CFD simulation of flow through

packed beds using the finite

volume technique

Submitted by Matthew John Baker to the University of Exeter as a

thesis for the degree of Doctor of Philosophy in Engineering in May
2011

This thesis is available for Library use on the understanding that it

is copyright material and that no quotation from the thesis may be
published without proper acknowledgement.

I certify that all material in this thesis which is not my own work
has been identified and that no material has previously been sub-
mitted and approved for the award of a degree by this or any other
University.
 Acknowledgements

And I would like to acknowledge my supervisor, Dr Gavin Tabor for

his continuous help and understanding.
 Abstract

When a disordered packed bed, or any heterogeneous media is studied

using computational fluid dynamics, the tortuous task of generating a
domain and creating a workable mesh presents a challenging issue to
Engineers and Scientists. In this Thesis these challenges are addressed
in the form of three studies in which both traditional and novel tech-
niques are used to generate packed beds of spheres and cylinders for
analysis using computational fluid dynamics, more specifically, the fi-
nite volume method. The first study uses a Monte-Carlo method to
generate random particle locations for use with a traditional CAD-
based meshing approach. Computational studies are performed and
compared in detail with experimental equivalent beds. In the second
study, where there is a need for actual, physical beds to be studied,
magnetic-resonance-imaging is used coupled with a novel approach
known as image based meshing. In parallel experimental studies are
performed on the experimental bed and compared with computational
data. In the third study, to overcome fidelity issues with the pre-
vious approaches, a physical packed bed is manufactured which is
100% geometrically faithful to its computational counterpart to pro-
vide a direct comparison. All three computational studies have shown
promising results in comparison with the experimental data described
in this Thesis, with the data of Reichelt (1972) and the semi-empirical
correlation of Eisfeld & Schnitzlein (2001). All experiments and com-
putational models were carried out by the author unless otherwise
stated.
Contents

1 Introduction 5

2 Literature Review 9
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Packed bed geometric properties . . . . . . . . . . . . . . . . . . . 10
2.2.1 Particle Equivalent Diameter . . . . . . . . . . . . . . . . 10
2.2.2 Packing Regimes . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Aspect Ratio . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.4 Porosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.5 The wall region . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 The Physics of Flow . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 Flow Regimes and Turbulence . . . . . . . . . . . . . . . . 19
2.3.2 Packed beds Reynolds Numbers . . . . . . . . . . . . . . . 23
2.3.3 Pressure drop . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Theoretical and Experimental Correlations . . . . . . . . . . . . . 28
2.4.1 Flow Models . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.2 Linear models . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.3 Non-linear models . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.4 Wall correction models . . . . . . . . . . . . . . . . . . . . 36
2.5 Computational literature . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.1 Structured Bed Approaches . . . . . . . . . . . . . . . . . 41
2.5.2 Simple Unit Cell Approaches . . . . . . . . . . . . . . . . . 48
2.5.3 Random unstructured Beds . . . . . . . . . . . . . . . . . 48
2.5.4 Scanned Geometries . . . . . . . . . . . . . . . . . . . . . 51
2.6 Literature Review Discussion . . . . . . . . . . . . . . . . . . . . 54

iv
 CONTENTS

3 Computational Fluid Dynamics 57

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2 The Finite Volume Method in CFD . . . . . . . . . . . . . . . . . 60
3.3 The pre-processor . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3.1 Mesh structure . . . . . . . . . . . . . . . . . . . . . . . . 61
3.3.2 Traditional CAD mesh generation . . . . . . . . . . . . . . 61
3.3.3 Image based mesh generation . . . . . . . . . . . . . . . . 64
3.3.4 Mesh quality analysis . . . . . . . . . . . . . . . . . . . . . 65
3.4 The solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.4.1 The Navier-Stokes Equations . . . . . . . . . . . . . . . . 70
3.4.2 Modelling turbulence . . . . . . . . . . . . . . . . . . . . . 72
3.4.3 Reynolds-Averaged Navier-Stokes (RANS) . . . . . . . . . 74
3.4.4 Turbulence models for RANS . . . . . . . . . . . . . . . . 76
3.4.5 Two equation models . . . . . . . . . . . . . . . . . . . . . 78
3.4.5.1 The k − εt model . . . . . . . . . . . . . . . . . . 78
3.4.5.2 The k − ω model . . . . . . . . . . . . . . . . . . 80
3.4.5.3 The k − ω SST model . . . . . . . . . . . . . . . 80
3.5 The Philosophy of CFD . . . . . . . . . . . . . . . . . . . . . . . 82

4 On creating packed beds using Monte-Carlo and traditional CAD

based approach 84
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.3 Computational Packing of Spheres . . . . . . . . . . . . . . . . . 89
4.4 Computational domain . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5 Computational fluid dynamics . . . . . . . . . . . . . . . . . . . . 100
4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.6.1 Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

v
 CONTENTS

5 Image based meshing of 3D MRI scans of packed beds 122

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2 Experimental data . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2.1 Magnetic Resonance Imaging . . . . . . . . . . . . . . . . 125
5.2.2 Porosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.3 Image Based Meshing . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.4 Computational Fluid Dynamics . . . . . . . . . . . . . . . . . . . 136
5.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

6 On creating geometrically faithful beds with RP 151

6.1 introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.2 Experimental method . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3 Rapid Prototyping and 3DP . . . . . . . . . . . . . . . . . . . . . 153
6.4 Particle roughness . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.5 Particle size distribution . . . . . . . . . . . . . . . . . . . . . . . 157
6.6 Computational Domain . . . . . . . . . . . . . . . . . . . . . . . . 159
6.7 Computational Fluid Dynamics . . . . . . . . . . . . . . . . . . . 160
6.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

