that

occur in the rate equation of reaction j). Without rigourous proof we propose
the following relation:

An1,i

=
i,s

Vp
Ap

2 
nr

i,j

j=1

nr

k=1

k,s


j 
zk zk 0

(2.64)

As can be seen, equation (2.64) reduces to the single reaction form (2.53)
when nr = 1. The derivatives of the reaction rates j to the z values are
weighed according to the inuence
 of each reaction in the overall bulk rate of
component i (note that i,s = j i,j j,s ).
To demonstrate the accuracy of equation 2.64, consider the following small
reaction system:
(a,1 )A + (b,1 )B C
(a,2 )A + (b,2 )B D

(2.65)
(2.66)

where
1 =
2 =

k 1 ca
k2 cb

For this system, equation 2.64 expands to

 2  

1,s 1 2,s 1
Vp
i,1
+
+
An1,i =
i,s z1
i,s z2
Ap


1,s 2 2,s 2
+
i,2
i,s z2
i,s z1

(2.67)
(2.68)

(2.69)

The derivatives can be calculated by substituting equation (2.35) in to the

kinetic equations
1
a,k k1
=
zk
Da
2
b,k k2
=
zk
Db
so
An1,i

38

=
i,s

(2.70)
(2.71)

Vp
Ap

2 


a,1 k1
a,2 k1
i,1 1,s
+
+ 2,s
Da
Da


b,1 k2
b,2 k2
i,2 1,s
+ 2,s
Db
Db

(2.72)

2.4 Approximating the eectiveness factor

Table 2.2: Parameter values used in the simple

reaction equations (2.65) and (2.66).
parameter
symbol
ca,s
bulk concentration A
cb,s
bulk concentration B
eective diusion coecient A Da
eective diusion coecient B Db
reaction rate constant
k1
reaction rate constant
k2
stoichiometry
a,1
stoichiometry
a,2
stoichiometry
b,1
stoichiometry
b,2

reaction network given by

value
unit
[mol/m3 ]
15
5
[mol/m3 ]
2.71e-6 [m2 /s]
3.14e-6 [m2 /s]
0.5657 [1/s]
0.12
[1/s]
-2
[]
-3
[]
5
[]
-7
[]

To demonstrate the accuracy of equation (2.64), the set of dierential equa

tions for the reaction system is solved using a 1-dimensional partial dierential
equation solver for a slab-shaped particle for several particle sizes between 0.01
and 50 mm. The (arbitrary) bulk and transport parameters used are listed in
table (2.2).
The values of An1 for both active species A and B calculated with equation
(2.64) are compared to the value of (1 2 ) calculated numerically (see gure
2.2). It can be seen that the equation for An1 is a good estimate for (1 2 ) in
the slow reaction (small particle size) regime. The equation has been compared
to numerical data for a wide range of the parameter values and was found to be
correct for the fast reaction regime for all parameter values, both for parallel
reactions (b,1 < 0) and for serial reactions (b,1 > 0).
As the rst Aris number is based on linearised kinetics at bulk conditions,
the validity of equation (2.64) can be extended from the simple rst order
kinetics used in the example to general kinetics. For rst order Arrhenius type
kinetics:


EA,j
ci
(2.73)
j = k0,j exp
RT
the cross derivatives can be calculated from



j
i,k r Hk EA,j
EA,j
= k0,j exp
+
ci
zk
RT
Di
RT 2

(2.74)

Similar equations can be written for n-th order or general power law ki
netics. However, as these expressions become more elaborate, it may be ad
vantageous (and even more accurate) to calculate the derivative by numerical
methods instead.

39

2. Chemical reaction and diusion in a single catalyst particle

(1- 2), An1 [-]

0.1

1 2a (numerical)
1 2b (numerical)
An1 estimate (a,b)

0.01

0.001

0.0001
0.01

0.1

dp [mm]

10

Figure 2.2: Value of An1 calculated by equation (2.64) compared to the value of
(1 2 ) calculated numerically for component A and B of the reaction system
given by reaction equations (2.65) and (2.66).

Interpolation
In the previous sections we have given a method to determine the value of
the zeroth Aris number for each reaction in a reaction network and we have
shown that we can use equation (2.64) to calculate the rst Aris number for
each component in a reaction network. The zeroth Aris number can give us
an estimate of the eectiveness factor for the fast reaction regime and the
rst Aris number can be used to calculate the eectiveness factor for the slow
reaction regime. The eectiveness in the intermediate regime is calculated by
an interpolation scheme.
To be able to do the interpolation, we need both Aris numbers either for
each reaction or for each component. The best results are obtained if the
interpolation is done for the reaction eectiveness, since the reaction rate for
a component may depend on several reactions that may be fast or slow.
We can calculate the slow reaction estimate of the eectiveness factor for
each reaction from the rst Aris number we calculated for each component

40

2.4 Approximating the eectiveness factor

by solving the mass balances for the slow reaction regime. As a rst step, he
slow reaction regime estimate of the eectiveness factor for component i can
be calculated from the rst Aris number. However, because we want to be
sure all reactions are well within the slow reaction regime, we write the Aris
numbers for a small particle. This particle size should be small enough that
the eectiveness factors are close to one, but not so small that round-o errors
become important. A factor 10-100 smaller than the actual particle size will
usually suce. The eectiveness factors can be estimated from:
1,i = 
1,i

1
1 + An1,i


= 1 An1,i

(An1,i 0)

(2.75)

(An1,i < 0)

(2.76)

where the star superscript denotes that these are values at decreased particle
size. The low rate estimate for the reaction eectiveness factors can now be
obtained from:
1,i i,s =

nr


1,j i,j j,s

(i = 1..nc )

(2.77)

j=1

which is always possible but has a slight complication that there are usually
more components than reactions so some equations of the system are not
independent. The rst Aris number An1,j for reaction j can then be calculated
from the eectiveness factor estimate
An1,j = 1 (1,j )2

(2.78)

Finally, the actual Aris numbers are retrieved by scaling the particle sizes
 2
Ap Vp

(2.79)
An1,j = An1,j
Vp Ap
The calculated eectiveness factors of the reactions of the simple reaction
network given above are compared with the low reaction rate region estimates
(based on the calculated value of An1 ) and the high reaction rate estimates
(based on the calculated value of An0 ) in gure 2.3. Initially, the eectiveness
factor for the second reaction increases when the particle size is increased, to
values above 1.5. The reason for this is that the bulk concentration of compo
nent B is relatively low; because of mass transfer limitation the concentration
of B inside the particle rises above the bulk value as it is produced by reac
tion 1. As a consequence, the rate of reaction 2 can exceed the rate at bulk
conditions.

41

2. Chemical reaction and diusion in a single catalyst particle

The low rate estimate for the eectiveness factor of reaction 2 correctly

follows the calculated curves. To calculate the eectiveness factor for the com
plete regime, an interpolation between the high and low reaction rate regions
is used.

10
Reaction 1 (numerical)
Reaction 2 (numerical)
Effectiveness factor [-]

High rate estimate

Low rate estimate

0.1
0.001

0.1

1
1
Particle size [mm]

10

Figure 2.3: Values of the eectiveness factor for reaction 1 and 2 of the reaction
network given by reaction equations (2.65) and (2.66) calculated by numerical
integration of the balance equations compared to the low reaction rate and
high reaction rate region estimates as a function of the particle size for a slabshaped particle.

For single reaction systems, Wijngaarden and Westerterp (1994) suggest

an implicit interpolation formula that can be solved iteratively:
1

1 + An1 + (1 )An0

(2.80)

Such a scheme can not be used for multiple reaction systems as it does not
scale correctly with the value of the reaction rate at the particle edge. Also, it
does not handle very high values of the eectiveness factor as the part under
the root sign may become negative.

42

2.4 Approximating the eectiveness factor

We will propose an alternative, explicit interpolation method. The correct

scaling of the predicted eectiveness factors will be ensured as long as a linear
combination of the fast and slow regime eectiveness factor estimates is used.
For a given reaction system and bulk conditions, the eectiveness factor
depends on the particle size only. For large particles, the reaction is fast
compared with mass transfer and the reactive zone is limited to a shell at
the outer edge of the particle. The eectiveness factor can be estimated by
an approximation based on An0 . When the particle size is decreased, the
thickness of the reactive shell remains nearly constant until the particle size
approaches the same order of magnitude as this penetration depth. Further
decreasing the particle size has a large inuence on the concentration proles
in the particle; this is the intermediate region where both the reaction rate
and the mass transfer rate are important for the overall conversion. When we
decrease the particle size further so that it is much smaller than the penetration
depth, the concentration proles become relatively at and the slow reaction
approximation based on An1 can be used.
Based on these considerations, it seems logical to use an estimate of the
penetration depth as a scaling factor for the interpolation between fast and
slow reaction regimes. For each reaction, in the fast reaction regime the pseudo
concentration z strictly increases from 0 at the particle edge to the equilibrium
value zeq at the end of the reaction zone. We will dene the penetration depth
j of reaction j as the distance from the edge of the particle where the value of
zj has reached 99% of its equilibrium value. This value can easily be obtained
from the proles calculated for the calculation of An0 .
We can then compare the actual particle size with the penetration depth.
If the particle size is much smaller than j , this reaction is in the slow regime,
if the particle size is much larger this reaction is in the fast regime. For
intermediate values of the penetration depth, an interpolation between the
two regimes is needed. A simple strategy that seems to work well is to take
the slow rate estimate for the eectiveness factor if the particle size is smaller
than and take the fast rate estimate elsewhere. The values of 1 and 2 are
shown for the example case in gure 2.3.
The eectiveness for components can be calculated from the eectiveness
factor for reactions 1 and 2 through
nr
j=1 i,j j j,s
i = nr
j=1 i,j j,s

(2.81)

The result for components A and B is shown in gure 2.4.

43

2. Chemical reaction and diusion in a single catalyst particle

1
estimate
estimate
a calculated
b calculated
a

Effectiveness factor [-]

0.8

0.6

0.4

0.2

0
0.001

0.01

0.1
Particle size [m]

10

Figure 2.4: Eectiveness for components A and B of the simple reaction net
work given by reaction equations (2.65) and (2.66) as a function of particle
size. The dots give the results of the numerical integration of the balance
equations; the lines give the estimated values.

2.5
2.5.1

Example: the Sohio process

Process description

The ammoxidation of propene to acrylonitril is selected to demonstrate the

method described above. In this process, acrylonitril (CH2 =CHCN) is pro
duced from propene (CH2 =CHCH3 ), oxygen and ammonia in a complex reac
tion network. As side-products, acetonitril (CH3 CN) and acrolein (CH2 =CH
CHO) are formed. In addition, the products can be oxidised further to carbon
dioxide and hydrogen cyanide (see gure 2.5).
The kinetics for this reaction system have been published for reaction over
a BiOMo catalyst (Hopper et.al., 1993 and Mleczko, 1996). The reaction
equations are:
1. CH2 =CHCH3 + NH3 + 1.5O2 CH2 =CHCN + 3H2 O
2. CH2 =CHCH3 + O2 CH2 =CHCHO + H2 O

44

2.5 Example: the Sohio process

acrolein
(C3O)

propene
(C3)

acrylonitril
(C3N)

hydrogen cyanide
(HCN)

acetonitril
(C2N)

Figure 2.5: Reaction network for the Sohio process (after Mleczko, 1996)

3.
4.
5.
6.

CH2 =CHCH3 + NH3 + 2.25O2 CH3 CN + 0.5CO2 + 0.5CO + 3H2 O

CH2 =CHCHO + NH3 + 0.5O2 CH2 =CHCN + 2H2 O
CH2 =CHCN + 2O2 CO2 + CO + HCN + H2 O
CH3 CN + 1.5O2 CO2 + HCN + H2 O

All reactions are of rst order in the organic reactant concentration:

rj = kj cCH,j

(2.82)

where the reaction rate coecient depends on the temperature according to

an Arrhenius expression:



EA,j 1
1
kj = k0,j exp

(2.83)
R
T
T0
Here T0 = 470 C is the temperature at which k0,j is determined. The ki
netic parameters for each reaction are listed in table 2.3. All six reactions are
exothermic. The heat of reaction depends only slightly on the temperature
and is therefore taken as a constant. It was calculated at the bulk tempera
ture (400 C) from data given by Hopper et.al. (1993).
In industrial practice, this this reaction is usually carried out in a uidised
bed reactor with particle sizes in the order of 0.1 mm where internal mass
transfer limitation is not an issue. Therefore, the published kinetic relations
and parameters can be assumed to approximate the intrinsic behaviour of the
catalyst.

45

2. Chemical reaction and diusion in a single catalyst particle

Table 2.3: Kinetic parameters for the ammoxidation of propene (Hopper et.al.,
1993 and Mleczko, 1996)
reaction
1
2
3
4
5
6
CH
C3
C3
C3
C3O
C3N
C2N
0.4056 0.0097 0.0174 0.6813 0.1622 0.073
k0 [1/s]
EA [kJ/mol]
79.5
79.5
29.3
29.3
82.9
29.3
r H [kJ/mol] -47.31 -332.37 -862.54 -178.99 -780.77 -590.78
Table 2.4: Assumed catalyst properties
parameter
particle diameter dp [m]
thermal conductivity s [W/m/K]
porosity  [-]
pore diameter dc [m]
tortuosity c [-]

(typical values)

values
0.01
0.1
0.65
1.0 106
3

For this example it is assumed that the reaction takes place in a spherical
porous catalyst particle. The catalyst properties are listed in table 2.4. The
eective diusion coecient inside the catalyst particles for each component
can be estimated from the binary molecular diusion coecients, Knudsen
diusion coecients and the tortuosity. Diusion coecients are calculated
at the bulk composition (table 2.5). The values for the eective diusion
coecients used are listed in table 2.5.

Table 2.5: Eective diusion coecient

component
De 106
propene (C3 )
2.61
oxygen (O2 )
3.21
ammonia (NH3 )
4.22
2.34
acrylonitril (C3 N)
2.28
acrolein (C3 O)
2.79
acetonitril (C2 N)
hydrogen cyanide (HCN) 2.79
3.39
carbon monoxide (CO)
2.70
carbon dioxide (CO2 )
4.24
water (H2 O)
3.34
nitrogen (N2 )

46

2.5 Example: the Sohio process

Table 2.6: Composition at the catalyst particle edge at reactor entrance con
ditions (after Hopper et.al., 1993)
component mole fraction
0.07
propene
0.18
oxygen
0.07
ammonia
0.68
nitrogen
0.01
others
The composition at the outer edge of the catalyst particle is taken as given
in table 2.6. The temperature at the outer edge of the particle is Ts = 500 C.
Note that external mass transfer limitation is not taken into account in the
model presented here.
The concentrations of the intermediate products (C3 N, C3 O and C2 N) in
the reaction feed are taken as zero in the literature. This means that initially,
the bulk rates of reactions 4, 5 and 6 are also zero, and therefore that the
eectiveness factor of these reactions is innitely large. In a real reactor this
condition could only occur in the rst layer of particles in the bed, and even
there, the value at the edge of the particles will be somewhat higher than the
bulk uid value because of external mass transfer limitation. Therefore, it
seems justied to put a lower limit on the concentrations of the intermediate
products; here we will take a value of 0.01 mole %.

47

2. Chemical reaction and diusion in a single catalyst particle

2.5.2

Numerical solution

The concentration and temperature proles inside the catalyst particle are
computed in two ways: once by integration of the full set of dierential equa
tions (2.18 and 2.19) and once by integration of the reduced set (2.37). For
the ammoxidation process the full set of equations is:



d2 cC3
2 dcC3

De,C3
+
= k1 cC3 k2 cC3 k3 cC3

dx2
1 x dx
 2

2 dcC3 N
d c C3 N
De,C3 N
+
= k 1 c C 3 + k4 c C 3 O k 5 c C 3 N
dx2
1 x dx
 2

2 dcC3 O

d c C3 O
De,C3 O
+
= k2 cC3 k4 cC3 O

dx2
1 x dx
 2

2 dcC2 N

d c C2 N
De,C2 N
+
= k3 cC3 k6 cC2 N

dx2
1 x dx
 2

2 dcO2

d c O2
De,O2
+
= 1.5k1 cC3 k2 cC3

dx2
1 x dx
2.25k3 cC3 + 0.5k4 cC3 O
2k5 cC3 N 1.5k6 cC2 N

d2 cN H3
2 dcN H3
De,N H3
+
= k1 cC3 k3 cC3 k4 cC3 O
(2.84)
1 x dx
dx2
 2

2 dcHCN

d cHCN
De,HCN
+
= k5 cC3 N + k6 cC2 N

dx2
1 x dx
 2

2 dcCO
d cCO
De,CO
+
= 0.5k3 cC3 + k5 cC3 N
dx2
1 x dx
 2

2 dcCO2
d cCO2
De,CO2
+
= 0.5k3 cC3 + k5 cC3 N + k6 cC2 N
dx2
1 x dx
 2

2 dcH2 O
d cH2 O
De,H2 O
+
= 3k1 cC3 + k2 CC3 + 3k3 cC3
dx2
1 x dx
+ 2k4 cC3 O + k5 cC3 N + k6 CC2 N


with the temperature equation



d2 T
2 dT
+
dx2
1 x dx


= r H1 k1 cC3 + r H2 k2 cC3
+ r H3 k3 cC3 + r H4 k4 cC3 O
+ r H5 k5 cC3 N + r H6 k6 cC2 N

48

(2.85)

2.5 Example: the Sohio process

Due in part to the simple form of the individual rate equations (2.82), it
may be suspected that the system could be solved without integrating all of
the equations (2.84). In fact, the rst four equations may be solved together
with the last one independently of the other six since only four concentration
variables appear on the right hand side of an equation. However, for prob
lems in which the individual reaction rates cannot be described by rst-order
behaviour, this may not be as clear. The procedure described above gives a
way to reduce the number of equations in a straightforward way, regardless of
the complexity of the rate equations. From equation (2.37) it follows that the
reduced set of equations is:
d2 z1
dx2
d2 z2
dx2
d2 z3
dx2
d2 z4
dx2
d2 z5
dx2
d2 z6
dx2

+
+
+
+
+
+

2 dz1
= k1 cC3
1 x dx

2 dz2

= k2 cC3
1 x dx

2 dz3

= k3 cC3
1 x dx

2 dz4

= k4 cC3 O
1 x dx

2 dz5

= k5 cC3 N
1 x dx

2 dz6

= k6 cC2 N
1 x dx

(2.86)

(2.87)
with (equation 2.35)
1

(z1 z2 z3 )

De,C3

cC3 O = cC3 O,s +

(z2 z4 )

De,C3 O

1
cC3 N = cC3 N,s +
(z1 + z4 z5 )
De,C3 N
1

cC2 N = cC2 N,s +

(z3 z6 )

De,C2 N

1
T = Ts
r Hj zj
j=1
cC3 = cC3 ,s +

(2.88)

The relatively compact representation of equations (2.86) and (2.88) com

pared to the full system (2.84) is a direct advantage of the introduction of the

49

2. Chemical reaction and diusion in a single catalyst particle

pseudo concentrations. Only the concentration proles that are required to

solve the system are shown in equation (2.88). Once the proles for the vari
ables zj are known, the other concentration proles can be calculated using
equation (2.35) as needed.
For the conditions given in table 2.6, the full system and the reduced system
are integrated using a nite dierence formulation. The proles for the vari
ables zj and the corresponding concentration proles are shown in gure 2.7
and 2.6, respectively. It can be seen that while the proles for the zj strictly
increase from the boundary value 0 to a value in the centre of the particle,
the concentration proles are more complex. The prole for the intermediate
products acrylonitril (C3 N) and acrolein(C3 O) even shows a maximum inside
the particle. Since it is mathematically equivalent, the solution of the full set
of equations naturally leads to the same set of concentration proles.

0.14

zj [mol/m3]

z1

z4

z2

z5

z3

z6

0.12
0.1

0.08

0.06

0.04

0.02

zj [mol/m3]

x [mm]

Figure 2.6: Proles inside the catalyst particle for the pseudo concentrations
zj for the Sohio reaction system in a 10 mm porous catalyst particle.

50

2.5 Example: the Sohio process

0.06
C3N

O2

NH3

C3O

C2N

0.05

0.04

0.03

0.02

0.01

ci [mol/m3]

ci [mol/m3]

C3

x [mm]

Figure 2.7: Concentration proles inside the catalyst particle for selected com
ponents for the Sohio reaction system in a 10 mm porous catalyst particle.

2.5.3

Approximation

Fast reaction regime

We will now use the approximation method to estimate the eectiveness factor
for the Sohio reaction in a porous catalyst particle. First we will calculate the
concentration proles for the fast reaction regime. Since the particle geometry
will not inuence the results, we will do this for a slab-shaped particle. To
ensure that the reaction only takes place in a thin region near the edge of the
particle, a large particle size is chosen for this calculation: 50 mm.
The kinetic equations given by (2.82) are valid only for an excess amount
of oxygen and ammonia. However, for the chosen conditions oxygen will be
depleted inside a large catalyst particle. Therefore, the reaction rate equation
is modied slightly:






1
EA,j 1
c O2
rj = k0,j exp

cCH,j 1 exp
(2.89)
T
T0
R
cO2 ,low
where cO2 ,low is taken as cO2 ,s /1000. This ensures that the reaction rate ap
proaches zero if the oxygen is depleted, but does not alter the kinetics at
normal oxygen levels.

51

2. Chemical reaction and diusion in a single catalyst particle

The resulting proles for the most interesting components are shown in

gure 2.9 and the z-value proles are shown in 2.8. From the latter we obtain
the gradients at the particle edge, the equilibrium value for z for each reaction
and the penetration depth . The resulting values are given in table (2.7).

0.15

0.1

3
2

0.05
z1
z2

1
0

zj [mol/m3] 106

zj [mol/m3] 106

10

15

z3
z4

20

z5
z6
0
25

x [mm]

Figure 2.8: Proles of the z values for each reaction of the Sohio process in a
very large porous catalyst particle

The eectiveness factor for reactions 4, 5 and 6 is considerably higher

than one, due to the fact that the bulk rates for these reactions is very low.
However, when the interpolation is done in a consistent way, the bulk reaction
rate is just a scaling factor and will not inuence the nal result. It also shows
that the chosen parameter values give something of a worst case situation for
the estimation of the eectiveness factor; as the reaction progresses, the bulk
concentrations of the (intermediate) products C3 N, C3 O and C2 O will increase,
the bulk rates will increase and the eectiveness values will decrease to values
closer to one.
Slow reaction regime
The next step is to calculate the value of An1,i for each component in the slow
reaction regime using equation (2.64). For this we need the derivative of each

52

2.5 Example: the Sohio process

0.06
C3N

O2

NH3

C3O

C2N

0.05

0.04

0.03

0.02

0.01

ci [mol/m3]

ci [mol/m3]

C3

10

15

20

0
25

x [mm]

Figure 2.9: Proles of the most important components of the Sohio process in
a very large porous catalyst particle

reaction rate j to each pseudo concentration zk , which for the given kinetics
(equation 2.82 and 2.83) can be calculated by equation (2.74). The estimates
of the rst Aris numbers for the components are now used to calculate the rst
Aris numbers for the reactions. Since the estimates are valid only in the slow
reaction regime, we decrease the particle size for which the Aris numbers are
determined until all component eectiveness factors are close to one; in this
case, decreasing the particle size by a factor 100 is sucient. We calculate
the slow reaction regime eectiveness factors from the Aris numbers using the
denition:
2
An1,i = 1 1,i

(2.90)

where the daggered variables are taken at the limit of small particle size and
the subscript 1 designates that this is the estimate for the slow reaction
regime. The eectiveness factor for the reactions in the slow reaction regime
is now calculated from the component mass balances:
i i,s

nr


j i,j j,s

(2.91)

j=1

53

2. Chemical reaction and diusion in a single catalyst particle

Table 2.7: Calculated values for the fast reaction regime for the Sohio process
for the given boundary conditions.
reaction zj,eq
zj /x|x=0
3
#
[mol/m ] [mol/m4 ]
[mm]
1
5.38e-06 3.04e-03
5.25
2
1.28e-07 7.26e-05
5.25
3
1.29e-07 8.02e-05
5.0
4
1.29e-07 4.44e-05
5.63
5
4.78e-06 1.35e-03
5.75
6
6.59e-08 1.73e-05
5.63
Table 2.8: Estimated values for the slow reaction regime for the Sohio process,
for a 10 mm slab-shaped particle.
reaction An1
1
1
19.481
0.221
2
19.577
0.22
3
18.505
0.226
4
-8.1
3.017
5
-1350.074 36.757
6
-33.539
5.877
and an equivalent expression for the temperature. These balances form a set of
nc +1 = 11 equations with nr = 6 unknowns (j ) so the set is over-determined.
This is of course caused by the fact that some equations are linear combinations
of two or more of the other equations. In this case, it is not very easy to spot
the dependencies. With some manipulation it can be seen that the equations
for C3 , C3 N, C3 O, C2 N, O2 and the temperature together form a complete set of
6 equations with 6 unknowns. The reaction eectiveness factors are obtained
from this set by matrix inversion. An alternative would be to use numerical
methods to nd the vector of j that best ts equation (2.91).
We now have the rst Aris number for an arbitrary but small sized particle.
To obtain the value for a given particle size is a simple scaling operation, since
the rst Aris number is proportional to the square of the particle volume to
area ratio (Vp /Ap ). The values are shown for a 10 mm particle in table (2.8).
Interpolation
From the values of the zeroth and rst Aris numbers we can calculate estimates
for the eectiveness factor for the fast reaction and slow reaction regimes,
respectively. The actual estimate of the eectiveness factor is an interpolation

54

2.5 Example: the Sohio process

between these values. In order to get a smooth transition between the slow

and fast reaction regimes, the following interpolation function is used:

d/j 1

f =1
2

f=

(d/j

)2

(d/j )
+ (j /d)2 1

d/j < 1

where the eectiveness factor is calculated from:

= (1 f )1,j + f 0,j

(2.92)

Figure 2.10 shows the estimated eectiveness values for each of the six reac
tions (for the reaction numbering, see gure 2.5) as well as the value calculated
by numerical integration of the set of dierential equations.
The rst three reactions show relatively standard behaviour, and the pre
diction of the estimation procedure is quite accurate (within a few percent
from the actual value). The low rate estimate overlaps the actual curve for a
large part of the domain. For reaction 4, eectiveness factors higher than one
are predicted, as the acrolein (C3O) concentration increases above the bulk
level because it is produced by reaction 2. The low rate regime curve now is
increasing at higher particle size, which is correct, although it gives slightly
too high values.
Reactions 5 and 6 show maximum eectiveness factors much higher than
one (up to about 30 for reaction 5, and 1.8 for reaction 6). The maximum is
predicted in the right order of scale, but there is a tendency to overestimate the
peak values. This may be improved by using a dierent type of interpolation
procedure.
For reaction 5 the maximum is nearly at the actual value, although this is
probably a coincidence. The curve for the low reaction rate regime is lower
than the actual values; this causes the estimated eectiveness factor for the
smaller particle sizes (0.5-2 mm) to be (10-30 %) lower than the actual values.
The deviation of the low rate curve is caused by the fact that the relation
used to calculate An1 (equation 2.64) only takes into account the direct inu
ence between reactions and not the secondary interactions. For instance, for
reaction 5, the fact that reaction 4 produces acrylonitril which will increase
the rate of reaction 5 is taken into account, but not the fact that reaction 2
produces acrolein which will speed up the rate of reaction 4. Especially when
the conversion in the bulk gas is low (i.e., at near the entrance of the catalyst
bed), as is the case for these simulations, the primary reactions 1, 2 and 3
will be relatively fast; the product concentrations will be low and therefore
the secondary reaction rates 4, 5 and 6 will be low, so the indirect inuence

55

2. Chemical reaction and diusion in a single catalyst particle

10

(1)

0.1

0.01
0.1

0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

100
(3)

0.01
0.1
100

0.01
0.1
10

10

100

10

100

10

100

(4)

0.1

numerical
estimate
low rate estimate
high rate estimate
1

0.01
0.1

10

numerical
estimate
low rate est.
high rate est.

(5)

10

1
0.1

numerical
estimate
low rate estimate
high rate estimate

(2)

10

0.1

10

10

numerical
estimate
low rate estimate
high rate estimate
1

(6)

10

100

0.1
0.1

numerical
estimate
low rate estimate
high rate estimate
1

Figure 2.10: Estimated eectiveness factor for the Sohio process as a function
of particle size (in mm). The estimated value is compared to the result of a
direct numerical solution of the set of dierential equations; also shown are
the low rate and high rate regime estimates between which is interpolated.

56

2.5 Example: the Sohio process

will be relatively high. A similar secondary inuence takes place through the
temperature: for each reaction the eect of every other reaction on the tem
perature is taken into account, but these reaction rates are taken at the bulk
temperature and not at the increased temperature inside the particle. To in
crease the accuracy of the estimation procedure in these conditions, it may be
possible to extend equation (2.64) to include the secondary eects (although
it appears that this will make it implicit). As we move down a catalyst bed
toward the exit, the reactant concentrations in the bulk will decrease and the
(intermediary) product concentrations will increase, so the dierence in reac
tion rates will decrease and the secondary interaction eects will become less
important.
Figure 2.11 shows the estimate as well as the value obtained from direct
numerical integration of the eectiveness factor for selected components as
function of the particle size for a slab-shaped particle at reactor entrance con
ditions. These curves are obtained by linear combination of the reaction eec
tiveness factors according to the stoichiometry. In contrast with the reaction
eectiveness actors, the component eectiveness factors can become zero or
negative. Therefore these plots are given on a linear scale instead of the cus
tomary log-log scales.
It can be seen that the eectiveness factor for the consumption of propene
(C3) and for the production of acrylonitril (C3N) and acetonitril (C2N) fol
lows more or less the standard curve: starting at one for small particle size
(slow reaction regime) and then gradually decreasing into an exponential de
cay towards zero. For these components (and others that are only produced
or consumed but are not shown: ammonia, oxygen and water) the estimation
is quite good (within 10 % of the true value). The largest errors occur at
intermediate particle sizes.
The curve for acrolein (C3O) shows peculiar results: the eectiveness factor
starts at one and with increasing particle size decreases, drops below zero and
then slowly increases towards zero. This is caused by the fact that C3O is
produced by one reaction (2) and consumed by another (4). At bulk conditions
there is a net production of C3O, but inside the particle the concentrations and
temperature change; apparently, the consumption reaction benets more from
these conditions than the production reaction. The result is a net consumption
of C3O above a certain particle size. It can be seen that the estimation shows
the correct trend, including the prediction of negative eectiveness factors in
the right order of scale, but the accuracy in the transition zone is poor.
A third class of curves is represented by the hydrogen cyanide plot but
also includes other components that are produced by follow-on reactions that
have a very low rate at bulk conditions (CO2 and CO). Consequently, the
eectiveness factor for these components can become much larger than one.

57

2. Chemical reaction and diusion in a single catalyst particle

C3

0.8

0.8

0.6

0.6

0.4
0.2
0
0.1

C3N

0.4

numerical
estimate
low rate estimate
high rate estimate
1

0.2

10

100

C2N

0.8

0
0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

100

numerical C3O
estimate
low rate estimate
high rate estimate

0.5

0.6
0
0.4
0.2
0
0.1

numerical
estimate
low rate estimate
high rate estimate
1

-0.5

10

50
40

100

HCN

numerical
estimate
low rate estimate
high rate estimate

-1
0.1

4
3

20

10

10

100

10

30

0
0.1

0
0.1

numerical
estimate
low rate estimate
high rate estimate

100

10

100

Figure 2.11: Estimated eectiveness factors for selected components and the
temperature for the Sohio process in a slab-shaped catalyst particle as function
of the particle size (in mm) for the given bulk conditions, compared with the
results obtained from numerical integration of the set of dierential equations.

58

2.5 Example: the Sohio process

Table 2.9: Assumed composition at the catalyst particle edge at 68 % conver

sion)
component
mole fraction
0.024
propene (C3)
0.084
oxygen (O2)
0.024
ammonia (NH3)
acrylonitril (C3N) 0.035
acetonitril (C2N) 0.026
0.001
acrolein (C3O)
In the case of HCN it is about 25 at its maximum in this case. The estimation
procedure also predicts this behaviour; for particle sizes near the maximum
and larger the accuracy is satisfactory (within 10-15 %), although on the small
particle size ank the accuracy is lower. This is caused by the inaccuracies in
the eectiveness curves for reactions 5 and 6.
The errors made in the eectiveness factor estimates at the entrance of the
bed (as given above) are worst case values as the eectiveness of the secondary
reactions will be much larger than one when the product concentrations in
the bulk gas are still low. In practice, the eect of the errors (e.g., on the
mass balance of a reactor) will be relatively small as the highest errors occur
for reactions with a very low absolute rate. As we move down the reactor,
the eectiveness factors will have values closer to one. As an example of this,
the eectiveness factors for the Sohio reaction system at 68 % conversion are
shown in gure 2.12. The temperature of the bulk gas depends on the cooling
rate; an arbitrary value of 560 C is taken for this plot. The bulk composition
also depends on the eectiveness factor upstream in the reactor; the values
taken are given in table 2.9. The estimate in this case gives excellent values
with errors less than one percent except in the transition area where maximum
the error is less than 8 % except for reaction 4 which has a maximum error of
about 15 %.
The maximum error occurs in the transition region between the fast and
slow reaction regimes. The maximum error may be further reduced with a
more elaborate interpolation method.
Complex geometry
We have shown that the approximation method developed here gives reason
able results for the Sohio reaction system in slab-shaped catalyst particles.
However, the advantage of the approximation lies primarily in the estimation
of the eectiveness factor for reaction networks in catalyst particles with a

59

2. Chemical reaction and diusion in a single catalyst particle

10

(1)

10

(2)

0.1

0.01
0.1

0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

10

100

(3)

0.1

0.01
0.1

10

10

100

(4)

0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

10

100

(5)

0.1

0.01
0.1

0.01
0.1

numerical
estimate
low rate estimate
high rate estimate

0.01
0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

10

100

(6)

0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

100

0.01
0.1

numerical
estimate
low rate estimate
high rate estimate
1

10

100

Figure 2.12: Eectiveness factor for the Sohio process as a function of particle
size (in mm) at a bulk gas composition corresponding to 68 % conversion (see
table 2.9). The estimated value is compared to the result of a direct numerical
solution of the set of dierential equations; also shown are the low rate and
high rate regime estimates.

60

2.5 Example: the Sohio process

more complex geometry. For one-dimensional particles like slabs, innite

cylinders and spheres, the numerical solution involves the integration of a set
of ordinary dierential equations. The approximation method also requires
the integration of a set of ODEs, so there is little computational advantage
(although one advantage could be that the approximation method gives results
for all particle diameters and shapes with just one integration of the ODEs at
given boundary conditions).
For complex geometry particles, the direct numerical solution requires the
integration of a set of coupled partial dierential equations. Here the ap
proximation method has a larger advantage because it still requires only the
integration of a set of ODEs. For a given particle shape, also the geometry
factor needs to be determined, but this depends on geometry only and needs
to be determined only once for a given geometry. As a demonstration, the
Sohio reaction system is modelled in a ring-shaped catalyst particle (see gure
2.13). Wijngaarden and Westerterp (1994) give the geometry factor for this
particle shape as a function of the inner diameter, outer diameter and length;
in this case = 1.12.
L = 16
Ro = 8
Ri = 4

Figure 2.13: Geometry of ring-shaped catalyst par

ticle

The computational eort needed to to calculate the eective reaction rate

for this ring-shaped particle using the approximation method is compared to
that needed to solve the full balances. For this, a general PDE solver (FlexPDE) with automatic grid adaptation is used. On a typical desktop computer,
this solver needs about 100 s and 1580 nodes to solve the full balances for one

61

2. Chemical reaction and diusion in a single catalyst particle

Table 2.10: Axial concentration proles for the Sohio process in a simple cooled

tube reactor packed with ring shaped catalyst particles (numerical solution).

C3N
C2N
C3O
O2
node conversion T
C3

C mol/m3 mol/m3 mol/m3 mol/m3 mol/m3

0
0.0%
400
2.207
0.032
0.032
0.032
5.674
1
3.4%
420
2.131
0.101
0.038
0.030
5.551
2
8.2%
440
2.025
0.196
0.044
0.028
5.375
3
14.7%
460
1.883
0.321
0.051
0.027
5.132
4
22.9%
480
1.701
0.472
0.058
0.026
4.805
5
32.9%
500
1.481
0.642
0.064
0.025
4.378
6
44.3%
520
1.229
0.810
0.069
0.025
3.840
7
56.3%
540
0.963
0.949
0.074
0.025
3.191
8
68.1%
560
0.704
1.031
0.076
0.023
2.447
9
78.5%
580
0.475
1.034
0.077
0.022
1.645
10
86.8%
600
0.292
0.954
0.076
0.019
0.837
particle to an accuracy of 0.01%. The approximation method for the same
conditions requires about 7 seconds and 93 nodes.
As an example, the approximation method is used to calculate the con
version in a wall-cooled ideal tube reactor packed with ring shaped catalyst
particles. The goal here is to show the accuracy of the approximation method
at dierent conditions, so the reactor model is relatively simple: plug ow is
assumed and the external mass transfer limitation is not taken into account.
The cooling of the tube is assumed to be arranged in such a way that the
temperature increases linearly from the entrance temperature (400 C) to the
exit temperature (600 C). This ensures that both low and high rate condi
tions occur in the reactor; in a real Sohio reactor the temperature will be kept
well below the latter value to limit oxidation of acrylonitril. The tube reactor
is modelled by a series of 10 sections in which the bulk conditions and eec
tiveness factor are assumed to be constant. The conditions at the inlet are
the same as those used previously, the concentrations of all components are
calculated from the reaction rates at bulk conditions and the calculated eec
tiveness factors. The resulting axial concentration and temperature proles
are given in table 2.10.
The numerical solution for the concentration and temperature elds inside
the catalyst particle are obtained by solving the partial dierential equations
using a general PDE solver (FlexPDE). The eectiveness factors calculated
from the numerical solutions are compared to the values obtained with the ap
proximation method in gure 2.14. It can be assumed that the error made in
the numerical method is much smaller than the errors made in the approxima

62

2.5 Example: the Sohio process

1.5

1.2

0.9

0.6

0.3

0
0

4
5
6
7
axial coordinate [arbitrary units]

Effectiveness factor [-]

Effectiveness factor [-]

tions. The curves for reactions 2 and 3 are very similar to that for reaction 1
and therefore are not shown. Because of the assumed temperature prole, the
reaction rates increase along the reactor and the eectiveness factors decrease.
The curves for the numerical solution and the values predicted by the approx
imation are reasonably close together for most of the domain. For reaction 5,
the dierence at the reaction entrance is relatively large (as was noted earlier),
but at higher conversions the dierence is quite small. For reaction 6 there is
a larger dierence at location 9; this is caused by the fact that in the approx
imation the transition between the high and low reaction rate regimes occurs
at those conditions and therefore the eectiveness factor is overestimated.

0
10

reaction 1 approx.

reaction 4 approx.

reaction 5 approx.

reaction 6 approx.

reaction 1 numerical

reaction 4 numerical

reaction 5 numerical

reaction 6 numerical

Figure 2.14: Eectiveness factors obtained from numerical integration of the

governing set of partial dierential equations compared to the results from the
approximation method for the Sohio reaction system in a ring shaped catalyst
particle

The most important results of a reactor simulation for reaction networks

is the selectivity. The selectivity determines the amount of reactants that is
converted to useful products and the amount of waste that is produced. In the
case of the Sohio process, the by-products are not easy to separate from the
product and hazardous; a careful reactor design can reduce waste production
and pollution (Hopper et.al., 1993). Figure 2.15 shows the conversion and

63

2. Chemical reaction and diusion in a single catalyst particle

the selectivity for the main product (acrylonitril) and by-products acetonitril
and cyanic acid as calculated using eectiveness factors determined from the
numerical integration of the full PDEs and using the approximation. The
selectivity for acrolein (C3O) is not shown as the amount produced is very
small (in fact, the assumed inlet bulk concentration is somewhat higher than
the equilibrium concentration and there is a small net consumption of acrolein).
It can be seen that the dierence between approximation and the numerical
is only a few percent point, which is acceptable for practical purposes. Note
that in practice a reactor for the Sohio process would never be designed like
this, as the reactant is converted for 50 % to the hazardous by-product cyanic
acid.

Selectivity [-], Conversion [-]

1
0.8
0.6
0.4
0.2
0
0

4
6
axial coordinate [arbitrary units]

10

S C3N approx.

S C2N approx.

S HCN approx.

X approx.

S C3N numerical

S C2N numerical

S HCN numerical

X numerical

Figure 2.15: Conversion of propene and selectivity towards products and byproducts in a simple 10-compartment tube reactor model for the Sohio pro
cess for ring-shaped particles; results using numerical solution of the PDEs to
obtain the eectiveness factor compared to results using the approximation
method.

64

2.6 Conclusions

2.6

Conclusions

We have introduced the pseudo concentration z of a virtual key component

in the balance equations that describe reaction and diusion in a catalyst
particle. For a single reaction system his simplies the balance equations to a
single ordinary dierential equation for z and a set of algebraic equations to
calculate the concentrations of all components involved and the temperature.
When applied to a reaction network, a straightforward method is found to
reduce the number of dierential equations from one for each component and
the temperature, to one for each reaction.
The Aris number approximation (Wijngaarden and Westerterp, 1994; Wijn
gaarden et.al., 1998) is rewritten here in terms of the pseudo concentration.
This gives one set of equations for the zeroth and rst Aris numbers for a single
isothermal or non-isothermal reaction with any number of components. With
these parameters, the eectiveness factor can be estimated for any kinetics and
particle shape.
For reaction networks, a new approximation method based on pseudo con
centrations and the Aris number approach was developed. In principle this
method is valid for general kinetics and general geometry. The method re
duces the computational load compared to a direct numerical integration of
the balance equations, especially for complex particle shapes. Instead of the
integration of a set of partial dierential equations, integration of a set of or
dinary dierential equations suces. It has been shown that the computation
time reduction using the approximation method can be in the order of a factor
10 using a general PDE solver.
The approximation method is tested for simple reaction networks in slab
shaped particles. It correctly describes the trend of the eectiveness factor
curves as a function of particle size (or, equivalently, zeroth Aris number).
The curves take the correct values in the limits of low and high reaction rate,
and show the correct trend in between. Eectiveness factors greater than
one and less than zero are predicted and in the right order of scale, although
somewhat larger errors are found in these cases.
It is shown that the error made in the approximation is in the order of 5 %
under most conditions and for the primary reactions, but may be as large as 50
% for the secondary reactions in exceptional cases where the eectiveness factor
is much larger than one. The large errors are probably caused by the fact that
the cross terms in the calculation of An1 are taken at bulk conditions. These
cross terms correct for the inuence of one reaction on the rate of another in
the slow reaction regime. The eect of the cross terms is only large if the rate
of the reaction is very low compared to the other reaction rates. Consequently,
these do not have a large eect on the mass balances. It may be possible to

65

2. Chemical reaction and diusion in a single catalyst particle

reduce the errors by modication of the estimation formula for the rst Aris

numbers, at the cost of some computational eort.

The eectiveness factor is calculated by an interpolation between the fast
and slow reaction regimes. A relatively simple interpolation strategy is used,
based on the penetration depth of the reaction system in a large particle. If
the equivalent particle size is smaller than the penetration depth, the slow
reaction regime estimate is taken, otherwise the fast reaction regime estimate.
By and large, this method works well for the entire range of reaction rates,
but it gives relatively large errors if the particle size and penetration depth
are comparable. The errors in the eectiveness factor values predicted by the
approximation method for particle sizes close to the penetration depth may
be reduced by using a more rened interpolation method.
In the model calculations, the maximum errors in the eectiveness factor
only occur when the reaction rate is very low. This is the case for reactions
whose rate depends on the concentration of a product component at the be
ginning of the reactor where the bulk concentration of product is low. Conse
quently, relatively large error is made in a small contribution. Therefore, these
errors have a very small impact on the overall results (predicted conversion and
selectivity) of the reactor model calculation. For the demonstrated cases, the
error made in the overall mass balance when using the approximation method
is in the order of a few percent.
It is concluded that for the systems that have been modelled, the approx
imation method can be used with sucient accuracy to calculate the eec
tiveness factor. In principle, the method works for any particle shape and for
general kinetics. However, only networks of rst order kinetics with Arrhenius
type temperature dependence were tested in this work. Extension to other
kinetics and particle geometries is left for future research.
With the extension of the Aris number approach to reaction networks, it
is possible to calculate the eective rate of reaction in catalyst particles with
general shape, for reaction networks and with general kinetics with a relatively
low computational eort but sucient accuracy so that it can be used in a
packed bed CFD model.

66

2.6 Conclusions

Nomenclature

Roman
Symbol
Ap
An0
An0
An1
An1
cCHj

units
m2
mol/m3

ci
ci,s

mol/m3
mol/m3

cp
dc
dp
De,i

J/kg/K
m
m
m2 /s

Dt,i

m2 /s

EA
r Hj
k0
L

J/kg
J/mol
(varies)
m

L
m

m
-

nc
nr
R
Ri
Ro
i
j


J/kg/K
m
m
mol/m3 /s
mol/m3 /s
mol/m3 /s

Variable
Particle outer surface area
Zeroth Aris number, as dened by equation (2.39)
Zeroth Aris number dened with linearised kinetics
First Aris number, as dened by equation (2.39)
First Aris number at the limit of small particle size
Concentration of the main organic reactant of reac
tion j in the Sohio process
Concentration of component i
Concentration of component i at the outer edge of
the particle
Specic heat
Diameter of the pores in a porous catalyst particle
Particle diameter
Eective diusion coecient of component i inside
the particle
Eective diusion coecient of component i in a
straight pore
Activation energy
Heat of reaction of reaction j
pre-exponential reaction rate constant
Distance from the particle edge to the symmetry
plane, axis or point
Length of ring shaped or cylindrical particle
Geometrical parameter (power with which surface
area increases with particle size)
Number of components in the reaction system
Number of stoichiometrically independent reactions
Universal gas constant (8.3144 J/kg/K)
inner radius of ring shaped particle
outer radius of ring shaped particle
Rate of consumption of component i
Rate of reaction j
Linearised reaction rate

67

2. Chemical reaction and diusion in a single catalyst particle

Symbol
t
T
Ts
Vp
x
X
y
yi
z
zj

units
s
K
K
m3
m
m
m
m
mol/m3

zeq

mol/m3

Variable
Time
Temperature
Temperature at the outer edge of the particle
Particle volume
Space coordinate
Large distance from particle edge
Space coordinate
Mole fraction of component i in the reaction mixture
Space coordinate
Pseudo concentration: independent variable in the
reduced set of equations
Pseudo concentration at thermodynamic equilibrium

Greek
Symbol
j
p

units
m
-

i
j
i

i,e

i,j

mol/m2 /s
W/m/K
kg/m3
-

c,act

68

Variable
Penetration depth of the reaction into a large particle
Porosity of the catalyst particle
Eectiveness factor: actual conversion divided by
conversion without mass transfer limitation
Eectiveness factor for component i
Eectiveness factor for reaction j
Eectiveness factor for component or reaction i at
the limit of small particle size
Geometry function dened in equation (2.52)
Molar ux of component i in coordinate direction e
Thermal conductivity inside the catalyst particle
Stoichiometry for component i in reaction j
Density of the catalyst
Tortuosity of the pores in a catalyst particle (equals
2
p,act
)
Geometric (actual) tortuosity of the pores in a cata
lyst particle

2.6 Conclusions

Literature
FlexPDE, Finite element software, web page: www.pdesolutions.com
Fuller, E.N., P.D. Schettler, J.C. Giddings (1966), A new method for the
prediction of gas-phase diusion coecients, Ind. Eng. Chem., 58, 19
Fogler, H.S. (1986), Elements of chemical reactor engineering, Prentice-Halll
Hopper, J.R., CL Yaws, T.C. Ho, M. Vichailak (1993), Waste minimisation
by process modication, Waste Management, 13, 3
Mleczko, L. (1996), Einuss der Hydrodynamik in einem
Wirbelschichtreaktor industriellen Mastabs auf Selektivitt und Ausbeute zu
Acrylnitril bei der Ammoxidation von Propen, Chemische Technik, 48 (3),
131
Reid, R.C. , Prausnitz, J.M., Poling, B.E. (1939), The properties of gases and
liquids, McGraw-Hill, 1987
Thiele, E.W. (1939), Relation between catalytic activity and size of particle,
Ind. Eng. Chem., 31, 916
Westerterp, K.R., R.J. Wijngaarden (1992), Principles of Chemical Reactor
Engineering, Ullmanns Encyclopedia of Industrial Chemistry, Vol. B4, 5
Wijngaarden, R.J., K.R. Westerterp (1994), Generalized formulae for the
calculation of the eectiveness factor, University of Twente, Enschede, the
Netherlands
Wijngaarden, R.J., A. Kronberg, K.R. Westerterp (1998), Industrial
catalysis: optimizing catalysts and processes, Wiley-VCH, Weinheim

69

Chapter 3
The wall eect
Summary
The structure of a packed bed near a wall diers from the structure far from any
wall. Therefore, the local bed parameters also dier. In this chapter, we will
look at the structure of the packed bed near a wall using an approach based
on separate particles. The bed structure is characterised by functions that
describe parameters such as porosity, particle outer surface area and tortuosity
as a function of the position in the bed. These functions can then be used in
a pseudo-continuous CFD model to describe the whole reactor.
The limited distance over which the porosity is inuenced by the wall
(about 5 particle diameters), suggests that the wall eect is only important for
very low vessel-to-particle-diameter reactors. However, due to the importance
of the wall zone in transfer of heat to or from the bed and because a relatively
large fraction of the ow may pass through the wall zone due to bypassing,
the wall eect is important for larger beds, up to D/d = 40. Hence, the wall
eect can have a considerable impact on the performance of many packed bed
reactors. To be able to make more ecient designs for these reactors, a model
is needed that predicts the inuence of the wall on the reactor performance.
Most of the discussion in this chapter is limited to packings of identically
sized, spherical particles. In real-life applications, spherical particles are not
very common. We use uniform spheres in order to be able to simulate and
understand the important features of the wall eect. For non-spherical parti
cles, the same mechanisms play a role but the ordering eect of the wall will
probably be less pronounced. In principle, the approach described here can
be extended to complex particle shapes. However, the computational eort
involved will be orders of magnitude higher.
In existing literature, the wall eect is commonly described in terms of the
bed porosity. The porosity proles have been measured directly by xating

71

3. The wall eect

particles in a bed and cutting and weighting parts of the bed. Fit curves have
been given for the oscillating porosity proles found as a function of the bedto-particle diameter ratio (e.g., Vortmeyer and Schuster, 1983, Mueller, 1991).
Critical inspection of the experimental data indicates that there is no evidence
that the shape of the porosity prole depends on the bed-to-particle diameter
ratio for D/d > 5.
Govindarao and Froment (1986) and Mariani et. al. (2001) tted the
experimental porosity proles using a particle centre distribution, i.e., the
density of particle locations as a function of distance from the wall. Mariani
et. al. (2002) claim that with this approach the particle outer surface area
prole near the wall can also be calculated. However, many dierent particle
centre distribution curves may give similar porosity proles but very dierent
particle outer surface proles, so this claim is not valid. Computer models have
been used to synthesise packed beds. Since as a result, all particle positions are
known, all bed parameters can be calculated for these simulated bed packings.
Spedding and Spencer (1994) give methods for these calculations, but do not
take into account the vessel wall. Mueller (1997) does take into account the
wall but only gives the porosity proles.
Two computer models are developed for the simulation of a random packing
of uniformly sized spheres: a numerical potential energy minimisation method
and a deterministic ball placement scheme. For the simulated packings, the
position of all sphere centres is readily available and parameters that would be
dicult or impossible to measure in a real packed bed can be computed with
relative ease. The bed packing simulation models are compared to existing
measurements to verify that they give realistic packed bed structures. The
simulated bed packings have a somewhat higher bulk porosity than experi
mental bed packings; this is probably caused by the fact that the particles are
not allowed to move under the weight of the particles that are placed on top
of them. From the simulated beds, a particle centre distribution is calculated;
this distribution corresponds to a realistic packing, and therefore can (in con
trast to a distribution that is tted to a porosity prole) be used to calculate
not only the porosity proles but also other parameters like the particle outer
surface area prole.
The porosity proles calculated form the simulated packings conform well
to the experimental porosity proles known in literature. The experimental
data suggests that in real packings, the wall eect is somewhat stronger than in
the simulated packings. This will at least in part be caused by the higher over
all porosity in the simulated packings, which allows more space for variations
in the particle layers near the wall.
The results show that the wall eect does not depend very strongly on
the vessel-to-particle diameter for the range considered (D/d > 5). This is

72

remarkable since most earlier researchers tried to quantify the wall eect as a
function of D/d. The average packing density will have more inuence on the
wall eect than the inuence of D/d.
It is obvious that the tortuosity is an important parameter in most trans
port processes in a packed bed, since it determines the length over which
transport through the continuous phase takes place. However, since it is hard
to measure tortuosity in a real packed bed, it is usually incorporated in more
or less lumped parameters such as eective transport coecients. It is clear
that the tortuosity is also strongly inuenced by the presence of a wall. The
tortuosity prole is calculated from simulated packings by dropping innitely
small particles through the packing using the potential energy minimisation
method. It is shown that, as a rst approximation, the local axial tortuosity
is inversely proportional to the local porosity.

73

3. The wall eect

3.1

Introduction

The goal of this work is to develop a tool to model chemical reactors with
solid catalysts using computational uid dynamics (CFD). A packed bed re
actor of industrial size can contain millions to billions of catalyst particles.
Therefore, it is in general feasible nor desirable to take into account each par
ticle separately in a CFD model of a complete reactor. In this chapter, we
will look at the structure of the packed bed, especially near a solid wall, based
on separate particles. This structure will be characterised by functions that
describe parameters such as porosity, particle outer surface area and tortuosity
as a function of the position in the bed. These functions can then used in a
pseudo-continous CFD model to describe the whole reactor.
It is well known that near a solid surface like the container wall, the particles
in a packed bed are no longer positioned randomly in space; due to the fact that
the particles cannot penetrate the wall surface, a (more or less) regular pattern
occurs near it. Starting at the wall, we nd a rst layer of particles that are
lined up against the wall. The next layers of particles will be increasingly less
ordered and far from the wall surface, the packing will be completely random.
This so-called wall eect occurs near the outer wall of a packed tube, but
also near inserts 1 , e.g., heat exchanger tubes used in highly exothermic or
endothermic processes. Measurements in cylindrical RPBs have shown that
the radial porosity prole shows complex behaviour; it starts at a value of 1
at the wall, drops to a minimum value of about 0.25 at approximately one
particle radius from the wall and oscillates to a mean value (the bulk porosity)
in about 5 particle diameters (Roblee et. al., 1958; Benenati and Brosilow,
1962; Goodling et.al., 1983).
The change in structure of the packed bed near the wall will inuence
the important transfer processes in this region of the bed. Not only will it
cause a change in the porosity near the wall, but it will also inuence other
bed parameters, e.g., specic particle outer surface area, gas path tortuosity,
number of particle to particle contact points. These parameters will have
their eect on the transfer processes. It is well known that the pressure drop
over a packed bed is very sensitive to the porosity, so the local velocity will
be inuenced strongly by the wall. Also mass transfer between the uid and
the particles and heat transfer between particles, uid and the wall will be
modied in the proximity of the wall. To be able to model ow and transport
of mass and heat in the region near the wall, we need to know parameters like
1

This is true even for the particles themselves; a radial porosity prole is observed around
the surface of each particle in the bed. (Spedding and Spencer, 1995)

74

3.1 Introduction

the porosity, particle outer surface area and tortuosity of the gas path as a

function of the distance from the wall.

The limited distance over which the porosity is inuenced by the wall
(about 5 particle diameters), suggests that the wall eect is only important
for very low vessel-to-particle-diameter reactors. However, since the ow re
sistance is very sensitive to the packing density, the eect causes bypassing of
ow along the wall. This may well inuence the ow in packed beds up to
40 of particles in diameter. In addition, especially for highly exothermic or
endothermic reactions like the Fischer-Tropsch, classical steam reforming or
ethene oxidation processes (that are typically carried out in relatively small
diameter packed tubes, the ow near the tube wall is very important for the
heat transfer from or to the catalyst particles inside. Hence, the wall eect can
have a considerable impact on the performance of many packed bed reactors.
To be able to make more ecient designs for these RPB reactors, a model
is needed that predicts the inuence of the wall on the reactor performance.
Therefore, not only the porosity prole but also the behaviour of the particle
outer surface area and tortuosity near the wall needs to be included in the
model.
Most of the discussion in this chapter is limited to packings of identically
sized, spherical particles. In real-life applications in packed bed reactors, all
sorts of particle shapes are usually applied for instance, cylinders, rings
or quadrulobes. Spherical particles are not very common since they are not
very easy to produce, and where they are used, they are usually not of a
uniform size. In this chapter, we restrict ourselves to uniform sphere packings,
in order to be able to understand the important features of the wall eect.
For non-spherical on non-uniform packings, the same mechanisms will play
a role, but the ordering eect of the wall will probably be less pronounced.
Therefore, the wall eect will in general be less important than for spherical
packings; however, there remain many cases in which detailed understanding
of the transfer processes near the wall is needed.
In principle, the approach described here can be extended to complex par
ticle shapes and (more easily) to spherical particles with a particle size dis
tribution. However, for less symmetrical particles, the number of degrees of
freedom is higher: a particle is not only characterised by its position in the
bed but also by its orientation. Therefore, the computational eort involved
will be orders of magnitude higher, not only because placing a single particle
is more dicult, but also because more repeated simulations would have to be
done to get statistically reliable data.
The wall eect has been included in RPB models in the past through an
exponential approximation of the porosity prole. Although it is convenient
in the computation, this prole does not capture the oscillating nature of

75

3. The wall eect

the actual prole, so these models cannot accurately predict the ow near
a wall in a packed bed of spherical particles. In addition, it is not only the
porosity near the wall that is of interest, but also the relative amount of particle
outer surface area and the tortuosity of the ow channels that determine the
transfer of momentum, mass and heat to the catalyst particles. Particle centre
distribution models (Govindarao and Froment, 1984; Delmas and Froment,
1988; Mariani et.al., 2002) are tted to actual measured porosity proles and
do capture the oscillating nature of the porosity near the wall. With this
approach, the particle outer surface area prole near the wall can also be
calculated, but it is questionable whether a particle distribution that ts the
porosity prole necessarily gives a correct particle outer surface prole. In
addition, the particle centre distribution does not allow the calculation of the
tortuosity prole.
In order to further study the wall eect, two numerical models are devel
oped for the simulation of a random packing of uniformly sized spheres. Since
of these simulated packings the position of all sphere centres is readily avail
able, parameters that would be dicult or impossible to measure in a real
packed bed can be computed with relative ease. The bed packing simulation
models will be compared to existing measurements and literature models to
verify that they give realistic packed bed structures. The models are used to
determine relations for the porosity, particle surface area and tortuosity as a
function of distance to a wall. These relations can be used to model ow, heat
and mass transfer in packed beds in general, without the need to simulate a
bed packing for each specic case.

3.2

A short review of existing literature

The wall eect has been known for quite some time and was frequently studied
using experimental techniques and dierent numerical models. The so-called
integral approach concentrates on the eect of the vessel diameter on the mean
bed porosity and ow resistance of the bed (e.g., Carman, 1937; Fand and
Thinakaran, 1990; Eisfeld and Schnitzlein, 2001), usually in cylindrical packed
beds of uniform spherical particles. This approach leads to expressions for the
mean porosity of the bed as a function of the vessel-to-particle diameter ratio:
b =

Vb Vp
= f (D/d)
Vb

(3.1)

Also, the ow resistance of the bed, usually described by some form of the
Ergun equation (Ergun, 1952) , is found to be a function of the vessel size:
dp
= Av + Bv 2
dz
76

A = f (D/d);

B = f (D/d)

(3.2)

3.2 A short review of existing literature

In addition to the eect of the bed diameter, the bed height is taken into
account in the so-called thickness eect (Zou and Yu, 1995). The inuence
of the wall eect on heat transfer between the uid and the wall was also
considered (e.g. Hennecke and Schl
under, 1973; Dixon, 1988). These models
are very useful for describing experimental data but they do not give a deeper
understanding of the wall eect. Consequently, application of these models to
packed beds with dierent geometries is dicult.
For very slender beds (D/d < 5), Computational Fluid Dynamics models
have been used to calculate the ow in the channels around a limited number
of spheres and near a wall, to study heat transfer between the wall and the bed
(Logtenberg and Dixon, 1998). The packing behaviour of these slender beds
where the wall zone thickness is of the same order of scale as the bed radius
is very dierent from that of larger beds and therefore the results cannot be
extrapolated.
In the so-called dierential approach, the structure of the bed and the ow
pattern near the wall is studied in more detail; here the porosity is treated as
a local variable inside the bed and is determined as a function of the distance
from the container wall.
The porosity prole near the wall has been measured directly by lling
the open space of a cylindrical packed bed with a xing agent that has a
density that diers from the density of the particles. After the xing agent had
solidied, successive layers of the cylinder are cut away. The mean porosity in
the removed volume can be computed from the weight of the removed material
through a simple material balance. Roblee et. al. (1958) were among the rst
who used this technique, using o.a. spherical particles made of cork and hot
wax as xing agent. Later Benenati and Brosilow (1962) used uniform lead
spheres and epoxy resin. Goodling et. al. (1982) used polystyrene spheres and
an epoxy resin that was mixed with nely ground iron to increase its density.
The results indicate that the porosity starts at a value of 1 at the wall, drops
to a minimum value of about 0.25 at approximately one particle radius from
the wall, and oscillates to a constant value after about 5 particle diameters for
uniform spherical particles. For packed beds consisting of spheres or dierent
sizes of non-spherical particles the eect is also found but not as pronounced.
Legawiec & Ziolkowsky (1994) measured the distance to the wall directly
for the rst layers of a packing using a depth gauge. They found that 98%
of the particles that are close to the wall (with their centres between 1 and
2 particle radii from the wall) actually touch the wall. About 10-12 % of the
analysed particles was found at a distance of 1.8 to 2.2 radii from the wall.
For use in analytical or numerical models, the porosity prole has been
simplied to a function that decreases exponentially from a maximum value at

77

3. The wall eect

the wall to a bulk porosity value at innite distance from the wall (Vortmeyer
and Schuster, 1983). To avoid numerical problems in some schemes, the poros
ity at the wall is set to a smaller value. A parameter C is introduced in the
equation to allow this; this parameter has to be adjusted to get the required
porosity at the wall for each value of the porosity at large distance of the wall
(b ):



x
(3.3)
(x) = b 1 + C exp 1
rp
Here x is the distance from the wall and rp the particle radius.
If the porosity is allowed to reach its real value at the wall, ((0) = 1) the
parameter C can be eliminated and the relation becomes:


2x
(x) = b + (1 b ) exp
(3.4)
dp
However, this prole does not take into account the oscillatory nature of the
porosity prole, e.g., the fact that the porosity has a minimum value at a
distance of about one particle radius near the wall. Therefore, the ow near
the wall in a packed bed cannot be accurately predicted when the wall eect
is modelled using equation 3.4. Mueller (1991) formulated an equation for the
porosity prole that does include the oscillating nature:
 = b + (1 b )J0 (ax)ebx

2.61 D/d

(3.5)

where
12.98
(D/d + 3.156)
2.932
a = 7.383
(D/d 9.864)
0.724
b = 0.304
(D/d)
a = 8.243

2.61 D/d 13.0

13 < D/d

He also proposes a correlation for the bulk porosity, that is the porosity far
from the wall, as a function of the vessel-to-particle diameter ratio:
b = 0.379 +

0.087
D/d 1.8

(3.6)

For beds with D/d > 10, this bulk porosity is about constant at a value of
0.38 to 0.39, as would be expected as the bulk porosity should be independent
of the vessel size. For lower vessel-to-particle diameter ratios, b increases to

78

3.2 A short review of existing literature

1
correlation D/d 13
correlation D/d 13

0.8

[-]

0.6

0.4

0.2

x/dp [-]

Figure 3.1: Porosity prole as predicted by the correlation of Mueller (1991),

using the two dierent expressions for a vessel-to-particle diameter ratio of 13.

a value of 0.475 at the lower limit. For D/d < 10, there is no zone in the bed
that is not aected by the wall; hence the dependence of b of D/d, which can
be seen as a correction for the inuence of the opposite wall on the porosity.
One obvious problem with these empirical correlation is that there is a
step in the value of a at D/d = 13, from 7.45 at the lower limit to 6.44 at the
higher limit, which causes a sudden change of the prole. Figure (3.1) shows
the prole for the two dierent values of a at D/d = 13. The predicted proles
in the high region (D/d > 13, gure 3.2) and, to a somewhat lesser extent, in
the low region (D/d between 4 and 13, gure 3.3) among themselves are quite
similar, especially in the rst few particle diameters from the wall. In light
of the error in the experimental data used (including the stochastic nature
of a packed bed) and the magnitude of the dierences between the empirical
relation and the experimental data (between 2.8% and 13.4%, as reported
by Mueller (1991)), the signicance of the dependency of the parameter a in
equation (3.5) on D/d can be put in doubt.
Govindarao and Froment (1986) proposed to describe the structure of the
bed near the wall by means of a particle centre distribution. It has been shown

79

3. The wall eect

1
D/d = 4
D/d = 6
D/d = 8
D/d = 10
D/d = 12

0.8

[-]

0.6

0.4

0.2

0
0

3
x/dp [-]

Figure 3.2: Porosity prole as predicted by the correlation of Mueller (1991),

for vessel-to-particle diameter ratios below 13.

that the packing properties near a wall are caused by ordering of the particles.
This ordering can be described by the particle centre distribution, that gives
the relative number of particle centres as a function of the distance to the wall.
None of the particles can have their centre in the zone between the wall and
one particle radius from the wall. At a distance of one particle radius from
the wall, there is a high concentration of particle centres, representing all the
particles that touch the wall. After this, there will be few particle centres until
somewhat less than three particle radii from the wall where the second layer
of particles is located. Since this layer is less ordered, the peak in the particle
centre distribution curve will be somewhat wider than that of the rst peak.
The third peak will be even less pronounced, until after about ten particle radii
the particle distribution will be approximately at. From the particle centre
distribution, other proles e.g., the porosity prole can be calculated
(Mariani e.a., 2001).
Govindarao and Froment (1986) used an approximate expression to derive
the particle centre distribution from measured porosity prole data from four
earlier sources. They divided the bed in cylindrical shells with a uniform width

80

3.2 A short review of existing literature

1
D/d = 14
D/d = 16
D/d = 18
D/d = 20
D/d = 22

0.8

[-]

0.6

0.4

0.2

0
0

3
x/dp [-]

Figure 3.3: Porosity prole as predicted by the correlation of Mueller (1991),

for vessel-to-particle diameter ratios above 13.

of an integer fraction of the particle radius (R = rp /m where m typically

has the value 3) and found that only the contributions of the rst two particle
layers, at a distance rp and approximately 3rp from the wall (shells m + 1
and 3m), depended on the bed-to-particle diameter ratio. Later, Delmas and
Froment (1988) improved the equations by taking into account the decreasing
volume of the cylinder shells ( for m = 3).
1.25
)
n4
(D/d)2
1.25
n9 = 4.65(1
)
n9
(D/d)2
7.4
)nj
nj = 2.1(1 +
(D/d)2

n4 = 5.25(1

(3.7)
(3.8)
j = 13, 14

(3.9)

Here, ni is the number fraction of particle centres in shell i (number of particles

in the shell relative to the total number of particles in the bed); n
i is the ratio
of the volume of shell i to the volume of the bed. Within each layer, the
particle centres are assumed to be distributed uniformly. The correlations are

81

3. The wall eect

said to be valid for D/d > 5. The curvature of the vessel wall is not taken
into account in these equations since its inuence is small (less than 5%) for
D/d > 5. In contrast to the earlier equations (Govindarao and Froment, 1986),
the relations of Delmas and Froment (1988) are valid only for m = 3, as can be
seen from the fact that nm+1 should be independent of m, while n
m+1 clearly
decreases with decreasing m.
The authors base their correlations on porosity prole data taken from
previous literature (Roblee et. al. (1958), Benenati and Brosilow (1962),
Ridgway and Tarbuck (1966) and Lerou et. al. (1980)). It has to be noted
that these measurements are all at higher D/d ratios (between D/d = 12.7 to
D/d = 20.3) except a single data point at D/d = 5.6. As can be seen, the
term that contains the contribution of D/d quickly decreases with increasing
D/d. In fact, for the rst two peaks, at D/d = 10, the contribution of this
term is only 1.25%. If the data point at D/d = 5.6 is removed from the data
set, a correlation where ni depends only on n
i and not on D/d may be just
as adequate as the inverse quadratic relation given by Delmas and Froment
(equation 3.9). In other words, it appears that there is little evidence that the
volume specic number fraction of particles in layer ni /n
i depends on D/d at
all. In fact, if we look at the factor before n
i in equation (3.7), the dierence
between its value at D/d = 5 (4.84) and at D/d = 20 (5.23) is less than
8%, so this decrease may well be caused at least in part by the fact that the
curvature of the vessel is not taken into account (an eect estimated by the
authors to be less than 5%). The dependence of the particle centre distribution
on the vessel-to-particle diameter ratio will become important only when D/d
becomes so small that the rst particle layers from both sides of the vessel
inuence each other, which will be the case if D/d < 5. In this range, one can
hardly speak of a randomly packed bed; the packing will become very sensitive
to the diameter ratio; these cases fall outside the scope of this chapter. For
D/d < 10, particle layers further from the wall will inuence each other, but
the importance of these layers on the bed properties is limited.
On the basis of these considerations, the particle density model by Delmas
and Froment (1988) can be simplied to a model without D/d dependence:
nm+1
n
m+1
n3m
n
3m
n4m+1
n
4m+1
n4m+2
n
4m+2

82

n4
= 5.25
n
4
n9
=
= 4.65
n
9
n13
=
= 2.1
n
13
n14
=
= 2.1
n
14
=

(3.10)
(3.11)
(3.12)
(3.13)

3.2 A short review of existing literature

Figure (3.4) shows that for the rst two particle diameters from the wall
there is only a small dierence in the predicted porosity proles for the full
model and the simplied equations above. For the full model, the last oscilla
tion between 2 and 2.5 particle diameters from the wall has a lower minimum
value than the minima closer to the wall, which is not in line with the ex
pected damped oscillatory behaviour. In the original publication (Govindarao
and Froment, 1986) this behaviour is also visible. In later publications (Del
mas and Froment, 1988, Papageorgiou and Froment, 1995) it appears that the
porosity prole is broken o just after the second peak (at about 2.1 particle
diameters from the wall) and set to decrease in a linear fashion to the bulk
porosity in one particle diameter.

1
Delmas & Froment
0.8

Delmas & Froment simplified

[-]

0.6

0.4

0.2

x/dp [-]

Figure 3.4: Calculated porosity prole (average porosity for concentric rings
with a width of 0.1 particle diameter) for D/d=8.56 for the full particle cen
tre density prole proposed by Delmas and Froment (1988) compared to the
simplied model (equations 3.10 to 3.13).

The model used by Delmas and Froment (1988) assumes that the particle
centres are distributed uniformly within each layer. Experimental results have
shown that for the rst layer, this may not be a valid assumption (Legawiec
and Ziolkowski, 1994): most of the near-wall particles actually touch the wall.

83

3. The wall eect

If a uniform distribution of the particle centres is assumed in the rst layer, in

fact none of the particles will touch the wall.

Mariani et. al. (2002) describe the particle centre distribution as a impulse
function at one radius distance from the wall (group 1), a small number of
particles uniformly distributed at a distance between 2 and 2.2 particle radii
from the wall (group 2), a larger group between 2.42 and 2.97 radii from the
wall (group 3) and the rest of the particles at a distance larger than 3.47
particle radii from the wall (group 4). The number of particles in the rst and
third group is tted to the porosity proles of Benenati and Brosilow (1962),
Ridgway and Tarbuck (1966) and Goodling et.al. (1983). The numbers are
given in terms of the maximum possible number of particles in a close packed
cylinder shell NC ; it is assumed that NC is a function of D/d, while the fraction
to which this close packing is realised in a bed is independent of the bed to
particle diameter ratio. A disadvantage of this approach is that the predicted
packing density near the wall does not depend on the bulk packing density,
i.e., the wall zone is not aected by loose or dense packing (compacting) of the
bed. Mariani et.al. nd that the rst layer contains 82.5% of the maximum
number of particles and the second layer contains 71.0% of the maximum value,
independent of bed porosity. The second group is assumed to contain 10.6%
of the number of particles in the rst group. The fourth zone contains the
rest of the particles, so that sum of the number of particles in the four groups
corresponds to the total number of particles in the bed.
The number of particles in the rst particle layer (at one particle radius
from the wall) and the second layer (at about 3 particle radii from the wall)
for the model by Delmas and Froment (1988) and the model by Mariani et.al.
(2002) is compared in gure 3.5. It can be seen that the dierence between
the two estimates is not very large, despite the considerable dierence in the
model layer structure and position.
Mariani et.al. (2002) claim that with this approach, the particle outer
surface area prole near the wall can also be calculated. However, there may
be many particle centre distributions that t the experimental porosity data
just as well as the one they propose, but give a dierent particle outer surface
distribution. In other words, it is not demonstrated that a particle distribution
that ts the porosity prole necessarily gives a correct particle outer surface
prole.
Several so-called sequential packing models have been used to create com
puter models of randomly packed beds. In these models, a packed bed is
simulated by adding new particles one by one at stable positions to a set of
previously determined particle positions. Generally, stable positions are con
sidered where a particle has three contact points, touching three other particles

84

3.2 A short review of existing literature

0.8
Layer 1; Mariani et.al.
Fraction particles in layer [-]

Layer 1; Delmas & Froment

0.6

Layer 2; Mariani et.al.

Layer 2; Delmas & Froment

0.4

0.2

0
5

10

15

20

D/d [-]

Figure 3.5: Number of particles (as a fraction of the total number of particles
in the bed) in the rst and second layer from the wall, as predicted by the
correlations of Froment and Delmas (1988) and Mariani et.al. (2002), as a
function of the bed to particle diameter ratio.

or two other particles and a container wall. The models dier in the way these
stable positions are found. A distinction can be made between stochastic
models in which a random number generator determines which of the stable
positions a particle will take, and deterministic models where the outcome is
predetermined (for a given initial or seed set of particles).
Spedding and Spencer (1994) have studied the structure of an innite
packed bed, but have not included wall eects in their simulations. They
used a dynamic model where each new particle was given a random velocity
and direction from a set position above the bed. The movement of the particle
was calculated taking into account the eects of impact, rolling and sliding of
the particle on the bed. They found that the resulting packing was not satis
factory and rejected the complex model in favour of models that did not take
into account momentum eects. In drop-and-roll models, a particle is released
from a random position just above the top of an existing bed. From there,
it moves straight down until it touches a particle or the bottom wall of the
container. In the rst case, the particle continues its way down by rolling along

85

3. The wall eect

the surface of the second particle until it reaches either a third particle or the
container wall or drops o. The simulation is continued in this way, calculating
the path of the particle along the surface of another particle, in the groove be
tween two other particles or between another particle and the container wall,
until it reaches a stable position. These models involve a relatively large num
ber of complex geometrical calculations to determine the particle path, which
leads to a complex and computationally intensive simulation procedure. The
particles are considered to be sticky i.e., a particle placed earlier will not move
when another particle is placed on top of it. Therefore, in theory, unphysical
(unstable) structures can occur in the bed and the packing density will usually
be less than that of real packings.
An alternative method to simulate a bed packing is the particle placement
model. Here, starting with an initial layer of particles, all possible stable
positions on top of the bed are calculated (e.g., Mueller, 1997). These positions
are either those where the new sphere touches three existing spheres or those
where the new sphere touches two existing spheres and the container wall.
The new sphere is placed at (e.g.) the lowest of these positions, after which
the list of stable positions is updated to reect the changes caused by the
new particle. This scheme will give the highest packing density possible. It
does need a seeding plane of particles to start the calculation. Mueller (1997)
compares dierent procedures for placing either wall spheres (touching the
wall) and inner spheres (not touching the wall) and compares the results to
measured porosity data. He nds the best t of the measurement data is the
packing method that gives the highest density; this is achieved by running the
model several times and nding the fraction of wall to inner spheres that gives
the highest packing density.

From the literature review, it can be concluded that there is no satisfactory

t function that accurately describes the porosity prole, particle outer surface
area prole and tortuosity near the wall. The particle centre density distribu
tion approach gives reasonable porosity prole predictions, but it is not certain
that the particle outer surface area is predicted correctly. Furthermore, there
is no method to predict the inuence of the wall on the tortuosity. Numerical
bed simulation models can give the opportunity to study all important param
eters: porosity, specic surface and tortuosity near the wall. However, none of
the publications on simulated beds give these parameters.
Therefore, two bed simulation models will be developed here. A potential
energy minimisation model and a particle placement model. The results of
the dierent models can be compared; it is expected that the former will give
somewhat higher porosity beds than the latter one.

86

3.3 Bed packing simulation models

3.3

Bed packing simulation models

In this section, two packed bed simulation models are developed; one stochastic
model where the stable position of a new particle is found by numerical min
imisation of the potential energy and a deterministic-stochastic model that is
a variant of the ball placement model. The purpose of this eort is to claculate
the porosity, outer particle surface area and gas path tortuosity as a function
of the distance from the wall, and if possible as a function of the bulk porosity
of the bed.

3.3.1

Potential energy minimisation model

In this computer model, equal-sized spherical particles are released with zero
speed at a random position just above the top of the bed in a cylindrical
container with radius R. The random positions are distributed uniformly
over the top surface of the bed. The movement of the particles is followed
as a function of time by minimisation of a function that is equivalent to the
potential energy. The spheres are allowed to overlap a distance l with other
spheres and the container wall. When a sphere overlaps with another sphere, it
will feel a force directed away from the other sphere, along the line connecting
both spheres. When a sphere overlaps with the side wall of the cylindrical
container, it will feel a force directed toward the axis of the container. When a
sphere overlaps with the at bottom of the container, it will feel a force directed
upward. The potential energy is the summation of the eect of gravity and
these repelling forces.
Each iteration, the particle is moved a short distance l in the direction
in which the potential energy decreases most rapidly, that is, in the direction
of the resultant force. The force that the new particle (n + 1) with radius rp
feels at position (x, y, z) is given by:
Fr = Fg + Fw + Fp

0

0
Fg =
-1



x
0
f
x2 + y 2 + r R
y

Fw = f (r z) 0 +
2 + y2
x
0
1

n
x xi

f (2r i )

y yi
Fp =

i
i=1
z zi

(3.14)
(3.15)

(3.16)

(3.17)

87

3. The wall eect

where i is the distance between the new particle and particle i


i = (x xi )2 + (y yi )2 + (z zi )2

(3.18)

The function f (x) is the force that occurs when a particle is compressed a
distance x; for simplicity a linear dependence was chosen:

0
x0
f (x) =
(3.19)
Kx
x>0
The constant K and the step size l are parameters of the model. When
the values are chosen so that Kl  1 (the contribution of gravity), the
spheres will not be compacted by more than a distance l at any time. The
step size should be chosen to be much smaller than the particle size. The
number of iterations or steps per particle is chosen large enough to ensure that
the particles reach a stable position. In a simulated packing, it is unlikely that
the distance between the lowest reachable position and the top of the highest
particle will be more than 5 particle diameters, hence a safe choice for the
number of steps is

nl > 10 2r
(3.20)
This algorithm can be programmed quite eciently; it only takes about
50 lines of code in the C programming language and simulating a bed of 5000
particles takes about 30 minutes on a standard personal computer.
Although it is implemented here for a uniform sphere packing, this method
could also be used to generate packings of non-uniform spheres.

3.3.2

Ball placement model

In the ball placement model, a list of all possible stable positions at the top
of the bed is maintained. Stable positions are positions where a sphere either
touches three other spheres or two other spheres and the container wall. If
a sphere with centre point p and radius r touches three other spheres with
centre points p1 , p2 and p3 , it must follow that point p is an intersection point
of three spheres at p1 , p2 and p3 with radius 2r. These points, if they exist,
are given by:

p p1  = 2r
p p2  = 2r
(3.21)

p p3  = 2r
The analytical solution is an unwieldy second order polynomial; the solu
tion can be simplied by a change of coordinate system so that sphere 1 is in

88

3.3 Bed packing simulation models

the origin, sphere 2 on the x-axis and sphere 3 in the x, y plane:




0
x3
x2





0
p1 =
p2 =
p3 =
0
y3
0
0
0

(3.22)

It is easy to see that x2 is simply the distance between spheres p1 and p2 and

that x3 is the distance between p3 and the line through p1 and p2 . Hence
 

x2 = d12 


 d .d


12
13



(3.23)
x3 =  d12 
d12 .d12
 

y3 = d12 
where
d12 = p2 p1
d13 = p3 p1

(3.24)

The two intersection points can be calculated easily in the local coordinate
system:


x = 2
2
 2


2
x x 2 x 3 + y3

y = 3

2y3


z = 1 x 2 y  2


(3.25)
(3.26)
(3.27)

Transformation to the global coordinate system gives the coordinates of

two intersection points of which only the one with the higher z coordinate is
a candidate for a stable position. We now have one sphere p that touches
three other spheres. The question remains whether this position will be stable
under gravity. The stability criterion may be formulated as follows: form a
triangle T by a projection of p1 , p2 and p3 on the x, y plane of the global axis
system (i.e, along the gravity vector). Position p is stable if (and only if) the
projection pxy of p on the x, y plane is inside the triangle T (see gure 3.6). It
can be easily veried that this criterion is met if the dot products of the vector
from a corner of T to pxy and from a corner to the next corner of T :

xy

(
pxy pxy
i pjxy
(i, j) = (1, 2); (2, 3); (3, 1)
(3.28)
i ). p
have the same sign for each given i, j.

89

3. The wall eect

Figure 3.6: Three spheres forming a stable (left) and unstable (right) position
for a fourth sphere. The projections on the (x, y) plane are drawn as well as
the triangle T and the projection of the fourth sphere (dashed) and its centre
point.

To further generalise the stability criterion, we realise that the three exist
ing spheres and the new sphere p form a tetrahedron with p at the top and p1 ,
p2 , p3 as base. If the vertical projection of p falls inside the vertical projection
T of the base triangle, the point p also falls inside the vertical projection of any
triangle that is formed by connecting any three points in each of the ribs p pi
of the tetrahedron. For instance, we could take the points where the sphere p
touches each of the other spheres (halfway each of the ribs of the tetrahedron).
The only important parameter for the stability of the sphere p is the location
of the three points on which it rests; the shape of the objects on which it rests
is not an issue. Therefore, we can state more generally that a sphere is in a
stable position under gravity if (and only if) the vertical projection of its cen
tre point falls inside the vertical projection of the triangle between the three
points on which it rests.
To nd the stable points that involve two spheres p1 and p2 and the wall, we
again take two spheres with double radius 2r and take a cylinder with radius
R r. The candidate for a stable point is the intersection point between both
spheres and the cylinder, if these points exist, with the highest z coordinate.

90

3.3 Bed packing simulation models

The intersection between both spheres will be a circle with radius rc in the
plane V halfway between both spheres and normal to the vector p1 p2 . Normal
projection of this circle on the x, y plane will give an ellipse with primary radius
rc and secondary radius rc cos() where is the angle between the plane V
and the x, y plane. If the spheres are smaller than the cylinder, which will
be the case in a packed bed, the ellipse will intersect the cylinder in at most
two points. These points can be found by rotation around the axis of the
cylinder to align the primary axis of the ellipse with the x axis, as given by
the equations:
 
2  

x xc + y yc = R 2
  
2

x
y
+
=1
rc
rc cos()




x1 + x 2
y 1 + y2
xc =
yc =
2
2

(3.29)

The resulting fourth order polynomial is solved numerically; after transfor

mation the x and y coordinates of the intersection points (if any) are found.
The z coordinates can be found by entering the values in the equations that de
scribe the spheres; for each sphere two roots are found, the intersection points
are given by the roots that are the same for both spheres. The candidate for a
stable position is the root with the highest z coordinate. The stability criterion
for a wall point is equivalent to that found for a sphere supported by three
other spheres.
The ball placement model must start with an initial layer of particles. This
can be either some regular packing or a layer of particles generated by another
packed bed simulation algorithm. A list of stable positions is calculated for this
initial layer; the number of stable positions is in the same order of magnitude
as the number of particles in the rst layer; for a hexagonal layer there are two
stable positions per particle, for a random layer less. A new sphere is placed
at the lowest stable position. If more than one stable position is at exactly
the same height (which will be very exceptional unless the existing bed has a
regular packing), one of these position is chosen at random. This will ensure
that the order in a bed will decrease with height in a random way, not in a
concidental way. After the particle has been placed, the list of stable positions
is updated by
1. deleting any stable positions that are less than 2r removed from the new
sphere and therefore no longer available, and
2. adding any new stable positions that involve the new sphere.

91

3. The wall eect

After this, a new position can be chosen et cetera. Since only stable positions
in the direct vicinity of the new sphere are aected, this is a very ecient
operation.
Since of all stable positions, the lowest one is chosen, the ball placement
algorithm will create random packings with the highest density. In order to be
able to create bed packings with dierent density, a parameter P is added to
the model. Instead of the lowest stable position, the model takes at random
one of the stable positions for which
z z0 + P (zmax z0 )

0P 1

(3.30)

For P = 0, the model is completely deterministic (for a random rst layer)

and will have a high packing density; for P = 1, the model is completely
stochastic and will have a relatively low packing density.

3.3.3

Modelling approach

Calculations were performed with both packing models for ve dierent vessel
over-particle diameter ratios. To rule out any inuence of the initial seed layer
of particles needed for the ball placement model, a seed layer was constructed
for each case as follows: a small number of spheres (50 for the smaller tubes,
100 for the largest) were dropped at random in an empty tube; then all balls
not touching the vessel bottom were removed, after which another batch of
spheres were dropped, et cetera. This procedure was repeated until, on visual
inspection, no sphere could be added so that it would touch the vessel bottom.
This seed layer was then used as initial condition for all further simulations.
In each case, beds were simulated by adding spheres to the existing number
in the bottom layer until the height of the bed was 110 particle diameters.
For the stochastic models, the results presented here are averaged results of 10
simulations starting with the same seed layer.
An example of the rst part of a simulated bed is shown in gure (3.7).

3.4
3.4.1

Bed packing simulation results

Mean bed porosity and thickness eect

One of the most obvious parameters that can be determined from a simulated
packed bed is the mean bed porosity. This is simply dened as the volume
of open space divided by the volume of the bed. The mean bed porosity will
depend on the size of the vessel and will be higher than the bulk porosity of the
packing far away from any walls, since it includes the eect of the wall. This in

92

3.4 Bed packing simulation results

Figure 3.7: A 3-D plot of the rst 200 particles in a ball placement simulation
(D/d=8.56).

principle includes the eect of the cylindrical wall as well as the bottom and top
boundaries of the bed. Since most literature data concerns only dependence on
the side wall, we will exclude the bottom and top sections of the bed. In order
to determine the extent of these sections, we will rst look at the inuence of
the top and bottom bed boundaries on the mean bed boundaries, the so-called
thickness eect.
For simulated as well as most real packings, there is a dierence between
the bottom of the bed and the top. The bottom will be conned by a solid
(albeit usually perforated) wall where the packing will be inuenced in (ap
proximately) the same way as near the side walls of the container. Near the
wall, the mean coordination number, i.e., the number of contact points per
sphere, will be approximately six. In other words, a particle will rest on three
contact points (either other spheres or the wall) and will on average support
three other spheres. The top of the bed however, is more or less a free surface,

93

3. The wall eect

where the coordination number is much lower. As such, the top of the bed
is not as well-dened as the other bed boundaries. The most convenient way
to dene it is as hmax , the top of the topmost particle. As there may well
be peaks and valleys on the top surface of the bed with amplitudes of several
particle diameters, the porosity of the bed dened this way increases rapidly
near the top of the bed. To study the thickness eect in the simulated packing,
the mean bed porosity between height h1 and h2 is dened as:
 n

Vi

m (h1 , h2 ) = 2 i=1
R (h2 h1 )

(3.31)

where Vi is the volume of particle i that lies inside the slice of the bed
considered. This includes sphere segments of particles whose centres lie outside
the range h1 , h2 but are closer than one particle radius to the edge and excludes
sphere segments of particles whose centres lie inside the range but are closer
than one particle radius to the edge (see gure 3.8)
Vi = 1/3 (r + zi h1 )2 (2r zi + h1 )

h1 r zi h1

Vi = 4/3r3 1/3 (r zi + h1 ) (2r + zi h1 )

Vi = 4/3r3

h1 zi h1 + r
h1 + r zi h2 r

Vi = 4/3r3 1/3 (r + zi h2 )2 (2r zi + h2 )

h2 r zi h2

Vi = 1/3 (r zi + h2 ) (2r + zi h2 )
Vi = 0

h2 zi h2 + r
elsewhere
(3.32)

The mean bed porosity is inuenced not only by the eect of the cylindrical
side-wall of the bed but also by the top and bottom boundaries. In order to be
able to exclude the eect of the top and bottom of the bed, the eect of these
boundaries on the mean bed porosity is investigated. The mean bed porosity
is calculated for a xed top cut plane of 5 particle radii from the top of the
bed and a bottom cut plane varying between 0 and 10 particle radii from the
bottom of the bed and vice versa. The result can be seen in gure (3.9) for a
bed to particle diameter of 8.56. From this gure it can be concluded that the
eect of the top and bottom bed boundary can be ruled out by disregarding
the part of the bed extending 5 particle diameters from the top and bottom.
Zu and You (1995) have given a formulation for the mean bed porosity,
based on their own experimental data and that of Carman (1937) and Dixon

94

3.4 Bed packing simulation results

zi

zi +r-h2

h2-zi +r

zi

h2

zi

zi +r-h1

h1-zi +r

zi

h1

Figure 3.8: The dierent contributing volumes to calculate the mean bed
porosity between two planes h1 and h2 .

(1988), for loose and dense packed beds as a function of the bed to particle
diameter that excludes the thickness eect:
(D/d) = 0.400 + 0.010(e10.68/(D/d) 1)

(D/d) > 3.91

(3.33)

(D/d) > 3.95

(3.34)

for the loose random packing and

(D/d) = 0.372 + 0.002(e15.30/(D/d) 1)

for the dense packing. Mariani et.al. (2002) propose a hyperbolic relationship:
(D/d) = 0.375 +

0.355
D/d

(3.35)

The mean bed porosity calculated from the simulated packings and the
values predicted by equations (3.33), (3.34) and (3.35) are compared in g
ure 3.10. The simulated packings are somewhat looser than the experimental
packings. This is a known feature of bed packing simulations that can prob
ably be ascribed to the fact that the particles are not allowed to move under
the weight of particles that are placed on top of them. In other words, the
fact that positions that are stable for a single sphere may become unstable

95

3. The wall eect

0.44
top cut
0.438
bottom cut

m [-]

0.436

0.434

0.432

0.43
0

10

Cut position [particle radii]

Figure 3.9: The eect of the top and bottom of the bed on volume average
bed porosity: the bulk porosity as a function of the distance of the cut plane
from the top of the topmost particle (top cut) and from the bottom of the bed
(bottom cut).

when other spheres are placed on top of it. The models follow the trend of
the experimental curves quite well. For the low range tube diameters, the
correlations by Zu and You predict a sharper rise in mean bed porosity than
is observed in the simulated packings.
It was expected that it would be possible to increase the porosity of the
simulated packings by changing the way particles are placed in the ball place
ment method (i.e., by not always choosing the lowest position available). It
was found that with the methods described it was not possible to increase the
porosity of the packing by more than 2 percent point.

3.4.2

Sphere centre density

The number of particles in the rst two layers near the wall of a number
of simulated ball placement packings is compared to the literature models
in gure (3.11). For the simulated bed, all particles between 2 and 3 radii

96

3.4 Bed packing simulation results

0.5
Yu & Woo loose
Yu & Woo dense
Mariani
Place
Drop

0.48

m [-]

0.46
0.44
0.42
0.4
0.38
0.36

10

12
D/d [-]

14

16

18

20

Figure 3.10: The mean bed porosity m as a function of the vessel-to-particle

ratio D/d as predicted by the deterministic ball placement model, the random
ball dropping model and the empirical relations for dense and loose packings
of Zu and You (1995) and Mariani et.al. (2002).

from the wall are included in the second layer. It can be seen that there is a
quite good agreement between the literature models and the ball placement
packings; although the rst layer of the simulated packings contains somewhat
less particles than the literature models predict.
The packing density of a typical packing from the ball placement model
of this work is compared to the packing density in the model of Froment and
Delmas (1988) and the model of Mariani et.al. (2002), for a bed to particle
diameter ratio of 10. The simulated packing contains about 4300 spheres
and has a height of 110 particle diameters; a slice with a height of 5 particle
diameters is removed from the top and bottom of the packing to remove any
bed thickness eects. The particle centre density was calculated with a uniform
bin size of 1/6 of a particle radius; the width of the rst layer of particles was
investigated separately and found to be very small (less than 0.01 particle
diameters wide).
The resulting accumulative particle density is shown in gure (3.12). The

97

3. The wall eect

0.8
Layer 1; Mariani et.al.

Fraction particles in layer [-]

Layer 1; Delmas & Froment

Layer 2; Mariani et.al.