7 Correlations and Conclusions 174

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.2 Porosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.2.1 Local Porosity . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.3 The dimensionless pressure drop model . . . . . . . . . . . . . . . 182
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

References 209

vi
List of Figures

2.1 Close hexagonal packing of spheres . . . . . . . . . . . . . . . . . 12

2.2 Face centred cubic packing of spheres . . . . . . . . . . . . . . . . 13
2.3 Simple cubic packing of spheres . . . . . . . . . . . . . . . . . . . 13
2.4 Random unstructured spherical pack . . . . . . . . . . . . . . . . 14
2.5 Left; Low aspect ratio bed. Right; High aspect ratio bed . . . . . 15
2.6 The wall region, to the right of the dashed line . . . . . . . . . . . 17
2.7 Laminar flow represented by streamlines . . . . . . . . . . . . . . 20
2.8 Turbulent flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.9 Tortuosity. Left; simple cubic arrangement (low tortuosity). Right;
a random disordered pack (highly tortuous). . . . . . . . . . . . . 27
2.10 Number of particles as a function of date in years . . . . . . . . . 44

3.1 Left; Structured grid, Right; unstructured grid . . . . . . . . . . . 62

3.2 Delaunay triangulation of a series of points . . . . . . . . . . . . . 63
3.3 Left; perfect tetrahedral cell, Right; highly skewed degenerate cell 66
3.4 Iterative loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.1 Chapter 4 methodology flow diagram . . . . . . . . . . . . . . . . 86

4.2 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3 Computational domain for 14mm spheres . . . . . . . . . . . . . . 93
4.4 Computational domain for 16mm spheres . . . . . . . . . . . . . . 93
4.5 Computational domain for 22mm spheres . . . . . . . . . . . . . . 94
4.6 Grid convergence study . . . . . . . . . . . . . . . . . . . . . . . . 95
4.7 Geometry and mesh of top of bed . . . . . . . . . . . . . . . . . . 96
4.8 Mesh cell equivolume skew for 14mm media . . . . . . . . . . . . 97

vii
 LIST OF FIGURES

4.9 Mesh cell equivolume skew for 16mm media . . . . . . . . . . . . 98

4.10 Mesh cell equivolume skew for 22mm media . . . . . . . . . . . . 99
4.11 Contours of cell squish index for 16mm media . . . . . . . . . . . 100
4.12 y+ values for the wall region for 14mm spheres . . . . . . . . . . 101
4.13 y+ values for the wall region for 16mm spheres . . . . . . . . . . 101
4.14 y+ values for the wall region for 22mm spheres . . . . . . . . . . 102
4.15 Contours of velocity at the near wall region . . . . . . . . . . . . . 103
4.16 Pressure drop per unit length as a function of the particle Reynolds
number (Redp ) for 14mm spheres. Experimental uncertainty ∆Redp =
±236, ∆(∆P/L) = ±4.167P a/m . . . . . . . . . . . . . . . . . . 106
4.17 Pressure drop per unit length as a function of the particle Reynolds
number (Redp ) for 16mm spheres. Experimental uncertainty ∆Redp =
±236, ∆(∆P/L) = ±4.167P a/m . . . . . . . . . . . . . . . . . . 107
4.18 Pressure drop per unit length as a function of the particle Reynolds
number (Redp ) for 22mm spheres. Experimental uncertainty ∆Redp =
±236, ∆(∆P/L) = ±4.167P a/m . . . . . . . . . . . . . . . . . . 108
4.19 Dimensionless pressure drop (φ) as a function of particle Reynolds
number (Redp ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.20 Porosity as a function of distance across the x-plane of packed bed 112
4.21 Contours of velocity magnitude for 14mm spheres (a,b), contours
of velocity magnitude for 16mm spheres (c,d), contours of velocity
magnitude for 22mm spheres (e, f). Fluid flow is in the +z direction113
4.22 Contours of turbulent kinetic energy for 14mm spheres (a, b) . . . 114
4.23 Dimensionless pressure distribution as a function of the distance
ratio across the bed for 14mm media . . . . . . . . . . . . . . . . 115
4.24 Dimensionless pressure distribution as a function of the distance
ratio across the bed for 16mm media . . . . . . . . . . . . . . . . 116
4.25 modified dimsensionless pressure drop (ψ ′ = φ′ = φε3 /1-ε) as a
function of the modified Reynolds number Re′ = Redp (1 − ε) (Eis-
feld & Schnitzlein, 2001) . . . . . . . . . . . . . . . . . . . . . . . 120

5.1 Chapter 5 methodology flow diagram . . . . . . . . . . . . . . . . 124

viii
 LIST OF FIGURES

5.2 Typical image x-axis slice produced by 3d MRI. Note the random
disordered nature of the of the packed bed from the cylinders being
effectively cut.ss . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3 Computational domain for cylindrical media generated from MRI
and IBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.4 Computational domain for spherical media generated from MRI
and IBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.5 Mesh produced using IBM from MRI data . . . . . . . . . . . . . 132
5.6 Close up of particle contact points for cylindrical media . . . . . . 133
5.7 Close up of particle contact points for spherical media . . . . . . . 133
5.8 Contours of cell squish index for cylindrical media. Note the struc-
tured grid in the free-stream. . . . . . . . . . . . . . . . . . . . . . 134
5.9 Mesh element volume distribution for cylinders . . . . . . . . . . . 135
5.10 Mesh element volume distribution for spheres . . . . . . . . . . . 135
5.11 Mesh y + values at near-wall region for cylindrical media . . . . . 137
5.12 Mesh y + values at near-wall region for spherical media . . . . . . 137
5.13 Cylinders cesh cell squish index . . . . . . . . . . . . . . . . . . . 138
5.14 Spheres cell squish index . . . . . . . . . . . . . . . . . . . . . . . 138
5.15 Pressure drop per unit length as a function of the particle Reynolds
number for cylindrical media. Experimental uncertainty ∆Redp =
+/ − 236, ∆(∆P/L) = ±4.167P a/m . . . . . . . . . . . . . . . . 140
5.16 Dimensionless pressure drop per unit length as a function of the
particle Reynolds number. . . . . . . . . . . . . . . . . . . . . . . 141
5.17 Pressure drop per unit length as a function of the particle Reynolds
number for spherical media. Experimental uncertainty ∆Redp =
±236, ∆(∆P/L) = ±4.167P a/m . . . . . . . . . . . . . . . . . . 142
5.18 Dimensionless pressure drop per unit length as a function of the
particle Reynolds number for spherical media . . . . . . . . . . . 143
5.19 Contours of velocity magnitude and pressure for cylindrical media
(a,b), contours of velocity magnitude and pressure for spherical
media (c,d). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
5.20 Contours of velocity magnitude (a) and pressure (b) for low Reynolds
(creeping flow regime) . . . . . . . . . . . . . . . . . . . . . . . . 145

ix
 LIST OF FIGURES

5.21 (a) Contours of turbulent kinetic energy, k. (b) Contours of tur-

bulent frequency, ω. . . . . . . . . . . . . . . . . . . . . . . . . . . 145

6.1 Chapter 6 methodology flow diagram . . . . . . . . . . . . . . . . 154

6.2 Photograph of rapid prototyped bed (Left) and computationally
recreated bed (Right) . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.3 modified dimensionless pressure drop (ψ ′ = φ′ = φε3 /1-ε) as a
function of the modified Reynolds number Re′ = Redp (1 − ε) (Eis-
feld & Schnitzlein, 2001) . . . . . . . . . . . . . . . . . . . . . . . 158
6.4 Computational domain . . . . . . . . . . . . . . . . . . . . . . . . 160
6.5 Cell squish index for computational domain . . . . . . . . . . . . 161
6.6 Velocity profiles for laminar and turbulent flow regimes. Dimen-
sionless velocity ratio is plotted as a function of distance across
pipe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.7 Pressure drop per unit length as a function of Reynolds number
for bi-distributed spheres of 16 and 10mm . . . . . . . . . . . . . 166
6.8 Dimensionless pressure drop as a function of volumetric flow rate
for bi-distributed spheres of 16 and 10mm where φ = ∆P dp /LρU 2 .
Note the little difference in φ irrespective of turbulent profile. . . 167
6.9 Dimensionless pressure drop as a function of Reynolds number for
bi-distributed spheres of 16 and 10mm . . . . . . . . . . . . . . . 168
6.10 Contours of velocity magnitude (a) and pressure (b) for mixed bed
of 16mm and 10mm beds of spheres with parabolic inlet profile.
Contours of velocity magnitude (c) and pressure (d) for mixed bed
of spheres with flat plane profile. Velocity magnitude and pressure
for a 1/7 law profile a, and b, respectively . . . . . . . . . . . . . 169
6.11 Vectors of velocity showing flow recirculation inside the bed at
3.0m/s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

7.1 Porosity correlations as a function of aspect ratio compared with

experimental data . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.2 Bed porosity as a function of distance from wall (x) non-dimensionalised
using the particle diameter (dp ) . . . . . . . . . . . . . . . . . . . 181
7.3 Standard deviation of porosity for packed beds . . . . . . . . . . . 183

x
 LIST OF FIGURES

7.4 All dimensionless pressure drop data presented in this thesis as a

function of particle Reynolds number . . . . . . . . . . . . . . . . 185
7.5 Proposed dimensionless pressure drop model compared with the
experimental data of Reichelt (1972) . . . . . . . . . . . . . . . . 187

xi
 LIST OF FIGURES

Publications
Baker, M. J., Young, P. G. Y., Tabor G. R. (2010) Image based meshing of
packed beds of cylinders at low aspect ratios using 3d MRI coupled with com-
putational fluid dynamics. Computers and Chemical Engineering. doi : 10.1016
j.compchemeng.2011.03.017
Baker, M. J., Tabor, G.R (2010) Computational analysis of transitional airflow
through packed columns of spheres. Computers and Chemical Engineering vol
34. pp 878 - 885.
Hu, Z., Notarberardino, B., Baker, M., Tabor, G., Hao, L., Turner, I and Yang,
L. (2009) On Modeling Bio-Scaffolds: Structural and Fluid Transport Character-
ization Based on 3-D Imaging Data.Tsinghua Science and Technology Vol 14. pp
20-23
Cox J, Abel R, Xuan VB, Cotton R, Young P, Baker M, Tabor G, Nickels T,
Fluid flow in and around the olfactory organ of a hammerhead shark, Annual
Meeting of the Society-for-Experimental-Biology, Glasgow, SCOTLAND, 28th
Jun - 1st Jul 2009. COMPARATIVE BIOCHEMISTRY AND PHYSIOLOGY
A-MOLECULAR INTEGRATIVE PHYSIOLOGY, vol. 153A, no. 2, 2009,
S68-S68
Cox J, Tabor G, Baker M, Young P, Cotton R, Xuan VB, Holmes W, Three-
dimensional structure and hydrodynamics of the nasal passageway of a hagfish,
Annual Meeting of the Society-for-Experimental-Biology, Glasgow, SCOTLAND,
28th Jun - 1st Jul 2009. COMPARATIVE BIOCHEMISTRY AND PHYSIOL-
OGY A-MOLECULAR INTEGRATIVE PHYSIOLOGY, vol. 153A, no. 2,
2009, S125-S125