0.6

Layer 2; Delmas & Froment

Layer 1; place model
Layer 2; place model

0.4

0.2

0
5

10

15

20

D/d [-]

Figure 3.11: Number of particles (as a fraction of the total number of particles
in the bed) in the rst and second layer from the wall, as predicted by the
correlations of Froment and Delmas (1988) and Mariani et.al. (2002) and as
calculated from the ball placement model as a function of the bed to particle
diameter ratio.

gure gives the number of particle centres as a fraction of the total number
of particles in the bed that can be found between the wall and the chosen
wall distance. Figure (3.13) gives the density function, i.e., the slope of the
curves in gure (3.12). The slope decreases to 0 at higher wall distances
because the radius and therefore volume of the remaining cylindrical bed core
becomes small. For the model of Mariani et.al. as well as for the simulated
bed packing, a fraction of the particles is at a position exactly one radius from
the wall, leading to a step change in the accumulated density function.
It can be seen that the number of particles in the rst layer near the wall
in the simulated packing is lower than for the literature correlations. Also, the
literature models expect no particles in the zones between layers, while in the
simulated packing the packing density never reaches zero. These two phenom
ena are connected as a lower packing density near the wall leaves more open
space for intermediate particles. Also, the notion that there are no particles in

98

3.4 Bed packing simulation results

fraction particles [-]

0.8

0.6

0.4

0.2

Delmas & Froment

Mariani et.al.
Place model

0
0

x/dp [-]

Figure 3.12: Cumulative particle density as function of the distance from the
wall for a simulated ball placement packing, the model of Delmas and Froment
(1988) and the model of Mariani et.al. (2002).

between the designated layers is an assumption that was used in the t proce
dures of both literature correlations, but that does not necessarily follow from
the experimental data. The main features of the literature correlations (very
narrow rst layer of particles at one particle radius from the wall, somewhat
wider layer of particles at about 3 particle radii from the wall) are observed in
the simulated packing.
When we consider the packing of particles in the rst two layers from the
wall, it is clear that the overall packing density is a much more important
parameter for the degree of order in the bed than the curvature of the vessel
wall. The higher the packing density, the more the packing will resemble a close
packing. Hence there will be less space in between the particles where other
particles belonging to intermediate layers t. Therefore, the peaks in the
density distribution will be sharper at higher densities. In the bed simulations,
the density is somewhat lower than that of the experimental beds, leading to
less close packing near the wall and less sharp peaks in the density distribution.

99

3. The wall eect

particle center density [-]

Delmas & Froment

Mariani et.al.

0.8

place model

0.6

0.4

0.2

0
0

x/dp [-]

Figure 3.13: Particle density as function of the distance from the wall for a
simulated ball placement packing, the model of Delmas and Froment (1988)
and the model of Mariani et.al. (2002).

3.4.3

Porosity prole

Mariani e.a. (2001) give analytical expressions to calculate porosity proles

from sphere centre distribution data. They are here rewritten to calculate the
proles for a simulated bed (i.e., e set of sphere positions) instead of a sphere
centre distribution.
The mean porosity in a cylindrical shell of the bed with width r can be
written as
 n
(r, r) = 1

vi (r + r) vi (r)
h ((r + r)2 r2 )
i=1

(3.36)

where vi (r) is the volume of particle i left inside a cylinder of radius r from
the centre of the bed:
4 3
3 rp
4 3
r (EV (r, ri ) + (r, ri ))
vi (r) =
3 p
0

100

ri < r rp
|ri r| < rp
ri > r + rp

(3.37)

3.4 Bed packing simulation results

where ri is the radial position of particle i


ri = x2i + yi2
and

if ri < r

/2 if ri = r
(r, ri ) =

0
if ri > r

The term EV is dened by:


1

EV =
(V + V B +
)F
2
A
(M B )

1

+ (A B )( 2 2 J + V D)
3
A k

(3.38)

(3.39)

(3.40)

where
A = min(1, S )

2
ri + r
S =
rp

2
ri r
B =

rp

M = max(1, S )
1
V = ( + 2 6 3)
3
1
V = (4 + 3 2)
3
r2 r2
= i 2
rp
=2

ri2 + r2
rp2

and the Carlson standard elliptic integrals (e.g., Press e.a., 1992):
F = RF (0, k 2 , 1)
D = RD (0, k 2 , 1)
J = RJ (0, k 2 , 1, p)

101

3. The wall eect

where

M A
M B

B
2
p=
k
A

k2 =

The use of equation 3.40 assures that the curvature of the cylindrical surface
is taken into account.
The porosity proles predicted by the particle placement model and the
particle drop model are compared to experimental data in gures 3.14 and
3.15. The model curves are for a bed-to-particle diameter ratio of 5, 10 and
15; the experimental data are for ve D/d ratios of 5.6 and higher. As can be
seen, neither the experimental data nor the model shows a clear dependence
of the porosity prole on D/d. The bulk porosity level of the experimental
data is lower than that of both models. Also, the amplitude of the oscillations
is slightly under-predicted by the model. These two observations are related,
since a higher porosity in a particle layer will create holes that can be partially
lled with particles that are somewhat oset from the layer.
The calculated porosity proles for the particle centre density distribution
proposed by Mariani et.al. (2002) and for the simplied model based on the
distribution proposed by Delmas and Froment (1988) are compared to exper
imental data and results of the particle placement model in gure (3.16). For
the model curves, the average porosity in concentric cylinder shells of 0.1 par
ticle diameter is plotted, since the experimental procedures also determine the
porosity in thin shells of the bed. For the ball placement model results, the
porosity proles for ten simulated beds with D/d from 5 to 20 were averaged.
It can be seen that there is quite good agreement between the computed and
measured porosity proles. Both particle centre distribution models predict
the porosity somewhat better than the ball placement model, but this is not
surprising since these are based on curve ts on the experimental data points,
while the ball placement model is built up from rst principles.

3.4.4

Particle surface area proles

Many processes in a packed bed, e.g., mass- and heat transfer or friction,
involve the outer surface of the particles. Often, for instance for chemical
reactors with cooled or heated walls, especially the rate of transfer near a wall
is important. Therefore, it is important to have a good estimate of the particle
surface area near the wall. In a randomly packed bed, the outer particle surface

102

3.4 Bed packing simulation results

1
Goodling, D/d=7.35
Goodling, D/d=8.56
Goodling, D/d=16.77
Benenati, D/d=14.1
Benenati, D/d=
Roblee, D/d=13.7
place model, D/d=10

0.8

[-]

0.6

Goodling, D/d=8.41
Goodling, D/d=10.7
Benenati, D/d=5.6
Benenati, D/d=20.3
Roblee, D/d=8.82
place model, D/d=5
place model, D/d=15

0.4

0.2

0
0

x/dp [-]

Figure 3.14: Porosity prole as predicted by the particle placement model (for
D/d=5,10 and 15) compared to experimental data (Goodling et.al. (1983),
Benenati and Brosilow (1962); Roblee et.al. (1958)).

area in a suciently large volume can be calculated from:

sv

Ap
6(1 )
=
V
dp

(3.41)

Near the wall, the volume of the particles in the ordered layers decreases quite
rapidly (leading to an increasing porosity to a value of 1.0 at the wall), while
the surface of the particles does not decrease at quite the same rate. In fact,
the particle outer surface per unit volume does not decrease to zero (as does
the particle volume per unit volume) , but to a nite value. Hence equation
3.41 can not be used to calculate the particle surface area from the porosity
prole near a wall.
Mariani et.al. (2001) devised a method to calculate the particle outer
surface area prole from a given particle centre density analytically. These
equations are rewritten here to allow calculation of the surface area prole
from a simulated bed packing. The particle out surface area per unit of volume
in a cylindrical shell between r and r + r can be written as:
n
Ai (r + r) Ai (r)
(3.42)
a
(r, r) = i=1
h ((r + r)2 r2 )

103

3. The wall eect

1
Goodling D/d=7.35
Goodling, D/d=8.56
Goodling, D/d=16.77
Benenati D/d=14.1
Benenati D/d=
Roblee, D/d=13.7
drop model, D/d=10

0.8

[-]

0.6

Goodling, D/d=8.41
Goodling, D/d=10.7
Benenati, D/d=5.6
Benenati, D/d=20.3
Roblee, D/d=8.82
drop model, D/d=5
drop model, D/d=15

0.4

0.2

0
0

x/dp [-]

Figure 3.15: Porosity prole as predicted by the particle drop model (for
D/d=5,10 and 15) compared to experimental data (Goodling et.al. (1983),
Benenati and Brosilow (1962); Roblee et.al. (1958)).

where Ai (r) is the outer surface of spherical particle i that is left inside a
cylinder with radius r:

ri < r rp
4Rp2
|ri r| < rp
4Rp2 (EA (r, ri ) + (r, ri ))
(3.43)
Ai (r) =

0
ri > r + rp
where ri is the radial position of particle i and the term EA is dened by:

1

(A + A B +
)F+
EA =
2
A
(M B )

(3.44)
1

(A B )( 2 2 J + A D)
3
A k
with the denitions given for equation 3.40 in the previous section and:
A = 1

A = 1

104

3.4 Bed packing simulation results

1
Goodling, D/d=7.35
Goodling, D/d=8.56
Goodling, D/d=16.77
Benenati, D/d=14.1
Benenati, D/d=
Roblee, D/d=13.7
Delmas & Froment simplified

0.8

[-]

0.6

Goodling, D/d=8.41
Goodling, D/d=10.7
Benenati, D/d=5.6
Benenati, D/d=20.3
Roblee, D/d=8.82
Mariani et.al.
place model

0.4

0.2

0
0

10

x/dp [-]

Figure 3.16: Measured porosity proles (Goodling et.al. (1983), Benenati

and Brosilow (1962); Roblee et.al. (1958)) compared to computed porosity
proles from particle density distribution models by Mariani et.al. (2002) and
equations (3.10 - 3.13) and results from the ball placement model.

The calculated particle surface area proles for the sphere dropping model
and the sphere placing method are shown in gure (3.17). The curve for the
sphere placing method gives the average curve of six such curves for simulated
beds of D/d=5 to D/d=19. There is a large dierence near the wall between
both literature correlations, even though the porosity proles that are pre
dicted by these models are very similar (see gure 3.16). The fact that in the
model by Delmas and Froment the particles in the rst layer are distributed
uniformly over al shell with a thickness of dp /6 causes the specic surface area
to drop to zero near the wall. In contrast, the specic surface predicted by
the distribution of Mariani et.al. is (nearly) constant near the wall since it
is inuenced only by the spheres with centres at one particle radius from the
wall. The ball placement model shows a more gradual decline near the wall,
due to the fact that in the simulated bed there are some particles close to,
but not touching the wall. Figure (3.17) shows clearly that the assumptions
made about the near-wall packing strongly inuence the specic surface area
near the wall, even for models that give similar porosity proles. Therefore, a

105

3. The wall eect

particle distribution tted to the porosity prole, as proposed by Delmas and

Froment and Mariani et. al. can not generally be used to calculate other bed
packing parameters.
2
Delmas & Froment
Mariani et.al.
Ball placement

sv/svb [-]

1.5

0.5

0
0

x/dp [-]

Figure 3.17: Specic surface proles (divided by the bulk specic surface) as
calculated from the computed porosity proles from particle density distri
bution models by Mariani et.al. (2002) and by Delmas and Froment (1988)
compared to the results from the ball placement model.

3.4.5

Tortuosity proles

The radial prole of the axial tortuosity can be studied using the potential
energy minimisation model described in section 3.3.1, when the diameter of a
tracer particle that is released is set to zero. Such a particle will not settle and
will instead nd a path through the ow channels to the bottom of the bed.
In fact, the path of such a particle is a streamline for creeping axial ow (ow
without recirculations, turbulence or inertia).
The radial tortuosity prole is generated from a large number (400) of such
tracer particle paths, starting at random positions just above a simulated bed.
Each tracer particle path is divided into stretches in with a height of of 1/2 bed
particle radius; for each of these stretches, the length of the path and therefore

106

3.4 Bed packing simulation results

the tortuosity is calculated. The bed is divided into a number of thin shells.
Each stretch of a path is attributed to a shell depending on its centre radial
position. For each shell, the average tortuosity of all stretches that have their
centre point at a radial position within that shell is calculated. The volume
average tortuosity in the bed, i.e.,
n
1  2
(ri ri21 )i
2
R i=1

(3.45)

is about 1.43, which is close to the expected value for randomly packed beds.
The inuence of the bed-to-particle ratio on the volume averaged tortuosity
is not very large because the volume of the near-wall zone is not very large
compared with that of the whole bed.
1.6

m [-]

1.4

1.2

10
D/d [-]

12

14

16

Figure 3.18: Calculated bed-average tortuosity for simulated beds with dif
ferent bed-to-particle ratios. For D/d=5 and D/d=8.56, points for multiple
dierent packings are shown.

The preference of the ow paths for the near-wall zone becomes visible
when the number of one radius particle path stretches in each cylindrical shell
is plotted (gure 3.20). It can be seen that there are many ow paths near the

107

3. The wall eect

wall and only a few ow paths that pass at a distance of one particle radius
from the wall. The reason for this is that the channels between the particles
in the rst layer are all approximately horizontal, so that there is often no
driving force for the tracking particle to cross the channel.

2.4

D/d=8.56(1)

D/d=8.56(3)
D/d=8.56(5)
D/d=7
D/d=11

2.2
2

D/d=8.56(2)
D/d=8.56(4)
D/d=5
D/d=9
D/d=15

[-]

1.8
1.6
1.4
1.2
1
0

2
x/dp [-]

Figure 3.19: Local tortuosity as a function of distance from the wall, for several
simulated packings with dierent bed to particle diameter ratio D/d.

There is quite a lot of spread in the raw tortuosity data. This is a result of
the stochastic nature of the data; the random factors in the bed packing and
the way the tortuosity is determined. The spread is especially large in regions
where the tortuosity is high, i.e., where the number of tracer particle stretches
in each shell is low.
From gure (3.19) and gure (3.14) it can be seen that the tortuosity prole
is roughly the inverse of the porosity prole: when the porosity increases, the
tortuosity decreases and vice versa. Therefore, it seems practical to try to
write the tortuosity as a function of porosity:
(x) = f ((x))

108

(3.46)

3.4 Bed packing simulation results

n/ntracks [-]

D/d=8.56(1)

D/d=8.56(2)

D/d=8.56(3)

D/d=8.56(4)

D/d=8.56(5)

D/d=5

D/d=7

D/d=9

D/d=11

D/d=15

0
0

x/dp [-]

Figure 3.20: Number of tracer particle path stretches as a function of distance

from the wall, divided by the number of tracks, for several simulated packings
with dierent bed to particle diameter ratio D/d. Note that the rst point (at
x=0) is not shown since it has a value 10 times the maximum of the graph.

where the function should have the following boundary behaviour:

1
 = b

w
= b

(3.47)
(3.48)

where b is the tortuosity far from the wall and w is the tortuosity at the
wall (which is close to one). The tortuosity at the wall is not exactly equal to
zero because a small volume around the point where the particles touch the
wall is not accessible for gas-phase transport since the space is too narrow. In
the model the size of this volume is determined by the step size with which the
ow paths are calculated. In reality there will also be excluded volume, but
the magnitude will depend on the transport process at hand. For instance, for
diusion processes the size will be in the order of the free path length of the
gas molecules. In general the excluded volume will be small.
It should be noted that in theory, tortuosity and porosity are independent
properties; if one would drill a ow path into a block of solid material, both

109

3. The wall eect

the porosity and the tortuosity can be chosen. However, in a xed randomly
packed bed, the properties of the ow channel are of course not independent,
e.g., compaction of the bed will not only lead to a higher porosity but also
to a change (increase) in tortuosity. In this case, the increase in tortuosity is
caused by ordering of the particles near the wall, which may not lead to the
same correlation between porosity and tortuosity as for instance compaction.
There are two classes of simple functions that comply to the restraints
listed above and give the observed inverse behaviour: a linear function and a
reciprocal function:
1 (x)
(b w )
1 b

1 1/(x)

(x) = w +
(b w )
1 1/b
(x) = w +

(3.49)
(3.50)

As can be seen in gure (3.21), both functions give quite similar results
except for low porosities, where the linear function has a maximum (1.66 for
the chosen bulk values) and the reciprocal function becomes innite, which
means that the ow paths are horizontal. The low porosity values near the
wall in a randomly packed bed are caused by the fact that the particles are
neatly stacked near the wall, all right on top of each other. Therefore, it is
probable that nearly horizontal ow paths and therefore very high values of
the local tortuosity to occur in this zone. The reciprocal correlation (3.50)
therefore is thought to describe the relation better.
In gure (3.22) the reciprocal correlation (3.50) is compared with the tor
tuosity prole. The latter was obtained by averaging the tortuosity proles of
ve dierent simulated beds (ball placement method). The bulk porosity was
taken as 1.43, the bulk porosity was 0.43; the error bars give plus and minus
one standard deviation of the ve averaged tortuosity values. The correlation
ts the calculated tortuosity prole quite well, although it seems that the anks
of the rst peak are a bit steeper than the correlation suggests. The peak of
the correlation is higher than most data points, the data set even shows a dip
in the centre of the rst peak. However, this is caused by the fact that the
high tortuosity ow paths are not taken by the tracing particles.

110

3.4 Bed packing simulation results

2.2
Linear
Reciprocal

[-]

1.8

1.6
b

1.4

1.2

b
0

0.2

0.4

0.6

0.8

[-]

Figure 3.21: Comparison of a linear and reciprocal function for correlating the
porosity  and the tortuosity

111

3. The wall eect

2
bed simulation
correlation

1.8

[-]

1.6

1.4

1.2

1
0

2
x/dp [-]

Figure 3.22: Local tortuosity as a function of wall distance; results from a

series of bed simulations (D/d=8.56) and correlation

112

3.5 Conclusions

3.5

Conclusions

Two dierent models were used to simulate a randomly packed bed of spherical
particles in a cylindrical container: a deterministic ball placement model and a
stochastic random ball dropping model. Both models were compared to each
other and to experimental data and models from literature concerning the
overall bed porosity, the particle centre density distribution and the porosity
prole.
Both models show very similar behaviour, despite the large dierences in
bed simulation algorithm. As was also found by other authors using similar
models (e.g., Spedding and Spencer, 1995), the simulated bed packings have
a somewhat higher bulk porosity than experimental bed packings. This is
probably caused by the fact that the particles, once they are settled, are frozen
in their position and are not allowed to move under the weight of the particles
that are placed on top of them.
The porosity prole at the wall of the vessel for both models is similar. The
experimental data suggests that in real packings, the wall eect is somewhat
stronger than in the simulated packings. This will at least in part be caused by
the higher overall porosity in the simulated packings, which allows more space
for variations in the particle layers near the wall. It cannot be ruled out that
realistic packings are not quite random in some way, e.g. because particles are
poured in at the centre of the bed, roll down a slope at some speed, skipping
positions that would be stable due to their momentum and not stopping until
they hit the wall. In this way, the wall positions could be lled preferentially,
leading to a higher density near the wall and a somewhat lower density further
from the wall.
Both models show that the wall eect does not depend very strongly on
the vessel-to-particle diameter for the range considered (D/d > 5). Critical
inspection of the experimental data found in literature (e.g., Goodling et.al.
(1986)) indicates that there is no evidence that the shape of the porosity prole
depends on the bed-to-particle diameter ratio for D/d > 5. This is remarkable
since most earlier researchers (Mueller, 1990; Delmas and Froment (1988))
tried to quantify the wall eect as a function of D/d.
It is expected that the packing density and the packing and compacting
method used to create a packed bed will have more inuence on the wall eect
than the inuence of D/d. The experimental conditions given in literature
are usually not sucient to calculate the bulk packing density with a good
accuracy. There is no dependable procedure to create random packings exper
imentally with a given porosity. On the other hand, it is not easy to control
the porosity of a simulated bed packing; for the ball placement method, the
addition of a stochastic parameter that causes a fraction of the particles to be

113

3. The wall eect

placed at random in one of the stable positions instead of in the lowest stable
position increased the bulk porosity only by about 2% point. It would be in
teresting for future research to modify the method (for instance, by allowing a
fraction of the particles to settle in metastable positions) to be able to simulate
a wider range of bed porosities and investigate the eect of the bulk porosity
on the wall eect.
Apart from the porosity, the simulated packing also allow the calculation of
other bed parameters. The fact that the position of all spheres in a simulated
packing is known, allows the direct calculation of the specic particle outer
surface as a function of the distance from the wall. Mariani et.al. (2001) give
a procedure for calculating this value based on a particle centre distribution;
however, this distribution cannot be measured directly in a real packing but
is determined by tting measured porosity proles. Therefore it depends on
the choices made in the tting procedure. For instance, the particle centre
distributions proposed by Mariani et.al. (2002) and Delmas and Froment give
quite similar porosity proles (which is not surprising since they are partly
based on the same porosity measurements), but very dissimilar particle surface
area proles.
A third parameter that is determined in the simulated packings is the
tortuosity. It is obvious that the tortuosity is an important parameter in most
transport processes in a packed bed, since it determines the length over which
transport through the continuous phase takes place. However, since it is hard
to measure tortuosity in a real packed bed, it usually does not appear explicitly
in the transport equations. Instead, it is incorporated in more or less lumped
parameters such as eective transport coecients. Although most literature
concerned with the wall eect focuses on the porosity prole or the particle
surface area, it is clear that the tortuosity is also strongly inuenced by the
presence of a wall. It has been shown that as a rst approximation, the local
tortuosity is inversely proportional to the local porosity.

114

3.5 Conclusions

Nomenclature
Roman
Symbol
A
Ap
Ai
a
B
b
C
D
dij
dp
EA
EV
Fg
Fp
Fw
f
hi
J0
k
l
n
ni
n
i
m
p
pi
pxy
r
rp
sv
sv,b
Vb
Vp
vi
x
y
z

units
Pa/s
m2
m2

(Pa/s)2

...
m
m
m

arbitrary
arbitrary
arbitrary
...
m

Pa

m
m
m2 /m3
m2 /m3
m3
m3
m3
m
m
m

Variable
constant in equation (3.2)
outer surface area of particles
outer surface area of particle i
t parameter
constant in equation (3.2)
t parameter
constant in exponential porosity prole
tube diameter
distance between centre points of sphere i and j
particle diameter
particle surface function dened by equation (3.44)
particle volume function dened by equation (3.40)
gravitational force vector
inter-particle repulsion force vector
wall repulsion force vector
generic function
vertical position of a cut plane
Bessel function of the rst kind and order 0
parameter in equation (3.40)
step size for sphere drop model
number of steps
fraction of particles in layer i
fraction of particles in layer i with no wall eect
number of concentric layers per particle radius
pressure; parameter in equation (3.40)
centre point of sphere i
projection of point p on the x, y plane
space coordinate in radial direction
particle radius
specic outer particle surface
specic outer particle surface in the bulk of the bed
volume of the bed (uid space + particles)
volume of the particles
volume of particle i
spatial coordinate; distance from the wall
spatial coordinate
spatial coordinate (in axial direction)

115

3. The wall eect

Greek
Symbol

A
V
A
V


m

units
-

b

i
m
w

Variable
angle

parameter in equation (3.44)

parameter in equation (3.40)

parameter in equation (3.44)

parameter in equation (3.40)

dierence, parameter in equation (3.40)

porosity
mean bed porosity (volume available for uid / bed
volume)
bulk porosity (porosity of the packing without wall
inuences)
parameter in equation (3.40)
tortuosity of the bed (path length / distance)
bulk tortuosity of the bed (tortuosity of the packing
without wall inuences)
average tortuosity in concentric layer i
average tortuosity of the bed
tortuosity at the wall
angle
parameter in equation (3.40)

Miscellaneous
Symbol
F
D
J

116

units
-

Variable
Carlson standard elliptic integral in equation (3.40)
Carlson standard elliptic integral in equation (3.40)
Carlson standard elliptic integral in equation (3.40)

3.5 Conclusions

Literature
Benenati, R.F., C.B. Brosilow (1962), Void fraction distribution in beds of
spheres, AIChE J. 8(3), 359-361
Delmas, H., G.F. Froment (1988), A simulation model accounting for
structural radial nonuniformities in xed bed reactors, Chem. Eng. Sci. 43,
2281-2287
Dixon, A.G. (1988), Wall and particle-shape eects on heat transfer in
packed beds, Chem. Eng. Commun. 71, 217-237
Eisfeld, B, K. Schnitzlein (2001), The inuence of conning walls on the
pressure drop in packed beds, Chem. Eng. Sci. 56, 4321-4329
Ergun, S. (1952), Fluid ow through packed columns, Chem. Eng. Progress
48(2), 89-94
Goodling, J.S., R.I. Vachon, W.S. Stelpug, S.J. Ying, M.S. Khader (1983),
Radial porosity distribution in cylindrical beds packed with spheres, Powder
Technol. 35, 23-29
Govindarao, V.M.H., G.F. Froment (1986), Voidage proles in packed beds of
spheres, Chem. Eng. Sci. 41 (3), 533-539
Govindarao, V.M.H., K.V.S. Ramrao (1988),Prediction of location of
particles in the wall region of a randomly packed bed of spheres, Chem. Eng.
Sci. 43 (9), 2544-2545
Govindarao, V.M.H., M. Subbanna, A.V.S. Rao, K.V. Ramrao
(1990),Voidage prole in packed beds by multi-channel model: eects of
curvature of the channels, Chem. Eng. Sci. 45(1), 362-364
Kubie, J. (1988), Inuence of containing walls on the distribution of voidage
in packed beds of uniform spheres, Chem. Eng. Sci. 43, 1403-1405
Legawiec, B., D. Zilkowski (1994), Structure, voidage and eective thermal
conductivity of solids within near-wall region of beds packed with spherical
pellets in tubes, Chem. Eng. Sci. 49, 2513-2520
Logtenberg, S.A., A.G. Dixon (1998), Computational Fluid Dynamics studies
of xed bed heat transfer, Chem. Eng. Proc. 37, 7-21
Mariani, N.J., G.D. Mazza, O.M. Martnez, G.F. Barreto (2000), Evaluation
of radial voidage prole in packed beds of low aspect rations, Can. J. Chem.
Eng. 78, 1133-1137

117

3. The wall eect

Mariani, N.J., O.M. Martnez, G.F. Barreto (2001), Computing radial packing
properties from the distribution of particle centres, Chem. Eng. Sci. 56,
5693-5707
Mariani, N.J., et.al. (2002), Packed Bed structure: evaluation of radial
particle distribution, Can. J. Chem. Eng. 80(2) 186-193
Mueller, G.E. (1991), Prediction of radial porosity distributions in randomly
packed beds of uniformly sized spheres in cylindrical containers, Chem. Eng.
Sci. 4692), 706-708
Mueller, G.E. (1992), Radial void fraction distributions in random packed
xed beds of uniformly sized spheres in cylindrical containers, Powder
Technol. 72, 269-275
Mueller, G.E. (1997), Numerical simulation of packed beds with monosized
spheres in cylindrical containers, Powder Technol. 97, 179-183
Papageorgiou, J.N., G.F. Froment (1995), Simulation models accounting for
radial voidage proles in xed-bed reactors, Chem. Eng. Sci. 50(19),
3043-3056
Roblee, L.H.S., R.M. Baird, J.W. Tierney (1958), Radial porosity variations
in packed beds, AIChE J. 4(4), 460-464
Ridgway, K., K.J. Tarbuck (1968), Voidage uctuations in randomly packed
beds of spheres adjacent to a contaning wall, Chem. Eng. Sci. 23, 1147-1155
Spedding, P.L., R.M. Spencer (1994), Simulation of packing density and
liquid ow in xed beds, Computers Chem. Engng 19(1), 43-73
White, S.M., C.L. Tien (1987), Analysis of ow channeling near the wall in
packed beds, Wrme- und Stobertragung 21, 291-296
Wijngaarden, R.J., K.R. Westerterp (1992), The statistical character of
packed-bed heat transport properties, Chem. Eng. Sci. 47(12), 3125-3129
Zou, R.P., A.B. You (1995), The packing of spheres in a cylindrical container:
the thickness eect, Chem.Eng.Sci. 50 (9), 1504-1507

118

Chapter 4
Flow resistance in randomly
packed beds
Summary
To calculate ow in a packed bed, a model is needed to relate the uid velocity
to the pressure drop. Traditionally, equations like the Ergun equation are used,
that relate pressure drop to the (mean) bed porosity and particle size. These
equations cannot bed used for a detailed ow model near the wall, because
they are based on the assumption of a uniform random packing.
One of the ways to look at ow through a packed bed is as a collection
of interconnected, twisted channels. In this chapter a detailed model for the
ow through such channels is presented that is based on rst principles and
engineering relations. With this model, the pressure drop can be written as a
function of velocity with the porosity , the specic particle outer surface area
sv and tortuosity as parameters. The result is:




2
! sv 1
p
2 f sv
3
f w0 + kt
f w0 |w0 |
(4.1)
= 2
3
43 2
z bed
where kt is a term that accounts for wall roughness, bends and expansion/
contractions. This equation shows the same dependency on the porosity and
particle size as the (primarily empirical) Ergun equation. In addition, when
the coecients are tted to the Ergun equation, very reasonable values are
obtained for the properties of the twisted channel (i.e., about one bend and
one expansion/ contraction for each particle passed, a tortuosity of 1.44 and a
wall roughness of 0.04).
In contrast with Ergun and similar empirical relations, the equation above
can be applied locally in the bed near a wall if the local porosity, particle
surface area and tortuosity are known.

119

4. Flow resistance in randomly packed beds

The equation found is applied in a two-dimensional model using the modi

ed Brinkman equation to calculate the radial velocity proles in packed tubes,
using the bed parameter proles from the previous chapter. They compare well
to experimental velocity proles from literature.

120

4.1 Introduction

4.1

Introduction

Randomly packed beds are used very widely in a broad range of equipment
for many types of solid-uid contacting processes. Consequently, there has
been a large interest in the modelling of the important processes in randomly
packed beds. Traditionally, randomly packed beds are applied as packed tubes.
Therefore, the majority of measurements have been carried out using this
cylindrical bed geometry. Flow resistance models for packed beds have been in
use for a long time and are very successful in describing these standard packed
beds. Even though no specic bed shape is implied, the experimental data on
which they are based is mostly for packed tubes. In the classical approach,
the eect of the tube wall is not taken into account; the tube is assumed to
be large compared to the particle diameter. The ow resistance equation of
Ergun (1952) is still very widely used and is relatively accurate in predicting
the pressure drop in packed tubes of small particles, even though there is a
signicant spread in the experimental values. In more recent years, the wall
eect has been taken into account in more detailed models (e.g., Vortmeyer
and Schuster, 1983; Delmas and Froment (1988), Bey and Eigenberger, 1996).
These models are in general limited to cylindrical beds.
Although the cylindrical packed tube is the most common geometry for a
packed bed reactor, other geometries may be more suitable for specic applica
tions. For low-pressure drop, shallow packed beds, the packed tube is usually
not the most ecient arrangement, and radial ow, lateral ow or parallel
passage reactors may be employed (Calis, 1995). Also for the development
of compact installations, non-standard geometries may be necessary for ease
of construction, accessibility and heat integration. An example of such an
application is a mobile reforming installation for fuel cell applications.
It has been shown that the presence of a wall has a large inuence on the
local structure of a randomly packed bed. The inuence of the wall extents up
to ve particle diameters into the bed. Therefore, unless the bed diameter is
very large compared to the particle diameter the wall will have an eect on the
ow resistance of the bed. As a rule of thumb, the wall eect needs to be taken
into account for beds with a diameter smaller than 40-50 particle diameters.
For cylindrical bed geometry, the wall eect may be taken into account by
adding a tube-to-particle diameter ratio dependence to the classical equations.
However, such an approach would only give a prediction of the pressure drop
over a packed tube, and not give any information about the ow near the
wall. For packed beds where the wall is important for the ow resistance,
the wall will also be important for other transport processes i.e., the transfer
of heat and mass. Detailed information about the ow near the wall will be
needed to predict the performance of such a packed tube as a heat exchanger

121

4. Flow resistance in randomly packed beds

or a chemical reactor. Furthermore, for non-cylindrical bed geometries, or if

there are internal walls in the bed (e.g., heat exchanger tubes), the wall eect
cannot be described by a simple tube-to-particle diameter ratio dependence.
Therefore, it is necessary to look in more detail to the processes that determine
the ow near the wall of a packed bed.
In the previous chapter relations were derived for the porosity, tortuosity
and specic outer particle area as a function of the distance to the wall. In
this chapter those relations are used to create a model that describes the ow
resistance of a randomly packed bed as a function of the distance from the
wall. Based on microscopic models for the ow in the packed bed, relations
are derived that describe the inuence of the porosity, specic outer particle
area and tortuosity on the ow resistance. This leads to a local-homogeneous
model, that is identical to the classical (e.g., Ergun) approach far from the
wall, but takes into account the change in bed structure near the wall.
In the following chapters, models will be derived for dispersion and chemical
reaction in randomly packed beds. These will be integrated with the results
of this chapter into a general Computational Fluid Dynamics based randomly
packed bed model.

4.2

A short overview of existing literature

Among the rst to acknowledge that the velocity prole in packed beds is not
uniform were Schwartz and Smith (1953). They measured the radial velocity
prole two inches downstream of primarily packed beds of cylinders and found
velocity proles with a single maximum near the wall. As at the time there were
no porosity prole measurements available, they calculated the porosity prole
from their velocity prole measurements and found proles that, starting from
the centre of the bed, are almost at until a distance of up to two particle
diameters from the wall and then increase monotonously towards the wall.
Later studies employed the reverse path. The porosity prole was measured
by direct methods (see chapter 3 for details), and this was used in various
models to calculate the velocity prole inside the bed. Although the porosity
measurements showed an oscillatory decrease of porosity from the wall towards
the centre of the bed, early and even some recent ow models employ an ex
ponential approximation of the porosity prole to simplify the equations (e.g.,
Vortmeyer and Schuster, 1983; White and Tien, 1987; Tsotsas and Schl
under,
1988; Cheng and Vortmeyer, 1988; Winterberg et al., 2000).
Detailed measurements of the radial velocity prole using optical techniques
(McGreavy et al., 1986; Giese et al., 1998) or using a monolith section to
conserve the ow prole downstream the bed (Bey and Eigenberger, 1997)

122

4.3 Flow resistance in innite packed beds

show oscillations of the velocity as a function of the radial position near the
wall. However, the amplitude of the oscillation and the distance in terms of
number of particle diameters into the bed in which it is damped out diers
considerably between authors.
Models to calculate the ow inside the bed are usually based on the Brinkman equation, that includes viscous friction in the uid as well as friction
caused by the particles of the bed (Vortmeyer and Schuster, 1983; White and
Tien, 1987; Delmas and Froment, 1988; Cheng and Vortmeyer, 1988; Giese et
al., 1998; Subayago et al., 1998; Winterberg et al., 2000). Some use a twozone model in which the bed is divided into a wall zone and a core zone. In
these models there is no viscous term and wall inuence is limited to the wall
zone (Schwartz and Smith, 1953; McGreavy et al., 1986). Cheng and Yuan
(1997) calculate the hydraulic diameter of ow channels as a function of the
distance to the wall. They use the Ergun equation to compute the ow prole
and therefore do not take into account radial (viscous) transport of momentum.
The surface of the vessel wall iss incorporated in the hydraulic diameter prole.
However, they do not take into account the dierence in distribution of volume
(porosity) and surface of the particles near the wall.
Also true 2-dimensional models are used, where the momentum balances
are solved in both the axial and radial directions (Papageorgiou and Froment,
1995; Bey and Eigenberger, 1996; Tierney et al., 1998). For all these mod
els, the porosity prole chosen has a large inuence on the resulting velocity
proles, and are thus a large factor in the comparison with measured proles.
Giese et al. (1998) used laser-Doppler velocimetry in a transparent packed
bed and determined the porosity from the locations where the velocity was
zero. Their model is very successful in describing their own measured velocity
proles, but there is a large dierence with other literature.
Many of these models use Ergun type friction coecients that are implic
itly based on the assumption that the bed is a uniform randomly packed with
spherical particles (e.g., Vortmeyer and Schuster, 1983; White and Tien, 1987;
Giese et al., 1998; Winterberg et al., 2000) although near the wall this as
sumption is not valid. Delmas and Froment (1988) take into account both the
specic surface area and the tortuosity prole; however, the way the tortuosity
is taken into account is based on empirical arguments and they do not take
into account the eect of the particle outer surface area prole correctly.

4.3

Flow resistance in innite packed beds

In a random packed bed, the void space in between the particles forms a con
tinuous network of twisted channels. When a uid ows through a randomly

123

4. Flow resistance in randomly packed beds

packed bed, these channels will cause a ow resistance. The pressure drop
across a packed bed has long been known to be proportional to the velocity
at low ow rates (Blake-Kozeny equation) and proportional to the square of
the velocity at high ow rates (Burke-Plummer equation). Ergun (1952) has
shown that an expression for the whole range of Reynolds numbers can be ob
tained by summation of the expressions for the low and high Reynolds number
range. This lead to his well-known formulation, which is adequate for dense
packings (0.2 <  < 0.7) of compact particles without holes:

p
(1 )2
1
= 150f 3 2 w0 + 1.75f 3 w0 |w0 |
 dp
z
 dp

(4.2)

For the purpose of application of a similar equation in a detailed packed bed

model, we will look into the basic assumptions underlying this equation (e.g.,
Bird et al., 2002). For the derivation of equation 4.2, we can either look at the
packed bed as a set of separate particles or as a set of twisted interconnected
ow channels. Since both approaches should lead to the same result, we follow
the latter path, which seems more physically correct for dense beds.
For low to intermediate velocity ow in a straight tube with constant crosssectional area, the pressure drop is caused primarily by wall friction. The wall
friction is characterised by the friction factor f , which is dened as the ratio
between the wall shear stress and the dynamic head of the uid:
w
f 1
(4.3)

v|v|
f
2
A force balance over a length z of a tube with cross sectional area A and
circumference W :
pA = w W z

(4.4)

leads to an expression for the pressure drop per tube length:

p
4f 1
(4.5)
=
( f v|v|)
z
dh 2
where dh is the hydraulic diameter:
4A
dh
(4.6)
W
which is dened so that for cylindrical tubes dh equals the tube diameter.
For laminar ow of a Newtonian uid in a smooth straight cylindrical tube
of constant diameter, the steady-state velocity prole can be solved analyti
cally: the friction factor f is found to be inversely proportional to the Reynolds
number (Poisseuille relation):
4f =

124

64
Re

(4.7)

4.3 Flow resistance in innite packed beds

For fully turbulent ow in not perfectly smooth tubes the friction factor is
independent of the Reynolds number and is a function of wall roughness.
The channels in the randomly packed bed dier from these perfect tubes
in a number of ways:
the combined cross sections of all channels is a factor  smaller than the
cross-section of the tube
the cross-section of the channels is not circular
the channels are not straight
the channel walls are not perfectly smooth
the channels are interconnected
Although the geometry of the channels is too complex to derive an expression
for the friction factor based on theoretical arguments only (parameters tted
to experimental data will be needed in the end), it does provide insight to
follow this road a little further.
In order to apply the equations derived for cylindrical tubes to the ow
channels in a packed bed, we need to determine the hydraulic diameter of the
tubes. For the twisted channel of a packed bed with porosity  and specic
surface sv , we can multiply numerator and denominator of equation (4.6) by
the channel length to obtain:
dh =

4
sv

(4.8)

This gives an idea of the size of the channels in a typical packed bed: for
random packings of spheres with a porosity of 0.37, the hydraulic diameter is
about 0.4 times the particle diameter.
The corresponding Reynolds number is calculated as:
Reh =

vdh
f

(4.9)

Here v is the mean velocity in the channel, so for a given mass ow, v increases
inversely proportional with the porosity.
The concept of a hydraulic diameter is based on the assumption that the
inuence of the channel wall is limited to a thin layer near that wall. Therefore,
the curvature of this layer can be ignored and the inuence depends only on the
size of the wall surface. For fully developed laminar ow, the velocity prole
spans the whole cross section, so the walls of a channel inuence the velocity

125

4. Flow resistance in randomly packed beds

in every point of the cross-section. Therefore, strictly spoken, the concept of a

hydraulic diameter can not be applied to laminar ow. However, in a packed
bed the changes in channel shape appear so often and suddenly that the ow
will not be fully developed anywhere in the bed (except for very small values
of the Reynolds number, say Re < 0.1). Therefore, we will use the hydraulic
diameter also in laminar conditions.
The tortuosity is used to characterise the measure of straightness of the
ow channels in a packed bed:
=

path length
length of channel
=
distance
height of bed

(4.10)

For randomly packed beds, the value is usually in the order of 1.4.
For a length l of a channel, the pressure drop can be written to depend
on the following factors:


1 2
p = () (f (Re))
v l
(4.11)
2
Here depends on the channel shape and f (Re) gives the dependence on ow
regime. The tortuosity inuences the pressure drop in the channels in three
ways:
As the tortuosity increases, so does the length of the channel (for a given
bed height), leading to higher pressure losses through a change in l
As the channel length increases with tortuosity, for a given mass ow
through the bed, also the velocity in the channel must increase, leading
to an increase in pressure loss through a change in v
A highly tortuous channel will (usually) have more bends and shape
changes, leading to an increase in pressure loss through a change in
The porosity, specic volume and tortuosity of a packed bed are independent
properties that can be varied separately from each other, as is shown in gure
4.1.
In the laminar regime, we can write for a slice of a (cylindrical) channel of
length l:


p
64 1
=
( f v|v|)
(4.12)
l l Reh dh 2
The average distance travelled in the axial direction is obviously l/ . The
velocity v in the channel is related to the supercial velocity w0 as:
w0
v =
(4.13)

126

4.3 Flow resistance in innite packed beds

= 0.25

sv = 2/d

=1

d
= 0.5
sv = 2/d

=1

sv

= 0.5
sv = 1/d
=1

= 0.5

sv = 2/d

=2

Figure 4.1: Independent variation of porosity, tortuosity and specic channel

surface area, illustrated for a square 2-dimensional grid.

Hence, with equation (4.8) and (4.9), the laminar term can be written as:


p
f s2
= 2 2 3 v f w0
(4.14)

z l
For packed beds with a uniform random packing of spheres, the specic outer
particle surface area can be written as a function of the porosity and particle
diameter:
sv =
so that


p
z

6(1 )
dp

=
l

2
2 f (1 )
72
f w0
dp2 3

(4.15)

(4.16)

Comparison with the rst term of theErgun equation (equation 4.2) reveals
that in the latter, has a value of 150/72 = 1.44, which is in excellent

127

4. Flow resistance in randomly packed beds

accord with the expected value. In other words, the pressure drop of an actual
packed bed under laminar conditions is quite similar to laminar ow through
a bundle of twisted cylindrical tubes.
Now consider the fully turbulent regime. As the channels cannot be con
sidered to be smooth, the friction factor of the channels will be independent
of the Reynolds number for turbulent ow. Therefore, in the fully turbulent
ow regime, equation 4.5 can be written as




p
4f 1
=
f v|v|
(4.17)
l t,wall
dh 2
where (4f ) is a value that will depend on the channel roughness. The lower
limit for (4f ) will be given by that for smooth tubes, about 0.01; for very rough
tubes, (4f ) increases to about 0.1 (e.g., Bird et.al., 2002). With equations
(4.8) and (4.9), we can write (4.17) in terms of the bed porosity, tortuosity
and specic surface:




p
3 sv 1
(4.18)
= 4f 3
f w0 |w0 |
2
z t,wall
4
To compare this result with the Ergun equation (4.2), we can substitute sv by
equation (4.15) which is valid for uniform randomly packed beds of spherical
particles:




p
3 3 (1 ) 1
f w0 |w0 |
= 4f
(4.19)
z t,wall
2
2 3 dp
To comply with the Ergun equation, 34 (4f ) 3 has a value of 1.75, so that has
a value between 3 and 6, depending on the value of 4f . Hence for turbulent
ow, the straight tube model leads to unphysically high values of . In other
words, the eective friction factor in a packed bed for turbulent ow (based on
the Ergun equation) is much higher than the wall friction factor for the straight
tube model (between 0.01 and 0.1). Apparently, in the turbulent region, the
contribution of friction at the particle wall to the pressure drop in packed beds
is small and pressure drop is mainly caused by other mechanisms, i.e., shape
drag due to bends and expansion/contractions in the channels.
The pressure drop caused by bends in the channel can be estimated when
we make a few assumptions. For turbulent ow in tube systems, the pressure
lost in one sharp (i.e., not smooth) bend is usually given as a ow resistance
coecient that is tabulated for dierent bend angles. The pressure loss in a
bend is then calculated as:
1
p = f v|v|
(4.20)
2
128

4.3 Flow resistance in innite packed beds

Here we need the pressure loss in the bend as a function of the bend angle
(see gure 4.2). From the literature values, the following behaviour can be
observed: for a straight channel ( = 0), the bend pressure loss is of course
zero, while for a 90 degrees bend, almost all kinetic energy of the ow is lost
( 1). For even sharper bends, the pressure drop increases to about twice this
value. A simple function that shows this behaviour is = 1 cos(). Without
proof it is postulated here that the bend pressure drop can be estimated by:
1
p1 = (1 cos()) f v|v|
2

(0 )

(4.21)

where is the angle of the bend. It can be seen (gure 4.3) that this function
ts the commonly used resistance factor values for sharp bends quite well,
while its mathematical form is very convenient.
To estimate the mean value of the bend angle, we can write:

1
sin( ) =
2

(4.22)

For equation (4.21) we need an expression for cos(). Since cos() = cos()
and cos() = 1 2 sin2 (/2),
2
cos() = 2 1
(4.23)

Let be the average distance between two bends in a channel; then the
pressure drop due to bends per length of bed is:




p
1
2
1
f v|v|
=
(1 2 )
(4.24)
z t,bends
2

The distance between bends depends of course on the particle size, but also on
the distance between two particles in the packing. Therefore, let us suppose
that the distance between two bends is proportional to the hydraulic diameter
dh :
= nb dh

(4.25)

so that nb is the number of hydraulic diameters of straight channel between

two bends. For a standard packed bed (porosity 0.37), the hydraulic diameter
is about 0.4 times the particle diameter so since we expect something in the
order of one bend per particle, the number of hydraulic diameters between
bends (nb ) should be between, say, 2 and 3. For the pressure drop due to
bends, in terms of the supercial velocity (equation 4.13), it follows that:




p
2( 3 ) 1
= 2
f w0 |w0 |
(4.26)
2
z t,bends
 nb dh

129

4. Flow resistance in randomly packed beds

dp

Figure 4.2: Schematic representation of a simple zigzag model to determine

the mean number of bends and bend angle in the ow channels of a packed
bed

or, eliminating the hydraulic diameter dh with equation (4.8):





p
( 3 )sv 1
=
f w0 |w0 |
2
z t,bends
23 nb

(4.27)

For a uniform randomly packed bed of spherical particles we can apply equation
(4.15):




p
3( 3 ) (1 ) 1
f w0 |w0 |
=
(4.28)
z t,bends
2
nb
3 dp
Hence, if the pressure drop in packed beds in the turbulent regime was
caused solely by the bends in the channels, the factor 3( 3 )/nb should have
the value 3.5 (compare the Ergun equation (4.2)). With the value of 1.44 for
found in the laminar regime for the standard packed bed, it follows that the
number of hydraulic diameters between bends nb is 1.33. Of course, as the
pressure drop is not caused by the bends alone, this number will be somewhat

130

4.3 Flow resistance in innite packed beds

2
1-cos( )
[1] smooth
[1] rough
[2]
[3]

1.6

[ ]

1.2

0.8

0.4

0
0

30

60

90

120

[]

Figure 4.3: Pressure drop coecient in sharp bends at high Reynolds num
bers as a function of bend angle (see gure 4.2) for dierent literature (hand
book) sources and the t function (equation 4.21). Literature data: [1] Leijen
dekkers et al. [red.] (1998); [2] Janssen and Warmoeskerken (1987); [3] Blerins
(1984).

higher in reality, but it is clearly in the right order of scale. The main result,
however, is that the pressure drop due to bends depends on the porosity and
particle diameter in a way that is consistent with the Ergun equation, and
increases linearly with 3 .
The pressure drop due to a sudden expansion from cross-sectional area A1
to A2 can be estimated from (e.g., Leijendekkers et.al., 1998)

p =

2
A2
1
1
f v22
2
A1

(4.29)

where the velocity v2 is taken at the downstream position. To estimate the

expansion in a packed bed we look at a close packing of four spheres with
radius R. The sphere centres form a regular tetrahedron with sides equal to
the sphere diameter. We dene a z axis perpendicular to one of the faces of
tetrahedron, so three of
the sphere centres have z = 0. The fourth, top sphere
4
has its centre at z = 3 3R. Now we look at the area occupied by the solids in

131

4. Flow resistance in randomly packed beds

cross sections perpendicular to the z axis, bounded by the prism that is formed
by the equilateral triangle that connects the bottom three sphere sections in
the plane z = 0 and the z-axis. This prism encloses a single vertical
segment of a ow channel; each channel is formed by a large number ofthese
segments in dierent directions. The cross sectional area of the prism is 3R2 .
The cross section Ab of the three bottom spheres consists of three 60 circle
segments:
 1
(R2 z 2 )
R z R
2
(4.30)
Ab (z) =
0
elsewhere
Likewise, the contribution of the top sphere to the cross section At is


4
3R R z 43 3R + R
(R2 (z 43 3R)2 )
3
At (z) =
0
elsewhere

(4.31)

3
2
1

Figure 4.4: Cross-sectional solid (white) and empty (black) space in a close
packing arrangement of four spheres. At plane 1, the three lower spheres are
cut through their centre points at a (local) maximum solid area; at cut plane
2, still only the lower three spheres are cut, at about the minimum solid area.
At plane 3, all four spheres are cut.