1
 LIST OF FIGURES

Nomeneclature
Sybmols
ap Specific particle surface area, Sp /vp (m−1 )

A Cross sectional area of tube (m2 )

Aratio Aspect ratio (Aratio = D/dp )
Aw Ergun semi-empirical coefficient (Eisfeld and Schnitzlein, 2001)

Bw Ergun semi-empirical (Eisfeld and Schnitzlein, 2001)

dp Particle diameter (m)

D Bed confining tube diameter (m)

E Energy (J)

f Fanning friction factor

F Force (N )
g Gravity (9.81 m/s/s)

h Height (m)
hL Head loss (m)

hr Distance moved (m)

Hratio Depth ratio (L/dp )
k Turbulent kinetic energy (J)

kp Permeability (m2 )
k1 Coefficient (Eisfeld and Schnitzlein, 2001)

k2 Coefficient (Eisfeld and Schnitzlein, 2001)

Kf Kinetic factor

L Packed bed length (m)

m Mass flow rate (kg/s)

2
 LIST OF FIGURES

p Probability
∆P Pressure drop (P a)

Q Volumetric flow rate (m3 /s)

r Pipe radius (m)

R Universal gas constant (8.314 J/molK)

R′ Gas constant for air (287 J/KgK)
Re Pipe Reynolds number (Re = ρU C/µ)

Redp Particle Reynolds number (Redp = ρU dp /µ)

ReD Pipe diameter Reynolds number (ReD = ρU D/µ(1 − ε))

Rep Packed bed Reynolds number (Rep = ρU dp /µ(1 − ε)

ℜ Reynolds stresses

S Source term
Sp Particle surface area (m2 )
t Time (s)

T Temperature K)
u Velocity (m/s)

U Superficial Velocity (m/s)

Ui Interstitial velocity/Pore velocity (Ui = U/ε)
vp Volume of particle (m3 )

V Volume of confining container (m3 )

Vv Volume of voids (m3 )

y Distance of cell node from wall (m)

y+ Dimensionless distance of cell node from wall

α Linear ’laminar’ coefficient of packed bed and fluid properties

3
 LIST OF FIGURES

β Non-linear ’turbulent’ coefficient of packed bed and fluid properties

Γ Turbulent diffusity

ℓ Turbulent length scale (m)

εt Turbulent kinetic energy dissapation

ε Porosity (Vv /V )
ρ Fluid density (kg/m3 )
µ Fluid dynamic viscosity (P a/s)

ν Fluid kinematic viscosity (m2 /s)

φ Dimensionless pressure drop (φ = ∆P dp /LρU 2 )

φe Ensemble
σ Route mean square (RMS)

τ Tortuosity (L0 /L)

τ0 Shear stress (P a)
θ Angle (rad)

ω Turbulent frequency (Hz)

4
Chapter 1

Introduction

It is not surprising that man has been observing the motion of fluids for centuries.
The Italian polymath Leonardo da Vinci drew some of the earliest sketches depict-
ing the chaotic motion of fluids in his quest for a better scientific understanding.
In the nineteenth-century, the physicist Osborne Reynolds conducted some of
the first experiments through die injection, in which he observed and quantified
the modes of fluid flow. Today, in the digital age, computational fluid dynamics
(CFD) is a popular approach used to analyse the complex behaviour of a fluid.
A computer, or more commonly computers are used to solve a large set of par-
tial differential equations to describe the motion of a fluid. CFD’s popularity
has stemmed from the relatively low set-up cost and man-hours compared to a
full empirical study, coupled with the generally accepted accuracy of CFD. In
addition, the amount of data which can be extracted from computational meth-
ods is far greater than its empirical counterpart without considerable on-cost. In
1991 D. R. Chapman postulated that CFD was en-route to succeeding empirical
methods and that

“wind tunnels were destined to become cabinets for computers”

This postulation is plausible for many simple fluid flow problems, but for complex
stochastic turbulent flows there is still no realistic model that provides a sufficient
level of accuracy. What is more, creating a working domain for complex geome-
tries with intricate curved surfaces can be challenging, if not near impossible for
some cases. The focus of this work is to study methods of generating packed bed

5
models for analysis using computational fluid dynamics. A packed bed can be de-
fined as “A fixed layer of small particles or objects arranged in a vessel to promote
intimate contact between gases, vapors, liquids, solids, or various combinations
thereof; used in catalysis, ion exchange, sand filtration, distillation, absorption,
and mixing” (McGraw-Hill, 2003). The focus of this work is to analyse three
different techniques, with particular focus on the drop in pressure caused by the
media when a fluid percolates through the media. In addition, this work aims
to study meshing techniques and associated fidelity issues with the generation of
computational models in comparison to experimental beds. Ultimately this work
leads to new correlations for pressure drop and porosity generated from the large
quantity of empirical and computational data generated in this study.
The first method presents the case study of three disordered beds of spheres
generated using a traditional CAD (computer-aided-design) based geometry defin-
ing approach. The beds are random unstructured and due to this, a Monte-Carlo
approach (repeat random-sampling) is used to provide a coordinate location for
each sphere comprising the bed. A computational domain is created using the
coordinates of each particle in conjunction with a tertiary CAD package, Au-
toCAD. The computational model is then imported into the commercial finite
volume meshing package, Gambit. Issues regarding meshing robustness are en-
countered using this technique in regard to a high level of cell skewness in at-
tempting to represent particle contact points. Despite localised skewness issues a
good computational data set was achieved. To validate the computational results,
a comparable physical model is created for experimental investigation in which
pressure drop per unit length is monitored. A physical representation cannot
be accurately recreated due to the beds’ stochastic disordered nature (each ran-
dom pack will be different). In this case, physically equivalent beds are created
and compared with computational data. An equivalent bed can be described as
having the same dimensional characteristics, such as bed length, particle diame-
ter and void volume, but may display a considerably different internal structure.
Both computational and empirical measurements are compared with the correla-
tion of Eisfeld & Schnitzlein (2001) and have shown promising results. Moreover,
this technique provides a good engineering approximation in regard to modelling
the pressure drop through random disordered beds. However, this technique

6
has proved limited in regard to successfully producing a workable computational
model due to mesh construction issues associated with the interface of highly
curved surfaces.
CAD based approaches coupled with an equivalent experimental bed have
yielded positive results in regard to predicting the expected pressure drop caused
by the media. However, in some cases it is desirable to compare the exact in-
ternal flow paths. This requires an identical experimental bed structure as the
computational domain, something not provided in the CAD based-equivalent
bed approach. To achieve this, non-invasive methods, such as magnetic reso-
nance imaging (MRI) and computed-tomography, used in the field of diagnostic
medicine have become a popular approach used by some researchers (Manz et al.,
1999; Sharma et al., 2001; Sullivan et al., 2005; Zhang et al., 2006) to determine
a packed bed’s internal structure. MRI or CT are used to determine the internal
structure from a grey-scale image, this is then segmented using a surface defining
algorithm, then a traditional CAD based approach is used to create a workable
mesh. Although possible, this method is tortuous and time consuming and often
requires considerable user intervention and simplification of the flow geometry.
To ameliorate this affect, the work described here uses MRI coupled with a tech-
nique known as image based meshing to recreate the internal bed structure and
create a suitable mesh for analysis. The technique is applied to MRI scans of
disordered packs of cylinders and spheres and the results have been shown to
be promising in regard to predicting the expected pressure drop caused by the
media. However, problems have arisen concerning geometric fidelity issues such
as particle necking as a result of the scan resolution.
Based on the knowledge gained and the restrictions concerning fidelity, this
Thesis moves onto replicating an exact geometry for experimental and computa-
tional analysis, where both experimental and computational beds are 100% geo-
metrically faithful and can be realistically compared. The method described here
utilises MacroPac and the Monte-Carlo algorithm to generate random stochastic
beds of spheres, bi-distributed and weighted 50:50 by volume. A 3d model is
created in a stereolithography (STL) file format from a simple C program suit-
able for import into the image-based surface defining and mesh creation software
ScanCAD/ScanIP/ScanFE (Simpleware, Exeter UK). In parallel a computational

7
model is created from the mesh using rapid-prototyping RP, more specifically 3d
printing (3dP), in which the computational model is geometrically faithful to
the experimental. This method ameliorates problems associated with comparing
computational results to equivalent beds and the associated problems with scan-
ning physical beds and replicating them using a surface defining algorithm. This
method produced the most promising results of the three methods in respect to
the closeness of experimental and CFD and the correlation proposed by Eisfeld
& Schnitzlein (2001).
The work as a whole has shown how three techniques can be applied to analyse
the flow structure and pressure drop through packed beds. The first method has
been shown to be relatively simple and efficient in respect to time, but the method
is burdened by the problem of highly skewed elements jeopardising localised flow
patterns. However it still provides a realistic engineering approximation for pres-
sure drop per unit length. The method is also limited in comparing equivalent
beds with the same dimensional characteristics, but different flow paths. The
second method has been shown to, again, produce a reasonable engineering ap-
proximation in regard to pressure drop, but is severely limited by geometric fi-
delity issues concerning the necking of particles caused by inefficiencies in the
scan resolution. In addition, the method is costly and time consuming and not
a realistic approach for industry based applications. The final method, as far
as the Author is aware, is the first application of rapid prototyping applied to
create a geometrically faithful bed for direct comparison with CFD. The method
is relatively costly, but due to its novel technique could used to investigate other
phenomena which is not possible with traditional methods and technique stands
on its own merits. In all three cases the results have yielded promising results
and provide the reader with an interesting insight into the creation and analysis
of packed beds using the three techniques.

8
Chapter 2

Literature Review

“An ordinary man can surround himself with two thousand books
and thenceforward have at least one place in the world in which it is
possible to be happy”
Augustine Birrell

2.1 Introduction
In this Thesis we are interested in the phenomena of gaseous fluid flow through
packed beds and how the bed geometry affects the flow. In addition, we are in-
terested in the techniques and methodologies by which this can be achieved with
accuracy and precision and the possibility of a viable alternative to empirical
techniques. In which case it is necessary to understand the previous literature
concerning the study of packed beds. The literature concerning the flow through
packed beds is vast and would constitute a whole thesis if discussed in its entirety.
To make the review logical and coherent, the packed bed and its geometric prop-
erties are introduced followed by the physics of fluid flux through such media.
A detailed review of the various correlations proposed to approximate pressure
drop through a packed bed are then discussed, followed by literature concerning
computational methods applied to packed beds.

9
 2.2 Packed bed geometric properties

2.2 Packed bed geometric properties

Packed beds have been used extensively in engineering processes such as filtration,
heat and mass storage, industrial stripping and catalysis. A packed bed can be
defined as a number of particles dumped into a container. The particles form
a matrix like structure which contains voids or pores in which a fluid is free to
percolate. A packed column is a specific case where the packed bed is confined
inside a cylindrical tube. The packed column can either be fixed or fluidized. As
the name suggests the particles in a fixed bed are static and fixed in position
and unable to move. A fluidized bed is the phenomena when the particles are
in motion, excited by the stream of fluid thus able to follow the random non-
deterministic movement of the fluid. In this work we are only concerned with the
study of fixed beds, their geometric properties and their influence on the flow.