132

4.3 Flow resistance in innite packed beds

A/R2 [-]

0
0

2
z/R [-]

Figure 4.5: Cross-sectional surface of four close-packed spheres and a plane

parallel to the plane through three sphere centres, as a function of the position
of the plane (black line). Dashed lines give the cross-section of the three
bottom spheres (Ab ) and of the top sphere At ; grey lines show surface that
falls outside the prismatic volume considered (see gure 4.4).

A plot of the total solid cross-sectional area is given in gure (4.5). Clearly,
the maximum cross sectional solid area occurs at z = 0. The minimum cross
sectional area of the channel is:



1
A1 =
3 R2
(4.32)
2

The minimum cross sectional solid area occurs at z = ( 43 3 1)R, so the

maximum cross sectional area of the channel is:



1
4
2
A2 =
3 (1 (
3 2) ) R2
(4.33)
2
3
Therefore, the expansion factor becomes:

( 43 3 2)2
A2

=1+
A1
2 3

(4.34)

133

4. Flow resistance in randomly packed beds

or

A2
1)2 =
ka = (
A1

2

( 43 3 2)2

0.87
2 3

(4.35)

If we assume that the interstitial velocity in the bed is the mean value
of the velocity upstream and downstream of the expansion, the downstream
velocity v2 equals 2/3v. If we now assume that there is one expansion for ne
hydraulic diameters of channel length, the pressure drop due to expansions
can be estimated as:




p
4ka 1
(4.36)
=
f v|v|
z t,expansion 9ne dh 2
where the factor appears because the pressure drop is given per unit bed
height z instead of per unit channel length. In terms of the supercial velocity
w0 and specic surface sv (equation 4.8, 4.13) this becomes:




p
k a 3 sv 1
(4.37)
=
f w0 |w0 |
z t,expansion
9ne 3 2
When there are expansions in the channel, there should also be contractions;
since the pressure drop in a sudden contraction is only a fraction of that of a
sudden expansion, and the dependence on bed parameters will be equal, the
contractions can be taken into account by a slight increase in the parameter ne .
For a uniform random packing of spheres, the specic surface can be eliminated
(equation 4.15):




p
2ka 3 (1 ) 1
(4.38)
=
f w0 |w0 |
2
z t,expansion
3ne 3 dp
For a standard packed bed (tortuosity 1.44), to comply with the Ergun equa
tion (equation 4.2), there should be 0.5 hydraulic diameters between two ex
pansions if the pressure drop was attributed to expansions alone. This is about
0.2 particle diameters for a standard packed bed, which is in the right order
of magnitude since obviously the pressure drop in a packed bed is not caused
channel expansions alone. The main result is that the pressure drop due to
expansion/contractions is proportional to to the third power.
For the pressure drop in the turbulent region, we can add the contribu
tions of bends and expansions/contractions together with the wall friction. To
reduce the number of degrees of freedom, we will assume that the number of
bends is equal to the number of expansions (nb = ne = n).



 

p
3 ( 3 ) ka 3 sv 1
+
f w0 |w0 |
(4.39)
= (4f )t +
z t
4
2n
9n 3 2
134

4.3 Flow resistance in innite packed beds

To determine the values for the remaining parameters n and (4f )t , we once
more compare this result to the Ergun equation (4.2). For uniform randomly
packed beds of spherical particles we can write:






p
3 3 3( 3 ) 2ka 3 (1 ) 1
f w0 |w0 | (4.40)
+
+
= (4f )t
z t
3 dp
2
n
2
3n
so that, to comply with the Ergun equation,


1
3 3 3( 3 ) 2ka 3
(4f )t
+
+
= 1.75
2
n
2
3n

(4.41)

If we take (4f )t to be 0.1 (for very rough walls) and = 1.44 as found for
laminar ow, it follows that n equals 2.1. If we take (4f )t to be 0.01 (for
not so rough walls), the value of n becomes 1.85. For ease of calculation, we
take n = 2. This means that there is one bend and one expansion for every 2
hydraulic diameters or about 0.8 particle diameters. With this value







p
2ka
sv 1
3

f w0 |w0 |
(4.42)
=
(4f )t + 1 +
9
43 2
z t
We will replace the factor before 3 by the single parameter kt , hence,




! sv 1
p
3
f w0 |w0 |
(4.43)
= kt
43 2
z t
If we require the equation to be identical to the Ergun equation (4.2) at =
1.44, kt has a value of 1.26 (this corresponds with a value of 0.0415 for the
wall friction factor (4f )t ).
The pressure drop in a packed bed over the full Reynolds number range
can now be written (equations 4.14, 4.43) as:




2
! sv 1
p
2 f sv
3
f w0 + kt
f w0 |w0 |
(4.44)
= 2
3
43 2
z bed
Note that since equation 4.15 has not been not used to eliminate the specic
outer particle surface, the assumption that the bed consists of uniform spheres
that is implied there, is no longer needed. The packing is completely charac
terised by the porosity, tortuosity and specic outer particle surface area, so
it can easily be applied to packings of non-spherical particles or particle size
distributions (the original restrictions to particle shape still apply). For the
special case of a uniform random bed with a tortuosity of 1.44, equation (4.44)
reduces to the Ergun equation (4.2) if an equivalent diameter based on the
particle outer surface area is used as the particle diameter:
dp,eqA =

6(1 )
sv

(4.45)

135

4. Flow resistance in randomly packed beds

If we look at two random packings of particles with the same size but with
a dierent bed porosity, it is clear that the tortuosity in the loose packing
will be lower than in the dense packing. Hence, tortuosity will increase with
decreasing porosity. This implies that the pressure drop in a packed bed, as
predicted by equation (4.44), depends more strongly on the porosity than is
suggested by Ergun (equation (4.2). However, it is important to realise that
the change in tortuosity between dense and loose random packings diers in
nature from the change in tortuosity near a wall due to the ordering of particles.

4.4

Flow resistance: the wall eect

For nite random packed beds, the porosity cannot be treated as a constant
value throughout the bed, but is a function of the distance x from the container
wall. The specic particle outer surface area sv is also a function of x. Equation
4.15 is not a very good estimate of the local specic particle surface area if the
packing is not fully random. Using the equations presented in the previous
chapter, both the particle volume prole and the particle surface prole can
be calculated from the particle centre distribution. In addition, the tortuosity
will change from a mean value far from the wall to a value of 1 at the wall;
a relation for the tortuosity prole was also presented in chapter 3. We can
write equation 4.44 as a function of the distance to the container wall x:
p
f [sv (x)]2
=2[ (x)]2
f w0
z
[(x)]3



 sv (x)
1
3
f w0 |w0 |
+ kt [ (x)] (x)
4[(x)]3 2

(4.46)

Near the wall, the porosity will increase and the tortuosity and specic
surface area will decrease. According to equation 4.46, the increase in porosity
and the decrease of tortuosity near the wall will have a strong diminishing
eect on the local ow resistance, while the decrease in specic surface area
will have a somewhat smaller eect. Figure (4.6) shows the proles for porosity,
tortuosity and specic area for a typical packed bed, according to the results
of the previous chapter.

4.5

Velocity prole

The one-dimensional ow in a packed bed is often described by the modi

ed Brinkman equation (e.g., Vortmeyer and Schuster, 1983), in cylindrical

136

4.5 Velocity prole

1.6
1.4

[-], [-], sv /sv,b [-]

1.2
1.0
0.8
0.6
0.4

sv /sv,b

0.2
0
0

x/dp [-]

Figure 4.6: Porosity , tortuosity and specic outer particle surface area sv
(normalized by its bulk value sv,b ) as a function of the distance to the wall
(in particle radii) for a typical randomly packed bed (based on the results of
chapter 3).

coordinates:


p
w0
e
2
= f1 w0 f2 w0 +
r
z
r r
r

(4.47)

with boundary conditions

wall : r = dt /2 w0 = 0
0
centre : r = 0 w
=0
r

(4.48)

Here f1 and f2 are the laminar and turbulent friction terms, often taken
from the Ergun equation. When the pressure gradient in the z direction is
xed, equation (4.47) reduces to an ordinary dierential equation and can
easily be solved to yield a radial (supercial) velocity prole.
When we take equation (4.46) to describe the ow resistance of a packed
bed as a function of the distance from the wall, it follows that
f1 (x) = 2[ (x)]2

f [sv (x)]2
f
[(x)]3

(4.49)

137

4. Flow resistance in randomly packed beds

Table 4.1: References and conditions

symbol w0,m [m/s] dp [mm]
Rep
8.0
268
0.5
1
4.5 76-450
0.25-1.5
2
9.8 165-987
0.25-1.5
3
7.0
235
0.5
4
9.8
658
1
5
8.6
4
0.004
6
8.6
77
0.076
7
8.6
103
0.102
8
8.6
532
0.526
9
12.7
10
0.014
12.7
100
0.14
12.7
1000
1.4

for data points in gure (4.7).

D/d note
a
6.3
a
11.1
a
5.1
a
7.1
a
5.1
b
9.30
b
9.30
b
9.30
b
9.30
c
9
c
9
c
9

a) Bey and Eigenberger, 1997. Fluid: air; velocity proles measured downstream of
a packed bed supported by a 3.5 mm long monolith to preserve the ow prole.
b) Giese et.al, 1998. Fluid: special liquid mixture (viscosity 8.5106 m2 /s); velocity
measured inside bed by Laser-Dopper velocimetry.
c) This work. Results of 1-D computational model for supercial velocity inside
the bed; uid properties correspond to air, eective viscosity is a function of wall
distance according to equation (4.51).

and
! sv (x)f
f2 (x) = kt [ (x)]3 (x)
8[(x)]3

(4.50)

The third term of equation (4.47) gives the dispersion of momentum in the
bed. The eective viscosity e is not known a priori and is usually a t factor
that is a function of the average Reynolds number. Here we have chosen a
prole based on the analogy between the dispersion of mass and momentum,
which will be described in more detail in the next chapter:

e
Rep 2 1
= 1 1 m +
(4.51)
8
f
The supercial velocity proles computed according to equation (4.47) are
compared to measured data from literature in gure (4.7) below.
It is clear that there is quite a large dierence between the two dierent
sources for experimental velocity proles. The measurements inside the cata
lyst bed show a much more pronounced oscillation than the measurements just

138

4.6 Conclusions

3.5
3

v0/vm [-]

2.5

Rep = 10

Rep = 100

Rep = 1000

2
1.5.
1
0.5
0
0

x/dp [-]

Figure 4.7: The radial supercial velocity prole as computed from the modi
ed Brinkman equation (4.47) with the porosity, tortuosity and specic outer
particle surface area shown in gure (4.6) for dierent values of the particle
Reynolds number (drawn lines), compared to literature data (dots, see table
4.1).

downstream of the bed. The data of Bey and Eigenberger show no evidence
of a dependence of the ow prole below the bed on the Reynolds number.
The calculated velocity prole ts the measured proles quite well, especially
considering that the model is based on physical arguments and reasonable
assumptions only.

4.6

Conclusions

The Ergun equation (4.2) for ow resistance in packed beds can be almost en
tirely be derived from a parallel twisted ow channel model. The laminar term
can be derived from Poiseuille ow in a twisted channel where the twistedness
of the channel is characterized by the tortuosity (length of the channel per
length travelled). The result is identical to the Ergun equation when a tortu

139

4. Flow resistance in randomly packed beds

osity of 1.44 is chosen, a value that corresponds very well to the values that are
usually proposed in literature. The turbulent term of the Ergun equation can
be derived from a model that includes wall friction at the channel walls and
pressure loss in bends, expansions and contractions of the channels. If the wall
friction factor for the ow channels is taken as 0.1 (for very rough walls) and
it is assumed that there is one bend and one channel expansion for roughly
each particle diameter of channel length (which seems reasonable), the model
(equation 4.44) is nearly identical to the Ergun equation for standard packed
beds (tortuosity 1.44). Since the model is not based on the assumption of
identical, spherical particles, it also applies to non-homogeneous packed beds
of non-spherical particles, provided the porosity, tortuosity and specic outer
particle area are known.
In contrast with the classical (empirical) equations, the relation derived
here can be used locally in a packed to calculate the ow near a wall based on
the porosity, particle outer surface area and tortuosity proles.
The ow resistance model was included in a modied Brinkman equation
(4.47) model to calculate velocity proles in a cylindrical packed bed. The
proles for the porosity, tortuosity and specic outer particle area were taken
from the numerical bed packing experiments described in an earlier chapter.
The calculated velocity prole ts the measured proles quite well, especially
considering that the model is based on physical arguments and reasonable
assumptions only.

140

4.6 Conclusions

Nomenclature

Roman
Symbol
A1
A2
Ab
At
d
dh
dp
dp,eqA
dp,eqV
dt
f
f1
f2
H
ka
kt

units
m2
m2
m2
m2
m
m
m
m
m
m
k/m3 /s
k/m4
m
-

L
n

m
-

nb
ne

p
r

Pa
m

R
sv
sv
sv,b
v
v2
w0
w0
x
z

m
m2 /m3
m2 /m3
m2 /m3
m/s
m/s
m/s
m/s
m
m

Variable
channel cross section upstream of an expansion
channel cross section downstream of an expansion
cross section area
cross section area
characteristic dimension
hydraulic diameter
particle diameter
particle surface equivalent diameter
particle volume equivalent diameter
tube diameter
friction factor
ow resistance factor in the laminar ow regime
ow resistance factor in the turbulent ow regime
height of bed
ow resistance constant for expansion eects
constant in turbulent ow term pressure drop equa
tion
length of tube
number of hydraulic diameters between two bends or
two expansions
number of hydraulic diameters between bends
number of hydraulic diameters between channel ex
pansions
(absolute) pressure
space coordinate in radial direction for cylindrical
coordinates
radius of a sphere
specic outer particle surface
mean specic surface in a subvolume of the bed
specic outer particle surface in the bulk of the bed
velocity
velocity upstream of a sudden expansion
supercial velocity
mean supercial velocity in a subvolume of the bed
spatial coordinate (perpendicular to the wall)
spatial coordinate (in axial direction)

141

4. Flow resistance in randomly packed beds

Greek
Symbol


b
mb

units
m
-

f
e
f
f

Pas
Pas
m2 /s
k/m3
k/m/s2

Variable
bend angle
bend angle
distance between two bends
(local) porosity (volume open to ow / total volume)
bulk porosity, porosity far from a wall
mean bed porosity (volume in bed open to ow /
volume of bed)
dynamic viscosity of the uid
eective dynamic viscosity
kinematic viscosity of the uid
density of the uid
tortuosity of the bed (path length / distance)
wall shear stress

Dimensionless groups
Symbol
Re
Reh
Rep

142

denition
vd
f

vdh

f
vdp
f

Variable
Reynolds number (inertia forces / viscous forces)
Reynolds number based on hydraulic diameter
Reynolds number based on particle diameter

4.6 Conclusions

Literature
Bey, O., G. Eigenberger (1996), Bestimmung von Stromungsverteilung und
Warmetransportparametern in sch
uttungsgef
ullten rohren,
Chemie-Ing.-Techn. 68(10), 1294-1299
Bey, O., G. Eigenberger (1997), Fluid ow through catalyst lled tubes,
Chem. Eng. Sci. 52(8), 1365-1376
Bird, R.B., W.E. Stewart, E.N. Lightfoot (2002), Transport Phenomena,
second edition, John Wiley and Sons, New York
Blevins, R.D. (1984), Applied Fluid Dynamics Handbook, Van Nostrand
Reinhold company, New York
Calis, H.P.A. (1995), Development of dustproof, low pressure drop reactors
with structured catalyst packings : the bead string reactor and the zeolite
covered screen reactor, PhD dissertation, Delft University of Technology,
Delft, the Netherlands.
Cheng, P., D. Vortmeyer (1988), Transverse thermal dispersion and wall
channeling in a packed bed with forced convective ow, Chem. Eng. Sci.
43(9), 2523-2532
Cheng, Z.-M., W.-K. Yuan (1997), Estimating radial velocity of xed beds
with low tube-to-particle diameter ratios, A.I.Ch.E. J. 43(5), 1319-1324
Delmas, H., G.F. Froment (1988), A simulation model accounting for
structural radial nonuniformities in xed bed reactors, Chem. Eng. Sci. 43,
2281-2287
Ergun, S. (1952), Fluid ow through packed columns, Chem. Eng. Progress
48(2), 89-94
Eisfeld, B., K.Schnitzlein (2001), The inuence of conning walls on the
pressure drop in packed beds, Chem. Eng. Sci. 56, 4321-4329
Epstein, N. (1989), On tortuosity and the tortuosity factor in ow and
diusion through porous media, Chem. Eng. Sci. 44(3), 779-781
Fand, R.M., Thinakaran, R. (1990), The inuence of the wall on ow through
pipes packed with spheres, J. Fluid Eng. 112, 84-88
Giese, M., K. Rottschafer, D. Vortmeyer (1998), Measured and modelled
supercial ow proles in packeds beds with liquid ow, A.I.Ch.E. J. 44(2),
484-490
143

4. Flow resistance in randomly packed beds

Givler, R.C., S.A. Altobelli (1994), A determination of eective viscosity for

the Brinkman-Forchheimer ow model, J. Fluid Mech. 258, 355-370

Goodling, J.S., R.I. Vachon, W.S Stelpug, S.J. Ying (1983), Radial porosity
distribution in cylindrical beds packed with spheres, Powder Technology 35,
23-29
Jakobsen H.A., H. Lindborg, V. Handeland (2002), A numerical study of the
interactions between viscous ow, transport and kinetics in xed bed
reactors, Comput. Chem. Eng. 26(3), 333-357
Janssen, L.P.B.M., M.M.C.G. Warmoeskerken (1987), Transport phenomena
data companion, Delftse Uitgevers Maatschappij, Delft
Leijendeckers, P.P.H., J.B. Fortuin, F. van Herwijnen, H. Leegwater [red.]
(1998), Polytechnisch zakboekje, Koninklijke PBNA b.v., Arnhem
MacDonald, I.F., M.S. El Sayed, K. Mow, F.A.L. Dullien (1979), Flow
through porous media the Ergun equation revisited, Ind. Eng., Chem.
Fundam. 18(3), 199-208
McGreavy, C., E.A. Foumeny, K.H. Javed (1986), Characterization of
transport properties for xed bed in terms of local bed structure and ow
distribution, Chem. Eng. Sci. 41(4), 787-797
Papageorgiou, J.N., G.F. Froment (1995), Simulation models accounting for
radial voidage proles in xed bed reactors, Chem. Eng. Sci. 50(19),
3043-3056
Puncochar, M., J. Drahos (1993), The tortuosity concept in xed and
uidized bed, Chem. Eng. Sci. 48(11), 2173-2175
Roblee, L.H.S., R.M. Baird, J.W. Tierney (1958), Radial porosity variations
in packed beds, A.I.Ch.E. J. 4(4), 460-464.
Subagyo, S., N. Standisch, G.A. Brooks (1998), A new model for velocity
distribution of a single-phase uid owing in packed beds, Chem. Eng. Sci.
53(7), 1375-1385
Schwartz, C.E., J.M. Smith (1953), Flow distribution in packed beds, Ind.
Eng. Chem. 45(6), 1209-1218
Tierney, M., A. Nasr, G. Quarini (1998), The use of proprietary
computational uid dynamics codes for ows in annular packed beds, Sep.
Purif. Techn. 13, 97-107

144

4.6 Conclusions

Tsotsas, E., E.U. Schl

under (1988), Some remarks on channelling and on

radial dispersion in packed beds, Chem. Eng. Sci. 43 (5), 1200-1203

Vortmeyer, D., J. Schuster (1983), Evaluation of steady ow proles in
rectangular and circular packed beds by a variational method, Chem. Eng.
Sci. 18(10), 1691-1699
White, S.M., C.L. Tien (1987), Analysis of ow channeling near the wall in
packed beds, Warme-Sto
ubertrag. 21, 291-296
Winterberg, M., E. Tsotsas, A. Krischke, D. Vortmeyer (2000), A simple and
coherent set of coecients for modelling of heat and mass transport with and
without chemical reaction in tubes lled with spheres, Chem. Eng. Sci. 55,
967-979

145

Chapter 5
Dispersion in randomly packed
beds
Summary
The goal of this chapter is to nd a method to model dispersion in a packed
bed in such a way that it can be used in a pseudo-continuous Computational
Fluid Dynamics (CFD) model for packed beds.
Dispersion determines, together with convection, the rate at which heat
and mass are transported in the packed bed and therefore is one of the most
important factors in a packed bed model. Historically, dispersion has been used
to describe the deviation of a real packed bed from a, usually one-dimensional,
ideal model. As a consequence, in such models, the dispersion coecient is a
container for all non-idealities, so its value and meaning depend strongly on
the model that is chosen. More recently, two-dimensional models have been
developed, among which the so-called standard or wall heat transfer (WHT)
model and the wall heat conduction (WHC) model are the best known. The
WHT model denes a heat transfer coecient at the wall to quantify a tem
perature jump at the wall boundary. Although this model is used with some
success, the physical correctness of the temperature jump is doubtful. The
WHC model uses the wall temperature as a more natural boundary condition.
The porosity prole is used (together with some parameters that were tted
to experimental data) to describe the heat transfer resistance at the wall.
None of these literature models is completely satisfactory for use in a CFD
code, because data is used that may not be dened or available in a generalgeometry bed (e.g., the Peclet number at the centre of the bed). Also it
is questionable whether these models can be extrapolated to non-cylindrical
beds. Therefore, an alternative model is developed here.

147

5. Dispersion in randomly packed beds

It is clear that the tortuosity of the ow channels will play an important

role in the mixing behaviour. Our model uses not only the porosity prole but
also the tortuosity prole to estimate the dispersion coecients. As it is based
on plausible assumptions and bed packing simulation results for the tortuosity
proles, the model does not have parameters that are tted to experimental
data. Therefore, it is by nature more general than literature models like the
WHC, WHT and similar empirical models. The result for radial dispersion of
heat is:
 2
a,r 1
1
 (1 + aR ) + (r2 ) as
=
+
(5.1)
2
r Pef,h
10
Per,h
This equation consists of two terms, the rst one for the stagnant contribution
to conduction and the second for the convective (high ow) contribution. Her
aR is the stagnant radiation contribution, as is the solid conduction contri
bution (as dened by Zehner and Sch
under, 1972); r is the bed tortuosity
in radial direction and a,r is the radial component of axial tortuosity. All
parameters are taken locally in the bed. For axial dispersion and dispersion of
mass and momentum, similar equations are given.
For cylindrical bed geometries, the results of our model are, as they should
be, comparable to the best literature models. The real power of our model is
in the modelling of non-standard, 2 and 3-dimensional bed geometries. These
cases cannot be calculated by a simple nite dierence method as was done
for the cases in this chapter, but require a nite volume (CFD) code. This will
be described in the next chapter.

148

5.1 Introduction

5.1

Introduction

Randomly packed beds are used for many types of solid-uid contacting pro
cesses. Consequently, there has been a large interest in the modelling of the
important processes in randomly packed beds. Traditionally, models for packed
beds are focused on packed cylindrical tubes. The ow in the tube is described
by a model that treats the bed as a continuum. The simplest of these models
is the ideal tube model that assumes a at velocity prole (plug ow), perfect
radial mixing and no axial mixing. More sophisticated models include axial
and radial dispersion coecients to account for non-ideal behaviour.
Although the cylindrical packed tube is the most common geometry for
a packed bed reactor, other geometries may be more suitable for specic ap
plications. For low-pressure drop, shallow packed beds, the packed cylinder
is usually not the most ecient arrangement, and radial ow, lateral ow or
parallel passage reactors may be employed (Calis, 1995). Also for the develop
ment of compact installations, for instance mobile reforming installations for
fuel cell applications, non-standard geometries may be necessary for, e.g., ease
of construction, accessibility or heat integration.
The goal of this work is to build a model for packed beds that can be used
in computational uid dynamics (CFD) code to simulate packed bed reactors
of any geometry. In a CFD model, the ow domain is divided into a large
number of subvolumes (cells). In a subvolume of a packed bed, the transfer of
momentum, heat and mass is determined only by local forces, gradients and
bed structure. Therefore, a physically correct model of a packed bed should use
only local information (temperatures, concentrations, porosity) in the trans
port balance equations. However, most of the generally accepted models for
heat conduction and dispersion in packed beds use global parameters such as
the Reynolds number at the bed inlet, the Peclet number at the inlet or at
the axis of the bed and the bed-to-particle diameter ratio. The goal of this
chapter is to create a model to calculate dispersion of mass and heat in packed
beds where only local information is used to calculate the mixing behaviour
locally in the bed. Such a model is by nature more generic than models that
are based on averaged parameters.
In the previous chapters, the structure of a randomly packed bed, especially
near a solid wall, was described by proles of porosity, specic particle outer
surface area and tortuosity. An expression was derived for the ow resistance
of a packed bed based on local bed properties. In this chapter, the model will
be extended to include axial and radial dispersion of momentum, heat and
mass inside a packed bed.
The general theory of dispersion in packed beds will be given and a brief
overview of the extensive existing literature is presented and discussed. It is

149

5. Dispersion in randomly packed beds

found that, although the models give a reasonable description of the exper
imental data, none of the literature models is satisfactory for application in
a CFD model. Therefore, a new dispersion model is developed that takes
into account the proles of porosity and tortuosity near the wall. This model
will be based on a physical description of the processes in the bed, using the
bed structure from simulated packings and will not be tted to experimental
data. For cylindrical beds, the model should give results that are comparable
to the existing literature models. Therefore, the results will be compared to
state-of-the-art models and published experimental data for cylindrical packed
beds.
The real power of our model is in the modelling of non-standard, 2 and
3-dimensional bed geometries. These cases cannot be calculated by a simple
nite dierence method as will be done for the cases in this chapter, but require
a nite volume (CFD) code. This will be described in the next chapter.

5.2

Dispersion in packed beds

One of the simplest models of a packed bed is the ideal tube model, where
there is no axial mixing and are no radial gradients. The value of a given
intensive quantity (i.e., heat or the concentration of a tracer) is given at any
time and place by the dierential equation:

= w0
t
z

(5.2)

where w0 is the supercial velocity of the uid. In real packed beds, the
assumptions of the ideal tube model are often not valid. Dispersion models can
be seen as a way to describe deviations from the ideal tube model. The mixing
behaviour of a packed bed can be considered at three levels of sophistication
(see gure 5.1). At the coarsest level, the question is in what degree the packed
bed works as a static mixer: how will temperature or concentration proles be
evened out as the ow passes the bed The bed itself is treated as a black box.
The corresponding model is a 1-dimensional tube model with axial dispersion:

2
Da,,1D 2
= w0
t
z
z

(5.3)

As a result, a step in the value of at the inlet will give a sigmoid response
at the outlet of the tube. One of the most common experimental methods to
determine the axial dispersion coecient is to t the axial dispersion model to
a recorded response curve. This dispersion coecient is a lumped parameter

150

5.2 Dispersion in packed beds

that contains every non-ideality in the bed, including real axial mixing of the
ow but also inuence of the solid phase and vessel wall (adsorption, heat
conduction, radiation), radial mixing and ow maldistribution (wall eects).
As a result, the dispersion coecient depends not only on operating conditions
(the velocity, uid properties) and packing geometry (packing density, particle
shape) but also on vessel geometry (tube-to-particle diameter ratio).
For non-adiabatic tubes, the heat transferred from or to the uid is written
as:
h = w,1D (T Tw )

(5.4)

where h is the heat ux per unit tube surface area and w,1D is the wall heat
transfer coecient. The latter is a lumped parameter that compensates for all
radial heat transfer limitations since in the plug ow model, there are no radial
temperature gradients. The wall heat transfer coecient can be determined
by tting the axial dispersion model to steady-state measurements on a wall
cooled or heated packed bed. In these procedures, the axial heat dispersion
coecient and the wall heat transfer coecient are determined simultaneously;
however, since transient eects are not part of the system, the axial dispersion
coecient determined in this way may be dierent from that determined from
a temperature response curve. In addition, any errors caused by the relative
coarseness of the wall heat transfer model will also nd their way into the axial
dispersion coecient.
The one-dimensional model does not give information about the radial
prole of inside the bed. To calculate this prole, a quasi two-dimensional
model can be dened, where variable is calculated as a function of both axial
coordinate x and radial coordinate r, but the parameters (w0 , Da,,2D , Dr,,2D )
are constant:

2
2
Da,,2D 2 Dr,,2D 2
= w0
r
t
z
z

(5.5)

When we compare the 1-dimensional model with the two-dimensional model,

it is clear that the physical meaning (and value) of the dispersion coecients
Da,,1D and Da,,2D is not the same:
2
2
2
Da,,1D 2 = Da,,2D 2 + Dr,,2D 2
z
r
z

(5.6)

The boundary conditions for quantities that cannot penetrate the wall are:

151

5. Dispersion in randomly packed beds

II

III

Figure 5.1: Dispersion models with dierent levels of detail: 1-D tube model
with axial dispersion, quasi 2-D model with constant parameters and real 2-D
model with bed characteristics.

z=0
z=L
r=0
r=R

= 0

Da,,2D

w z

=0
z

=0
r

=0
r

(5.7)

The boundary condition for the temperature equation at the vessel wall is
somewhat dierent since there can be heat transport through the wall:
r=R

e,r

T = w,2D (T Tw )
r

(5.8)

The wall heat transfer coecient w,2D now does not correct for the existence
of radial temperature proles, since these are taken into account in the model.

152

5.2 Dispersion in packed beds

However, it is still needed to correct for radial proles of bed packing and
velocity that are caused by the reactor wall. The boundary condition at the
wall causes a jump in the temperature at the wall from the uid temperature
to the wall temperature; this temperature jump is not physical. Since the wall
eect also inuences ow in the core of the bed, the dispersion coecients are
still a function of the tube-to-particle diameter ratio.
To accurately describe the packed bed as a chemical reactor, the mixing
behaviour has to be known at each point in the bed, taking into account
porosity, particle surface area and tortuosity proles near the wall. In empty
tubes, the heat transfer resistance near the wall is associated with a stagnant
uid layer. In packed tubes, as we have seen in chapter 4, the heat transfer
resistance near the wall coincides with a high-velocity region. There will of
course be a stagnant uid layer at the wall, but the thickness of this layer
will be small compared to the particle diameter; the heat transfer resistance
caused by this layer will be in the same order as the particle-to-uid heat
transfer resistance. The inuence of this layer on the temperature prole is
small enough to disappear in the experimental error, which is considerable as
the temperature proles inside packed beds are by nature not easy to measure
(Wijngaarden and Westerterp, 1992). Therefore, the heat transfer resistance in
packed beds at a wall must be caused by a decrease in the dispersion coecient
near the wall, caused by the structure of the bed.
w0

Dr,,2D
=
Da,,2D
t
z
z
z
r
r

(5.9)

When the dispersion coecient is taken as a function of the distance from the
wall, the wall heat transfer coecient (and temperature jump near the wall) is
no longer needed to describe measured temperature proles, and the physically
correct boundary condition T = Tw can be used at the wall. A disadvantage of
this model is that there are many possibilities for the choice of the dispersion
coecient (and velocity) proles.
Dispersion is not only a way to model deviation from ideal behaviour, but
also a physical phenomenon. A random packed bed can be described as a
network of twisted ow channels with uneven cross-section, where the ow of
neighbouring channels is repeatedly mixed and split. Even though the bed
packing may be regarded as isotropic (if averaged over a large enough volume
and far from a wall), dispersion is not an isotropic phenomenon. Two directions
of dispersion can be distinguished: radial and axial dispersion, i.e., in the radial
and axial directions of a packed cylindrical tube. For general bed geometry, it
would be better to relate the dispersion to the direction of the net ow instead
of to the tube geometry, but since they are so generally used, these terms

153

5. Dispersion in randomly packed beds

are retained here. However, one should bear in mind that radial dispersion
is really dispersion normal to the direction of the ow and axial dispersion is
parallel to the ow. As a consequence, close to a wall, radial dispersion will
be perpendicular to the wall and axial dispersion parallel to the wall.
Close to a wall, the bed packing will not be isotropic: there will be a
dierence between the tortuosity of the uid channels parallel to the wall
and perpendicular to the wall. Therefore, there will be three dierent types of
dispersion near a wall: axial dispersion which is parallel to the ow and parallel
to the wall; radial dispersion perpendicular to the ow and perpendicular to
the wall and what we will call transversal dispersion parallel to the wall, but
perpendicular to the ow.
Axial dispersion (parallel to the ow) is caused by dierences in residence
time of dierent ow paths in the bed, i.e., by proles of the axial velocity.
Three dierent mechanisms for axial dispersion can be distinguished. The rst
mechanism is molecular transport (conduction and diusion) through the uid
and through the solid phase. The second mechanism operates on the particle
scale: the axial velocity prole in the channels will not be at due to the fact
that the velocity at the particle edge will be zero; since the channels are not
straight but twisted and of uneven cross-sectional area. In addition, due to
the random character of the bed packing, local dierences in packing density
on the scale of several particles can lead to preferential ow paths and thus to
local dierences in axial velocity. The third mechanism operates on the scale of
the bed: axial velocity proles on the bed scale are caused by the eect of the
reactor wall on the packing. Only the rst two mechanisms should be included
in a true 2-dimensional (or 3-dimensional) local dispersion model. The third
should be modelled as part of the geometry, not as part of the dispersion.
Radial dispersion (perpendicular to the wall and the direction of the ow)
is caused by exchange between dierent ow lines. This is caused by molec
ular transport (diusion, conduction) and by convective mixing due to the
interconnection of the ow channels. The bed structure inuences the radial
dispersion through porosity and the tortuosity of the uid channels. Since ra
dial dispersion takes place in channels perpendicular to the wall, the tortuosity
to be used diers from that used for axial dispersion.
Transversal dispersion (perpendicular to the ow but parallel to the wall)
is caused by the same mechanisms as radial dispersion: splitting and mixing
of ow channels. However, for transversal dispersion these ow channels will
be parallel to the wall, so the tortuosity of these channels is the same as that
used for axial dispersion. Far from a wall, where the bed structure is isotropic,
transversal and radial dispersion will be identical.

154

5.2 Dispersion in packed beds

At low ow rates, both axial and radial dispersion are important factors
for the transfer of mass and heat; conductive mechanisms prevail. As the ow
rate approaches zero, the dierence between axial and radial direction vanishes
and both will be equally important. In this conduction mode (Balakrishnan
and Pei, 1978b), mass is transported by diusion though the uid phase only;
diusion through the particle phase is either non-existent (in the case of solid
particles) or very small (in the case of porous catalyst particles). Heat, how
ever, can be transported by conduction through the stagnant uid, through
the solid phase and also by radiation, so that at low ow rates, the eective
heat conduction term diers from the eective mass dispersion term.
As the ow rate increases, the mixing will become more intense. The bal
ances of mass and heat become dominated by convective transport (caused by
the net ow) and convective mixing (random exchange of uid where two adja
cent channels are mixed and split). In axial direction, the convective transport
term will increase faster than the mixing term, so that the relative importance
of axial mixing will decrease. In radial direction the net ow is zero by de
nition, so the radial transport will be dominated by radial convective mixing.
In the high ow rate regime, heat and mass will be transported by the same
mechanism, so that heat conduction can be treated completely analogously to
mass dispersion.
For packed bed ow systems, the dispersion is usually given in terms of the
dimensionless particle Peclet number: the ratio of convective transport over
dispersive transport, where the particle diameter dp is taken as a characteristic
length:
Pei, =

w0 dp
Di,

(5.10)

Here w0 is the supercial velocity, Di, is the dispersion coecient of quan

tity and subscript i designates the axial (in the direction of w0 ) or radial
(perpendicular to w0 ) direction. In addition, a stagnant (conduction) Peclet
number can be dened:
Pe0i, =

w0 dp
D0

(5.11)

where D0 is the dispersion coecient at stagnation conditions (i.e., the eective

diusion coecient) and a turbulent Peclet number
Pe

i =

w0 dp
D
i

(5.12)

where D
i is the dispersion coecient in the limit of innite Reynolds numbers.

Note that this depends on the ow only so the subscript has been dropped.

155

5. Dispersion in randomly packed beds

Apart from these three Peclet numbers that depend on bed structure, the

molecular Peclet number can be dened:

Pef, =

w0 dp
Df,

(5.13)

where Df, is the molecular, unconned diusion coecient (or heat conduc
tion coecient) for quantity .

5.3
5.3.1

A short overview of existing literature

Experimental methods

From the early 1930s on, a large amount of research has been performed re
garding the transport of mass and heat in randomly packed beds. The main
purpose of the bulk of this research is to determine relationships for the dis
persion uxes in the bed. Thereto, models of dierent sophistication are de
veloped in which the dispersion coecients are adjustable parameters. These
are determined by tting the model to data from dierent experimental set
ups. Balakrishnan and Pei (1978a,b,c) and Tsotsas and Martin (1987) give
extensive reviews of earlier literature.
The model and the experimental method are usually closely linked. In one
type of experiment a signal (step, pulse or wave) is set on the inlet tracer
concentration or temperature and the response (mixing cup tracer concentra
tion or temperature) is recorded at the outlet of the bed. The axial dispersion
coecient is then calculated by tting the model results to this prole. More
detailed experiments inject a tracer at the centre of the bed and record the
concentration prole at dierent radial positions downstream of the injection
point (Foumeny et al., 1992), thus determining the axial and radial dispersion
coecients simultaneously. Votruba et.al. (1972) determined the axial disper
sion coecient alone, in the absence of radial temperature proles by heating
a packed bed at the downstream side with an infrared lamp and recording
the steady-state axial temperature prole. A more or less standard setup is a
packed bed with a heated or cooled wall, where the results include not only
dispersion in both directions, but also transfer of heat from the vessel wall.
Yagi and Kunii (1960) studied an annular bed between a cooled and a heated
tube, while Bey and Eigenberger (1996) studied a packed bed between two
at walls with each a dierent temperature. More common setups of this type
use a single cooled or heated wall (Olbrich and Potter, 1972a; Specchia et.al,
1980; Dixon, 1988; Tsotsas and Schl
under, 1990; Vortmeyer and Haidegger,
1991; Borkink and Westerterp, 1992; Freiwald and Paterson, 1992; Martin and
Nilles, 1993; Dixon and Van Dongeren, 1998). Equivalent methods for mass

156

5.3 A short overview of existing literature

transfer use for instance porous walls from which water evaporates (Hennecke
and Schl
under, 1973) or walls from which mercury sublimates (Olbrich and
Potter, 1972b). Tsotsas (1992) measured the average outlet concentration of
a compound that sublimates from the surface of the bed particles, thus com
bining mass dispersion and particle to uid mass transfer.
It is clear that the dierent experimental procedures that include transient
eects (heat capacity of particles, absorption of tracer on or in particles), wallto-uid transfer eects or particle-to-uid transfer eects, combined with the
models with dierent levels of sophistication lead to dierent correlations for
(and in fact, meaning of) the dispersion parameters. Experimental errors as
well as modelling errors nd their way in to the values of the tted parameters.
As an example, Foumeny et. al. (1992) tted the same model to transient con
centration injection responses and steady-state values. They obtained dierent
correlations for the dispersion coecients for the transient and steady cases,
indicating that some transient eects that were present in the experiments,
were not taken into account in the model and ended up in the value of the
dispersion coecients.

5.3.2

The standard model

Radial and (sometimes) axial dispersion are taken into account in a pseudohomogeneous two-dimensional plug ow model that is commonly called the
standard or wall heat transfer (WHT) model (equations 5.5, 5.7 and 5.8). This
model is most widely used in engineering practice and recommended in hand
books (e.g., Tsotsas, 1997). Here, the velocity is taken as constant while con
centrations and temperature have proles in axial and radial directions (e.g.,
Hennecke and Schl
under (1973), Specchia et.al., 1980; Dixon (1988), Freiwald
and Paterson (1992), Tsotsas (1992), Dixon and Van Dongeren (1998)). The
reduced transport of heat and mass near the wall is described by a wall heat
transfer coecient W , that is commonly calculated from a Nusselt relation,
e.g, Tsotsas (1997):
n
Nuw = Nuw,0 + aRem
0 Pr

(5.14)

For low ow conditions, the temperature drop near the wall will become less
steep; the prole will extend further into the bed and the assumption of a
temperature jump at the wall will become less accurate. Therefore, the wall
heat transfer models tend to become less accurate for lower Reynolds numbers
(Re0 < 100). The minimal Nusselt number Nuw,0 is used to improve the
predictions in the low-ow regime; it usually depends on the bed-over-particle

157

5. Dispersion in randomly packed beds

diameter ratio D/dp , e.g., Martin and Nilles (1993)


 0
5

N uw,0 = 1.3 +
D/dp f

(5.15)

As a renement of the standard model, a heterogeneous model is presented

by some authors, where the particle temperature is solved as a separate vari
able from the uid temperature (Dixon and Cresswell, 1979; Wijngaarden and
Westerterp, 1993; Westerterp e.a., 1993). However, it seems that this is only
necessary when there is reason to expect a considerable temperature dierence
between the solid and the uid, such as in the case of an exothermal chemical
reaction inside a porous catalyst particle. Since the reaction does not inu
ence the dispersion itself (Wijngaarden and Westerterp, 1989), this should not
alter the meaning or value of the dispersion coecients. Hence, the (quasi)
homogeneous model can be used to determine the dispersion coecients in the
absence of chemical reaction, and these dispersion coecients can later be used
to model packed beds with reaction. Due to the added complexity (chemical
reaction kinetics, particle-to-uid heat and mass transport), it does not seem
practical to employ a reactive system to determine the dispersion coecients.

5.3.3

The wall heat conduction model

The standard or WHT model was extended by taking into account the ra
dial porosity and velocity proles (Delmas and Froment, 1988; Vortmeyer and
Haidegger, 1991; Winterberg and Tsotsas, 2000; Winterberg et.al, 2000). In
this so-called wall heat conduction (WHC) model, the radial eective heat
conductivity varies with the distance to the wall, and the need for a wall heat
transfer coecient w and wall temperature jump is removed. The boundary
condition for the temperature at the wall becomes: T (r = R) = Tw . The ra
dial variation in radial eective heat conductivity is calculated by taking into
account the radial proles of porosity and velocity. Winterberg et.al (2000)
have compared the predictive performance of the WHC model with the wall
heat transfer model and conclude that the WHC mode is more accurate than
the WHT model. Therefore, it can be concluded that the WHC model is one
of the best currently available models for heat and mass transfer in packed
beds with uid ow. Our modelling eorts will be compared to the WHC
model according to Winterberg et.al (2000), which will be described here in
some more detail.
In the wall heat conduction model, the wall eect is taken into account
through the porosity prole. The assumed porosity prole is usually an ex
ponential expression, however the parameters chosen vary considerably. For

158

5.3 A short overview of existing literature

example, Winterberg et.al. (2000) choose an expression attributed to Giese

(1998):



Rr
(r) = b 1 + 1.36 exp 5.0
(5.16)
dp
while an earlier expression proposed by Vortmeyer and Schuster (1983) is also
widely used (e.g. Vortmeyer and Haidegger, 1991):



Rr
(r) = b 1 + 0.55 exp 1 2
dp



(5.17)
Rr
= b 1 + 1.495 exp 2.0

dp

Both these expressions are claimed to be valid for packed beds of particles
with small deviations from the spherical shape. The main dierence between
the expressions is the factor 5 respectively 2 inside the exponent. In equation
(5.16) the porosity decreases to within 1% of its bulk value at one particle
diameter from the wall, while in equation (5.17) this value is not reached until
2.5 particle diameters distance from the wall (5.2). Both equations have the
disadvantage that the predicted porosity at the wall depends on the value of
the bulk porosity, so strictly they are valid only for a given bulk porosity value
(for the latter: 0.4).
The radial velocity prole of the WHC model is calculated from the ex
tended Brinkman equation


p
f e
w0 (r)
2
r
(5.18)
= f1 w0 (r) f2 [w0 (r)] +
z
r r
r
where f1 and f2 are the laminar and turbulent factors from the Ergun equation
(see the previous chapter for details). The eective viscosity e is approxi
mated by:
e
= 2.0 exp(CRe0 )
f

(5.19)

where C = 2.0 103 . This equation is attributed to Giese et.al. (1998), but
there a value of C = 3.5103 is proposed. The Reynolds number Re0 is based
on the average supercial velocity in the bed. It seems somewhat inconsistent
that a constant value is used for the eective viscosity (dispersion of momen
tum) while considerable eort is put into the development of a relation for the
eective conductivity (dispersion of heat) that depends on the wall distance.
The mechanism for eective transport of momentum in a packed bed is very

159

5. Dispersion in randomly packed beds

1
Winterberg et.al.
0.8

Vortmeyer and Schuster

This work

[-]

0.6

0.4

0.2

2
x/dp [-]

Figure 5.2: Porosity prole used in the WHC model by Winterberg et.al.
(2000), equation (5.16), compared to that proposed earlier by Vortmeyer and
Schuster (1983), equation (5.17) and a typical porosity prole determined from
a bed packing simulation in this work.

similar to that for eective transport of heat and it also depends on local bed
structure.
The temperature proles are calculated from a 2-dimensional quasi-homo
geneous heat transport model. In steady-state and without chemical reaction,
this reads:


1
T
2T
T
(5.20)
+ e,a (r) 2 w0 rf cp,f
0=
e,r (r)r
z
r
z
r r
The eective radial heat conduction e,r is calculated from
e,r (r) = 0 (r) + K1,h Pef,c f (R r)f

(5.21)

where 0 and the function f depend on the distance from the wall. The
stagnant eective heat conductivity is calculated according to Zehner and
Schl
under (1970, 1972) with the radial porosity prole given by equation (5.16).
For the convective mixing term of the heat conductivity, a parabolic prole near

160

5.3 A short overview of existing literature

the wall is proposed:


2
Rr
f (R r) =
K2,h dp

0<

Rr
< K2,h
dp

(5.22)

At the wall, the heat conductivity is equal to the conductivity of the uid; it
increases quadratically until a distance K2,h . At a distance further than K2,h
particle diameters from the wall, f is taken as 1.0, so the bulk value for the
eective heat conduction is used. Hence, there is a sharp transition between
the wall zone and the bulk zone. An empirical correlation is proposed for the
parameter K2,h :


Re0
K2,h = 0.44 + 4 exp
(5.23)
70
so that the distance over which the wall inuences the heat conductivity de
pends on the Reynolds number, from 0.44 particle diameters at high Reynolds
numbers to 4.44 particle diameters at low Reynolds numbers. Later (Winter
berg and Tsotsas, 2000), a dependence on the Peclet number instead of the
Reynolds number was proposed, in order to describe liquid/solid packed bed
systems in addition to gas/solid systems:


Pef
(5.24)
K2,h = 0.44 + 4 exp
50
A similar model is used for mass transfer in packed beds. However, for mass
transfer, the value of K2,m is the constant value 0.44. In other words, for
mass transfer the thickness of the wall zone with decreased convective mixing
is always 0.44 particle diameters, while for heat transfer it depends on the
Reynolds number and can be between 0.44 and 10 times that value. The
physical background of this is not clear, because at high Reynolds numbers,
the mechanism for mass transfer and heat transfer is the same. It appears that
there must be a physical phenomenon related with heat conduction through
the solid or radiation that is not taken into account correctly in this model.
The error in the heat conduction term is then corrected for in the convective
mixing term.

5.3.4

Other models

Although the majority of literature follows the pseudo-homogeneous approach,

there are some other approaches that are worth mentioning. K
ufner and Hofman (1990) developed a packed bed model based on a mixing cell model with
a porosity prole near the wall. Logtenberg and Dixon (1997) made a model

161

5. Dispersion in randomly packed beds

of a small bed of eight spheres in a commercial CFD code and determined the
velocity and temperature elds. The wall heat transfer coecient (or NuW )
was calculated by tting the CFD results to a formulation of the WHT model.
The number of spheres in the CFD model and the bed-to-particle diameter
(D/dp = 2.86) are too low to allow extrapolation of their results to a general
randomly packed bed. Furthermore, it seems that they use a relatively low
number of computational cells in the CFD model, and they did not investigate
the eect of mesh renement on their results. However, they nd some inter
esting phenomena. In the ow eld, they observe the development of eddies
between the spheres at higher Reynolds numbers (Re > 39). Also, they nd
that the qualitative features of the ow eld do not change when laminar ow
equations are used instead of turbulence models.

5.4

Dispersion modelling

Although a number of dispersion models can been found in literature, none of

these is completely satisfactory for use in a CFD code. None of the models
adheres to the rule of locality: that the ow parameters should depend on local
parameter and variable values only and not on far removed or average data
(like the average bed porosity or the Peclet number at the centre of the bed).
Furthermore, all models use parameters that are tted to experimental data;
as the experimental data is almost exclusively for tubular geometry, this limits
the general applicability of the modes. For these reasons, a new dispersion
model will be developed here. An attempt is made to derive this model as
much as possible from physical mechanisms based on local bed structure.
Despite the dierences in models and experimental setups, there is a rea
sonable agreement about the form of the correlation for the axial and radial
Peclet number in unconned packed beds (Tsotsas and Schl
under, 1988). The
basic concept is that the stagnant contribution and the convective mixing con
tribution can be added together to get the Peclet number at intermediate
conditions:
1
1
1
= 0 +
Pei,
Pe Pei

(5.25)

for direction i (radial or axial) and quantity (e.g., temperature or a con

centration). Therefore, the dispersion at intermediate ow conditions can be
calculated if the dispersion at stagnant conditions and high-ow conditions are
known. Equation (5.25) will be the basis for the dispersion model developed
here.

162

5.4 Dispersion modelling

5.4.1

Correlations for Pe at stagnant conditions

At stagnant conditions, a dierence must be made between the dispersion

(conduction) of heat and the dispersion (diusion) of mass. It is assumed
that the contribution of the solid phase in mass transport is negligible. For
non-porous particles this is a natural assumption; for porous catalyst particles,
mass may diuse through the particles as well as through the uid in the voids
between the particles. However, since the eective diusion coecient in a
typical porous catalyst particle is at least an order of scale lower than that in
an unconned gas space, the contribution of the solid phase will not be taken
into account for mass transport on the bed scale. On the other hand, in most
cases the heat conduction through the particles will be of the same order of
scale as the heat conduction through the uid, so the solid phase needs to be
taken into account in the heat transport term..
Dispersion of mass at stagnant conditions is caused by diusion. The eec
tive diusion coecient of a component in the stagnant uid in a packed bed
(D0 ) diers from the (molecular) diusion coecient in an unconned stag
nant uid (Df, ) because mass cannot pass through the particles. It depends
on bed porosity and the tortuosity (mean path length through the uid per
net distance) of the channels in the bed packing. If we look at the bed as a
bundle of parallel channels, we can dene three dierent diusion uxes. The
diusion ux inside the channels can be given by:
p = Df

d
d

(5.26)

where  is the length coordinate along the channel and Df the molecular un
conned diusion coecient (assuming the diameter of the channels is much
larger than the mean free path length of the gas molecules). In order to trans
port a molecule a distance x the distance travelled through the channel is
 = x. Hence, the interstitial diusion ux inside the bed is given by:
i =

1
Df d
p =

d

(5.27)

The supercial diusion ux is based on the empty tube area instead of the
cross-sectional area of the pores, hence it diers a factor . Therefore, with
d = dx,
s = i =

Df  d
b2 dx

(5.28)

If we now dene
0
s De

d
dx

(5.29)

163

5. Dispersion in randomly packed beds

it follows that the eective diusion coecient in a stagnant packed bed is

related to the molecular diusion coecient as:

0
=
De

Df
2

(5.30)

The bed tortuosity appears in the denominator to the second power, and not
to the rst power as is sometimes suggested (Foumeny et.al., 1993; Berger
et.al., 2002, see also Epstein, 1989).
It should be noted that the tortuosity used here is the tortuosity of the
channels of the bed (unrelated to the ow), in the direction of the transport.
Therefore, even though there is no dierence in mechanism between axial and
radial dispersion under stagnant conditions, there is a dierence between dis
persion perpendicular and parallel to the wall because the bed packing is not
isotropic near the wall. This radial tortuosity, perpendicular to the wall, is
quite dierent from the axial tortuosity, parallel to the wall. The radial tor
tuosity will be one up to a distance of at least rp , since the uid
is always in
direct view of the wall in the wall zone. In the zone from rp to 3rp , a decreas
ing fraction of the uid is in direct view of the wall, while further away, the
tortuosity will be approximately equal to the bulk porosity, see gure 5.3. The
axial tortuosity prole will be that obtained from particle tracking simulations
in chapter 3.
For the radial tortuosity, we will assume the following radial prole:
x < rp
rp < x < 2rp
x > 2rp

r = 1
r = 1 + (x/rp 1)(r,b 1)
r = r,b

(5.31)

The value of the tortuosity far from the wall (b ) depends on the bulk
porosity of the bed: a lower porosity will (for random packings of sphere-like
particles) give a higher tortuosity. Zehner and Schl
under (1970), based on
experimental data for the eective diusion coecients for several types of
packing and porosities, propose the following correlation
0

De
= 1 1 b
Df

(5.32)

This correlation is very widely used (Tstotsas and Martin, 1987; Bauer, 1988;
Vortmeyer and Haideggeer, 1991; Winterberg et.al., 2000; Jakobsen et.al.,
2002). The corresponding tortuosity according to equation (5.30) can be found:


b = 1 + 1 b

164

(5.33)

5.4 Dispersion modelling

1
1

x/rp

Figure 5.3: Sketch of the radial tortuosity near the wall. Nearly all of the
particles in the rst layer touch the wall (Legawiec and Ziolkowski, 1994), so
that the radial tortuosity
is exactly one up to one particle radius from the
wall. Between 1 and 3 to 2 radii from the wall, some of the uid is in the
channel perpendicular to the wall (dashed area). Beyond this area the radial
tortuosity is assumed equal to the bulk value.

For standard packed beds ( 0.4), this equation gives values for the tortuosity
that are somewhat
lower than the expected value of about 1.4 (it reaches a

maximum of 2 as b 0). This is even more apparent when it is compared

with other models for the bed tortuosity, e.g. according to Puncochar and
Drahos (1993), see gure (5.4):
1
b =
b

(5.34)

It is important to realise that these correlations (equations 5.33 and 5.34)

give values for the bulk or average tortuosity of packings with dierent porosi
ties, which is not necessarily the same as the tortuosity in a given packing as
a function of the distance to the wall. Winterberg et.al. (2000) propose to use

165

5. Dispersion in randomly packed beds

[-]

1.5

1
0.2

0.4

0.6

0.8

1.0

[-]

Figure 5.4: Comparison of the bed tortuosity b as given by Puncochar and

Drahos (1993) and as implied by the model of Zehner (1970) as a function of
the bed porosity .

the local porosity in equation (5.32), which implies that the (ordered) packing
with high porosity near the wall in a dense bed is equivalent to the (random)
packing far from the wall in a less dense bed; it is obvious that this is not
true. It is clear that in the shell up to one particle radius from the wall, the
eective diusion coecient should be proportional to the porosity only, not
proportional to the root of the porosity.
As equation (5.32) describes the measured data quite well, the model se
lected here should tend towards this equation far from the wall. However, near
the wall, the eective diusion coecient changes according to equation (5.30)
with the radial tortuosity prole given by (5.31). Hence,
Pe0m (x) =

w0 dp
r (x)2
w0 dp r (x)2
=
Pe
=
f,m
0
De
Df (x)
(x)

(5.35)

For the conduction of heat, the situation is more complex than that for
the diusion of mass. Several mechanisms contribute to the transport of heat:

166

5.4 Dispersion modelling

conduction through the uid, conduction from the uid to a particle, conduc
tion inside a particle and between two particles (although this mechanism is
disputed since the contact between two particles is a point unless the particles
can be deformed appreciably; Wijngaarden and Westerterp, 1989) and radia
tion between particles and between uid patches. However, for non-conducting
solids and in the absence of radiation, the heat conduction should be equivalent
to the mass diusion in equation (5.35).
Yagi and Kunii (1957) proposed the following relation (if we disregard
radiation eects):
0e
(1 )
=

f
+ fs

(5.36)

where f is the heat conductivity of the uid, s that of the solid. The param
eter is the dimensionless conduction length between two particle centres,
which is usually taken as 1.0. This relationship does not take into account
conduction of heat through the uid phase; if the solid phase conductivity
becomes zero, the overall heat transfer coecient will also approach zero.
Specchia et al. (1980) elaborate on a similar correlation by Kunii and Smith
(1960)
0e
(1 )
=+

f
+ fs

(5.37)

Here and are dimensionless parameters describing the packing. Specchia

et.al. suggest a value of 2/3 for and 0.222 for . Note that in the limit
of non-conducting solid (which is equivalent to the mass transfer situation),
comparison of equation (5.37) with equation (5.30) leads to the conclusion that
equals 1, i.e., it does not take into account the tortuosity of the path along
which heat is transported in this situation.
Zehner and Schl
under (1970) base their correlation on a two-dimensional
cylindrical cell model where the heat transport takes place along parallel lines.
This model is widely accepted in literature (e.g., Bauer and Schl
under, 1978;
Tsotsas, 1997). In this model, the shape of the particles is modied in the
model to account for the fact that in reality the heat ow lines are not parallel.
The shape of the particles is described by the relation
r2 +

z2
=1
(B (B 1)z)2

(5.38)

Here B is an adjustable shape parameter; for B = 1, the shape is a sphere (see

gure 5.5). For random packings, the parameter B depends on the porosity

167

5. Dispersion in randomly packed beds

and can be estimated as:


10/9
1
B = 1.25


(5.39)

Particle 2

B=0.01
B=0.1
B=1.0
B=2.0

Particle 1
dp

Figure 5.5: Shape of the contact area between two particles as a function of
shape parameter B according to equation 5.38.
Part of the heat conduction goes through the uid phase (with conductivity
f ) only; the rest goes through the solid phase at some stage. The relative size
of these parts must be such that if the solid conductivity is zero (eectively
blocking any heat ow lines that pass through the solid), the result must be
the same as for mass diusion (equation 5.35):
0e


= 2 (1 + aR ) + (1 2 )as
f
r
r

(5.40)

Where as is the eective heat conductivity factor of ow lines that pass through
the solid and aR the radiation contribution. Note that this formulation diers
from that of Zehner and Schl
under (1970):

0e
= (1 1 ) (1 + aR ) + ( 1 )as
f

168

(5.41)

5.4 Dispersion modelling

Of course, the value is the same for bulk conditions provided equation (5.33)
is used to estimate the radial tortuosity r . Winterberg et.al. (2000) use
equation (5.41) with the local value of the porosity to estimate the eective
heat conductivity near the wall. For non-conducting solids and in the absence
of radiation, the heat conductivity should be equivalent to the eective mass
diusion. Therefore, in the shell up to one particle radius from the wall, the
eective heat conductivity should dier from the molecular heat conductivity
by a factor  only. It is clear from gure (5.6) that our equation (5.40) with
radial tortuosity prole (5.31) shows this behaviour while equation (5.41) does
not.
The eective heat conduction through the solid is calculated by integration
over the particle cross section normal to the direction of the conduction. The
result is (Zehner and Schl
under, 1970):


2 B(as 1)  as  B 1 B + 1

as =
ln

+
(5.42)
N
as N 2
B
N
2
where
N = 1 B/as

as =

s
f

(5.43)

For higher temperatures (usually above approximately 200 C), radiation

becomes a factor in the transfer of heat. Since many heterogeneous chemical
reactors operate at temperatures well above this value, radiation of heat needs
to be taken into account. This is done by adding an additional (parallel) heat
transfer term through the uid phase (Zehner and Schl
under, 1973):
R =

4T 3 dp
2/R 1

(5.44)

Here is the Stephan-Boltzmann constant (56.7051 109 W/m2 /K4 ) and R

the emissivity of the particles (a value between 0.5 and 1.0 for many materials).
With this additional term, equation (5.40) becomes
0e


= ( 2 ) (1 + aR ) + (1 2
)as

r
r
f
with
as




2 B(as + aR 1)
B1
as + aR
=
ln

2
B
N
N
as N

B+1

(aR B)
2B

(5.45)

(5.46)

169

5. Dispersion in randomly packed beds

where
N = 1 (B aR )/as

aR =

R
f

(5.47)

Therefore, the stagnant Peclet number for heat can be written as:
 2

r
w0 dp
r2
0
Peh = 0
= Pef,h
(1 + aR ) + (1 )as


e /(cp )

(5.48)

1
Winterberg et.al.

kbed [-], [-]

0.8

This work

Porosity

Porosity

Conductivity

Conductivity

0.6

0.4

0.2

0
0

x/rp [-]

Figure 5.6: Stagnant bed conductivity factor kbed = e,bed /f in radial direc
tion for a typical packed bed (D/dp = 8.65) without radiation and solid heat
conductivity according to the literature model and according to this work .
For the latter, the conductivity equals the porosity for x/rp < 1, which is the
correct behaviour. Far from the wall, both approaches give similar results.

5.4.2

Correlations for Pe at high Reynolds numbers

At high values of the Reynolds number (or, more accurately, at high values
of the stagnant Peclet number Pe0 ), dispersion of mass and heat in a packed
bed is dominated by convective mixing. Therefore, there is no dierence in the

170

5.4 Dispersion modelling

treatment of heat and mass. However, there is a dierence between dispersion

in the axial and radial direction.
The value of the Peclet number in the radial direction at high Reynolds
numbers can be derived theoretically from a mixing cell model (Schl
under,
1966). In two dimensions, this model is quite straightforward (gure 5.7).
Two types of mixing cells can be distinguished: axial nodes above each particle
(points) and radial nodes between two particles (open circles).
x
2'
2
1'
1

j=0

flow
i= 1

1'

2'

3'

Figure 5.7: Rectangular grid mixing cell model of

a packed bed in two dimensions, after Schl
under
(1966).
The theory is based on the stationary convection-diusion equation:
w0

2
= Dr, 2
z
x

(5.49)

In terms of the discrete mixing points:

w0 x2
(i,j+1 i,j ) = (i1,j i,j ) (i,j i+1,j )
Dz

(5.50)

If the ow is equally distributed over the mixing points, it follows that

i,j+1 =

i1,j + 2i,j + i+1,j

4

(5.51)

Hence,
i,j+1 i,j =

i1,j 2i,j + i+1,j

4

(5.52)

171

5. Dispersion in randomly packed beds

or, with (5.50):
w0 x2
=4
Dz

(5.53)

If the assumption is made that x = z = dp , then it follows that for this

2-dimensional geometry:
Pe
r,2D = 4

(5.54)

For a rectangular 3-dimensional network, Schl

under assumes that at the
radial nodes (between two particles) two ows join while at the axial nodes
(nodes above each particle) four ows from the closest radial nodes join. Ac
cordingly, with x = y = z = dp it is found that:
Pe
r,3D = 8

(5.55)

As, far from the wall, each particle in a packed bed of spheres rests on three
others, a rectangular network as used by Schl
under does not seem the to be
the best representation of a real random packed bed. In fact, the bed could
equally well or better be represented by a hexagonal close packed structure.
For such a structure (gure 5.8), ow from three sources is mixed at each node
that is located above each particle (see gure 5.9).

Figure 5.8: Sphere packing corresponding to the

mixing cell structure in gure 5.9.

172

5.4 Dispersion modelling

B
A

C
d

c
b

Figure 5.9: Mixing cell structure for a hexagonal

close packed structure. The top cell P depends on
three lower mixing cells A,B,C that in turn depend
on the seven cells at the bottom (numbered a-f in
a hexagonal arrangement with p at the centre).

It is clear that the concentration i at a node i is the mean value of the

concentrations in the three nodes below it. Hence, we can write

P p
1
=
= (a + b + c + d + e + f 6p )
z
3d0
9 3dp

(5.56)

In order to avoid complications, we assume that the concentration gradient is

parallel to the line e-b in gure 5.9. Hence,
1
a = c (b + p )
2

f = d (e + p )
2
and

1
= (2b + 2e 4p )
z
9 3dp

(5.57)

The second derivative of concentration expressed in the node concentrations

is:


2
1 b P
P e
=

(5.58)
dp
dp
dp
x2

173

5. Dispersion in randomly packed beds

and therefore

Pe
r,3D = 6 3 ( 10.4)

(5.59)

If, however, we assume that the concentration gradient is perpendicular to the

line e-b in gure (5.9), the result is
Pe
r,3D = 9

(5.60)

For concentration gradients that do not happen to be parallel or perpendic

ular to one of the lines (d-a),(e-b) or (f-c), the corresponding Peclet number
will be in between these values. For randomly packed beds, this number will
dier from these values, but will probably be somewhat higher than the value
suggested by Schl
under. In literature models, values between 7 (Bauer, 1988)
and 11 (Gunn, 1987) are used. The values found experimentally are in this
same range: Borkink and Westerterp (1992) nd values between 8.8 and 10.9
for spheres and lower values for cylinders and Raschig rings (7.6 and 4.2 re
spectively); Dixon (1988) nds values between 7 and 10 for spheres (between
4 and 8 for cylinders); Foumeny et al. (1992) found a value of 12.5. A value
of 10 seems to be an acceptable average for random packings of spheres (e.g.,
Dixon and Cresswell, 1979; Delmas and Froment, 1988).
The derivation above is based on the assumption that the packing is regular.
Although it is shown that there is some dependence on the structure of the
bed, it can be accepted that it is approximately correct for a random packing.
Near the wall the bed is not random, and therefore the mixing behaviour will
be dierent. The change in bed structure will have an eect on the distribution
of mixing points in space and on the ow rates between nodes.
The assumption in e.g., equation (5.54) that x = z implies that the gas
paths make an angle
of 45 with the axis of the bed, or in other words, that
the tortuosity equals 2. For other values of the tortuosity, we can write (see
gure 5.10):
x 2
= 1
z

(5.61)

Then if we take x = dp , (5.54) becomes

Pe
r,2D =

4
2

(5.62)

For the three-dimensional case we need to take into account the radial
and tangential dispersion near the wall. We dene a radial and tangential

174

5.4 Dispersion modelling

flow

Figure 5.10: Graphical representation of the rela

tion between x, z and the tortuosity in a rect
angular mixing cell model of a packed bed in two
dimensions

component for the tortuosity of the axial ow paths:


la2 + lr2
a,r
l
 a
la2 + lt2
a,t
la

(5.63)
(5.64)

where la , lr and lt are the axial, radial and tangential components of a section
of a ow path. Of course, the three-dimensional tortuosity is dened as

la2 + lr2 + lt2
(5.65)
a
la
The structural correction for the bed tortuosity prole near the wall be
comes:
Pe
r,3D = 

10
2
a,r

(5.66)

for the radial dispersion. The tortuosity that needs to be used in this equation
is the radial tortuosity component of the axial ow paths, where only ow
variations in radial direction (i.e., perpendicular to the wall) are taken into

175

5. Dispersion in randomly packed beds

account. Near the wall, the ow will be parallel to the wall, which means
that the tortuosity has a value of 1. This will be the case at distances up
to one third particle radius from the wall. Fluid in this region will ow in a
zigzag manner around the particles, but the zigzags will be primarily in the
tangential plane at constant wall distance. In contrast, at one particle radius
from the wall, the uid will ow almost exclusively in radial direction, so the
radial component of the axial tortuosity will be approximately equal to the 3D
tortuosity. Therefore, we will assume that the a,r will be 1.0 up to a distance
of rp /3 from the wall, and then increase to the maximum value of the 3D axial
tortuosity at one particle radius from the wall (see gure 5.11).
Since it is quite exceptional for strong temperature and concentration gra
dients to occur parallel to a wall, the eect of the tangential dispersion will
usually be negligible. Therefore, we will, use the same value for the radial and
tangential tortuosity.
The model is quite sensitive to the shape of the tortuosity prole used to
calculate the radial dispersion term, especially in the rst particle radius from
the wall. The curve used is an estimate of the actual prole. Since near the
wall the bed is quite ordered, it would not be very dicult to build a CFD
model for the rst layer of particles near the wall and calculate the ow and
(through a tracer particle simulation) the actual tortuosity prole at high ow
rates. However, this is outside the scope of this work.
For axial dispersion at high ow rates, there is not as much data available
because axial dispersion in unconned packed beds becomes a less important
transport mechanism at higher ow rates. The value used in most previous
investigations (e.g. Dixon,1988, Jakobsen et.al.,2002) is
Pe
a = 2

(5.67)

Foumeny et.al. (1992) found values of about 1.9 in mass transfer experiments;
Gunn (1988) found a value of 2.0 for spheres and values between 1 and 2 for
cylinders, based on older experimental data.

5.4.3

Dispersion of momentum

The ow pattern inside the packed bed not only changes the transport of heat
and mass, but also of momentum. Again, we can make a dierence between
axial and radial dispersion. Axial dispersion of momentum will lead to a force
exerted by high-velocity uid on low-velocity uid in the axial direction on
a streamline. This force will be negligible compared to the force exerted on
the uid due to the friction in the bed. Therefore, only radial dispersion of
momentum will be taken into account. At very low velocities, the transfer takes

176

5.4 Dispersion modelling

2
r [-]

[-]

1.8

a [-]

1.6

a,r [-]

[-], [-]

1.4
1.2
1
0.8
0.6
0.4
0.2
0

x/rp [-]

Figure 5.11: Proles of the three dierent tortuosities as a function of the

distance from the wall; the 3-dimensional axial tortuosity a that is used in
ow resistance calculations, the radial component of axial tortuosity a,r that
is used in radial convective mixing calculations and the radial tortuosity r
that is used in stagnant radial dispersion calculations

place on a molecular scale. Momentum is transferred when a high velocity

stream and a low velocity stream share a single channel, by viscous eects. In
this regime, transfer of momentum between two neighbouring streams may be
less intense than in an open space since they are separated by solid particles
for a signicant fraction of the way. In the high-velocity regime, we can look
at the packed bed as a number of mixing volumes with connecting channels.
The balance of momentum must then be satised for each mixing volume.
Therefore, if a high velocity stream enters from one channel and a low velocity
stream from another, the exit streams will have an intermediate velocity. In
eect, the stream on the low-velocity side will be accelerated at the cost of
the high-velocity side momentum. Therefore, high velocity gradients (on a
macroscopic scale) inside the bed will be suppressed.
Not much literature is available on the subject of dispersion of momentum
in packed beds. The mechanisms for transfer of momentum are very similar to

177

5. Dispersion in randomly packed beds

the mechanisms for the transport of mass, so an analogous model is proposed

here. Increased transport of momentum in turbulent ow is often represented
as an increase in eective viscosity. We will here introduce an equation for
the eective viscosity in radial direction instead of the equivalent of a Peclet
number for momentum (which would be an eective Reynolds number).

5.4.4

Dispersion model

The overall dispersion model is formed by adding together the contributions

for stagnant conditions and at high Reynolds numbers (equation 5.25). For
dispersion of mass:
1

1
= 2
+
r Pef, 2
Pea,

1

= 2
+
r Pef,
Per,

(5.68)

2 1
a,r
10

(5.69)

For dispersion of heat:

1
 (1 + aR ) + (r2 ) as 1
=
+
r2 Pef,h
Pea,h
2
1
 (1 + aR ) + (r2 ) as
=
+
Per,h
r2 Pef,h
For radial dispersion of momentum:
 2
Rep a,r
1

e,r
= 2+
r
10
m

 2
a,r 1
10

(5.70)

(5.71)

(5.72)

In these equations, Re and Pe are based on the local supercial velocity and
properties.

5.5

Results

In order to compare the dispersion model presented here with literature data
and models, a 2-dimensional heat transfer model is presented for ow and heat
conduction in a wall-heated packed tube. The porosity, tortuosity and specic
surface proles are given in chapter 3. The velocity prole inside the packed

178

5.5 Results

bed is calculated using the relations given in chapter 4, with the eective
viscosity according to equation (5.72). The transfer of heat is calculated from
the partial dierential equation 5.20:


1
T
2T
T
+ e,a (r) 2 w0 rf cp,f
0 =
e,r (r)r
r r
r
z
z
with boundary conditions
r=0

T
=0
r

z = 0 T (r) = T0

r = R T = Tw
z=1

T
=0
z

The eective radial and axial heat heat conductivity are calculated as a func
tion of the wall distance using the equations (5.68)-(5.71). Physical properties
of air are calculated as a function of the local temperature. Figure (5.12) gives
the results of this model, compared with the results of the model of Winterberg et.al. (2000) and measured temperature prole points given by the same
authors. The parameters for the calculations are summarised in table 5.1.
Table 5.1: Values of the parameters used for gure 5.12
Quantity
symbol value
D
Bed diameter
0.075 m
L
Bed length
0.2 m
dp
Particle diameter
9.5 mm
Bed-to-particle diameter ratio D/dp
8
Bed bulk porosity
0.38
b
Bed bulk tortuosity
1.41
b
Axial wall tortuosity
1.1
w
Particle material
ceramic
Particle conductivity
s
1.2 W/m/K
Particle emissivity
0.85
R
Fluid
air
M
Molar mass
0.028 kg/mol
p
Pressure
105 Pa
Re0
Reynolds number at inlet
480
Figure 5.13 compares the results of the model with the wall heat conduction
model for dierent ow conditions. Note that the results of both models are
quite similar. In Winterberg e.a. (2000) temperature proles for these cases are
also shown and compared to measured values after Dixon (1988). The curves

179

5. Dispersion in randomly packed beds

100
whc model
this work
experimental

90

T [C]

80
70
60
50
40
0

x/rp [-]

Figure 5.12: Temperature prole for a steam-heated tube lled with ceramic
particles; comparison of model results of this work, results from the wall heat
conduction (WHC) model according to Winterberg et.al. (2000) and measured
data after Dixon (1988), as presented by Winterberg e.a.; see table 5.1 for the
parameter values used.

appear to t the measured values quite well. However, with our models (both
the detailed prole model and the WHC based model) we cannot reproduce
the curves presented by Winterberg et.al. without signicantly changing the
parameters (e.g., the conductivity of the ceramic material) compared to the
case presented in gure 5.12.
In the original article by Dixon, the raw data for the temperature proles
is not given. Instead, the author presents derived parameters, like the radial
Peclet number for dierent packings of nylon, ceramic and steel particles and
for dierent D/d ratios. The packings with 6.3 and 9.5 mm particles give simi
lar eective radial Peclet numbers for each material. However, for the packing
with 12.7 mm particles which corresponds to the cases in gure 5.13, anoma
lous behaviour is found. The ceramic packing gives a higher conductivity than
expected, which even roughly coincides with the steel particle packing, whereas
the nylon particle packing gives a much lower conductivity than expected. The
steel packing seems to be in line with the other measurements.
As the radial thermal conductivity is strongly dependent on the solid con

180

5.6 Conclusions

ductivity (at least for Re < 1000), it is hard to understand how a ceramic
packing (s = 1.2 W/m/K) could have the same eective radial conductivity
as an identical steel packing (s = 50 W/m/K). Therefore, we must conclude
that the data for the 12.7 mm ceramic spheres contains some unknown eect
and therefore cannot be used to validate the dispersion model.
Table 5.2: Values of the parameters used for gure 5.13
Quantity
symbol value
D
Bed diameter
75 mm
L
Bed length
200 mm
d
Particle diameter
12.7 mm
p
Bed-to-particle diameter ratio D/dp
5.9
Bed bulk porosity
0.38
b
Bed bulk tortuosity
1.41
b
Axial wall tortuosity
1.1
w
Particle material
ceramic
Particle conductivity
s
1.2 W/m/K
Particle emissivity
0.85
R
Fluid
air
M
Molar mass
0.028 kg/mol
p
Pressure
105 Pa
Re0
Reynolds number at inlet
165, 270, 730

5.6

Conclusions

It can be seen that the model given in this work follows the wall heat con
duction (WHC) model proposed by Winterberg et.al. quite closely. This
is remarkable since the WHC model contains several parameters that have
been tted to a large number of experimentally determined temperature pro
les, while the model in this work was based on physical properties of random
packed beds obtained from bed packing simulation results. Figure (5.13) shows
that this model is able to describe the temperature prole for dierent values
of the Reynolds number as well as the WHC model.
The decrease in thermal conductivity near the wall of a packed bed is
attributed here to the bed structure and more specically to the bed tortuosity.
Since there is a high velocity zone between the wall and one particle radius from
the wall, a high level of convective mixing would be expected there. However,
since the tortuosity is also close to one in this region, the high velocity does
not lead to intensive mixing. Due to these two strong opposing eects, the

181

5. Dispersion in randomly packed beds

100
whc model
this work

90

Re0 = 165

T [C]

80

70
Re0 = 270
60

50
Re0 = 730
40
0

x/rp [-]

Figure 5.13: Temperature proles for a steam-heated tube lled with ceramic
particles; comparison of model results of this work with results from the wall
heat conduction (WHC) model according to Winterberg et.al. (2000); see table
5.2 for the parameter values used.

model is quite sensitive to the value of the tortuosity near the wall. A CFD
calculation for the rst layer of particles near the wall could give more detailed
information about the actual tortuosity in the ow paths in this zone.
Our model is based on physical mechanisms and estimates for the structure
of the bed; therefore it is by nature more general than literature models like
the WHC, WHT and similar, empirical models. For 2-dimensional cases, the
results of our model will (and should) be comparable to the best literature
models. The real power of our model is in the modelling of non-standard,
3-dimensional bed geometries. In fact, there are many 3-dimensional cases
for which models like the WHC cannot be used, for instance because the
centre of the bed (at which the Peclet number is needed) is not dened for
an asymmetrical bed. These 3-dimensional cases cannot be calculated by a
simple nite dierence method as was done for the 2-D cases in this chapter,
but require a nite volume (CFD) code. This will be described in the next
chapter.

182

5.6 Conclusions

Nomenclature
Roman
Symbol
as
aR

units

as

...

B
cp , f
d
dp
D
D0

J/kg/K
m
m
m
m/s2

Di,

m/s2

Da,,1D

m2 /s

Da,,2D

m2 /s

Dr,

m2 /s

Df,

m2 /s

f
f1
f2
kbed

...
...

K1,h

K1,m

K2,h

K2,m

Variable
solid phase relative heat conductivity (=s /f )
relative contribution of radiation to heat transfer
(=R /f )
auxiliary solid phase heat transfer variable in the
transfer equation by Zehner (1970)
parameter in particle shape function
uid heat capacity at constant pressure
distance to the wall
particle diameter
diameter of the bed
eective diusion coecient inside the bed for in
the stagnant uid
eective diusion coecient for in direction i at
high Re numbers
axial dispersion coecient (parallel to the ow) for
quantity , in 1-dimensional model
axial dispersion coecient (parallel to the ow) for
quantity , in 2-dimensional model
radial dispersion coecient (perpendicular to the
ow)
molecular diusion coecient for in a stagnant un
conned uid
function
Ergun coecient for laminar ow
Ergun coecient for turbulent ow
relative contribution of stagnant bed conductivity
(=0 /f )
Constant in high Reynolds heat dispersion coecient
correlation
Constant in high Reynolds mass dispersion coe
cient correlation
Constant in high Reynolds heat dispersion coecient
correlation
Constant in high Reynolds mass dispersion coe
cient correlation

183

5. Dispersion in randomly packed beds

Symbol

L
M
N

units
m
m
kg/mol
...

p
r

Pa
m

rp
R
sv
sv
t
T
Tw
w0
x
z

m
m
m2 /m3
m2 /m3
s
K
K
m/s
m
m

184

Variable
coordinate along length of channel
length of tube
molecular mass
auxiliary variable in heat transfer equation by Zehner
(1970)
(absolute) pressure
space coordinate in radial direction for cylindrical
coordinates
particle radius
radius of tube
specic surface
mean specic surface in a subvolume of the bed
time
temperature
wall temperature
supercial velocity
spatial coordinate (perpendicular to the wall)
spatial coordinate (in axial direction)

5.6 Conclusions

Greek
Symbol
w,1D
w

units
W/m2 /K
W/m2 /K


b
R
e,q
e,r
f
s
0
R
f
e

W/m/K
W/m/K
W/m/K
W/m/K
W/m/K
W/m/K
m2 /s
m2 /s

[]

h
f

a
a , r
b
r

[]/m2 /s
W/m2
kg/m3
W/m2 /K4

Variable
wall heat transfer coecient, 1-dimensional models
wall heat transfer coecient 2-dimensional models
dimensionless conduction length between two parti
cle centres according to Yagi and Kunii (1957)
parameter in heat transfer relation of Yagi and Kunii
(1957)
(local) porosity (volume open to ow / total volume)
bulk porosity, porosity far from any walls etc.
particle emissivity
eective axial heat conductivity
eective radial heat conductivity
heat conductivity of the uid
heat conductivity of the solid
heat conductivity of the bed at stagnant conditions
heat conductivity contribution of radiation
molecular kinematic viscosity of unconned uid
eective kinematic viscosity of uid
dimensionless parameter describing the packing in
heat transfer relation of Yagi and Kunii (1957)
generic intensive quantity (momentum, heat, concen
tration)
ux of quantity
heat ux
density of uid
Stephan-Boltzmann constant (56.7051 109 )
tortuosity (path length / distance)
tortuosity for paths in axial direction
tortuosity in axial direction, radial component
tortuosity at bulk conditions (far from the wall)
tortuosity of the bed for paths in radial direction

185

5. Dispersion in randomly packed beds

Dimensionless groups
Symbol
Nuw

denition

Nuw,0

w,0 dp
f
w0 dp
Df

Pef

w dp
f

Pef,c

w0,c dp
Df,c

Pei,

w0 dp
Di,

Pe0
Pe
i
Pr

w0 dp
0
D
w0 dp
Di0

Re0

w0 dp
f

186

Variable
Nusselt number (total heat transfer/diusive heat
transfer) at the wall
Minimum value of Nusselt number at the wall
molecular Peclet number (convective transfer / dif
fusive transfer)
molecular Peclet number at conditions at the centre
of the bed
Particle Peclet number for direction i and quantity

Stagnant particle Peclet number for quantity

Turbulent particle Peclet number for direction i
Prandtl number (hydrodynamic / heat transfer
boundary layer)
Reynolds number (inertia forces / viscous forces)

5.6 Conclusions

Literature
Balakrishnan, A.R., D.C.T. Pei (1978a), Heat transfer in gas-solid packed
bed systems: 1. A critical review, Ind. Eng. Chem. Process Des. Dev. 18(1),
30-40
Balakrishnan, A.R., D.C.T. Pei (1978b), Heat transfer in gas-solid packed
bed systems: 2. The conduction mode, Ind. Eng. Chem. Process Des. Dev.
18(1), 40-46
Balakrishnan, A.R., D.C.T. Pei (1978c), Heat transfer in gas-solid packed
bed systems: 3. Overall heat transfer rates in adiabatic beds, Ind. Eng.
Chem. Process Des. Dev. 18(1), 47-50
Bauer, R., E.U. Schl
under (1978), Eective radial thermal conductivity of
packings in gas ow. Part II. Thermal conductivity of the packing fraction
without gas ow, Int. Chem. Eng. 18(2), 189-204
Berger, R.J., J. Perez-Ramrez, F. Kapteijn, J.A. Moulijn (2002), Catalyst
performance testing: Radial and axial dispersion related to dilution in
xed-bed laboratory reactors, Appl. Catal. A: General 227, 321-333
Bey, O., G. Eigenberger (1996), Bestimmung von Stromungsverteilung und
Warmetransportparametern in sch
uttungsgef
ullten rohren,
Chemie-Ing.-Techn. 68(10), 1294-1299
Borkink, J.G.H., K.R. Westerterp (1992), Determination of eective heat
transport coecients for wall-cooled packed beds, Chem. Eng. Sci. 47,
2337-2342
Calis, H.P.A. (1995), Development of dustproof, low pressure drop reactors
with structured catalysts packings, dissertation Delft University of
Technology, Delft, the Netherlands
Cheng, P., D. Vortmeyer (1988), Transverse thermal dispersion and wall
channelling in a packed bed with forced convective ow, Chem. Eng. Sci.
43(9), 2523-2532
Delmas, H., G.F. Froment (1988), A simulation model accounting for
structural radial nonuniformities in xed bed reactors, Chem. Eng. Sci. 43,
2281-2287
Dixon, A.G., Cresswell, D.L. (1979), Theoretical prediction of eective heat
transfer parameters in packed beds, A.I.Ch.E. J. 25(4),663-676

187

5. Dispersion in randomly packed beds

Dixon, A.G. (1988), Wall and particle-shape eects on heat transfer in

packed beds, Chem. Eng. Commun. 71, 217-237
Dixon, A.G., J.H. van Dongeren (1998), The inuence of the tube and
particle diameters at constant ratio on heat transfer in packed beds, Chem.
Eng. Proc. 37, 23-32
Epstein, N. (1989), On tortuosity and the tortuosity factor in ow and
diusion through porous media, Chem. Eng. Sci. 44(3), 779-781
Foumeny, E.A., M.A. Chowdhury, C. McGreavy, J.A.A. Castro (1992),
Estimating of dispersion coecients in packed beds, Chem. Eng. Technol.
15, 168-181
Freiwald, M.G., W.R. Paterson, Accuracy of model predictions and reliability
of experimental data for heat transfer in packed beds, Chem. Eng. Sci.
47(7), 1545-1560
Giese, M., K. Rottschafer, D. Vortmeyer (1998), Measured and modelled
supercial ow proles in packed beds with liquid ow, A.I.Ch.E. J. 44(2),
484-490
Givler, R.C., S.A. Altobelli (1994), A determination of eective viscosity for
the Brinkman-Forchheimer ow model, J. Fluid Mech. 258, 355-370
Gunn, D.J. (1987), Axial and radial dispersion in xed beds, Chem. Eng.
Sci. 42(2), 363-373
Hennecke, F-W, E.U. Schl
under (1973), Warme
ubergang in beheizten oder
gek
uhlten Rohren mit Sch
uttingen aus Kugeln, Zylindern und
Raschig-Ringen, Chemie-Ing.-Techn. 45(5), 277-284
Jakobsen H.A., H. Lindborg, V. Handeland (2002), A numerical study of the
interactions between viscous ow, transport and kinetics in xed bed
reactors, Comput. Chem. Eng. 26(3), 333-357
Kunii, D., J.M. Smith (1960), Heat transfer characteristics of porous rocks,
A.I.Ch.E.J. 6(1), 71-78
K
ufner, R., H. Hofmann (1990), Implementation of radial porosity and
velocity distribution in a reactor model for heterogeneous catalytic gasphase
reactions (TORUS model), Chem. Eng. Sci. 45(8), 2141-2146
Legawiec, B., D. Ziolkowski (1994), Structure, voidage and eective thermal
conductivity of solids within near-wall region of beds packed with spherical
pellets in tubes, Chem. Eng. Sci. 49(15), 2513-2520

188

5.6 Conclusions

Logtenberg, S.A., A.G. Dixon (1998), Computational Fluid Dynamics studies

of xed bed heat transfer, Chem. Eng. Proc. 37, 7-21

Martin, H., Nilles, M. (1993), Radiale Warmeleitung in durchstromten
Sch
uttungsrohren, Chem.-Ing.-Tech. 65(12), 1468-1477
McGreavy, C., E.A. Foumeny, K.H. Javed (1986), Characterization of
transport properties for xed bed in terms of local bed structure and ow
distribution, Chem. Eng. Sci. 41(4), 787-797
Olbrich, W.W., O.E. Potter (1972a), Heat transfer in small diameter packed
beds, Chem. Eng. Sci. 27, 1723-1732
Olbrich, W.W., O.E. Potter (1972b), Mass transfer from the wall in small
diameter packed beds, Chem. Eng. Sci. 27, 1733-1743
Puncochar, M., J. Drahos (1993), The tortuosity concept in xed and
uidized bed, Chem. Eng. Sci. 48(11), 2173-2175
Schl
under, E.U. (1966), Warme und Sto
ubertragung zwisschen
durchstromten Sch
uttungen und darin eingebetteten Einzelkorpen.
Chem.-Ing.-Techn. 38(9), 967-979
Specchia, V., G. Baldi, S. Sicardi (1980), Heat transfer in packed bed
reactors with one phase ow, Chem. Eng. Commun. 5, 361-380
Tsotsas, E., H. Martin (1987), Thermal conductivity of packed beds: a
review, Chem. Eng. Process. 22, 19-37
Tsotsas, E., E.U. Schl
under (1988), Some remarks on channelling and on
radial dispersion in packed beds, Chem. Eng. Sci. 43 (5), 1200-1203
Tsotsas,E., E.U. Schl
under (1990), Heat transfer in packed beds with uid
ow: remarks on the meaning and the calculation of a heat transfer
coecient at the wall, Chem. Eng. Sci. 45(4), 819-837
Tsotsas, E. (1992), On mass transfer, dispersion, and macroscopical ow
maldistribution in packed tubes, Chem. Eng. Proc. 31, 181-190
Tsotsas, E. (1997), Warmeleitfahigkeit von Sch
utschichten, VDI Warmeatlas,
8. Afulage, Section Dee
Vortmeyer, D., E. Haidegger (1991), Discrimination of three approaches to
evaluate heat uxes for wall-cooled xed bed chemical reactors, Chem. Eng.
Sci. 46(10), 2651-2660

189

5. Dispersion in randomly packed beds

Vortmeyer, D., J. Schuster (1983), Evaluation of steady ow proles in

rectangular and circular packed beds by a variational method, Chem. Eng.
Sci. 18(10), 1691-1699
Votruba, J., V. Hlavacek, M. Marek (1972), Packed bed axial conductivity,
Chem. Eng. Sci. 27, 1845-1851
Westerterp, K.R., W. de Jong, G.H.W. van Benthem (1993), Comments on
discrimination of three approaches to evaluate heat uxes for wall-cooled
xed bed chemical reactors, Chem. Eng. Sci. 48(14), 2669-2670
Winterberg, M., E. Tsotsas (2000), Modelling of heat transport in beds
packed with spherical particles for various bed geometries and/or thermal
boundary conditions, Int. J. Therm. Sci 39, 556-570
Winterberg, M., E. Tsotsas, A. Krischke, D. Vortmeyer (2000), A simple and
coherent set of coecients for modelling of heat and mass transport with and
without chemical reaction in tubes lled with spheres, Chem. Eng. Sci. 55,
967-979
Wijngaarden, R.J., K.R. Westerterp (1989), Do the eective heat
conductivity and the heat transfer coecient at the wall inside a packed bed
depend on a chemical reaction? Weaknesses and applicability of current
models, Chem. Eng. Sci. 44(8), 1653-1663
Wijngaarden, R.J., K.R. Westerterp (1992), The statistical character of
packed-bed heat transport properties, Chem. Eng. Sci. 47(12), 3125-3129
Wijngaarden, R.J., K.R. Westerterp (1993), A heterogeneous model for heat
transfer in packed beds, Chem.Eng.Sci. 48(7), 1273-1280
Yagi, S., D. Kunii (1957),A.I.Ch.E.J. 3, 373
Yagi, S., D. Kunii (1960), Studies on heat transfer near wall surface in
packed beds, A.I.Ch.E. J. 6(1), 97-104
Zehner, P., E.U. Schl
under (1970), Warmeleitfahigkeit von Sch
uttungen bei
massigen Temperaturen, Chemie-Ing.-Techn. 42(14), 933-941
Zehner, P., E.U. Schl
under (1972), Einuss der Warmestrahlung und des
Druckes af den Warmetransport in nicht durchstromten Sch
uttungen,
Chemie-Ing.-Techn. 44(23), 1303-1308

190

Chapter 6
Modelling of packed bed
reactors using Computational
Fluid Dynamics
Summary
In this chapter, the models developed in the previous chapters are implemented
in a Computational Fluid Dynamics (CFD) code. The CFD tool used in this
work is Comow, which is a code that is developed as a tool for the process
engineer, in contrast to many larger CFD systems that require a ow dynamics
expert to be operated. For the current work, it was decided to create a new
version of Comow based on Dolfyn (www.dolfyn.net), which is an open source,
single phase ow code. Dolfyn does not include the models needed to simulated
packed bed reactors. Therefore, the code was extended with models for variable
physical properties of ideal gas mixtures and the packed bed model including
ow resistance, dispersion and heterogeneously catalysed chemical reactions.
The models used for the simulation of ow and reaction in the catalyst bed
are those that were developed in the previous chapters. The porosity and the
specic outer particle surface area of the packed bed are treated as constant
(but not uniform) elds to enable the calculation of ow resistance, dispersion
and chemical reaction rates locally in the bed.
The equations used to evaluate gas mixture properties are generally those
advised by Reid et.al. (1987). The density is calculated according to the ideal
gas law, with local composition, temperature and pressure. It is usual in CFD
practice to treat the uid as incompressible when the velocities are low (well
below the speed of sound). It is shown that this is only true for systems where
the mechanical energy is (mainly) conserved. For highly dissipative systems

191

6. Modelling of packed bed reactors using CFD

like packed beds, the density needs to be calculated as a function of local

pressure.
The dispersion coecients are calculated according to the theory developed
in chapter 5. In the standard CFD codes, dispersion is usually treated as an
isotropic phenomenon, while in a packed bed it depends on the direction. In
the direction of the ow there is axial dispersion and perpendicular to the ow
there is radial dispersion, and the dispersion coecients for these directions
dier. Relations are developed to calculate the dispersion coecient for each
face of a cell, taking into account the orientation of each face with respect to
the local ow.
The model is completed with (Maxwell-Stefan based) relations to calculate
the mass transfer limitation from the bulk of the uid to the edge of the
particle, and the chemical reaction model described in chapter 2.
The CFD code is veried by comparing simulations of simple tubular ge
ometries with hand calculations. In addition, the sensitivity to the grid density
is assessed; it is found that the number of radial cells needs to be more than
about 50 to have sucient resolution of the wall zone. The calculated radial
velocity proles are compared with literature data and the correspondence is
found satisfactory. The heat transfer model is validated by comparison of the
model result with experimental literature data for a steam-heated tube; the
model ts the measurements as well as may be expected. Finally, a laboratory
reactor for the catalytic oxidation of ethane is modelled; good correspondence
is found between the axial and radial temperature and concentration pro
les predicted by the model and those measured by Vortmeyer and Haidegger
(1991).
To demonstrate the capabilities of the CFD code, two model reactions
are simulated: the catalytic reduction of NO with ammonia and the catalytic
oxidation of SO2 . The deNOx reactor is usually placed in the exhaust of a
process (e.g., a burner or an engine) where little pressure drop can be allowed.
The performance of a tubular reactor is compared with that of a radial ow
reactor. The pressure drop of the radial ow reactor is much lower than that of
the simple tube, but there is some reduced conversion due to maldistribution
of ow over the bed. As a rst step, measures are taken to improve the ow
distribution and increase the conversion. With these improvements, the radial
ow reactor shows a conversion that is three times as high as the tubular
reactor, at one third of the pressure drop. The SO2 oxidation reaction is an
exothermic equilibrium reaction with a relatively complex dependence of the
reaction rate on temperature and concentrations. Industrially, this process is
usually performed in a staged packed bed reactor with interstage cooling. Two
alternatives are simulated: a wall-cooled tubular reactor and a tube-in-tube

192

6.1 Introduction

design where the heat removed from the reaction zone is used to heat up feed
stream. The results of the simulation of the tubular design are compared to a
1-dimensional literature model. It is believed that in this case, the CFD model
gives better results than the simple model. The tube-in-tube design could be
an interesting energy and space saving alternative for the conventional staged
bed design. The performance of such a reactor would be hard to predict, let
alone optimise, without the aid of a relatively complex CFD model, due to the
strong coupling between the heat transfer from the reaction zone to the feed
ow and the kinetics of the oxidation reaction.
The data generated by the simulations is relatively detailed, giving pres
sure, velocity, temperature and concentrations at every point of the grid.
Therefore, it is quite a task to duplicate this resolution in experiments and
measurements, and validation of the model on a detailed level in a complex
simulation is not possible at this time. This is valid for this work in particu
lar, but also more in general as the possibilities for detailed modelling rapidly
improve. It is a challenge for future experimentalists to make similar improve
ments in measurement techniques for chemical reactors.