2.2.1 Particle Equivalent Diameter

A packed bed consists of particulate matter which is influential in determining
the properties of the bed in regard to the flow properties. The bed particles can
be uniform in size and shape, such as mono-sized spheres, or can be irregular
such as coal or gravel. In order to study a packed bed mathematically, the bed
has to be described by a number of parameters. In the case of a bed particles,
the parameter of interest is the particle diameter. When the bed particles are not
spherical they are represented as an equivalent sphere of the same volume (Sissom
& Pitts, 1972) which can be characterised by the equivalent diameter or by the
sphericity (a particles deviation from a perfect sphere), ψ. When the particles
are spherical the equivalent diameter simply reduces to the particle diameter. To
get a mean representation of bed equivalent diameter, it is coupled with the bed
surface area per unit volume giving the relationship

ap πd2p
=
(1 − ε) π/6d3p

10
 2.2 Packed bed geometric properties

which re-arranging for dp gives

6(1 − ε)
dp =
ap

where ap is the specific particle surface area, Sp /vp (m−1 ) where Sp is the particle
surface area and vp is the particle volume. ε is the porosity, given as the volume
of voids or pores (Vv ), divided by the total volume, V (ε = Vv /V ). However
the particle equivalent diameter does not completely eliminate the effect of the
particle shape (Aris, 1957). For example, a particle may have the same equivalent
diameter, but may be geometrically completely different, changing the flow path-
ways through the bed considerably. In the case of beds with varying diameter
particles a mean average can be introduced. However, this does not give a realis-
tic representation of the beds properties as the assumption is the bed packing will
behave as if the bed particles are mono-sized. In reality, bed particles of varying
sizes will pack more closely than mono-sized particles.
In this work we investigate the packing and flow phenomena of spheres and
cylinders. A sphere can be described as an infinitely sided volume, or as a single
continuous surface which is perfectly round. In reality a perfectly round sphere
does not exist and the closest to being ‘perfectly round’ is Einsteins fused quartz
gyroscope with a sphericity of less than 40 subatomic layers, which is considered
by some as being the world’s roundest object. The fused quartz gyroscope was
manufactured to measure the effect of frame dragging and geodetic effect of space
time form Einstein’s 1916 general theory of relativity.
Cylinders are described by their length to width ration, lc /w. A short stubby
sphere would have a low length to width ratio, a long cylinder, would have a large
length to width ratio.

2.2.2 Packing Regimes

In addition to the particle geometry, the arrangement in which they are ordered
within the bed is highly influential in regard to the ease in which a fluid can
pass through the media. In the case of random irregular size/shape particles
(coal, gravel etc), the packing regime will always be random and unstructured

11
 2.2 Packed bed geometric properties

with no known mathematical description to describe the particles orientation and

position. Beds can be loosely packed or densely packed, based on the number of
particles packed into a known volume, and this is sometimes referred to as packing
efficiency. In addition, the packing can be structured or random unstructured.
If a packing regime is considered to be structured, the coordinates of each par-
ticle have a full mathematical description and constant respective porosities and
packing densities.
The astronomer Johannes Kepler (1611) studied the packing of mono sized
spheres to determine which was the most efficient way of packing in regard to
packing densities and efficiency. He questioned, should the spheres be packed
directly on top of one another, known as simple cubic packing, or should the
sphere layers be staggered so the spheres on the second layer sit within the hollows
formed by the first layer, known as face centred cubic packing. Kepler conjectured
that face centre cubic packing was the densest possible packing scenario, but could
not prove this mathematically.
For a simple cubic arrangement the packing density is ρ = π/6, approximately
√
0.52. For a face centred cubic arrangement the packing density is ρ = π/3 2
approximately 0.74. In addition, if the central coordinate of the first sphere in
a pack is 0, 0, then the next spheres centroid in the series will be one diameter
(dp ) width away in the x, y and z directions. The relative ease in which simple

Figure 2.1: Close hexagonal packing of spheres

mathematical relations can be applied to describe the packing regime makes the

12
 2.2 Packed bed geometric properties

structured packing regimes appealing to many researchers (Dalman et al., 1986;

Delele et al., 2008; Lloyd & Boehm, 1994; Logtenberg et al., 1999).

Figure 2.2: Face centred cubic packing of spheres

Figure 2.3: Simple cubic packing of spheres

When a packing regime is considered unstructured, the location of each parti-

cle is random and stochastic and there is no deterministic mathematical algorithm
to describe the particle location. A unstructured pack can be created physically
by simply dumping a number of particles into a container. However, due to the
stochastic nature of unstructured packing, it is virtually impossible that an un-
structured pack with the same disordered structure can be replicated. In the case
of a few particles (< 5) there is a chance statistically that the same bed structure
may be replicated. In the case of several thousand particles the chance of repli-
cating a packed bed by dumping particles is near impossible. This phenomena
has posed a problem to researchers when trying to separate different phenomena
from the packing regime as the same bed structure cannot be replicated. This
issue further poses a problem in respect to reinforcing CFD data with empiricism.
Unstructured beds can be generated computationally, but there currently exist no

13
 2.2 Packed bed geometric properties

method in which they can be reproduced with geometric fidelity. The other op-
tion available to researchers is to create a random unstructured pack physically
and use non-invasive methods, such as MRI or CT to create a computational
model.