6.1

Introduction

In this chapter, the models developed in the previous chapters are implemented
in a Computational Fluid Dynamics (CFD) code. The CFD package used in
this work is Comow, which is a code that is developed as a tool for the pro
cess engineer, in contrast with many larger CFD systems that require a ow
dynamics expert to be operated. Although there are certainly cases where the
ow in e.g. a chemical reactor is very critical and a highly accurate solution of
the ow eld is needed, in many (if not most) situations, a faster, less expensive
and less accurate solution is just as useful. This is especially the case for inter
nal ows that are often determined in large part by the ow channel geometry
(as opposed to free streams around an object), and if there are obstructions
in the ow (e.g., tube bundles or packed beds). Also, in many situations even
the purpose of the simulation is to compare two cases, e.g., to see if a change
in geometry is an improvement and the absolute quantitative accuracy is less
important as long as the solution is qualitatively correct. For this purpose,
Comow uses only a subset of the general CFD functionality: 2-dimensional
mostly hexagonal grids and only very simple turbulence modelling. On the
other hand, there are comparatively extensive models for physical properties
and ow restrictions, of which the packed bed module developed in this work
is the most elaborate example.

193

6. Modelling of packed bed reactors using CFD

Strictly speaking Comow is just a pre- and postprocessor for an external

CFD solver, although the solver needs to be extended with a library of routines
to support the additional models used by Comow. As a solver, Phoenics
(http://www.cham.co.uk) has been used in the past, as well as a custom solver
based on the (now historic) champion code (see e.g., Pun and Spalding, 1976).
For the current work, it was decided to create a new version of Comow
based on a more modern CFD solver. The requirements for the solver, that it
should be lightweight and that the source code should be available, lead to the
selection of Dolfyn (www.dolfyn.net). This is an open source, unstructured
CFD code written in Fortran 90 and based on the numerical method described
by Ferziger and Peric (2002). Dolfyn is a plain, single phase ow code and does
not include many of the the models needed to simulated packed bed reactors.
Therefore, the code was extended with models for variable physical properties
of ideal gas mixtures and the packed bed model including ow resistance,
dispersion and heterogeneously catalysed chemical reactions.

In this chapter, the CFD packed bed model will be described, validated and
demonstrated. The basic ow equations solved by Dolfyn will be presented,
followed by the equations used to calculate the physical properties of the gas
mixture as a function of temperature, pressure and composition; these are
based on the relations recommended by Reid, Prausnitz and Poling (1987) but
some specic issues for CFD codes are discussed as well. Then the specic
packed bed models for ow resistance, dispersion and chemical reaction and
transport of heat and mass in the catalyst particle are described. The full
details of these models are given in the previous chapters of this work, therefore
in this chapter the focus is on the relation between the CFD code and the
packed bed models.
The CFD packed bed code will be veried using a basic geometry that
can be easily computed with hand calculations; the eect of the density of
the computational grid is assessed. Subsequently the code will be validated by
comparison to literature models and experimental data. Again, the underlying
packed bed models were veried and validated in the previous chapters so the
main focus in this chapter will be on the eect of combination of the eects in
a 2-dimensional model.
Finally, the use of the code will be demonstrated for two relevant chemical
processes: the selective catalytic reduction of nitric oxide by ammonia and
the catalytic oxidation of SO2 . For these reactions, packed tube designs will
used to compare the packed bed CFD code with conventional 1-dimensional
calculations. Subsequently, the CFD code will be used to simulate and improve
a more advanced reactor concept for both demonstration processes.

194

6.2 Models

For this work, the Comow preprocessor is used throughout, so only sta
tionary calculations have been performed using 2D Cartesian grids. However,
the same code can be used for 3D calculations and with some restrictions1
for transient calculations. Unstructured, arbitrary cell shape grids can be
used in principle (as far as Dolfyn can handle them), although the interpola
tion routines for the bed packing proles near the wall will be somewhat less
accurate for non-cartesian grids.

6.2

Models

6.2.1

Basic ow and transport equations

The basic ow equations and the balance equations for concentrations and
temperature are solved using the Dolfyn solver. Dolfyn is a plain, single phase
ow code based on the numerical method described by Ferziger and Peric
(2002). A short description of the method will be given below so that we are
able to show where the packed bed model intervenes in the Dolfyn code. A
full description of the ow model can be found in the original text.
The Dolfyn code uses a nite volume method to solve the ow equations.
The basis of this method is an integral representation of the generic balance
equation for a intensive quantity , for a nite (sub-) volume bounded by
the surface S:

d + v ndS =
ndS + q d
(6.1)
t
S
S

where n is the normal of surface S. The terms are easily identied as the
(transient) accumulation term, the convective term, the diusive term with
diusivity and a volumetric source term q. Note that each term in this
equations has units []kg/s.
The equations are solved on a colocated unstructured grid of volume ele
ments made up of an arbitrary number of faces:

d +


k

v ndS =
Sk


k

ndS +
Sk

q d

(6.2)

The enthalpy and material balances inside a catalyst particle are not solved in the time
domain, but the (macroscopic) ow phenomena can be calculated as a function of time. This
is valid provided that the transport processes on the particle scale are much faster than the
macroscopic processes

195

6. Modelling of packed bed reactors using CFD

Fluxes of quantities through each face Sk are approximated using the mid
point rule, where the face value fk is calculated by linear interpolation.

f dS fk Sk
(6.3)
Sk

and the volume integrals are approximated by the cell volume multiplied
by the cell centre value qp :

qd qP
(6.4)

so that for each cell with centre P and faces k




(k kvk nk Sk ) =
(,k ()k nk Sk )+q,P (6.5)
(P P )+
t
k
k
where P is the value of at the cell centre, k is the value at the centre of
cell face k, vk is the velocity vector at the face centre, nk is the unit normal
vector of face k, Sk is the surface area of face k and is the volume of the
cell. ,k is the (eective) diusion coecient for quantity at cell face k.
The face value k is estimated by linear interpolation between the two
bordering cells (designated P and N ), where k is the relative distance from
cell centre P to face centre k on the line P N :
k = k P + (1 k )N

(6.6)

The gradient at the face centre ()k is calculated from the cell centre gradi
ents in the same way.
k = k ()P + (1 k )()N

(6.7)

The gradient at the cell centre is calculated from the cell centre value dierence
between the current cell centre and that of all neighbours by solving for ()P
the set of equations:
N P = ()P (xN xP )

(6.8)

where there is one equation for each neighbour N . Since there are usually more
than three neighbour cells and only three coordinate directions in , the set
is usually over-specied; the gradient is calculated by using a least squares
method.
In addition, there are corrections for non-orthogonal cells (i.e., where the
centre of the face between cell P and cell N does not lie on the line P N ), but
as in this study only orthogonal grids are used, these are not described here.

196

6.2 Models

6.2.2 Porosity, specic particle outer surface area and

tortuosity
The porosity and the specic outer particle surface area of the packed bed are
treated as constant (but not uniform) elds to enable the calculation of ow
resistance, dispersion and chemical reaction rates locally in the bed.
The porosity is calculated at initialisation time from a particle centre dis
tribution as a function of the distance to the closest wall, according to the
equations developed in chapter 3 of this work. For simplicity it is assumed
that cell faces and walls are aligned with the co-ordinate axes. Catalyst beds
can have wall eects activated for one or more of their sides, since for sides
where there is no solid wall and for symmetry boundaries, no wall eect is
expected. The porosity of a single cell is calculated by a weighed average of
the porosities expected for each of the distances to active walls alone. The
weighing factor is the inverse of the wall distance, i.e.

i /xi,p
(6.9)
 =  i
i (1/xi,p )
where xi.p is the distance from the cell centre to the wall and i loops over
all walls with wall eect enabled. The term i is the cell porosity that would
arise if only wall i would be active; this is calculated from the cell dimensions
and centre location. For instance, for a wall on the low x (west) side of the
cell, it is calculated by integration of the particle centre distribution between
the west and east side of the cell (x = xi,w and x = xi,e ), see Figure (6.1).
This approach will ensure that there is a smooth transition from the porosity
prole at one wall to the porosity prole at another (opposite or adjoining)
wall. However, it should be kept in mind that the approximations are valid
only in regions where walls are further apart than 5 particle diameters.
The particle outer surface area is calculated in an analogous way as the
porosity prole from the particle centre distribution, i.e.

sv /xi,p
sv =  i i
i (1/xi,p )

(6.10)

There are three dierent tortuosities that are needed to calculate the bed
ow resistance and the eective dispersion coecients. For the ow resistance
we need the tortuosity of a ow path, i.e., in axial direction. This was shown
in Chapter 5 to be

197

6. Modelling of packed bed reactors using CFD

xW,w

W,p

xW,e

xE,w

xE,p
xE,e

Figure 6.1: Dierent cell wall distances for calculation of porosity proles

3D,a = w +

(1/ 1)
(1/b 1)(b w )

(6.11)

We also need the radial tortuosity, which is assumed to be equal to the

bulk tortuosity in most of the bed, and decreases to 1.0 between 2 and 1 radii
from the wall. The wall distance used here is the distance from the cell centre
to the closest active wall. The tortuosity that needs to be used to calculate the
radial dispersion coecient is the radial (2D) tortuosity of ow paths in axial
direction (2D,a ). This tortuosity diers from the 3D axial tortuosity that is
used for the ow resistance calculations in that only ow variations in radial
direction (i.e., perpendicular to the wall) are taken into account; tangential
(parallel) ow does not lead to additional dispersion. Near the wall, the ow
will be parallel to the wall, which means that the tortuosity has a value of
about 1. This will be the case at distances up to one half particle radius from
the wall, where the velocity peak is located. Fluid in this region will ow in
a zigzag manner around the particles, but the zigzags will be mainly in the
tangential plane at constant wall distance. In contrast, at one particle radius

198

6.2 Models

from the wall, the uid will ow primarily in radial direction and hardly in
tangential direction, so the 2-D tortuosity will be approximately equal to the
3D tortuosity. Therefore, we will assume that the 2-D tortuosity will be one
up to a distance of rp /3 from the wall, and then increase to the maximum
value of the 3D tortuosity at one particle radius from the wall.

6.2.3

A note on discretisation

In a CFD simulation, the packed bed will be divided into a large number of
computational cells. Since we use a pseudo-homogeneous approach, each cell
will get average values for the porosity, tortuosity and other bed parameters.
Strictly speaking, these average values are only valid if each cell contains many
catalyst particles. However, in many cases, the cells may have a size that is
comparable to the particle size or even smaller. This is still correct if the
parameter and variable values do not change rapidly in any direction. Of
course, we cannot expect to nd details of the ow on the particle scale in the
simulation, but the simulation will give an accurate estimate of the average
ow eld.
The actual ow eld in a packed bed depends on the random position of
the particles. Therefore, if we were able to measure the ow variables locally
in the bed, the measurement values at each point would be inuenced by the
coincidental positions of the particles near that point. Each time we reshue
the particles in the bed, we will get a (very) dierent measurement value at
the same point. If we reshue the particles many times, and average the
measurement values for each bed realisation, we can suppress the stochastic
nature of the randomly packed bed and get a clearer picture of the actual
process. The packed bed model predicts precisely these averaged values.
Near the wall, the situation is slightly dierent. There the cell size will
usually be smaller than the particle size, at least perpendicular to the wall,
to catch the ow pattern in the wall zone. Here the parameters and variables
change rapidly within one particle diameter, so strictly we cannot use averaged
values. However, since the ow is almost parallel to the wall, we can expect
to get on average correct values over longer stretches of the wall. Of course,
this is only true if we use the actual particle outer surface area and tortuosity
near the wall, taking into account the wall eect.
For the calculation of chemical reaction in the particles, there is an addi
tional complication. The calculation of the conversion in the catalyst particles
is based on the assumption that the temperature and concentrations in the
gas phase at the edge of the particle is constant. Far from the bed, this is

199

6. Modelling of packed bed reactors using CFD

not problematic since the concentration and temperature gradients are usu
ally not steep compared to the particle size. However, near the wall we have
many particles that have concentration and temperature gradients along their
surface. For these particles, the assumption of constant surface conditions are
not valid and hence we strictly cannot use the calculation methods presented
in chapter 2.
Luckily the error made is not as large as it may seem. To calculate the
reaction rate in the bed, we calculate the eectiveness factor for a particle
assuming it has a constant particle edge (or actually, hydrodynamic boundary
layer edge) concentration that corresponds to the local values in the cell. Using
the eectiveness factor, we can calculate the uxes into the pellet per unit of
pellet surface. The source term for the conservation balances in the bed is that
number multiplied by the amount of pellet outer surface area in the cell, not
the pellet volume in the cell. In this way, we will get approximately correct
values e.g., near the wall. The error made is because the eectiveness factor
used is based on the assumption of symmetric concentration proles in the
particle. In reality, if e.g. the concentration of a reactant is low on one side
of a particle an high on another, the high side ux will increase compared to
the uniform case since some reactant will diuse through the particle to the
low-concentraion side. Therefore, our model will slightly underestimate the
concentration proles caused by the wall eect.

6.2.4

Properties

In a chemical reactor, the temperature, pressure and composition can vary

signicantly from point to point. Therefore, to get accurate ow simulation
results, the uid properties should be calculated as a function of the local tem
perature, pressure and composition. Here the uid properties are recalculated
at the beginning of each outer iteration loop (i.e., with the variable values from
the previous iteration). This leads to approximately (rst order) correct re
sults even though this approach may not be correct in a strict (mathematical)
sense, because for instance gradients of transport coecients are not taken
into account. For the cases evaluated here, the accuracy is sucient and it
is certainly an improvement over the mathematically correct but physically
incorrect constant property assumption. The equations used to evaluate prop
erties are generally those advised by Reid et.al. (1987); these are reproduced
here for sake of completeness.

200

6.2 Models

Density
To calculate the density of the uid as a function of the local composition,
temperature and density, an equation of state is needed. For systems where
the uid is a liquid, the density can (usually) be taken independent of the
pressure. For (real) gas ow systems, specic equations of state have been de
veloped that can be used e.g., at elevated pressures; such powerful but complex
models fall outside the scope of this work. For many processes that operate
at intermediate pressure and temperature, the ideal gas law can be used with
sucient accuracy:

f =

(xi Mi ) p
RT

(6.12)

In CFD calculations, usually two ow regimes are discerned: compressible

and incompressible. Compressible ow simulations take into account the eect
of the pressure on the density, while in incompressible simulations, the density
is not a function of pressure (but could be a function of temperature and
composition); in the latter case a constant, reference pressure should be used
in equation 6.12 instead of the local static pressure. The general rule is that
for low velocity ows (Mach number much smaller than 1), incompressible ow
can be assumed. For process engineers, this is somewhat counter-intuitive as
considerable pressure loss usually is found in process equipment. Therefore,
we will look into this issue at some more detail.
There are two mechanisms by which the static pressure can decrease,
through dynamic eects (i.e., acceleration or stagnation of the ow) where
mechanical energy is conserved and through dissipation (e.g., friction at walls)
where mechanical energy is dissipated. The dynamic eects become impor
tant for high-speed (near sonic) ows. For isentropic ow, the conservation of
mechanical energy leads to the following well-known relation between pressure
and velocity (e.g., Shapiro, 1953):


p
1 2 1
= 1+
M
p0
2

(6.13)

where M is the Mach number, p0 the total (stagnation) pressure and the
specic heat ratio. For most gases, has a value between 1 and 1.6. For low
velocities, the energy balance simplies to the Bernouilli equation
1
p0 = p + v 2
2

(6.14)

201

6. Modelling of packed bed reactors using CFD

The relation between the velocity (Mach number) and the gas density for an

isentropic ideal gas can be written as:


 1

1 2 1
= 1+
M
0
2

(6.15)

For M < 1 the inuence of is small. The dependence of the density on the
Mach number is shown in Figure 6.2. It can be seen that for Mach numbers
below 0.3, the decrease in density is less than 5%. Note that the error made
in neglecting the eect of static pressure on the density will lead to changes
in the velocity eld; the mass balance will be kept for each cell. Therefore,
for low speed ows (less than 100 m/s for atmospheric air), the decrease in
density due to dynamic eects can be neglected unless accuracies higher than
95% in the velocity eld are desired. For process equipment, the gas ows
rarely exceed values above 10-20 % of the speed of sound.
1

0.8

[-]

0.9

= 1.6
0.7
= 1.01
0.6

0.5
0

0.2

0.4

0.6

0.8

M [-]

Figure 6.2: Relative change in density as a function of Mach number for isen
tropic ow, for dierent values of .

However, this result alone does not mean that the gas density can be taken
as independent of pressure. There will also be a pressure loss due to dissipation
of kinetic energy (either in turbulent eddies or at the wall). This is especially
important for ows found in the process industry, where gas passes through

202

6.2 Models

restrictions like for instance tube bundles or packed beds. Due to friction,
mechanical energy is not conserved, but some is turned into thermal energy.
Therefore, the total pressure will decrease (and the total temperature will rise).
In non-isentropic, subsonic ows, the relative decrease in pressure can easily
be much larger than the dynamic pressure eects. The decrease in pressure
will lead to a decrease in density, which at the same mass ow will lead to
an increase in velocity. The higher velocities will lead to dierent values of
the mass, heat and momentum transport coecients. As an example, gure
6.3 gives ow values for a simple packed bed in a tube where the variable
density is taken into account. The density at the outlet is lower than that
at the inlet due to the decrease in pressure and increase in temperature. The
superposition principle for packed beds indicates that if the pressure drop is not
small compared to the inlet pressure, the Ergun equation should be integrated
over the height with variable density. In gure 6.4, the dierence between a
constant density and a variable density calculation is shown for the tube of
gure 6.3; it shows that in this case the pressure drop increases by almost
10%. I can be concluded that the eect of non-isentropic pressure loss on the
density of the uid should be taken into account for ow simulations of process
equipment in general and packed beds in particular.

pin = 1.6 bar

Tin = 293 K
3
in = 1.9 kg/m
vin = 1 m/s

dp = 1 mm
= 0.4

pout = 1 bar
Tout = 335 K
3
out = 1.04 kg/m
vout = 1.8 m/s

= 1.8e-5 Pa s
cp = 1005 J/kg/K

Figure 6.3: Pressure drop over a typical packed bed with variable density.
Outlet pressure is assumed atmospheric, inlet pressure is adjusted to get 1
m/s supercial velocity. Note that this example is based on the traditional
Ergun equation, integrated over the bed height.

203

6. Modelling of packed bed reactors using CFD

2.0

p [bar], v [m/s]

1.5

1.0

0.5

p constant density
p variable density
v

0.0
0.0

0.2

0.4

0.6

0.8

1.0

h/H [-]

Figure 6.4: Axial pressure and velocity prole for the typical packed bed of
gure 6.3 with constant and variable density.

In the CFD code, the fact that the density is a function of local static pres
sure does not necessarily change the ow from incompressible to (completely)
compressible. The inuence of the pressure on the density is not as direct as
in truly compressible ow. To prevent a strong coupling between the density
and the pressure, which could lead to instabilities, the density eld is updated
only at the end of an iteration step and is kept constant during an iteration
step. For higher velocity incompressible ows, it may be needed to base the
density calculation on the local stagnation pressure instead of the static pres
sure to prevent instabilities, but for the ows considered here, the dynamic
head is always much smaller than the total pressure so this would not make a
dierence.

204

6.2 Models

Viscosity

The pure component gas viscosity is estimated at a function of temperature

using the approach of Lucas (as described in Reid et.al., 1987)

= 0.807Tr0.618 0.357 exp [0.449Tr ]

+0.340 exp [4.058Tr ] + 0.018)

Fp0

(6.16)

with


Tc
= 0.176
M 3 p4c
2 pc
r = 52.46 2
Tr

1/6
(6.17)
(6.18)

and

1 + 30.55(0.292 Z )1.72
c
FP0 =

1 + 30.55(0.292 Zc )1.72

|0.96 + 0.1(Tr 0.7)|

where
Tc
pc
M

K
bar
kg/kmol
D

0 r < 0.022
0.022 r < 0.075

(6.19)

r 0.075

critical temperature
critical pressure
molar mass
dipole moment

According to Reid et.al. (1987), the expected errors are between 0.5 and
1.5 % for non-polar compounds and 2-4 % for polar compounds. The correla
tion is not suitable for highly associated gases like acetic acid and is only valid
at low pressure (well below the critical pressure). For exibility, in the com
puter code the pure gas viscosity is calculated by partwise linear interpolation
between tabulated values (see gure 6.5).
The viscosity of a gas is caused by mechanical energy transfer through
collisions between gas molecules. In a gas mixture, these molecules can have
dierent masses, while in a pure gas, only collisions between similar molecules
occur. Therefore, the mixture viscosity can not be calculated by simply adding
together the relative contributions of all species. Instead, the mixing rule of

205

6. Modelling of packed bed reactors using CFD

5.0E-05
N2
NO
NH3
H2O
O2

4.0E-05

[Pa s]

3.0E-05

2.0E-05

1.0E-05

0
200

300

400

500

600

700

800

T [K]

Figure 6.5: Dynamic viscosity of pure gases for some gas components, accord
ing to equation 6.16

Herning and Zipperer is used (as given in Reid et.al, 1987).

 yi i (T )

F (i)
#

Mj
yj
F (i) =
Mi
j

(6.20)

(6.21)

Here yi is the mole fraction of component i, Mi is the molar mass in

(kg/mol) and i (T ) the dynamic viscosity in (Pa s) at the local temperature
of pure i. Note that the dynamic gas viscosity is a very weak function of the
pressure but the kinematic viscosity is coupled with the local pressure through
the density. The shear stress depends on the dynamic viscosity, and therefore
the inuence of the pressure on the viscous term in the Navier-Stokes equation
will be small.

206

6.2 Models

Thermal conductivity
The pure gas thermal conductivity can be estimated by the method of Chung,
as described by Reid et.al. (1987)
M
3.75
=
cv
cv /R

(6.22)

where M is the molar mass (in kg/mol), R is the universal gas constant and
cv is the heat capacity at constant volume. For an ideal gas
cp c v = R

(6.23)

The value of parameter is calculated from

0.215 + 0.28288 1.061 + 0.26665Z
0.6366 + Z + 1.061
cv 3
=

R 2
= 0.7862 0.7109 + 1.3168 2
Z = 2.0 + 10.5Tr2

= 1+

(6.24)
(6.25)
(6.26)
(6.27)

As the thermal conductivity is estimated using viscosity data, the error

in the former is usually larger than that made in the viscosity correlation;
according to Reid et.al. it is about 5-7 % maximum for non-polar compounds.
For exibility, the temperature dependence is represented by a piecewise linear
approximation in the computational code, see gure 6.6.
The heat conduction in a gas is caused by thermal energy transfer through
collisions between gas molecules. Therefore, the molecular thermal conductiv
ity of the gas mixture is calculated using an approach that is analogous to the
viscosity mixing rule
f =

 yi i (T )
i

F (i)

(6.28)

where i (T ) is the pure gas thermal conductivity of species i at the local

temperature T (in W/m/K).
Heat capacity
Pure component heat capacity at constant pressure is estimated using third
order polynomial interpolation functions from the physical properties database
in Reid et.al. (1987). These functions are very widely used, even though no
accuracies or temperature boundaries are given by the authors. For exibility,

207

6. Modelling of packed bed reactors using CFD

0.12

0.10

[W/m/K]

0.08

N2
NO
NH3
H2O
O2

0.06

0.04

0.02

0
200

300

400

500

600

700

800

T [K]

Figure 6.6: Thermal conductivity of pure gas species as a function of temper

ature for several gas components

the curves are represented by piecewise linear approximation in the computa

tion code, see gure 6.7.
The mixture heat capacity is calculated by simply adding the contributions
of all components.

cp =
yi cp.i
(6.29)
i

where cp,i is the heat capacity in J/mol/K of pure species i.

There is a complication when using a composition dependent heat capacity
in CFD simulations. The general transport equation (6.1) for scalar quantities
needs to be modied slightly when it is written for the enthalpy. This is caused
by the fact that convective transport of heat is governed by the enthalpy of the
ow, while the diusion of heat (conduction) is governed by the temperature
gradient dT /dx, not the enthalpy gradient dH/dx. If the heat capacity is not
a function of temperature, the enthalpy and temperature are related by
H = cp (T T0 )

208

(6.30)

6.2 Models

3500
N2
NO
NH3
H2O
O2

3000

cp [J/kg/K]

2500
2000
1500
1000
500
0

200

300

400

500
T [K]

600

700

800

Figure 6.7: Piecewise linear interpolation scheme of the pure species heat
capacity of several gas components

where T0 is a chosen reference temperature. If we take T0 = 0, the (stationary)

enthalpy balance can be stated in the standard way ( = cp T in equation 6.1):

(cp T )v ndS =

T (cp T ) ndS +

qT d

(6.31)

If there is a gradient in the heat capacity (due to a change in gas composition),

the gradient of cp T in the rst right hand term can be nonzero even if the
temperature is constant. Therefore, heat could be conducted from a high cp
zone to a low cp zone and a temperature gradient would appear spontaneously,
which is in violation of the second law of thermodynamics. It is obvious that
it is no solution to choose = T , since that would introduce a similar error in
the convective term.
Another complication occurs when the heat capacity is a function of the
temperature. In this case equation (6.30) is not valid, and the denition of the

209

6. Modelling of packed bed reactors using CFD

specic heat should be used to relate enthalpy and temperature:

T
H(T ) =
cp (T )dT

(6.32)

T0

2
cp T

H, cp T [MJ/kg]

1.6

N2

N2

H2O

H2O

NH3

NH3

1.2

0.8

0.4

0
200

300

400

500

600

700

800

900

T [K]

Figure 6.8: Comparison of the enthalpy calculated using the integral (6.32)
and approximation (6.30) for some gas compounds.
The dierence between the simple approach (equation 6.30) and equation
6.32 is shown in gure 6.8. For many gas compounds (in this example oxygen,
nitrogen, nitrous oxide), the heat capacity is a weak function of temperature,
and the approximation is quite satisfactory. However, for other compounds
(e.g., water and ammonia), the dierence is larger. As can be seen in gure
6.8, the error in the enthalpy can be quite large (about 20%). If one would
calculate the temperature from an estimated enthalpy for ammonia at the
higher range of the graph, the error in the temperature value would be around
80 K.
Of course, gure 6.8 was calculated with a reference temperature of 273 K.
The error that is made can be reduced signicantly if the reference temperature
is taken at an average of the temperatures found in the ow. For instance, for

210

6.2 Models

ammonia if the temperature of the uid is between 273 and 873 K and T0 is
set to 573 K, the error is reduced by about a factor 2 compared with the case
above, where T0 = 273K.
It can be concluded that the approximate equation (6.30) can be used for
ow problems where the temperature dierences in the domain are not too
large and where the temperature dependence of the specic heat of the main
gas compounds is not too strong, i.e.,



cp (T0 )(Tmax Tmin ) 

(6.33)
1
 < 5%
Tmax


c
(T
)dT
p
Tmin
Here a fairly large error of 5 % is proposed since usually the specic heat
data calculated from engineering correlations will not be much more accurate.
The correct steady state enthalpy balance can be written as:



Hv ndS =
T cp T ndS + qT d
(6.34)
S

in discrete form:


(k Hkvk nk Sk ) =
(T,k cp,k Tk nk Sk ) + qT,P
k

(6.35)

To solve this equation, we can either use the enthalpy as a variable and calcu
late the temperature when the coecients are needed, or vice versa. Since it
is easier to calculate the enthalpy if the temperature is known than the other
way around, the temperature is chosen as a eld variable.
The face value of the enthalpy Hk is estimated by linear interpolation be
tween the two bordering cells:
Hk = k HP + (1 k )HN

(6.36)

or, when the approximation for the enthalpy is used,

Hk = k cp,P (TP T0 ) + (1 k )cp,N (TN T0 )
= k cp,P TP + (1 k )cp,N TN cp,k T0

(6.37)

The temperature gradient at the face centre Tk is calculated from the cell
centre temperature gradients in the normal way:
Tk = k TP + (1 k )TN

(6.38)

The temperature gradient at the cell centre is calculated from the temperature
dierence between the current cell centre and that of all neighbours by solving
TN TP = (T )P (xN xP )

(6.39)

211

6. Modelling of packed bed reactors using CFD

Of course, we can divide everything by the reference heat capacity cp,0 ,

 

  k Hk
cp,k
qT,P
T,k

(6.40)
vk nk Sk =
Tk nk Sk +
cp,0
cp,0
cp,0
k
k
Hence, compared to the constant cp case, we need a modied face centre tem
perature
k HP + (1 k )HN
Tk =
(6.41)
cp,0
and a modied diusivity
cp,k
T,k
(6.42)
T,k =
cp,0
Diusion coecients
The binary diusion coecients for all combinations of gas species are calcu
lated using the semi-empirical method of Fuller (Reid et.al (1987))
4.522 108 T 1.75
DAB = 

2
1/3
1/3

p MAB VA + VB
AB =
M

2
1/MA + 1/MB

(6.43)

(6.44)

where
DAB
T
p
Vi
MAB
MA

m2 /s
K
Pa
m3 /mol
kg/mol
kg/mol

Binary diusion coecient of A in B (or vice versa)

Local temperature
Local pressure
Molecular diusion volume of compound i
Average molar mass
Molar mass

The diusion coecient of each component in the mixture is then calculated

according to Blancs law (Reid et al (1987))
 yj
1
=
(6.45)
D
Di,m
ij
j
This is only accurate for diusion of a dilute component in a homogeneous
mixture, but there seems to be no better simple way to estimate gas mix
ture diusion coecients. The most correct way would be to implement the
Stefan-Maxwell diusion equation (instead of Ficks law) in the CFD code, but
that would increase the complexity and computational intensity of the calcu
lation considerably, while the accuracy of the ow solution would not increase
dramatically in any but a few exceptional cases.

212

6.2 Models

6.2.5

Flow resistance

The ow resistance of packed beds is computed according to the equations

developed in Chapter 4 and is introduced in the ow equations as a momentum
source term. In general the ow resistance is given by an equation of the form
dp
= Av + Bv|v|
(6.46)
dh
where A and B depend on local porosity, tortuosity, specic surface, density
and viscosity. The force exerted by the uid on the particles (and vice versa)
in a subvolume of the bed with cross sectional area A and length h is
therefore
Fd = A h (Av + B|v|v) = (Av + B|v|v)

(6.47)

The stationary momentum equation has the form (in direction i)

viv ndS =
ijij ndS pi ndS + qvi d
S

(6.48)

where each term has the units of force. After discretisation, the source term
has the form


kg

Q = S + S
[]
(6.49)
s
where the rst term contains the explicit part of the source term and the second
term the implicit part. The friction term is written as an implicit source:
Qui = (A + B|vk1 |) vk

(6.50)

where vk1 is the velocity from the previous iteration and vk for the current
iteration. The coecients A and B are calculated from
2
23D,a
s2v
(6.51)
3


3

kt 3D,a 3D,a sv
B=
(6.52)
3
Note that, although the ow resistance is introduced as a source (sink) term
in the momentum equations, the inuence of a homogeneous packed bed on
the velocity eld in e.g. a simple tube is small. The major inuence of the
resistance will be on the pressure eld. The dissipated kinetic energy must be
balanced by a transformation of potential energy.
The kinetic energy that is dissipated in the bed is of course turned into
heat, so there should be a source term for the enthalpy balance equal to

A=

Qh = vk2 (A + B|vk1 |)

(6.53)

213

6. Modelling of packed bed reactors using CFD

6.2.6

Dispersion

The dispersion coecients are calculated according to the theory developed in

chapter 5. In the CFD code, the dispersion coecients enter the equation in
the rst right hand term of equation 6.1. However, in the standard approach,
dispersion is treated as an isotropic phenomenon, while in a packed bed it
depends on the direction; in the direction of the ow there is axial dispersion
and perpendicular to the ow there is radial dispersion, and the dispersion
coecients for these directions dier.
It should be noted that in the majority of packed bed reactor designs, axial
dispersion is not an important factor and the solution of a computation would
not be dierent if the axial dispersion coecient would be set to zero or set to
be equal to the radial dispersion coecient. This would simplify the equations
in this section rather drastically. However, there are cases (for instance at low
ow rates or in low ow rate regions of the domain) where axial dispersion
does become important.
The task at hand is to calculate an average dispersion coecient for the
centre of face S of the control volume, for a given face average velocity vector
and axial and radial dispersion coecient.
The dispersion term in integral form for a face of a control volume is:

ndS

(6.54)
S

which can be written in Cartesian (global) coordinates as:

ndS
=
x nx + y ny + z nz dS
x
y
z
S
S
or, alternatively in the local coordinate system of face S:

ndS
=
1
n1 + 2
n2 + 3
n3 dS
x1
x2
x3
S
S

(6.55)

(6.56)

where x1 is the coordinate parallel to n

and x2 and x3 are parallel to face S.
Since obviously n2 = 0 and n3 = 0 and n1 = 1 by denition,

dS
(6.57)
ndS
=
1
x1
S
S
Therefore, what we need to know is the dispersion coecient parallel to n

(normal to face S). This dispersion coecient depends on the angle between
the velocity vector and the face normal. If the velocity is normal to the face,
the dispersion coecient is equal to the axial dispersion coecient; if the

214

6.2 Models

velocity if parallel to the face, the dispersion coecient is the radial dispersion
coecient. Furthermore, the dispersion coecient should always be between
the values of the radial and axial dispersion coecients. If Dr = Da , then the
dispersion coecient should be independent of the direction of the velocity.
Take velocity vector v. Then the axial dispersion vector will be:
v
da =
Da
|v|

(6.58)

Now dene two vectors perpendicular to v:

v y vz
v,1 = vz vx
vx v y

vy (vx vy ) vz (vz vx )
v,2 = v v,1 = vz (vy vz ) vx (vx vy )
vx (vz vx ) vy (vy vz )

(6.59)

(6.60)

Then we have two radial dispersion vectors,

v,1
Dr
d r,1 =
|v,1 |

v,2
dr,2 =
Dr
|v,2 |

(6.61)

and the dispersion in each coordinate direction can be calculated from the
dispersion in the local coordinate system (da , d,1 d,2 ):


2
2

+
d
+
d
d2
,1,x
,2,x
x
 a,x

2
2
2

y
(6.62)
=
= da,y + d,1,y + d,2,y


z
2
2 + d2
da,z
,1,z + d,2,z
Some calculus shows that this is equal to:
 2

2
 Avx2 + Dr2
Da2 Dr 2

+
D
Av
A=
=
r
 y
|v|2
Avz2 + Dr2

6.2.7

(6.63)

Pellet scale models

Chemical reaction
Chemical reaction in the catalyst pellets give rise to source terms for compo
nent mass balance equations and the enthalpy balance in the CFD simulations.
These sources are calculated based on the local porosity, specic outer particle

215

6. Modelling of packed bed reactors using CFD

surface area, temperature, composition and gas velocity using the approach

presented in Chapter 2.
The concentrations are calculated in terms of mass fractions. The source
term q in equation 6.1 has units kg/m3 /s []. Therefore for component i
the source term is specied in kg[i]/m3 /s.
Transfer of mass and heat between catalyst particles and uid
For packed bed chemical reactors, the transfer of heat and mass from the uid
to the particles of the bed is an important mass transfer step. For strongly
endothermic or exothermic reactions, the temperature of the particle can be
quite dierent from the temperature of the uid in its vicinity. Likewise, for
fast reactions, there can be a dierence between the concentration of reactants
and products near the particle surface and in the surrounding uid. This
dierence is caused by the fact that the uid velocity at the particle edge is
zero, so convective transport of heat and mass is limited in the neighbourhood
of the particle; this can be seen as a boundary layer around the particle where
heat and mass transfer is by molecular transport (diusion, conduction) only.
The rate of conversion in the bed can be determined by this transport
depending on the reaction and mass transfer parameters so it has to be
taken into account in a xed bed reactor model (as was shown explicitly by
Wijngaarden and Westerterp, 1989 and 1993).
Heat and mass transfer between the particle and the uid is governed by
the particle Sherwood (for mass transfer) and Nusselt (for heat transfer) di
mensionless numbers:
Shp, =

kp, dp
Dm,

Nup =

kp,h dp
f

(6.64)

where kf p, is the heat transfer coecient between the uid and the particle for
quantity , dp the particle diameter and Dm, the molecular diusion coecient
of component . The Sherwood and Nusselt numbers depend on the ow
conditions and are usually determined from empirical correlations as a function
of the particle Reynolds and Schmidt respectively Prandtl numbers.
Since the ow and the boundary layer around a spherical particle is dis
turbed by neighbouring particles, the Sherwood number for a particle in a
packed or uidised bed diers from that of a single sphere in an innite uid.
However, for high porosities ( 1), the packed bed relation should be equal
to the single-sphere relation:
1

lim Shp = 2.0 + 0.66Re 2 Sc

3
1

216

(6.65)

6.2 Models

For more realistic packed bed porosities, the value of Sherwood at low Reynolds
numbers will go to a nite value that is not necessarily equal to 2.
lim Shp = f ()

Re0

lim f () = 2

(6.66)

1

where f is a function of the bed packing.