Figure 2.4: Random unstructured spherical pack

2.2.3 Aspect Ratio

In order to compare similar scenarios with varying dimensional values it is often
desirable to use a dimensionless property. The primary dimensionless property
used to characterise a packed bed is the ratio between the equivalent diameter
(dp ) of the particle and the container diameter (D). This is referred to as the
aspect ratio given as
D
Aratio =
dp
Packed beds are categorised as being low or high aspect ratio; a tube of sand
would be considered high aspect ratio, a tube snooker balls would be considered
low aspect ratio. As we might expect, the velocity profile through the high
aspect ratio bed would be fairly uniform across its profile and can be considered
as a pseudo-homogeneous network of capillaries with uniform flow rates. The
network of flow through a low aspect ratio bed would be highly disordered due
to the varying sizes of the voids as a result of the inhomogeneities in the packing

14
 2.2 Packed bed geometric properties

leading to a variation in local porosity. The exact value at which a low aspect ratio

Figure 2.5: Left; Low aspect ratio bed. Right; High aspect ratio bed

bed becomes a high aspect ratio bed or vice versa is not well defined. However,
Aratio = 50 is often suggested as a reasonable value to distinguish between a low
and high aspect ratio bed.
In addition to the aspect ratio, the bed length can be defined as the dimen-
sionless bed depth ratio, which describes the ratio of bed depth to the particle
equivalent diameter given as
L
Hratio = (2.1)
dp

2.2.4 Porosity
A packed bed can be described as a restriction or a partial blockage in a pipe
which interacts with the fluid in some way. The ease with which the fluid can pass
through the medium is determined by the volume of voids within the bed, Vv and
is governed by the area available for the flow to pass. This can be represented in
a dimensionless form as porosity (ε) or void fraction given as the volume of voids
divided by the total volume.

15
 2.2 Packed bed geometric properties

Vv
ε= (2.2)
V
Because regular packings have a complete mathematical description their
porosity can be represented by a simple geometric relation. For a face centred
cubic pack this is given as  
π
ε=1−
6
and for close hexagonal as  
π
ε=1− √
32
The porosity of most concern to packed beds is the bulk average porosity as
described. We have discussed the fact that a high aspect ratio bed is random and
disordered in structure, and due to this its porosity can vary quite considerably
throughout the bed. Conversely, in a high aspect ratio bed there is little variation
in porosity throughout the bed.
Due to the porosity’s simple volume fraction relation, determining an exper-
imental value for porosity is relatively simple. Water displacement is the tra-
ditional method, where the volume of the confining container (V ) is measured.
Water is added to the container filled with the particulate matter filling the void
volume and as a result the volume of water entrained is equal to Vv . However, this
method is susceptible to inaccuracies from entrained air and wetting issues associ-
ated with porous particles. Other methods are also possible, such as non-invasive
methods such as magnetic resonance imaging and X-ray computed tomography
(Mantle et al., 2001; Nguyen et al., 2005; Sederman et al., 2001). However, these
methods suffer from resolution issues resulting in questionable values of void vol-
ume, Vv .
Empirical correlations exist to predict porosity in disordered beds such as
those of de Klerk (2003); Dixon (1988); Jeschar (1964) and Zou & Yu (1995).
However, these approaches are regressions based on empirical data, rather than
mathematical reasoning and can vary quite considerably.

16
 2.2 Packed bed geometric properties

2.2.5 The wall region

When particles such as spheres, are packed into a container, they usually pack
together in a random disordered fashion (Fig 2.4). The particles directly adjacent
to the confining wall will not pack as efficiently as those in the core of the bed due
to the containers flat surface. As a result of this there is an increased porosity
next to the wall in comparison to the bed core. This region of increased porosity

Figure 2.6: The wall region, to the right of the dashed line

is refers to as the wall region, whilst the region not affected by the confining wall
is regarded as the core region. In very low aspect ratio beds (1.2 < Aratio < 3),
such as a cluster of a few spherical particles, the wall region is dominant across
the whole bed. In high aspect ratio beds the wall region is still present, but
statistically has no measurable effect on the motion of the fluid. In the case of
spherical particles uniform in shape and size the wall region can be quantified
as dp /2 (Martin, 1978), which holds well in most cases. For other particles such
as cylinders, the exact value can vary quite considerably due to the particles
orientation and packing density, but a reasonable first assumption of the wall
region thickness would to employ the assumption made by Martin (1978).

17
 2.3 The Physics of Flow

2.3 The Physics of Flow

The nature of the flow through packed beds is complex and notoriously difficult to
study due to the medias opaque, disordered characteristics. In a high aspect ratio
bed the flow is disordered but statistically can be considered homogeneous due to
the relatively uniform size of the voids. In a low aspect ratio bed the bed structure
is considerably more disordered and interconnecting large voids increase velocity
channelling. In some cases an individual flow path can be observed known as the
pathway of mainstream velocity in which the majority of the fluid flux is being
channelled as a result of the inhomogeneities in the packing.
However, a single flow path may not guaranteed and may change transiently
and polyfurcate into other regions. Complicating matters further, the wall region
is likely to be highly influential. We know that in a low aspect ratio bed, at the
wall the porosity is likely to be considerably higher, therefore this becomes the
pathway of least resistance and increased channelling can be observed. The effect
of the wall region has spurred much debate, especially in regard to the resistance
to the flow caused by the medium. In high velocity flow the increased porosity at
the wall provides less resistance to the flow therefore the fluid flux is transferred
through the bed in an easier fashion (Foumeny et al., 1993; Stanek, 1994). In
low velocity flow, the increased surface area results in an increase in the frictional
forces exerted on the flow, thus resulting in increased resistance as observed by
Mehta & Hawley (1969).
The velocity profile may also attribute to the affect of the wall region. The 1/7
power law turbulent velocity profile associated with high speed flow (Re < 1500)
allows more fluid flux across the cross sectional area allowing more fluid to be
channelled into the wall region (the boundary layer thickness is much less in a
turbulent regime) in constrast to the parabolic velocity profile associated with
fully developed ’low speed’ laminar flow (Re < 1500)where the boundary layer
thickness is much greater.
The carriage of flow through a porous medium obeys the same relationships
as for basic fluid mechanics. The volume of fluid per unit time or fluid flux
transported through the bed is described by the volumetric flow rate, Q (m3 /s).
This is related to the superficial velocity, U , from simple dimensional reasoning