Commonly used functions for the particle Sherwood number (e.g. Dixon,
1979 and 1988; Papageorgiou and Froment, 1995) are of the form
Shp =

0.255 2/3 1/3

Re Sc


Re > 100

(6.67)

This equation has the disadvantage that it is not valid for lower Reynolds
numbers and also does not conform to equation (6.65).
Gunn (1978) derived a relationship for the Sherwood number in packed
beds based on a statistical model of the bed. The result of this model is that
lim Shp = 2.36

Re0

1


(6.68)

It is clear that this result does not follow the limiting behaviour for high
porosities. This may be caused by the fact that the packed bed model used
in the derivation is not valid for highly porous systems. As a relation that is
valid for a wider Reynolds range, Gunn postulates:
Shp =(7 10 + 52 )(1 + 0.7Rep0.2 Sc1/3 )
+ (1.33 2.4 + 1.22 )Re0p.7 Sc1/3

Re < 105

(6.69)

The polynomial dependency on the porosity is somewhat arbitrary, and is

chosen such that as Reynolds goes to zero, equation 6.68 is approximately
followed for common bed porosities in the neighbourhood of 0.4, while for
porosities of 1.0, the single-sphere value is found. The limiting relations are:
lim Shp = 7 10 + 52

(6.70)

lim Shp = 2 + 1.4Re0.2 Sc1/3 + 0.13Re0.7 Sc1/3

(6.71)

Re0
1

The latter relation is not equal to equation (6.65), but is very similar for
Re < 2000.
Gnielinski (1988) gives:



2
2
Shp = (1 + 1.5(1 )) 2 + Shlam + Shturb
(6.72)

217

6. Modelling of packed bed reactors using CFD

with

Shlam = 0.644Re1/2 Sc2/3

(6.73)

0.8

Shturb =

0.037Re Sc
1 + 2.443Re0.1 (Sc2/3 1)

(6.74)

The limiting relations are:

lim Shp = 2 + 3(1 )

2
lim Shp = 2 + Sh2lam + Shturb

Re0

1

(6.75)
(6.76)

The latter is very close to equation (6.65); less than 10% dierence for Re <
2000. However, equation (6.72) can only be used for Sc > 1 (which is often
the case since molecular diusion coecients are usually low compared to the
uid viscosity).
For heat transfer, the same relations are used as those for mass trans
fer, with the dimensionless numbers replaced by their heat transfer analogies
(Sh Nu, Sc Pr). This analogy is valid as long as the heat transfer mecha
nism is the same as the mass transfer mechanism (i.e., diusion and conduction
through the boundary layer). At higher temperatures, heat will also be trans
ported by radiation. However, as high temperature applications are almost
always gas-solid systems, heat transfer between the particle and the uid by
radiation will be less important; most radiative heat transfer will be from one
particle to its neighbour. This eect is already taken into account in the axial
and radial heat dispersion terms.
The dierent models are compared in gure 6.9. As can be seen, the
results for the models of Gunn and Gnielinski are very similar (for Re > 10),
so the choice between the two is somewhat arbitrary. The former more closely
corresponds to the high-Reynolds model used by Dixon (1988). The model of
Gunn is chosen here because of its somewhat simpler formulation.
To calculate the mass transfer between the solid particle and the bulk of the
uid, often the assumption is made that the transport term for dierent com
ponents are independent. For instance, Papageorgiou and Froment (1995)
write
m,i = kp sv (ca,i cb,i ) = (1 )Re,i

(6.77)

where m,i is the mass ux of component i (mol/m3 /s), kp is the mass trans
fer coecient (m/s), sv the specic particle surface area (m2 /m3 ), cs,i the
concentration of component i at the particle edge (mol/m3 ), cb,i the bulk con
centration (mol/m3 ) and Ri,e the eective reaction rate per volume catalyst

218

6.2 Models

1000

Sh [-], Nu [-]

100

Gunn (1978)

Gnielinski (1988)

Dixon (1988)

Single sphere

10

0.1
0.1

10

100

1000

10000

Re [-]

Figure 6.9: Dierent literature models for the particle Sherwood number (or
Nusselt number for heat transfer) as a function of the particle Reynolds number
for a packed bed porosity of 0.4 and Sc(= Pr) = 10.

material (mol/m3 /s). However, this approach leads to problems for multi
component reaction systems. As an example, consider the following simple
isothermal oligomerisation reaction
A B
Let us assume that the uid is a gas that consist solely of A, B and inert
material N. Since the pressure gradient over the boundary layer is negligible,
it follows for the partial pressures of A and B:
p pN = pb,A + pb,B = ps,A + ps,B

cb,A cs,A = cs,B cb,B (6.78)

In steady state, the mass of A that is consumed and therefore transported to

the particle is equal to the mass of B that is produced. Hence
m,A = m,B

(6.79)

219

6. Modelling of packed bed reactors using CFD

But also, according to 6.77

m,A = kp,A sv (cs,A cb,A )

m,B = kp,B sv (cs,B cb,B )

(6.80)

Hence, the mass transfer coecients for A and B must be related by:
kp,A = kp,B

(6.81)

This is unlikely as the mass transfer coecients can be calculated indepen

dently from (for instance) equation 6.69. In reality, the stream of molecules A
will cause a drift ux toward the particle that will hinder molecules B to diuse
outward. In this simple system, it is quite easy to correct for the drift ux,
but for non-isothermal multi-component reaction systems, another approach
must be taken.
A more fundamental way to approach the mass transfer problem is by the
Maxwell-Stephan equation:
1 dxi  vj vi
xj
=
(6.82)
Di,j
xi dz
j
where xi is the mole fraction of component i, z is the distance into the bound
ary layer, vi is the velocity of component i (equal to the ux divided by the
concentration) and Di,j the binary Maxwell-Stephan diusivity of components
i and j. This equation states that the driving force (the chemical potential gra
dient on the left hand side) is in equilibrium with the resistance that molecules
i feel from all other components in the boundary layer. The equation is some
what cumbersome to solve, and since it is only a small step in a complete
packed bed reactor model, a simplied approach (Wesselingh and Krishna,
1990) is followed here. The dierential equation (6.82) is approximated by its
dierence form:
xi  vj vi
=
xj
(6.83)
ki,j
xi
j
Here ki,j is the binary mass transfer coecient for species i and j; the velocities
vi and vj are calculated at the mean boundary layer concentration and tem
perature. This equation is accurate enough for most engineering applications,
especially in the light of the uncertainties in mass transfer and other param
eters in the model, while retaining the basic features of the full dierential
equation. Non-equimolar diusion will also have an eect on the heat transfer
through the boundary layer, causing a thermal drift:

i
h = kp,h T +
i
H
(6.84)
i

220

6.3 Model verication and validation

i is the enthalpy of species i. Equations (6.83) and (6.84) can be

where H
solved simultaneously with the diusion-reaction equations inside the catalyst
particles.

6.3

Model verication and validation

The code that is validated is a combination of the Dolfyn ow solver and the
models for packed bed reactors developed in this work. Since the ow regime
and geometry are relatively simple (low velocity, mostly cylindrical geometry,
rectangular grid), the accuracy of the ow solution itself is quite sucient
(much higher than the accuracy of the catalyst bed model and the accuracy
in which parameter like bed porosity are known). Therefore, simulations done
with the complete CFD model (Dolfyn and packed bed code) can be used to
verify or validate the packed bed code.
The verication of the model focuses on the correctness of the results in
numerical sense, for ow in a simple cylindrical packed bed. The validation
of the sub-models (porosity proles, velocity proles, dispersion of mass and
heat) has been done in the previous chapters. Here this will repeated for the
full 2-dimensional CFD model, comparing results to literature data for steam
heated tubes (ow and heat transfer) and a laboratory-scale ethane oxidation
reactor (ow, mass and heat transfer).

6.3.1

Flow in a packed tube

To establish a base case, the ow resistance in a cylindrical, isothermal tube

was calculated using the CFD model as well as a hand calculation using the
classical (dierential) Ergun equation. The pressure at the outlet of the tube
was taken as atmospheric and the inlet pressure was increased to set a given
ow rate. A symmetrical boundary condition was applied at the wall (slip wall)
and no radial particle distribution proles were taken into account. Therefore,
the geometry and boundary conditions closely approximate the assumptions
of the hand calculation. The results are shown in Figure 6.10. It can be seen
that, as expected, the CFD model and the hand calculation give very similar
results in this case (the root mean square of the dierence is about 1.6 %).
The data deviates a little from a straight line in the log-log plot because of the
eect of decreasing density with decreasing gas pressure.
For the one-dimensional case, the number of cells in the CFD model does
not inuence the results very strongly. There need to be enough cells in the
axial direction to capture the pressure prole so that the density variation with
pressure is accurately represented. Since there are no proles in radial direction

221

6. Modelling of packed bed reactors using CFD

10000
No wall
Hand calculation

p [Pa]

1000

100

10

0.1
1.0E+01

1.0E+02

1.0E+03
Re [-]

1.0E+04

1.0E+05

Figure 6.10: Comparison between a hand calculation based on the classical

Ergun equation with variable density and a CFD model that is a close approx
imation of the assumptions of the hand calculation (see table 6.1)

Table 6.1: Calculation parameters

number of cells in bed zone (x y)
number of iterations
uid
wall condition
porosity prole
inlet condition
temperature
N2
H2 O
NO
NH3
outlet condition
temperature
chemical reaction

222

for a one-dimensional isothermal tube

70 30
2500
ideal gas mixture
symmetry
none
specied velocity and composition
20 C
0.80 kg/kg
0.10 kg/kg
0.05 kg/kg
0.05 kg/kg
absolute pressure 101325 Pa
constant
none

6.3 Model verication and validation

Table 6.2: Calculation parameters for an isothermal tube

number of cells in bed zone (x y) 70 ny

number of iterations
2500
uid
ideal gas mixture
wall condition
no slip adiabatic
porosity prole
yes
bed diameter
1m
D/d
12.5
inlet condition
specied velocity and composition
velocity
0.5 m/s
temperature
20 C
N2
0.80 kg/kg
H2 O
0.10 kg/kg
NO
0.05 kg/kg
NH3
0.05 kg/kg
outlet condition
absolute pressure 101325 Pa
temperature
constant
chemical reaction
none
for any variable, the number of cells in radial direction does not inuence the
results.
When the wall eect is taken into account, the number of cells in radial
direction becomes important because the radial proles of porosity, tortuosity
and specic area needs to be captured. Also, the velocity prole at the wall
needs to be captured to get a good estimate of the wall friction. In order
to increase the number of cells near the wall, a variable grid spacing is used
according to:
 f
rj
j
=
R
n

(6.85)

where ri is the radial position of division line j, R is the radius of the tube, n
the number of divisions in radial direction and f the spacing factor. Here, a
value of 1.25 is used for f .
A number of simulations were performed with dierent numbers of cells
in radial direction for the conditions given in table 6.2. The radial velocity
prole in the bed is shown as a function of the number of cells in the radial
(y) direction is shown in gure 6.11, and the deviation from the last solution
is shown in 6.12. It can be seen that in this case at least 40 cells are needed
in radial direction to approach the nal solution to 1-2 %.

223

6. Modelling of packed bed reactors using CFD

1.2
ny
10
20
30
40
50
60

v [m/s]

0.8
0.6
0.4
0.2
0
0.2

0.25

0.3

0.35
r [m]

0.4

0.45

0.5

Figure 6.11: Radial velocity proles in an isothermal packed tube for dierent
radial mesh densities. Simulation parameters according to table 6.2.

The pressure drop over the bed is shown in gure 6.13 as a function of the
number of radial cells. It can be concluded that with a radial cell count of 50,
the radial proles are well represented for this case (less than 1% dierence in
pressure drop). For higher tube to particle diameter ratios a higher resolution
near the wall may be required to capture the radial proles, but the relative
eect of the wall zone on the ow in the tube becomes smaller for D/d > 12.5.
Conversely, for lower ratios, the radial proles are more easily captured in good
resolution, but the eect of the wall zone on the ow in the tube is larger. In
this chapter, we will look at beds with D/d < 12.5 and use ny = 50.
The calculations are performed using a rst order upwind discretization
scheme. The case with ny = 50 was recomputed using a second order central
dierence scheme. The velocity prole in the bed was nearly identical to the
rst order case (the root mean square of the dierence was about 5e-5 m/s).
The pressure drop over the bed was also very similar (183.5 versus 183.3 Pa).
Therefore, it is shown that for macroscopic ow through packed beds like these,
rst order calculations can be used without signicant loss of accuracy.

224

6.3 Model verication and validation

Deviation from final solution [m/s]

0.04

0.03

0.02

0.01

0
0

10

20

30

40

50

60

Number of radial cells

Figure 6.12: Mean root square of the deviation of the velocity prole and the
nal solution (60 radial cells) as a function of the number of radial cells. For
the simulations with high ny (marked *), the number of axial cells (nx) in the
bed was increased to 120 to improve cell shape.

The pressure drop over the bed described by table 6.2 with D/d = 12.5
and the same model with a larger particle size so that D/d = 5 is calculated at
dierent inlet velocities. The results are compared with the earlier simulations
without wall eect (which is the same as the result of the classical Ergun
equation) in gure 6.14. Obviously, the pressure drop of the low bed-to-particle
diameter ratio decreases drastically due to the bypassing of ow through the
wall zone. In fact, the ratio between the pressure drop in a tube with xed,
nite D/d to that of an identical tube with no wall eect taken into account
is constant (not a function of the ow rate), at least at higher ow rates.
The radial velocity proles for the case of table 6.2 for dierent inlet ve
locities are shown in gure 6.15. The same proles are compared to the 1
dimensional simulations done in chapter 4 in gure 6.16, where they are nor
malised with the inlet velocity. It can be seen that for the lower ow cases
(Re 100), the shape of the proles changes with the ow rate; the maximum
velocity peak becomes more pronounced. At higher ow rates (Re > 100),

225

6. Modelling of packed bed reactors using CFD

200

10
0

p [Pa]

-10
160
-20
140
-30
120

-40

100
0

20

40

60

80

deviation from final value [%]

*
180

-50
100

ny [#]

Figure 6.13: Pressure drop over the packed bed as a function of the number of
radial cells, simulations according to table 6.2. For the simulations with high
ny (marked *), the number of axial cells (nx) in the bed was increased to 120
to improve cell shape.

the turbulent term is dominant and the shape of the proles remains about
the same. It can be seen that there is a good correspondence between the
1-dimensional model and the 2-dimensional CFD model. The CFD results are
compared to the same literature points used for the 1-dimensional model in
chapter 4 in gure 6.17. The correspondence of the model with the measured
data is as good as may be expected; the oscillations in the velocity prole
damp out at a somewhat smaller distance to the wall, which is inherited from
the porosity prole used.

6.3.2

Heat transfer in a steam heated packed tube

In order to verify the transport of heat through dispersion, a steam heated

packed tube is calculated. The geometry is basically the same as in the previous
section, except that the wall is not taken as adiabatic but has a constant
temperature of 100 C. The temperature of the inlet stream is 20 C. As

226

6.3 Model verication and validation

1000
D/d = 5
D/d = 12.5

p [Pa]

800

D/d =

600

400

200

0
0

0.2

0.4

0.6

0.8

v0 [m/s]

Figure 6.14: Pressure drop as a function of inlet velocity for the isothermal
packed bed of table 6.2 with dierent bed-to-particle diameter ratios.

Table 6.3: References and conditions for data points in gure (6.17).
symbol w0,m [m/s] dp [mm]
Rep D/d note
a
6.3
8.0
268
0.5
1
a
4.5 76-450 11.1
0.25-1.5
2
a
5.1
9.8 165-987
0.25-1.5
3
a
7.1
7.0
235
0.5
4
a
5.1
9.8
658
1
5
b
8.6
4 9.30
0.004
6
b
8.6
77 9.30
0.076
7
b
8.6
103 9.30
0.102
8
b
8.6
532 9.30
0.526
9
a) Bey and Eigenberger (1997). Fluid: air; velocity proles measured downstream
of a packed bed supported by a 3.5 mm long monolith to preserve the ow prole.
b) Giese et al. (1998). Fluid: special liquid mixture (viscosity 8.5 106 m2 /s);
velocity measured inside bed by Laser-Doppler velocimetry.

227

6. Modelling of packed bed reactors using CFD

2.1

v0 [m/s]

1.8

v [m/s]

1.5

0.01

0.1

0.25

0.5

1.0

1.2
0.9
0.6
0.3
0
0.2

0.25

0.3

0.35
r [m]

0.4

0.45

0.5

Figure 6.15: Velocity proles for simulations based on table 6.2 with dierent
inlet velocities.

the ow eld is now no longer isothermal, the physical properties (density,

viscosity, conductivity and specic heat) vary with position.
Grid density variation calculations were performed for beds with bed-to
particle diameter ratio of 5 and 12.5. The resulting temperature proles are
shown in gures 6.18 and 6.19. It can be seen that for a radial cell count of 50
or more, the temperature proles are not inuenced by the grid density.
To show the inuence of the ow rate, calculations were performed at
dierent inlet Reynolds numbers. The parameters for these calculations are
shown in table 6.4.
As expected, it can be seen that the inuence of the wall region increases
with increasing Reynolds numbers. Figure (6.21) gives the radial thermal con
ductivity as a function of the distance to the wall and gure (6.22) gives the
axial thermal conductivity. These gures show the transition from molecular
conduction to conduction by convective mixing. It can be seen that for low
Reynolds numbers, the solid conductivity (which is much higher than the con
ductivity of the stagnant uid) is dominant, so the conductivity increases with

228

6.3 Model verication and validation

1D Re= 10
1D Re= 1000
2D Re= 4.0
2D Re= 39.6
2D Re= 395.9
2D Re= 989.9

v/v0 [-]

0
0.2

0.25

0.3

0.35
r [m]

0.4

0.45

0.5

Figure 6.16: Velocity proles from 2-D CFD simulations compared to earlier
1-D simulations for dierent values of the Reynolds number

Table 6.4: Values of the parameters used for ow rate dependency calculations
Quantity
symbol value
D
0.075 m
Bed diameter
L
0.2 m
Bed length
dp
7.5 mm
Particle diameter
10
Bed-to-particle diameter ratio D/dp
0.38
Bed bulk porosity
b
1.41
Bed bulk tortuosity
b
1.1
Axial wall tortuosity
w
Particle material
ceramic
Particle conductivity
s
1.2 W/m/K
Particle emissivity
0.85
R
Fluid
air
M
Molar mass
0.028 kg/mol
p
Pressure
105 Pa
Re0
Reynolds number at inlet
0.1 - 1000

229

6. Modelling of packed bed reactors using CFD

4.0
3.5
3

Rep = 4

Rep = 39.6

Rep = 989.9

v0/vm [-]

2.5
2
1.5.
1
0.5
0
0

5
x/rp [-]

Figure 6.17: Velocity proles from 2-D CFD simulations compared to literature
data, see table 6.3.

decreasing porosity. For higher Reynolds numbers, the uid conductivity (due
to mixing) is dominant, so the conductivity increases with uid velocity. As
the velocity is (in general) higher where the porosity is higher, the conductivity
plot is inverted: where there is a maximum conductivity for low ow (e.g., at
0.7 particle diameters from the wall), there is a minimum conductivity for high
ow. Directly at the wall, radial mixing is absent, so the conductivity there is
not inuenced much by the Reynolds number.
To our knowledge, the radial prole of axial and radial conductivity has
not been measured directly. This would require a large number of temperature
measurements at a large number of radial positions in the packed bed. The
temperature proles that can be found in existing literature do not have a
sucient resolution to compute these proles. It would be interesting to see if
the proles predicted by the packed bed model (and especially the predicted
inversion of the peaks) could be reproduced experimentally.

230

6.3 Model verication and validation

100

ny
20

30

80

50

T [C]

100

60

40

20

x/rp [-]

Figure 6.18: Temperature prole for D/d=5, for dierent radial cell counts

231

6. Modelling of packed bed reactors using CFD

100

ny
20

30

80

50

T [C]

100

60

40

20

6
7
x/rp [-]

10

11

12

13

Figure 6.19: Temperature prole for D/d=12.5, for dierent radial cell counts

232

100

100

99

80

98

60

Re=0.1
Re=1.0
Re=10
Re=100
Re=1000

97

T [C]

T [C]

6.3 Model verication and validation

40

96

20
0

x/rp [-]

Figure 6.20: Temperature proles in a steam-heated tube for dierent inlet

Reynolds numbers. Parameters for these calculations are given in table 6.4.

233

6. Modelling of packed bed reactors using CFD

aR [cm2/s]

24

Re=0.1
Re=1.0
Re=10
Re=100
Re=1000

20

16

12

aR [cm2/s]

0
0

x/rp [-]

Figure 6.21: Radial thermal conductivity in a steam heated tube for dierent
inlet Reynolds numbers. Parameters as given in table 6.4. Note that the lines
for Re=0.1 and Re=1 almost coincide.

234

6.3 Model verication and validation

aA [cm2/s]

500

Re=0.1
Re=1.0
Re=10
Re=100
Re=1000

400

300

200

100

aA [cm2/s]

0
0

x/rp [-]

Figure 6.22: Axial thermal conductivity in a steam heated tube for dierent
inlet Reynolds numbers. Parameters as given in table 6.4.

235

6. Modelling of packed bed reactors using CFD

To validate the results of the CFD model, it is compared to the so-called

wall heat conduction (WHC) model developed in recent years by Winterberg
et.al. (2000). The WHC model is presented as the new standard model to
compute heat transfer between the wall and a packed bed, and is tuned to
and validated with a broad range of literature data for dierent systems and
boundary conditions in axisymmetric geometries (Winterberg and Tsotsas,
2000). The detailed dispersion model was already compared to the WHC
model using a 2-dimensional, constant property model in Chapter 5. Here it
is compared to the results of the full CFD solution. For this comparison the
WHC model is also implemented in the CFD code.
Evolving from a lumped parameter model, the WHC model has some prac
tical disadvantages when applied in a CFD code. For instance, the velocity in
the centre of the bed is needed to calculate the eective radial conductivity in
at any location in the cell. In a CFD code, this velocity usually not available
at a cell level. More fundamentally, information that cannot be deduced from
the state of a cell and its neighbours in a CFD simulation, but depends on
cells that are far removed from that cell, can also in reality not have a direct
inuence on the cell itself. Therefore, all important parameters should depend
only on local variable values. Therefore, the fact that in the WHC model ther
mal conductivity in each cell depends on the Peclet number in the centre of
the bed cannot be physically correct. In fact, it may well be that the geometry
of the bed is such that it is hard to determine where this centre is.
In the current implementation, as a workaround, the velocity in the centre
of the bed is estimated in the CFD code at the start of each iteration by
calculating the average velocity in the bed with the third power of the wall
distance as weighing factor.
The detailed model gives similar results as the WHC model for steam
heated tubes, as can be seen in gure 6.23, although for higher values of the
Reynolds number the temperature in the centre of the bed seems to be under
estimated somewhat by the detailed model. The behaviour of the CFD model
is not dierent from the 2-dimensional nite dierence model shown in the
previous chapter. Again, it should be noted that the WHC model was tuned
to the experimental data while the detailed model was derived and based on
numerical simulation results for packed beds of identical spherical particles.
Some dierences could arise from the fact that the data set contains measure
ments for (somewhat) non-spherical and/or non-homogenous packings.

236

6.3 Model verication and validation

A comparison with a dierent set of experimental data is shown in gure

6.24. It can be seen that the t in this case is quite good. The temperature
jump in the near bed zone, which is caused by a combination of the ow
prole near the wall and the radial eective conductivity prole, appears to
correspond well to the measured prole. The temperature in the centre of the
bed seems to be slightly underestimated.

100
WHC model

90

Detailed model

80

Measured values
T [C]

70
60
50
40
30
20

4
x/rp [-]

Figure 6.23: Comparison between the WHC model of Winterberg et.al. (2000)
and the detailed model developed in this work for a steam heated tube, ce
ramic particles (s = 1.2 W/m/K) and air (Re=480). The large symbols are
measured values after Dixon (1988), as published by Winterberg et.al. (2000).
Note that the WHC model was tted to this data and the detailed model was
not.

237

6. Modelling of packed bed reactors using CFD

120

Calculated
100

Measured

T [C]

80

60

40

20

4
x/rp [-]

Figure 6.24: Results of the detailed model compared with experimental data
given by Derckx and Dixon (1996) for a heated wall 2 tube, 1/4 ceramic
spheres with air (Re=550, L=0.1175 m).

238

6.3 Model verication and validation

6.3.3

Catalytic oxidation of ethane

The catalytic oxidation of ethane at intermediate temperatures is used by Vort

meyer and Haidegger (1991) to validate dierent approaches to heat transport
calculations for wall-cooled reactions. The catalytic reaction has an acceptable
rate at temperatures that can be achieved in laboratory conditions, and where
the rate of the homogeneous reaction is negligible.
C2 H6 + 3.5O2 2CO2 + 3H2 O

(6.86)

Vortmeyer and Haidegger give overall reaction rate equations for this re
action in a packed bed of 4 mm diameter, PdO coated, -alumina particles
with a porosity 0 of 40 %. The original article contains a number of printing
errors, so the corrected equations are given here:

!  Tb1 +T1 9  y n
EA

exp
fs
T 597.36K
k
I
RT
Tb1
yb




kII exp (0.05407T ) y

fs
597.36K > T > 626.03K
yb
$

r0 =



2
3

kIII 1000
3.62 1000

T
T

%  n

1000 3

+3.72 T
0.77 yyb fs
T 626.03K
(6.87)
where
&
n=
T1
T2
Tb1
Tb2
kI
kII
kIII
EA
yb
fs

(0.024T 13.63)
1

Tb2 +T2
Tb2

6.231

T 658.75K
T > 658.75K

= |Tb1 T |
= |Tb2 T |
= 570.0 [K]
= 581.3 [K]
= 3.0 106 [mol/m3 /s]
= 22.5 1017 [mol/m3 /s]
= 10.0 [mol/m3 /s]
= 94470 [J/mol]

= 0.002

= 1.067

239

6. Modelling of packed bed reactors using CFD

The reaction rate at a given location in the bed with porosity  is given by
r() = r(0 )

1
1 0

(6.88)

The reactor consists of a 40 mm diameter tube, lled with 4 mm particles.

The active length of the bed is 16 mm, but it is sandwiched in between two
sections of inactive particles in order to get a well-established ow prole.
The concentration of ethane at the inlet is 0.5 mass % in air. The physical
properties of the packing were taken from the original publication. Two cases
given by Vortmeyer and Haidegger were simulated in the detailed CFD model:
at a mass ow rate of 0.2 and 0.25 kg/m2 /s. This gives particle Reynolds
numbers (at inlet conditions) of 25 and 32. The inlet temperature was set to
603 Kelvin and the wall temperature was 601 K. The calculated temperature
eld clearly shows the hot spot in the rst part of the catalyst bed. The
results of the calculations are compared to the measured data by Vortmeyer
and Haidegger in gures 6.25 to 6.28.
740

Rep = 25

Measured (r=0)
Measured (r=R)

720
700

Calculated (r=0)
Calculated (r=R)

T [K]

680
660
640
620
600
580

0.1

0.2
0.3
Axial coordinate [m]

0.4

0.5

Figure 6.25: Axial temperature prole in the centre and at the wall of a cat
alytic ethane oxidation reactor. CFD simulation results compared with mea
sured data by Vortmeyer and Haidegger (1991) for a particle Reynolds number
of 25.
It can be seen that the measured proles are reproduced quite well; about
as good as the model proposed by Vortmeyer and Haidegger and much better
than the alternative models presented by them (notably the wall heat transfer
model). The hot spot temperature seems to be slightly underestimated by our

240

6.3 Model verication and validation

Mass fraction Ethane []

Rep = 25

Measured (r=0)
Measured (r=R)

Calculated (r=0)
Calculated (r=R)

4
3
2
1
0

0.1

0.2
0.3
Axial coordinate [m]

0.4

0.5

Figure 6.26: Axial ethane mass fraction prole in the centre and at the wall of
a catalytic ethane oxidation reactor. CFD simulation results compared with
measured data by Vortmeyer and Haidegger (1991) for a particle Reynolds
number of 25.

740

Rep = 32

Measured (r=0)
Measured (r=R)

720
700

Calculated (r=0)
Calculated (r=R)

T [K]

680
660
640
620
600
580

0.1

0.2
0.3
Axial coordinate [m]

0.4

0.5

Figure 6.27: Axial temperature prole in the centre and at the wall of a cat
alytic ethane oxidation reactor. CFD simulation results compared with mea
sured data by Vortmeyer and Haidegger (1991) for a particle Reynolds number
of 32.

241

6. Modelling of packed bed reactors using CFD

Mass fraction Ethane []

Rep = 32

Measured (r=0)
Measured (r=R)

Calculated (r=0)
Calculated (r=R)

4
3
2
1
0

0.1

0.2
0.3
Axial coordinate [m]

0.4

0.5

Figure 6.28: Axial ethane mass fraction prole in the centre and at the wall of
a catalytic ethane oxidation reactor. CFD simulation results compared with
measured data by Vortmeyer and Haidegger (1991) for a particle Reynolds
number of 32.

model. Especially for the higher ow case, the hot spot location is shifted one
or two centimetres upstream compared to the measurement data. It might be
that the axial dispersion is slightly overestimated by our model. However, it
could also be that the dierence is caused by a measurement error (for instance
because of the inuence of the measuring probe on the ow in the bed).

242

6.4 Demonstration

6.4

Demonstration

6.4.1

Reduction of NO in a packed tube

As a rst model reaction, the selective catalytic reduction (SCR) of nitrogen

oxide with ammonia is selected. This process is often used to remove NOx
from ue gas streams to reduce air pollution. A well-known catalyst for these
reactions is vanadium oxide. The NO reduction reaction can be described as:
4NO + 4NH3 + O2 4N2 + 6H2 O

(6.89)

As ammonia is at least as harmful as nitric oxide, neither NH3 nor NO will

be present in a large excess. Under these conditions, it is expected that the
rate of reaction will depend both on the NO concentration and on the NH3
concentration. The following reaction rate equation is used:
rN O = k0 e

EA
RT

cnN O cm
N H3

(6.90)

with
k0
EA
n
m

0.5301
12240
0.5284
0.6354

(mol/m
3 )1nm /s
J/kg

In the ue gas of a gas engine, the NO concentration could be as high

as 6000 ppm. For the model reaction a concentration of 0.5 mole% will be
used, with a stoechiometric amount of ammonia, in air. The reaction takes
place at temperatures of about 250 C. Due to the relatively low reactant
concentrations, the eect of the heat of reaction can be neglected. Therefore,
the reactor will be isothermal (provided of course it is well insulated).
In general, the design of a tubular reactor (i.e., tube diameter, particle di
ameter) will be a trade-o between the pressure drop, the catalyst bed volume
needed and heat transfer to or from the reactor. For an isothermal (or au
tothermal) tubular reactor, a low bed-to-particle diameter design will usually
not be the optimal solution. If a high pressure drop is permitted, a small cat
alyst particle size will usually be chosen. On the other hand, if a low pressure
drop is required, a large diameter bed with a short length becomes attrac
tive. For very shallow beds, an equal distribution of the ow over the bed will
become dicult and an alternative geometry (for instance the Radial Flow
Reactor discussed below) may have advantages.
For demonstration purposes, we will design a (relatively) low pressure drop
isothermal tubular packed bed reactor. The design of the model reactor is

243

6. Modelling of packed bed reactors using CFD

based on the following considerations: the basic reactor design is a cylindrical

xed bed, packed with spherical catalyst particles. The ue gas ow is 525
Nm3 /h, and the maximum pressure drop over the catalyst bed is 100 mbar.
The amount of catalyst that is used to treat the ue gas is 70 kg and the
catalyst particles have a density of 920 kg/m3 . The design that follows from
these assumptions is summarised in table 6.5. The pressure drop requirement
leads to an unusually large catalyst particle size, and therefore a low catalyst
usage and low conversion. The model clearly demonstrates the inuence of the
wall on the packed bed performance in this case.
Due to the relatively simple design, the reactor performance can be pre
dicted quite easily using traditional methods. We will do this here to compare
the results with the CFD model. The pressure drop can be calculated from the
Ergun equation, using the bulk porosity and gas properties at inlet conditions.
This gives a pressure drop of about 5300 Pa. If we assume that the diusion
rates of nitric oxide and ammonia do not dier too much, we can simplify the
rate equation to a simpler n-th order scheme:
(n+m)

rN O = kcN O

(6.91)

For the simplied kinetics, the eectiveness factor can be estimated from
the Thiele modulus, which can be calculated from (Fogler, 1986):
'
(n+m1)
dp kp cN O
=
(6.92)
2
De
As the catalyst particles are large, it can be expected the value of will also
be large, i.e., the reaction takes place in the diusion limited region where
the eectiveness factor is inversely proportional with the Thiele modulus., i.e.

(6.93)

Using this approach, the calculated conversion is 37.8 %.

For the CFD model, a two-dimensional axisymmetric geometry was used
with 120 x 60 rectangular cells. The simulation was stopped when the residuals
had decreased by four orders of scale, which was after about 2000 iterations.
The results are compared to the results of the simple model in table 6.6.
In this case, the eectiveness factor is dened as the actual local rate of
conversion of NO as a fraction of the rate of conversion that would be reached

244

6.4 Demonstration

Table 6.5: Values of the parameters used for deNOx reaction calculations

Quantity
symbol value
D
0.3 m
Bed diameter
L
1.1 m
Bed length
dp
30 mm
Particle diameter
10
Bed-to-particle diameter ratio D/dp
0.41
Bed bulk porosity
b
1.41
Bed bulk tortuosity
b
1.1
Axial wall tortuosity
w
Fluid
air, 0.5 % NO, 0.5 % NH3
M
Molar mass
0.028 kg/mol
p
Outlet pressure
101325 Pa
w0,in
Inlet velocity
4 m/s
Table 6.6: Overall results of isothermal tubular model reactor simulations

Quantity
simple model CFD model
4500 Pa
5300 Pa
pressure drop
34.7 %
37.8 %
conversion
8.3 %
average eectiveness factor 8.2 %
if there were no (internal and external) mass transfer limitation, i.e., the con
centration everywhere in the catalyst particle was equal to the local bulk gas
concentration. This number depends on the local concentrations, ow rate,
catalyst volume (porosity) and surface area,. Therefore, it depends on the
location in the bed; both on the axial position and on the wall distance. The
number given in the table above is a volume average over the entire bed. The
eectiveness factor proles 100 mm from the inlet side and 100 mm from the
outlet side of the bed are shown in gure 6.29.
The concentrations of the components are also a function of both the axial
and the radial position in the bed. Figure 6.30 shows the radial NO concentra
tion proles at dierent axial positions. The concentration increases slightly
from the centre of the bed towards the wall; there is a steep jump at a distance
of about a half particle radius from the wall. It shows that there is a quite
well-dened wall zone where the velocity is relatively high and conversion low.
The mass transfer between the wall zone and the bulk zone is limited because
the low-porosity zone at one particle radius from the wall acts as a barrier.
In this case, the distance over which a dramatic eect on the concentration
can be seen (0.5rp ) is much smaller than the distance over which the porosity
prole is important (5 10rp ). However, due to the cylindrical geometry, the
0.5rp thick wall zone still is over 10% of the bed volume. In addition, since the

245

6. Modelling of packed bed reactors using CFD

velocity in the wall zone is high, about 15% of the air passes through the wall

zone.
The local values of the NO concentration, NO conversion rate and eec
tiveness factor are shown in gure 6.31.

0.12

0.1

[-]

0.08

0.06

y=1.0

0.04

0.02

y=0.1

0
0

10

x/rp [-]

Figure 6.29: Radial proles of the eectiveness factor for the deNOx reaction
near the bed entrance and near the bed exit of at tubular reactor

As expected, the dierence between the traditional approach and the CFD
model is not very large in this case. The traditional approach was made as
simple as possible in this case, which required a simplication of the reaction
rate equation. Of course, the simple model could be extended, for instance
to take into account the actual rate equation (e.g., by using the Aris number
approach described in chapter 2 of this work) or external mass transfer limita
tions. In this approach it is up to the engineer to know when more details still
need to be taken into account and when the model is complex enough to give
results with a sucient accuracy. In the CFD model, the majority of these
eects is taken into account by default, whether they are important in a given
case or not.

246

6.4 Demonstration

0.55
y=1.0
NO concentration [mass%]

0.5

y=0.7
y=0.4

0.45

y=0.1
0.4

0.35

0.3
0

5
x/rp [-]

10

Figure 6.30: Radial NO concentration proles (mass fraction) in a tubular

reactor at dierent axial positions in the catalyst bed

6.4.2

Reduction of NO in a radial ow reactor

With a CFD code, we are of course not limited to tubular geometries. As

an example, gure 6.32 shows the geometry of a simple Radial Flow Reactor
(RFR) design, using the same amount of the same catalyst particles as the
tubular design. The ow enters axially through the inner channel, then ows
radially outward through the packed bed that is xed between two concentric
gauze cylinders, and nally ows out in axial direction through the annulus
between the outer gauze cylinder and the reactor wall. As a rst guess, the
RFR is designed so that the cross sectional area of the inner tube is equal to
the cross sectional area of the annulus, so that the average inlet and outlet
velocity are equal. The remaining parameters are the same as those given in
table 6.5; the RFR contains the same amount of catalyst as the tubular design
given above. The additional parameters for the RFR are given in table 6.7.
The resulting velocity eld is shown in gure 6.33. The RFR design suers
from the fact that in the central channel as the velocity decreases, the pressure
increases (approximately) according to Bernoullis law. In the annulus, the

247

6. Modelling of packed bed reactors using CFD

NO CONCENTRATION
MASS FRACTION
5.30e-3
5.07e-3
4.84e-3
4.51e-3
4.38e-3
4.15e-3
3.92e-3
3.69e-3
3.46e-3
3.23e-3
3.00e-3

EFFECTIVENESS
0.110
0.098
0.088
0.077
0.066
0.055
0.044
0.033
0.022
0.011
0.00

REACTION RATE
[MOL/M3/S]
0.240
0.216
0.192
0.168
0.144
0.120
0.069
0.072
0.048
0.024
0.000

Flow

Figure 6.31: NO concentration, eectiveness and NO consumption rate in the

catalyst bed in a tubular reactor. Axis of symmetry is on the lower side of
each eld, the inlet is on the left side, the outlet to the right.

Table 6.7: Additional radial ow reactor design parameters

Quantity
symbol value
Inlet channel diameter R0
0.15 m
Bed outer diameter
0.25 m
R1
Reactor diameter
0.29 m
R2
Bed height
0.6 m
H
Inlet velocity
4 m/s
w0,in
pressure decreases with increasing velocity. As a result, the pressure across

the bed increases in axial direction, causing a maldistribution of the ow over

the catalyst bed (see gure 6.34). The remedy for this would be to change the

248

6.4 Demonstration

RFR design so that the inlet and outlet are on the same side (although this
is somewhat more dicult to construct), or to design inserts in the central
channel and annulus to create a more even pressure prole.
On a more detailed level, it is interesting to see that no boundary layer
is built up along the gauze layer in the central channel. This is caused by
the radial component of the ow, which pushes the boundary layer into the
bed. In the annulus, the reverse is true: the radial ow streaming out of
the bed causes an additional build-up of the boundary layer, forcing the main
ow outward in the direction of the reactor wall. As the boundary layer does
not contribute (much) to the axial ow, the eective annulus cross sectional
area decreases and the maximum velocity increases. The geometry could be
modied to correct for this.
The maldistribution is increased because the maximum value of the pres
sure drop across the bed, near the end of the central channel, is located at a
wall in the bed, causing a relatively large velocity in the wall zone.

R2
R1
R0

Figure 6.32: Radial ow geometry for the deNOx reactor

Some overall performance parameters of the simple RFR design are given
below.

249

6. Modelling of packed bed reactors using CFD

Velocity
[m/s]

NO
[kg/kg]

6.0

5.0e-3

5.0

4.0e-3

4.0

Rate
[mol/m3/s]

3.0e-3

3.0
2.0e-3
2.0
1.0
0.0

1.0e-3

0.0e-0

Effect.
[-]

0.20

0.10

0.16

0.08

0.04

0.07

0.12

0.06

0.10

0.05

0.08

0.04

0.06

0.03

0.04

0.02

0.02

0.01

0.00

0.00

Figure 6.33: Velocity eld, NO concentration, reaction rate and eectiveness

factor for NO in the radial ow deNOx reactor

pressure drop
14.5
conversion
31
average eectiveness factor 8.6

Pa
%
%

If we compare these results to the packed tube design, it is clear that the
RFR design shows only a small decrease in conversion (from 34.7 to 31 %),
due to additional ow maldistribution, but at a fraction of the pressure drop
required for the tube design (14.5 instead of 4500 Pa). Of course, the radial
ow reactor is less compact then the tubular reactor; the volume of the RFR
reactor is about twice the volume of the packed bed. However, if one were to
construct a shallow packed bed with a larger diameter to decrease the pressure
drop, a diuser section would be required to distribute the ow evenly over
the bed. In practice, a radial ow reactor is more compact than a shallow
bed reactor with the same pressure drop. Figure 6.33 shows the eld values
for the NO concentration, eectiveness factor and source for the RFR design.
Obviously, the design is far from optimal, because of the choice to use the same
particles as in the tubular reactor of the previous section.

250

6.4 Demonstration

25
central
annulus
20

bed
pressure
drop

p [Pa]

15

10

0
0.2

0.3

0.4

0.5
h [m]

0.6

0.7

0.8

Figure 6.34: Axial static pressure prole in the central channel and in the
annulus of the RFR

Both the eectiveness factor and the pressure drop are low, so the particle
diameter can be reduced to increase the conversion. A conical insert in the
central channel can help improve the equal distribution of ow over the bed.
As a third measure, the annular outlet channel can be enlarged to allow for
the area taken by the additional boundary layer thickness. As an example, all
three these modications have been made to the model. The particle size was
reduced to 1 mm; this increases the pressure drop to 960 Pa which is much
higher but still low compared to the pressure drop of the tube reactor. The
eectiveness factor increases to 90-95 %. The NO concentration eld is given
in gure 6.35. As can be seen, the conversion in the main part of the bed is
about 90 %, but there is some slip of NO through the wall zone. The region
in the outlet channel where the NO concentration is higher because of the slip
through the bottom wall zone of the bed is relatively large, but the velocity
in that region is low because of the (extended) boundary layer. Therefore the
overall decrease in conversion because of the wall slip is only one percent point
from 90 to 89 %.
Note that although the change in geometry of the axial channel and the

251

6. Modelling of packed bed reactors using CFD

annulus certainly has inuence on the ow for the original, low pressure drop

bed, the impact is not so large for the smaller particle size bed as in that case
the pressure drop over the bed is much larger than the pressure drop in these
channels.
NO
[kg/kg]
5.0e-3

4.0e-3

3.0e-3

2.0e-3

1.0e-3

0.0e-0

Figure 6.35: Concentration of NO in the modied design radial ow deNOx

reactor

6.4.3 Non-isothermal chemical reaction in a packed tube:

oxidation of SO2
The oxidation of SO2 is a well-known and industrially important process. The
reaction of SO2 with oxygen is catalysed by vanadium oxide
V2 O5

SO2 + 1/2O2 SO3

(6.94)

The equilibrium reaction is relatively slow and exothermic. Therefore, in the

design of a reactor, two counteracting mechanism are important: for a high
reaction rate the temperature needs to be as high as possible, while for a
complete conversion the temperature should be low. Below about 425 C the

252

6.4 Demonstration

equilibrium lies almost completely toward the left side of equation 6.94. The
conventional reactor design (e.g., Fogler, 1986) consists of three or four packed
beds in series, with interstage cooling, see gure 6.36.
SO2
Temperature [C]
400

450

500

550

600

25
50
75
Conversion [%]

100

cooling
cooling
cooling

SO3
0

Figure 6.36: Typical layout, temperature and conversion prole of a conven

tional SO3 reactor (based on data given by Fogler, 1986)

The equilibrium constant for reaction (6.94) can be calculated from ther
modynamic data (JANAF, 1971):

Kp =

pSO3

5129/T 4.869

= 10

1/2

pSO2 pO2

eq


11810
= exp
11.21
T

(6.95)

with T in K and Kp in bar1/2 . The intrinsic reaction rate is given by Harris

and Norman (1972) as
1/2

rSO2 = 

pSO2 pO2
1/2

A + BpO2 + CpSO2 + DpSO3

(
2 1

pSO3
1/2

(6.96)

Kp pSO2 pO2

253

6. Modelling of packed bed reactors using CFD

where Pi is the partial pressure locally at the active surface of a catalyst

particle.


4960
A = exp
6.8
T

B = 0


7350
C = exp
+ 10.32
T


6370
D = exp
7.38
T
Although this is not usually done in the industrial practice, it would be
possible to carry out the SO2 oxidation in a cooled tube reactor instead of the
staged bed concept shown in gure 6.36. In order to remove the reaction heat
eciently, tubes with a relatively low diameter should be used. In addition, the
diameter of the catalyst particles used is relatively large (8 mm is a common
value). Therefore, the bed-over-particle diameter ratio will be quite small and
the wall eect will be important.
The CFD approach developed in this work is compared with the conven
tional engineering method. The latter is taken from Fogler (1986), who uses
the SO2 oxidation process as an example calculation in his text book. He uses
an overall kinetic equation ascribed to Eklund, which gives the overall rate for
a bed packed with cylindrical, vanadia loaded pumice particles with a diameter
of 8 mm and a length of 8 mm. This overall kinetic equation includes eects
of internal and external mass transfer limitation, and therefore can only be
used for these specic particles and for a limited range of ow conditions. It
was found that the overall kinetics could not be reproduced using the intrinsic
kinetics give above. Apparently, the reaction rate for the vanadia catalyst used
by Fogler depends on the temperature and concentrations in a dierent way
than the vanadia catalyst described by Harris and Norman (1972).
To make an honest comparison between the CFD approach and the engi
neering approach, an overall kinetic equation was derived from the intrinsic
kinetics given above for 8 mm spherical particles. The overall reaction rate is
given as a function of conversion and particle edge temperature; this function
was tted to a large number of microbalance solutions. Although somewhat
unconventional, the best t is obtained using polynomial expressions. Since
the gas composition is given in terms of the conversion, this model is only valid
for the reactor inlet conditions chosen (see table 6.9).
The overal reaction rate is given by
(
)

2

pSO3
(6.97)
rSO2 = a1 + a2 + a3 1
1/2
Kp pSO2 pO2
254

6.4 Demonstration

where is the conversion of SO2 and the coecients ai depend on temperature

through a third order polynomial (in K):
ai = bi,1 T 3 + bi,2 T 2 + bi,3 T + ci,4

(6.98)

The values of the coecients are given in table 6.8. With these coecients,
the overall kinetic equation describes the simulation results with a RMS error
less than 1% for temperatures between 400 and 600 C and for conversions
from 0 to 90 %.
Table 6.8: Fitted coecients bi,j for the overall kinetics for an 8 mm vana
dium oxide particle (equation 6.97) according to the intrinsic kinetics given in
equation (6.96).
j/i
1
2
3
1
2.316e-6 -1.715e-6 -0.4634e-6
2
-5.495e-3 4.385e-3
0.811e-3
3
4.293
3.666
-0.413
4
-1104
999.0
54.18
The overall kinetics equation (6.97) is substituted in the engineering model
given by Fogler. The solution of this simpler model still involves the inte
gration of three coupled dierential equations, for the conversion, temperature
and pressure along the length of the reactor tube, so some programming is re
quired. As a one-dimensional model, the radial dispersion taken into account
through an assumed value of the heat transfer coecient between the uid
and the wall. This parameter should also include wall eects and therefore
will depend on the tube to particle diameter ratio.
The CFD simulation takes into account the wall eect and calculates the
internal diusion limitation of the particles at local bed conditions. There
fore, the model can be used for other particle and reactor geometries. The
parameters used for both simulations are shown in table 6.9.
The results of the CFD simulation are shown in gure (6.37) and compared
to the results of the engineering model in gure (6.38). What can be seen is
that the SO2 concentration drops steadily; the conversion at the end of the
bed (l = 2.8m) is about 36 % for the CFD model and 40 % for the engineering
model.
In the CFD results, the exothermic reaction leads to a hot spot in the tube,
at 1 m from the entrance. The simple model predicts a less well-dened
hot spot one meter further downstream. This is probably caused by the fact
that the wall eect is not taken into account in the 1D model: due to the wall
eect, the velocity in the centre of the bed decreases, so that the hot spot
occurs at a more upstream location.

255

6. Modelling of packed bed reactors using CFD

SO2
[kg/kg]
0.200
0.192
0.185
0.178
0.170
0.162
0.155
0.148
0.140
0.132
0.125

T
[C]
476
458
441
424
407
390

S(SO2)
[mol/m3/s]
1.60
1.44
1.28
1.12
0.96
0.80
0.64
0.48
0.32
0.16
0.00

E
[-]
0.75
0.60
0.45
0.30
0.15
0.00

Figure 6.37: Simulation results for the SO2 oxidation tubular reactor model.
Fields for the SO2 mass fraction, the temperature, the reaction rate and cat
alyst eectiveness. The axial length is scaled by a factor 1/10 compared with
the radial distance.

256

6.4 Demonstration

Table 6.9: Values of the parameters used

tions
Quantity
symbol
xSO2 ,0
Inlet SO2 mole fraction
Inlet SO3 mole fraction
xSO3 ,0
Inlet oxygen mole fraction
xO2 ,0
Bed diameter
D
Bed length
L
Particle diameter
dp
Bed-to-particle diameter ratio D/dp
Bed bulk porosity
b
Bed bulk tortuosity
b
Axial wall tortuosity
w
Particle material
Particle material density
kg/m3
Particle conductivity
s
Particle emissivity
R
Outlet pressure
p
Inlet velocity
w0,in
Inlet temperature
T0
Wall temperature
Tw

for SO2 oxidation reaction calcula

value
0.11
0.001
0.1
0.07 m
2.8 m
8 mm
8.75
0.41
1.41
1.1
ceramic
600
1.2 W/m/K
0.85
101325 Pa
4 m/s
393 C
400 C

The maximum temperature in the bed is a useful value to know since it

will be needed to predict e.g., the degradation of the catalyst material. The
maximum temperature predicted by the 1-D model is close to the maximum
average value from the CFD model (430 and 435 C respectively). However,
the maximum temperature at the hot spot in the bed is much higher than the
average temperature: about 470 C. It is obvious that a 1-dimensional model
can not be used to predict the highest temperature in the bed with a high
accuracy.
As Vortmeyer and Haidegger (1991) have shown, the location and magni
tude of the hot spot depend strongly on the heat transfer model. The CFD
model is a useful tool to estimate such parameters.

6.4.4

Oxidation of SO2 : alternative reactor design

As an alternative design for the SO2 oxidation reactor, a tube-in-tube concept

is used. This concept, a sketch of which is shown in gure 6.39, is well known
especially for steam reformers. The reaction takes place in reaction tubes,
that consist of an inner and an outer tube. In the reactor there are two tube

257

6. Modelling of packed bed reactors using CFD

TCFD average

TCFD max.

X1D

XCFD

480

460

0.8

440

0.6

420

0.4

400

0.2

380

0.5

1.5

2.5

Conversion [-]

T [C]

T1D

Length [m]

Figure 6.38: Axial conversion (X) and temperature (T) proles in a tubular,
wall-cooled SO2 oxidation reactor. CFD model results (maximum and owaveraged values) compared to the results of the 1-dimensional model given by
Fogler (1986) for the same geometry and ow.

plates. The inner tubes are connected to the top tube plate and the outer
tubes are connected to the bottom tube plate. In this way, the SO2 rich
feed that enters at a relatively low temperature at the top can ow into the
centre tubes. The centre tubes are lled with a packed bed of inactive, but
relatively good heat conducting pellets. At the bottom end of the tube the ow
reverses and ows up through the annulus that is lled with catalyst pellets.
Here the oxidation reaction takes place. The excess heat is taken up by the
feed stream in the centre tubes, which is heated up to the desired reaction
temperature. In order to get sucient exchange of heat between the feed and
the reaction zone, the reactor tube assemblies need to be small in diameter.
Therefore, the wall eect will be important in the design of this reactor. To
obtain the desired temperature prole in the reaction zone, that is needed to
have both a suciently large reaction rate and a sucient conversion at the
outlet, the heat transfer to the feed ow needs to be engineered carefully. One
of the advantages of such a design is that the thermal expansion of the reactor

258

6.4 Demonstration

tubes can be easily accommodated (only the dierential axial expansion of the
inner and outer tube needs to be allowed). If needed, additional cooling or
heating could be applied through the space in between the tube assemblies
(for instance at startup). Another advantage is that any heat taken up by
the feed ow reduces both the need to preheat the feed and the need to cool
down the reaction mixture as is done in the conventional process (gure 6.36).
Of course, there are disadvantages (more expensive construction, possibly a
higher pressure drop) and an technical/economic study should show whether
this concept has a better overall performance over any other designs, but this
is outside the scope of this work.
SO2

400

Temperature [C]
450
500
550

600

SO3

25
50
75
Conversion [%]

100

Figure 6.39: Demonstration system reactor for SO2 oxidation with approxi
mate desired conversion and temperature proles.

For the demonstration system, a single reaction tube assembly is consid

ered. The geometry and bed parameters for single tube assembly are sum
marised in table 6.10. The outer wall of the assembly is assumed to be ther
mally insulated. The inner tube wall is assumed to have a low heat resistance
compared with the resistances of the bed-wall interfaces on both sides. The
kinetics of the reaction are taken as before (equation 6.96). The computational

259

6. Modelling of packed bed reactors using CFD

model is a 2-dimensional, axisymmetric section of the tube assembly. Due in

part to the large dierence between the radial and axial dimensions, the re
quire numbers of computational cells is rather large. The coarse grid model
consist of 295 44 cells, the ne grid model 530 90.
Table 6.10: Values of the parameters used for demonstration model calcula
tions
Quantity
symbol value
Inner tube
Di
Diameter
140 mm
Bed length
5m
L
Particle diameter
2 mm
dp
Bed-to-particle diameter ratio D/dp
70
Bed bulk porosity
0.40
b
Particle material
ceramic or stainless steel
Particle conductivity
20 W/m/K
s
Outer tube
Do
Diameter
100 mm
Annulus width
30 mm
da
Particle diameter
2.5 mm
dp
Bed-to-particle diameter ratio D/dp
12
Bed bulk porosity
0.40
b
Particle material
ceramic
Particle conductivity
3 W/m/K
s
Fluid (mass %)
79 % N2 , 11 % O2 , 10 % SO2
Mass ow
m
0.462 kg/s
Inlet temperature
Tin
400 C
Outlet pressure
p
101325 Pa
The pressure drop over the complete assembly is 2.5 bar, which is due to
the fact that relatively small particles are used both in the central channel and
in the catalyst bed. The pressure dierence has an eect on the gas density
(see gure 6.40), and through this on the velocities, so that even at the same
temperature level, the inlet and outlet velocity will be dierent. Figure 6.42
shows the velocity proles at dierent axial positions in the tube assembly.
The temperature eld is shown in gure 6.41. It shows that the feed ow
is heated up from 400 C to 600 C in the central tube. In the catalyst bed,
the oxidation reaction rate at rst is high (about 3 mol/m3 /s), causing the
temperature to rise to a maximum of 680 C. As equilibrium is approached,
the reaction rate decreases (to about 1 mol/m3 /s) and also the temperature
slowly decreases because heat is given o to the feed stream. The decrease in

260

6.4 Demonstration

catalyst
inert packing

DENSITY

[kg/m3]

2.10

1.93

1.76

1.59

1.42

1.25

1.08

0.91

0.74

0.57

0.40

Figure 6.40: Sketch of the reaction tube assembly geometry (left) and the gas
density prole (right). Note that the length of the tube has been scaled by a
factor 1/10 for clarity.

temperature shifts the reaction towards the product side so the reaction rate
does not drop o to zero and the conversion still increases.
Figure 6.41 shows the SO2 concentration in the bed, and gure 6.43 shows
the conversion in the centre line of the catalyst bed. As can be seen, the
conversion in this rst attempt at a design is still too low (about 60 %).
However, the conversion is still increasing almost linearly up to the end of the
bed, so the conversion can be increased by increasing the bed length.
Also shown in gure 6.41 is the eectiveness factor for the oxidation reac
tion. This is quite high (nearly 100 % at the bottom half of the bed to about
60 % at the top), which indicates that it is possible to increase the particle
diameter to reduce the pressure drop without having increase the total catalyst
amount. Of course, this would also decrease the bed-to-particle diameter ratio
and thereby increase the wall eect. For optimal conversion and pressure drop
it may be advantageous to use large particles in the bottom part of the bed
and (increasingly) smaller particles in the top part.
As a rst optimisation step, the particle size is increased from 2.5 to 10 mm

261

6. Modelling of packed bed reactors using CFD

TEMP

[C]

680

652

624

596

568

540

512

484

456

428

400

SO3

[mol/m3]

0.199

0.187

0.175

0.163

0.151

0.139

0.127

0.115

0.103

0.091

0.079

RATE

[mol/m3/s]

3.484

3.136

2.787

2.439

2.090

1.742

1.394

1.045

0.697

0.348

0.000

EFFECT.

[-]

1.00

0.90

0.80

0.70

0.60

0.50

0.40

0.30

0.20

0.10

0.00

Figure 6.41: Proles of the temperature, the SO2 concentration, reaction rate
and the eectiveness factor for the tube assembly (see gure 6.40 for a sketch
of the model geometry). Note that the length of the tube has been scaled by
a factor 1/10 for clarity.

and the bed length is increased to 8.5 m. The amount of heat transferred to
the unreacted gas in the inner tube will increase both because of the increase in
length and because of the (desired) increase in conversion (i.e., heat released by
the reaction). Therefore, the inlet temperature can be decreased; in this case
from 400 to 250 C. The resulting axial temperature and conversion prole
is show in gure 6.44. The overall conversion has increased to nearly 70 %.
Further simulations could be performed to nd the optimal inlet temperature,
tube and particle diameters but this is outside the scope of this demonstration.

262

6.4 Demonstration

x=1

x=2

x=3
x=4
x=6

Axial velocity [m/s]

2
0
-2
-4
-6
-8
-10
-12
0

0.01

0.02

0.03

0.04

0.05
r [m]

0.06

0.07

0.08

0.09

0.1

Figure 6.42: Velocity proles at dierent axial positions in the tube assembly
for the coarse grid model. Positive velocity is downward.

263

6. Modelling of packed bed reactors using CFD

700

0.7
Conversion
Temperature

680

0.5

660

0.4

640

0.3

620

0.2

600

0.1

580

Temperature [C]

SO2 conversion [-]

0.6

560
0

3
x [m]

Figure 6.43: Conversion of SO2 and the temperature in the centre line of the
catalyst bed as a function of axial position.

264

6.4 Demonstration

0.8

650

0.7
600

0.5
550

0.4
0.3
0.2

Temperature [C]

SO2 conversion [-]

0.6

500

0.1
450

0
0

x [m]

Figure 6.44: Conversion of SO2 and the temperature in the centre line of the
catalyst bed as a function of axial position, for the modied reactor.

265

6. Modelling of packed bed reactors using CFD

6.5

Conclusions

An open source, unstructured, nite volume CFD code (Dolfyn) has been ex
tended with the sub-models developed in previous chapters of this work so
that it can be used to simulate packed bed reactors. This combined code has
been veried, validated and its use to design and optimise packed bed chem
ical reactors has been demonstrated. Validation was based on experimental
data known from literature, and therefore not very extensive and limited to
cylindrical packed beds.
The models for the bed packing properties were developed in Chapter 3
of this work. Under the restriction that hexagonal, axis-aligned cells are used
in the catalyst bed(s), the equations derived there are easily adapted to be
implemented in a CFD code. In discretization of the proles, care has been
taken to ensure that the total mass of catalyst in the reactor does not depend
on the chosen grid density.
Apart from the catalyst bed properties, it is also important to take into
account the physical properties of the gas phase as a function of composition,
temperature and pressure. Contrary to common CFD practice, the density
cannot be taken independent of the pressure when packed bed reactors are
simulated even though the velocities are well below the speed of sound. It has
been shown that the (non-isentropic) pressure loss that takes place in packed
beds can and should be taken into account when the density is calculated.
The specic heat of most gases is a relatively weak function of the temper
ature. Therefore, if the temperature dierences in a model are not too large,
for most gases it is allowed to approximate the enthalpy as cp (T T0 ) instead
of the integral over cp dT , provided T0 is chosen wisely. It is important to re
alise that the enthalpy balance diers from a general scalar balance since the
enthalpy is the leading variable for convective transport while the temperature
is the leading variable for diusive transport (conduction). Failure to modify
the balance for the temperature eld in this sense can lead to solutions that
violate the second law of thermodynamics.
The thermal conduction near the wall is inuenced strongly by the wall
eect. A striking feature of the eective thermal conductivity proles is that
with increasing Reynolds number, an inversion of the proles is predicted. For
laminar ow, the gas phase conductivity is (in general) lower than the solid
phase conductivity, while for high Reynolds numbers, there is intense mixing
in the gas phase and therefore it can have higher conductivity than the solid
phase. Consequently, the situation can arise that at low Reynolds numbers the
thermal conductivity is lowest for high porosity zones, while at high Reynolds
numbers the conductivity is highest for high porosity zones.

266

6.5 Conclusions

The results of simulations of a steam-heated tube using the packed bed CFD
model were compared to experimental data by Dixon (1988) as presented by
Winterberg et.al. (2000) and of Derckx and Dixon (1996). The radial temper
ature prole was seen to correlate well with the measured data. Especially the
temperature drop in the wall zone is accurately predicted by the packed bed
model.
For further validation, a laboratory scale ethane oxidation reactor was
modelled and the results were compared to measurements by Vortmeyer and
Haidegger (1991). Again, the model results correspond very well to the mea
surements, without using any tting parameters.
As a demonstration, a reactor for the catalytic reduction of NO by ammonia
was modeled using the combined code. For a tubular reactor design, the model
results are compared with hand calculations. For this specic design, the hand
calculation underestimates the conversion by 10 % because it does not take
into account hte wall eect.
The tubular design is compared to a radial ow reactor design using the
same amount of the same particles. The results showed that the RFR has a
much lower pressure drop (14.5 Pa instead of 4500 Pa) and only a little lower
conversion (31 instead of 35 %). The CFD model shows that this decrease
in conversion is mainly caused by ow maldistribution: the axial pressure
gradient in the inlet and outlet channels is comparable to the pressure drop
over the bed. It also shows that the eectiveness factor is low, so the particle
size could be reduced to increase the conversion further, at the cost of some
extra pressure drop.
A modied design for the RFR was made based on the results of the rst
simulation. An insert was placed in the inlet channel and the outlet channel
was widened to make up for the boundary layer. The particle size was reduced
to 1 mm. The new simulation shows a much improved ow pattern and an
increase in conversion to nearly 90 %. Although the importance of the wall zone
is decreased with the decreased particle size, some slip of reactants through
the wall zone can still be observed.
In order to demonstrate the code for a industrially relevant and interesting
case, an SO2 reactor was simulated. The SO2 oxidation takes place in a rela
tively slow equilibrium reaction, where at temperatures where the reaction is
suciently fast, the equilibrium does not allow full conversion of SO2 . First a
cooled tube design was calculated and compared to results of a 1-dimensional
model given by Fogler (1986). Due to the fact that the 1-dimensional model
does not take into account the wall eect, the hot spot location is predicted
further downstream by this model. It is believed that the CFD model in
this case gives the more correct prediction. Furthermore, it is clear that a

267

6. Modelling of packed bed reactors using CFD

1-dimensional model can not accurately predict the peak temperature in the
reactor since there is a quite dened radial temperature prole.
As a more practical alternative, a tube-in-tube concept is simulated. The
advantage of the tube-in-tube concept over conventional SO2 oxidation reactors
would be that the reaction heat is removed by the feed ow, so both the need
for cooling and for feed preheating is reduced. This type of reactor is dicult
to design from scratch because of the strong two-way interaction between the
preheating of the feed and the reaction. It is shown that a CFD code is a
useful tool to model and optimise such a system.
It has been shown that a CFD tool has been developed which can be
used successfully in the modelling for design and optimisation of packed bed
chemical reactors with complex reactions, non-standard geometry and high
interaction of ow and heat transfer. Several simpler cases have been simulated
for which literature models were available and the accuracy of the code was
found to be satisfactory. The data generated by the simulations is relatively
detailed, giving average pressure, velocity, temperature and concentrations at
every point of the grid. Therefore, it is quite a task to duplicate this resolution
in experiments and measurements, and validation of the model on a detailed
level in a complex simulation is not possible at this time. This is valid for this
work in particular, but also more in general as the possibilities for detailed
modelling rapidly improve. It is a challenge for future experimentalists to
make similar improvements in measurement techniques for chemical reactors.

268

6.5 Conclusions

Nomenclature
Roman
Symbol
A
B
ci
cp
cp
d, dp
di
D
Da
DAB
Di,m
Dr
EA
h
Fd
H
H
k0
Kp
M
Mi
n
p
pc
p0
pi
q
ri
R
sv
S

units
Pa s/m2
Pa s2 /m3
mol/m3
J/kg/K
J/kg/K
m
m
m
m2 /s
m2 /s
m2 /s
m2 /s
J/mol
m
N
m
J/mol
...
bar1/2

kg/mol
m
Pa
Pa
Pa
bar
[]kg/m3 /s
mol/m3 /s
J/kg/K
m2 /m3
m2

Variable
Coecient for laminar part of friction equation
Coecient for turbulent part of friction equation
Concentration of component i
Heat capacity (at constant pressure)
Heat capacity (at constant volume)
Particle diameter
Distance to wall i
Bed diameter
Axial dispersion coecient
Binary diusion coecient of A in B
Diusion coecient of compound i in a mixture
Radial dispersion coecient
Arrhenius activation energy
Co-ordinate along length of bed
Drag force
Length
Enthalpy
Pre-exponential reaction rate constant
Equilibrium constant for SO2 oxidation reaction
Mach number
Molecular mass for component i
Normal vector
Pressure
Critical pressure
Total (stagnation) pressure
Partial pressure of component i
Source term for variable
Reaction rate for component i
Universal gas constant (8.3144 J/kg/K)
Specic outer particle surface
Surface

269

6. Modelling of packed bed reactors using CFD

units
s
K
K

m/s
m3
m
m

Variable
Time
Temperature
Critical temperature
Reduced temperature (= T /Tr )
Velocity vector
Molecular diusion volume of compound i
Position vector
Distance from wall

Symbol

b
R

f
k
s

a
a , 2D

units
kg m/s
...
kg/s
Pa s
W/m/K
W/m/K

W/m/K
D
m2 /s
kg/m3
kg/m3
-

m3
-

Variable
Porosity (volume open to ow / total volume)
Bulk porosity, porosity far from any walls
Emissivity
Specic heat ratio
Diusivity coecient
Arbitrary intensive variable
Mass ow
Dynamic viscosity
Thermal conductivity
Thermal conductivity of uid
Relative distance from cell centre to cell face k
Thermal conductivity of solid
Dipole moment
Kinematic viscosity
Density
Stagnation density
Tortuosity (path length / distance)
Tortuosity for paths in axial direction
Tortuosity for paths in axial direction, radial compo
nent
Tortuosity of the bed at bulk conditions (far from
the wall)
Tortuosity of the bed for paths in radial direction
Volume
Conversion

Symbol
t
T
Tc
Tr
v
Vi
x
x

Greek

270

6.5 Conclusions

Literature

Derckx, O.R., Dixon, A.G. (1996), Determination of the xed bed wall heat
transfer coecient using computational uid dynamics, Num. Heat Transfer
part A: Applications, 29(8), 777-794
Dixon, A.G., Cresswell, D.L. (1979), Theoretical prediction of eective heat
transfer parameters in packed beds, A.I.Ch.E. J. 25(4),663-676
Dixon, A.G. (1988), Wall and particle-shape eects on heat transfer in
packed beds, Chem. Eng. Commun. 71, 217-237
Ferziger, J.H., Peric, M, Computational methods for uid dynamics, 3rd Ed.,
Springer Verlag, 2002
Fogler, H.S., Elements of chemical reaction engineering, Prentice-Hall, 1986
Gnielinski, V. (1988), Warme
ubertragung Partikel-Fluid in durchstromten
Haufwerken, section Gh in VDI-Warmeatlas, 5. Auage, VDI Verlag,
D
usseldorf
Gunn, D.J. (1978), Transfer of heat or mass to particles in xed and uidised
beds, Int. J. Heat Mass Transfer 21, 467-476
Harris, J.L., Norman, J.R., Ind. Eng. Chem. Process Des. Dev. 11 (1972),
564
JANAF Thermochemical tables, 2nd ed., D.R. Stull, H. Prophet, Project
Directors, NSRDS-NBS 37, Washington D.C.: U.S. Government Printing
Oce, 1971), as quoted by Fogler (1986)
Papageourgiou, J.N., G.F. Froment (1995), Simulation models accounting for
radial voidage proles in xed bed reactors, Chem. Eng. Sci. 50(19),
3043-3056
Pun, W.M., D.B. Spalding, A General Computer Program for Twodimensional Elliptic Flows, HTS 76/2, CHAM Ltd. (1976), as described on
http://www.simuserve.com/cfd-shop/hts76-2.htm
Reid, R.C. , Prausnitz, J.M., Poling, B.E., The properties of gases and
liquids, McGraw-Hill, 1987
Shapiro, A.H. , The dynamics and thermodynamics of compressible uid
ow, The Ronald press company, new york, 1953

271

6. Modelling of packed bed reactors using CFD

Vortmeyer, D., Haidegger, E. (1991), Discrimination of three approaches to

evaluate heat uxes for wall-cooled xed bed chemical reactors, Chem. Eng.
Sci. 46(10), 2651-2660
Wesselingh, J.A., R. Krishna (1990), Mass Transfer, Ellis Horwood series in
chemical engineering, Ellis Horwood, Chichester, UK
Wijngaarden, R.J., K.R. Westerterp (1989), Do the eective heat
conductivity and the heat transfer coecient at the wall inside a packed bed
depend on a chemical reaction? Weaknesses and applicability of current
models, Chem. Eng. Sci. 44(8), 1653-1663
Wijngaarden, R.J., K.R. Westerterp (1993), A heterogeneous model for heat
transfer in packed beds, Chem.Eng.Sci. 48(7), 1273-1280

272

Chapter 7
Conclusions
The central hypothesis of the present work stated that in the modelling of
packed bed reactors, an alternative approach is possible that is a combination
of the engineering and scientic approaches and that combines the advantages
of both. Four underlying sub-hypotheses were given as necessary conditions
for the main hypothesis. We will rst consider each of the sub-hypotheses and
then return to the main hypothesis.
The topics that have been covered, ranging in scale from a single catalyst
particle to a complete reactor are:
Reaction and diusion in a catalyst particle
Structure of a packed bed, especially near a wall
Flow and pressure loss in a packed bed
Dispersion on the bed scale
CFD model of a packed bed reactor
In each of these topics, we made contributions to modelling methods known
from literature.
In chapter 2, we built on methods to calculate and estimate diusionlimited reaction inside porous catalyst particles. The mass and heat balances
in this case lead to a set of dierential equations: one for each component in
the reaction mixture plus one for the temperature. It is known that the actual
number of degrees of freedom is equal to the number of independent reactions.
This is always less than or equal to the number of components. We presented
a straightforward method to reduce the number of dierential equations to the
number of reactions, based on pseudo concentrations.

273

7. Conclusions

The computation time needed to solve the dierential equations depends

strongly on three factors: the particle geometry, the form of the kinetic equa
tion and the number of simultaneous reactions taking place. For simple par
ticle geometries and single reactions with simple kinetics, analytic solutions
are available for the concentration proles. These can be evaluated with little
computational eort. For single reaction systems with general kinetics and
particle geometry, a numerical solution is required that is more computationally intensive. Westerterp and Wijngaarden (1992) gave an approximation
method (the so-called Aris number approach) that greatly reduces this eort.
However, their method can not be used for cases with multiple simultaneous
reactions.
For multiple reactions with general kinetics, numerical integration of the
set of dierential equations was formerly the only approach. Especially for
complex particle geometries, where the equations are partial dierential equa
tions, the numerical solution is computationally intensive. We presented an
approximation method based on the Aris number approach for general par
ticle geometry and general reaction network kinetics. This greatly reduces
the computation eort involved in the solution of this type of problem. The
method has been demonstrated for the Sohio process for the ammoxidation of
propene, described as a network of six rst order, non-isothermal reactions.
For this process, the solution of the full system with a general PDE solver
required more than 1500 nodes and 100 seconds. The approximation method
for this system used 93 nodes and took just 7 seconds.
As a result, we can conclude that it is possible to develop a method to sim
plify the calculation of chemical reaction with diusion limitation in a catalyst
particle for general kinetics and particle shape so that it requires a limited
amount of computation time and is accurate enough to use in a reactor model.
Therefore, the rst sub-hypothesis can be accepted.
Chapter 3 focused on the eect of a wall on the structure of a packed
bed. It is known that near the wall, the distribution of the particles is no
longer random. In literature, this wall eect is often described by a porosity
prole near the wall. However, for the detailed modelling of ow and transfer
of mass and heat near the wall, not only the porosity but also other bedstructure related parameters like the specic particle outer surface area and
the tortuosity of the ow channels are important. Mariani et. al. (2002) and
Delmas and Froment (1988) used a tting procedure to derive a particle centre
distribution that matches measured porosity proles. From this distribution,
the particle outer surface area distribution can be calculated. However, we
have shown that there are many particle centre distributions that give (more
or less) the same porosity prole; the resulting distribution depends strongly

274

on the tting procedure used. To overcome this, we used numerical techniques

to simulate bed packings of mono-disperse spherical particles in cylindrical
vessels. The particle centre distributions from these simulations have a physical
basis and therefore they describe the actual bed structure. In contrast with
existing literature, we found that there is no evidence, neither in the literature
experimental data base nor in the bed simulation results, that the bed structure
depends on the vessel diameter for D/d > 5.
Apart from the porosity and the particle outer surface area proles, the
most important factor that governs transport in a packed bed is the tortuosity.
Starting with the simulated packings, tortuosity proles were determined using
a particle tracking procedure. It was shown that as a rst approximation, the
local tortuosity near a wall is inversely proportional to the local porosity.
We can conclude that the structure of a packed bed near a wall can be
characterised so that the main parameters that inuence transport processes in
the bed (porosity, particle outer surface area and tortuosity) can be described
as a function of the location in the bed. Therefore, the second sub-hypothesis
can be accepted.
The ow resistance of packed beds is usually described in terms of the
particle size and the bed porosity. However, equations such as the well-known
Ergun equation that describe a large experimental data base (with a variation
of 15 %) are valid for random packings in cylindrical tubes. Therefore, they
cannot be used directly for detailed models near the wall.
In chapter 4, a model for the ow resistance in a packed bed was devel
oped where the bed is described as a set of parallel twisted ow channels. The
resistance of the ow channels was written as a function of local porosity, spe
cic particle outer surface area and tortuosity, using engineering correlations
for pressure loss in ow channels. It was shown that when the wall eect is
ignored, this model reduces to the Ergun equation. The values of the tted
coecients of the Ergun equation are reproduced when plausible estimates are
used for physical properties of the ow channels (e.g., wall roughness).
When the model is used with the near-wall proles from chapter 3, veloc
ity proles near the wall are obtained. These were compared with measured
velocity proles known from literature. The model results tted the experi
mental proles quite well, especially when considering that the model does not
contain any t parameters and the fact that measuring velocity proles inside
packed beds is notoriously dicult and error prone.
The existing engineering rules to calculate ow resistance of packed beds
have been adapted and rewritten in terms of the main parameters (porosity,
particle outer surface area and tortuosity) so they can be used on a local scale

275

7. Conclusions

in a CFD-based packed bed model. In this way, the wall eect is taken into
account in a natural way through the proles of the main parameters near the
wall. Therefore, we can accept the third sub-hypothesis.
Next to convection, dispersion is the most important transport mechanism
in packed beds. In the last few decades, there has been a trend towards packed
bed dispersion models that take into account the porosity prole near the
wall. One of the most prominent of these is the Wall Heat Conduction (WHC)
model described by Winterberg et.al. (2000). The WHC model can not be
used directly in a CFD calculation because it uses global information like the
Peclet number at the centre of the bed. In chapter 5 an alternative model
was presented that takes into account the tortuosity as well as the porosity
prole. It was shown that for heat transfer to packed tubes, the results of our
model are very similar to the WHC model, even though, in contrast to the
WHC model, our model was not tted to this data. Also, as it uses only local
information, our model can be implemented in a CFD code where a cylindrical
(or even symmetrical) geometry can not be assumed.
It can be concluded that a dispersion model has been developed that de
scribes heat and mass transport in a packed bed based on the main bed pa
rameters in such a way that it can be used locally in a CFD-based packed bed
model. Therefore, the last sub-hypothesis can be accepted.
In chapter 6, all submodels presented in previous chapters were combined
into a single packed bed reactor model that is implemented in a CFD code.
The packed bed model has been validated (in a limited way) by comparing
results with literature data for cylindrical tubes. Both for a steam-heated tube
(Dixon, 1988) and for a bench-scale ethane oxidation reaction (Vortmeyer and
Heidegger, 1991), the model results were found to correspond very well to the
measurements. It should be noted that these results were obtained without
any tting parameters.
The code was demonstrated by the simulation of two examples of reactors
where the ow in the bed is important: a low pressure drop, radial ow reactor
for the selective catalytic reduction of NOx by ammonia in ue gasses and a
SO2 oxidation reactor where the exchange of heat through the reactor wall and
hence through the wall zone is important.
As the four sub-hypotheses have been accepted, we will now discuss the
main hypothesis. We will rst determine if the packed bed model fulls the
main requirements.
The rst requirement is that the model should have a higher accuracy than
the engineering approach. We have shown that for standard (cylindrical) ge
ometry the model gives comparable results as the engineering rules. This is

276

correct behaviour because engineering rules are tted to experimental data for
this geometry. For many cases, for instance wall-cooled exothermic reactors,
our model gives more accurate results than a hand calculation because it takes
into account the wall eect. For non-standard geometry, engineering rules can
not be used directly. Therefore, in the engineering approach the non-standard
geometry needs to be approximated by an equivalent standard geometry. For
our model, complex geometries can be modelled directly without simplica
tion. Therefore we state that for these cases our model will be more accurate
than the engineering rules. However, due to the lack of experimental data
available, this has not been proven yet.
The simulation time should be short enough for practical purposes. The
computation time needed to solve a model depends strongly on the complexity
of, for example, the bed geometry and the reaction kinetics. For a simple case
like a deNOx reaction in a tubular reactor, the computation time on a standard
personal computer was found to be in the order of 10-20 minutes. For more
complex cases like the SO2 oxidation in concentric tubes, the computation
time was in the order of a few hours. This is deemed acceptable for day-to
day use in engineering practice. In future the calculation time can decrease
signicantly because neither the CFD solver nor the packed bed model code
have been optimised for computation speed.
When used in combination with the Comow preprocessor, the level of
CFD expertise needed to set-up a simulation is limited. Due to the limitation
to 2-D geometry, the generation of a grid is easy and not time-consuming.
Consequently, the time required to set-up a simulation is very short, typically
less than one hour. The CFD solver and the packed bed model also support 3
dimensional grids, but this requires more expertise in pre- and postprocessing.
A further requirement is that the packed bed model should be generic in
the sense that it can handle a wide range of bed geometries, particle geometries
and reaction kinetics. Any bed geometry that can be dened in the CFD grid
can be used in the packed bed model. The porosity, tortuosity and particle
outer surface area proles are based on packings of uniform spherical particles.
For dierent particle shapes, this results in an overestimation of the wall eect.
In principle, any particle shape can be used if the particle centre distribution
near a wall is known for that particle geometry. This would require a relatively
small change in the code.
For the estimation of the reaction rate, any particle shape can be used
provided the geometry factor for that shape is known. This value needs to
be calculated only once for a given particle shape. The kinetics can currently
be given as networks of general power law equilibrium reactions with Arrhenius
type temperature dependence. More complex kinetics can be added to the code
with a small programming eort.

277

7. Conclusions

The source code of the packed bed model will be released under the GNU
Public Licence. Also, the solver (Dolfyn) and the pre- and postprocessor (Com
ow) that are used in this work have been released to the public as open source
software. Therefore, we have a complete CFD tool for the modeling of packed
bed reactors that is available freely to chemical reactor engineers worldwide.
A CFD tool has been developed which can be used for the design and
optimisation of packed bed chemical reactors with complex reactions, non
standard geometry and high interaction of ow and heat transfer. Engineering
rules have been adapted and are applied in this tool on a local scale, using
a relatively simple CFD solver as the computational engine. We have shown
that this model is accurate, fast, easy to use and generic. Therefore we can
conclude that the main hypothesis can be accepted.

278

Literature
Dixon, A.G. (1988), Wall and particle-shape eects on heat transfer in
packed beds, Chem. Eng. Commun. 71, 217-237
Vortmeyer, D., Haidegger, E. (1991), Discrimination of three approaches to
evaluate heat uxes for wall-cooled xed bed chemical reactors, Chem. Eng.
Sci. 46(10), 2651-2660
Westerterp, K.R., R.J. Wijngaarden (1992), Principles of Chemical Reactor
Engineering, Ullmanns Encyclopedia of Industrial Chemistry, Vol. B4, 5
Winterberg, M., E. Tsotsas, A. Krischke, D. Vortmeyer (2000), A simple and
coherent set of coecients for modelling of heat and mass transport with and
without chemical reaction in tubes lled with spheres, Chem. Eng. Sci. 55,
967-979

279

280

Samenvatting
In het overgrote deel van de industriele chemische processen spelen katalysa
toren, stoen die een reactie versnellen zonder zelf te worden omgezet, een
belangrijke rol. Meestal is er sprake van heterogene katalyse; de katalysator
is hierbij een vaste stof, terwijl de reactanten als vloeistof en/of gas door de
reactor stromen. Een voorbeeld van heterogene katalyse is de katalysator die
in de uitlaatsystemen van autos wordt toegepast. Hierbij is de katalysator
aangebracht op een keramische honingraatstructuur (monoliet). Voor indus
triele chemische processen die op een veel grotere schaal plaatsvinden zijn
gestructureerde pakkingen zoals deze monolieten relatief duur. In de meeste
chemische reactoren wordt daarom gebruik gemaakt van gestorte pakkingen
van poreuze keramische extrudaten.
Voor de werking van reactoren met heterogene katalyse speelt een aan
tal transportprocessen een rol. De uitwisseling van reactanten en produkten
tussen de processtroom en de katalysator is uiteraard van cruciaal belang. In
veel gevallen moet ook warmte worden toe- of afgevoerd om een snelle en zo
volledig mogelijke omzetting van de uitgangsstoen in produkt te bewerkstel
ligen. De uitwisseling van warmte tussen het gepakte bed en de reactorwand is
dan ook belangrijk. De transportprocessen worden benvloed door het macro
scopische stromingsproel in het gepakte bed. De invloed van de wand op de
stroming is hiervoor een belangrijke factor.
Om tot een optimaal ontwerp van een dergelijke reactor te komen is het
nodig om een model op te stellen waarin al deze processen worden meegenomen.
Hiervoor is een scala aan mogelijkheiden, die verschillen in complexiteit en
benodigde rekentijd. Als uitersten zijn er ruwweg twee aanpakken te onder
scheiden, de engineeringaanpak en de wetenschappelijke aanpak.
Bij de engineeringaanpak wordt gestreefd naar een zo eenvoudig mogelijke
beschrijving van de reactor, op basis van empirische modellen. De nauw
keurigheid moet voldoende zijn om het ontwerp en de optimalisatie van de
reactor te ondersteunen. Het resultaat is typisch een model dat inlaat- en
uitlaatcondities aan elkaar relateert. Deze beschrijving is speciek voor een

281

bepaald reactorontwerp. Het voordeel is dat de rekentijd meestal beperkt is;

een nadeel is dat het model alleen geldt voor geometrieen waarvoor empirische
relaties beschikbaar zijn.
Bij de wetenschappelijke aanpak, wordt zoveel mogelijk uitgegaan van natu
urwetten op microschaal. Voor een gepakt bed reactor kunnen de stroming en
transportprocessen op deeltjesschaal bijvoorbeeld worden doorgerekend met
Computational Fluid Dynamics (CFD) technieken. Dit levert generieke en
gedetailleerde modellen op waaruit veel inzicht kan worden verkregen in de op
tredende processen. Deze zijn echter meestal niet geschikt als ontwerpgereed
schap vanwege de hoge mate van expertise die nodig is om het model op te
stellen en te gebruiken. Bovendien maakt de lange rekentijd bij grotere hoeveel
heden deeltjes het toepassen van deze modellen voor reactoren op industriele
schaal onmogelijk.
De centrale hypothese van dit werk is dat er tussen de engineeringaanpak en
de wetenschappelijke aanpak een middenweg is, waarbij de nauwkeurigheid van
de resultaten hoger is dan die van de engineeringaanpak, terwijl de rekentijd (en
de inspanning die nodig is om het model op te stellen) voldoende klein is zodat
het model kan worden gebruikt voor het ontwerpen van chemische reactoren.
Het hiervoor ontwikkelde ontwerpgereedschap bestaat uit een bibliotheek met
routines die gekoppeld kan worden aan een CFD code. Dit maakt het mogelijk
om een gepakt bed reactor door te rekenen met behulp van engineeringrelaties
maar binnen het CFD domein. Om dit mogelijk te maken, moesten er op een
aantal gebieden modellen worden ontwikkeld.
De belangrijkste processen in een gepakt bed reactor zijn reactie, convectie
(stroming en drukval) en dispersie (menging, stof- en warmteoverdracht).
De reactie vindt plaats op de schaal van de katalysatordeeltjes. Voor het
berekenen van reactie en diusie voor eenvoudige gevallen (enkele reactie, een
voudige deeltjesgeometrie en/of eenvoudige reactiekinetiek) zijn berekenings
methoden beschikbaar die accuraat genoeg zijn en niet teveel rekentijd kosten.
In veel praktische gevallen spelen echter meerdere reacties een rol. Er was geen
geschikte methode bekend om deze situaties te berekenen. Daarom is er een
methode ontwikkeld om de zogenaamde Arisgetal benadering uit te breiden
voor reactienetwerken met willekeurige kinetiek in poreuze katalysatordeeltjes
met willekeurige geometrie.
Convectie en dispersie vinden plaats op de schaal van het gepakte bed. Deze
processen worden bepaald door de structuur van de pakking. In de buurt van
wanden is de structuur van een willekeurige pakking anders dan ver van elke
wand. Hierdoor zal ook het convectie- en dispersiegedrag in de buurt van een
wand anders zijn dan ver er vandaan. In een CFD model van een gepakt bed
moeten naar onze mening de transporttermen worden berekend op basis van de

282

lokale bedstructuur. De engineeringmodellen voor het berekenen van convectie

en dispersie in gepakte bedden maken echter zonder uitzondering gebruik van
gemiddelde waarden voor de karakteristieken van het bed. Daarom was het
nodig nieuwe modellen te ontwikkelen voor stroming en menging in gepakte
bedden.
De structuur van het gepakte bed kan worden gekarakteriseerd door lokale
parameters: porositeit, tortuositeit en speciek uitwendig deeltjesoppervlak.
De lokale waarden van deze parameters in de buurt van een wand werden
bepaald uit een groot aantal gesimuleerde willekeurige stapelingen van uni
forme bolvormige deeltjes. Deze stapelingen zijn gegenereerd met een hiervoor
geschreven computerprogramma.
De stroming in een gepakt bed kan worden gezien als stroming door een
groot aantal parallele kanalen die telkens gesplitst en samengevoegd worden.
Er is een model ontwikkeld dat de stroming door deze kanalen beschrijft als
functie van de lokale bedparameters. Voor standaardwaarden van de para
meters bleek dit model identieke resultaten te geven als de algemeen gebruikte
engineeringrelaties. Dit is opmerkelijk omdat het hier ontwikkelde model niet
get is aan experimentele drukvaldata voor gepakte bedden. Bovendien kan
dit model, in tegenstelling tot de engineeringrelaties, worden gebruikt om de
stroming in de buurt van wanden te beschrijven.
Het dispersiegedrag van gepakte bedden werd eveneens beschreven als func
tie van de lokale bedparameters (porositeit, tortuositeit en speciek uitwendig
deeltjesoppervlak). De engineeringrelaties gebruiken gette parameters (zoals
de warmteoverdrachtscoecient naar de wand) of proelen die niet overeen
komen met de structuur van het bed. De waarden van deze parameters zijn
over het algemeen bepaald aan de hand van meetwaarden voor cylindrische
buisreactoren. Hierdoor kunnen deze relaties strikt genomen alleen gebruikt
worden voor cylindrische bedden. Het hier ontwikkelde model gaat direct uit
van de structuur van het bed. Omdat het model geen tparameters bevat is
het algemener dan de literatuurmodellen. De resultaten komen overeen met
de literatuurmodellen en meetwaarden voor eenvoudige (cylindrische) gevallen.
In principe kan het model echter ook worden gebruikt voor niet-cylindrische
reactoren.
Als laatste stap zijn de verschillende modellen samengevoegd en gecombi
neerd met een vrij beschikbare, open source CFD code. Zon CFD bereke
ning levert als resultaat de snelheid, druk, temperatuur en concentraties op
ieder punt in het katalysatorbed. Experimentele gegevens waarmee zulke ge
detailleerde resultaten kunnen worden gevalideerd, zijn in de literatuur nauwe
lijks voor handen. Een beperkte validatie is uitgevoerd aan de hand van gepu
bliceerde metingen aan een reactor voor de oxydatie van etheen. De berekende

283

temperatuur- en concentratieproelen kwamen goed overeen met de gemeten

waarden.
Om de bruikbaarheid van het model als ontwerpgereedschap te demonstre
ren is een tweetal voorbeeldsystemen doorgerekend: de katalytische reductie
van NO met ammoniak, zoals wordt toegepast in rookgasreiniging, en de oxy
datie van zwaveldioxide voor de produktie van zwavelzuur. Verschillende re
actorontwerpen werden vergeleken. Hiermee werd aangetoond dat het CFD
model geschikt is om dit soort systemen te simuleren en te optimaliseren.
In de praktijk wordt het gebruik van CFD voor het ontwerpen van chemi
sche reactoren beperkt door de hoge kosten van CFD software en de expertise
die nodig is om deze te kunnen bedienen. Het gepakt bed model dat in het
kader van dit onderzoek ontwikkeld is, wordt gedistribueerd als onderdeel van
het pakket Comow. Met dit pakket kunnen ingenieurs met basiskennis op het
gebied van stroming en transportprocessen maar zonder specialistische CFD
expertise gepakt bed reactoren gedetailleerd doorrekenen. Comow wordt als
open source software gratis beschikbaar gesteld aan de reactor-engineering
gemeenschap.
We kunnen concluderen dat er een simulatietool is ontwikkeld dat gebruikt
kan worden als ontwerpgereedschap voor gepakt bed reactoren met complexe
reacties, niet-standaard geometrie en een hoge mate van interactie tussen stro
ming, reactie en warmteoverdracht. Het model levert resultaten met meer
detail dan de standaard engineeringaanpak, zodat het mogelijk is om alter
natieve reactoren met betere prestaties te ontwerpen. De centrale hypothese
van dit werk kan dus worden geaccepteerd.

284

Curriculum vitae
Bouke Tuinstra was born on December 8th, 1969 in Cleveland, Ohio, United
States of America. Before he was one year of age, his family repatriated to
the Netherlands. He passed secondary education (V.W.O.) in 1988 at the
Alfrink College in Zoetermeer. From September 1988 to April 1996 he studied
Chemical Engineering (Scheikundige Technologie) at the Delft University of
Technology where he graduated in the chemical reactor engineering group of
prof. Van den Bleek.
The research described in this thesis was started in December 1996 as a
part-time occupation. Bouke Tuinstra was employed by Randstad Polytech
niek (later called Yacht) and stationed at Comprimo Computational Engi
neering b.v. in Schiedam. He divided his time equally between commercial
CFD projects and his PhD research. After less than one year, Comprimo
Computational Engineering was merged into Stork Comprimo Schiedam and
then discontinued. Bouke was stationed at the Stork Comprimo headquarters
in Amsterdam, working part-time as an engineer. In the years that followed,
Stork Comprimo was reorganised, renamed Stork Engineers & Contractors b.v.
and nally sold o to Jacobs Engineering at the end of the year 2000.
In December 2000, Bouke Tuinstra started working as an R&D engineer at
Stork Product Engineering in Amsterdam, while continuing his PhD work for
2 days a week. In November 2005 Stork Product Engineering was dissolved
into the Stork Fokker AESP organisation. Bouke Tuinstra moved to Stork
Inoteq, working full-time as Systems Engineer. In January 2007, Stork Inoteq
was taken over by Stork FDO to form Stork FDO Inoteq b.v. At this company
Bouke is employed to date.

285

286

Dankwoord
In de jaren die ik met dit onderzoek bezig ben geweest, hebben veel mensen
op de een of andere manier bijgedragen aan dit werk. Een aantal mensen wil
ik niet ongenoemd laten. Henri Roelofs, Cor Merks en Hans de Lathouder
waren namens Comprimo, Kiwa en DSM betrokken bij de totstandkoming van
het project en de nanciering voor de eerste 4 jaar. Met Henri heb tot 2002
samengewerkt bij verschillende Stork bedrijven en tijdens de eerste vier jaar
was hij een onmisbare schakel tussen mij en de Stork organisatie. Hans heeft
zich eerst namens DSM en daarna op persoonlijke titel ingezet voor het be
houd van het Comow pakket in woelige jaren. Met Cor heb ik gewerkt aan
het modelleren van korrelbedreactoren voor leidingwaterontharding. Sjoerd
Romkes was vanuit de TU Delft betrokken bij dit interessante stuk onderzoek.
Helaas had het te weinig raakvlakken met het huidige werk om in deze disser
tatie te worden opgenomen. Stan Vermeulen heeft als afstudeerder metingen
voor mij gedaan aan een industriele schaal Radial Flow deNOx reactor; dat
de resultaten achteraf onbruikbaar bleken door een beschadiging aan het in
wendige van de reactor doet niets af aan zijn werk. De toenmalige Proeabriek
staf en collega promovendi wil ik bedanken voor ondersteuning en de goede
sfeer in de tijd dat ik daar (on-)regelmatig aanwezig was.
De leden van de vereniging INUDENT hebben zich steeds ingezet voor het
voortbestaan van het programma Comow. Dankzij deze inspanningen, die
soms niet door puur bedrijfseconomische overwegingen waren ingegeven, heb ik
het programma Comow jarenlang kunnen onderhouden. Bij het opheen van
de vereniging hebben de leden Comow als open source software beschikbaar
gesteld. Mede hierdoor kon ook de in dit werk ontwikkelde code op deze wijze
vrij gegeven worden.
Verder wil ik de stichting Dolfyn en met name Henk Kr
us bedanken voor het
ontwikkelen en vrijgeven van de Dolfyn solver en de vruchtbare samenwerking.
Bijzondere dank ben ik verschuldig aan mijn promotor Cock van den Bleek,
die mijn begeleiding jarenlang in zijn vrije tijd heeft voortgezet nadat hij was
gestopt met werken. Zelfs tot in zijn huis in Frankrijk kon ik hem met mijn
manuscript lastigvallen. Bijzondere dank ook voor mijn begeleider HP Calis.

287

Al jaren geleden heeft zijn carri`ere buiten de TU Delft voortgezet. Desondanks

was hij steeds beschikbaar om tijdens een vrije middag het concept van een
hoofdstuk door te nemen.

288