18
 2.3 The Physics of Flow

in the form of
Q
U= (2.3)
A
where A (m2 ) is the cross sectional area of the pipe. The superficial velocity is
termed ‘superficial’ as it is the velocity that would be present in the absence of
the media. For instance, a flow measuring device placed immediately before the
media would measure the superficial velocity.
Another term to used describe the velocity is the interstitial velocity, Ui , which
is the average velocity within the pores. This takes into account the bed porosity
through the relation
U
Ui = (2.4)
ε
where ε is the porosity. This forms the average pore velocity, as there is no
guarantee of homogeneity within the pores in a disordered bed and it is derived
from the global property of average porosity.

2.3.1 Flow Regimes and Turbulence

The flow of a fluid can be characterised by a selection of dimensionless numbers,
such as the Froude number, Strouhol number, Prandtl number, but none of which
are of more importance to this work than the dimensionless number proposed
by Reynolds (1883). The Reynolds number is characterised by the ratio of the
inertial and viscous forces and is defined as

ρU C
Re = (2.5)
µ

where ρ is the fluid density (kg/m3 ), U is the velocity (m/s) (for packed beds
the superficial velocity), µ is the dynamic viscosity (P a.s) and C is an appro-
priate scaling length (Reynolds, 1883). In the case of pipes the scaling length
would simply be the diameter, in the case of aerofoils the chord length. In fact,
any sensible scaling length can be used providing consistency is followed when
matching the Reynolds number for equivalent experiments. Alternatively, we can
define the Reynolds number as

UD
Re = (2.6)
ν

19
 2.3 The Physics of Flow

where the density and dynamic viscosity is replaced by a single function known as
the kinematic viscosity ν (m2 /s), given as the ratio of the dynamic viscosity and
the density, ν = µ/ρ. Reynolds (1883) conducted a series of experiments in which
he injected dye into a clear glass pipe carrying water. Reynolds (1883) observed
the motion of the dye and concluded that at low velocities the fluid shears in a
laminated fashion and the streak-line remains unbroken (stream-flow). At high
velocities the streak-line is broken and disperses and mixes into the fluid in a
sinuous fashion. Today, we categorize the modes of flow observed by Reynolds
(1883) as being either laminar, transitional or turbulent.

Figure 2.7: Laminar flow represented by streamlines

Laminar flow is the motion of a fluid where fluid layers shear over each other
in a orderly parallel (laminated) fashion as shown in figure 2.7. Viscosity effects
are dominant and the flow exhibits no instabilities and no naturally induced
vortices. In the case of pipe flow, the shear stresses and strains as a result of
the fluids interaction with the pipe wall produce a parabolic velocity profile.
If the Reynolds number is increased, laminar flow develops into a transitional
regime. The lamination has now become unstable and bursts of chaotic behaviour
(turbulence) is observed, which is then damped out by viscous effects. If the
Reynolds number is further increased the unstable bursts of chaos become more
frequent ultimately breaking up the lamination into a full turbulent flow regime.
Turbulent flow is chaotic with random fluctuations in velocity and formation of
vortices (Eddies) which are random in space, time (Landahl & Mollo-Christensen,
1986) and of varying frequencies. Thus, the flow is highly rotational, diffusive and
dissipative and it is impossible to derive an complete model of the phenomena.
The largest vortices are associated with low frequency fluctuations in which the
size is governed by the confining geometry. They gain their kinetic energy from
viscous stresses and strains from the fluids interaction with its surroundings. The

20
 2.3 The Physics of Flow

Figure 2.8: Turbulent flow

large vortices break up into smaller vortices in which kinetic energy is dissipated,
referred to as the energy cascade (Lien & Leschziner, 1994). The smaller vortices
break up further into the smallest possible eddies which are viscosity dependant
and governed by the Kolmogorov length scale (ν/εt )0.25 and represent high fre-
quency turbulence, where ν is the kinematic viscosity and νis the average rate
of energy dissipation per unit mass. The transition from small vortices to the
smallest possible vortices is where the majority of the kinetic energy is dissipated
in the form of heat and less so, sound energy. A poetic description of the tur-
bulent cascade was described by Lewis Fry Richardson’s adaptation of the poet
and author, Jonathan Swifts “Ad infinitum” (A flea hath a smaller flea) with

“Great whirls have little whirls, which feed on their velocity. And
little whirls have lesser whirls, and so in to viscosity.”

We often think of vortices as being helical or spiral. In the free-stream this may
often be the case, but the fluids interaction with its surroundings can often distort
the shape of the vortex due to shear stresses and strains from the fluid itself or
the wall region.
As the viscosity is the main contributor to the damping of vortices in turbulent
flow, turbulence is more prevalent in low viscosity flow such as air. Laminar flow
is more prevalent in high viscosity flows, such as oil or polymers.
We can use the original equation proposed by Reynolds (1883) to make an
engineering approximation to the likely mode of the fluid flow. The low Reynolds
Stokes flow regime usually exists when Re << 1. The critical Reynolds number
where transition occurs from laminar to turbulent is often quoted as Recrit = 2000

21
 2.3 The Physics of Flow

or more commonly Recrit = 2300 (White, 2003) for circular pipe flow based on
the pipe diameter as the suitable scaling length. However this approach of one
single value for the onset of turbulence in misleading. A more realistic approach
is to suggest a range of values where Laminar flow usually exists in a pipe below
Re = 2300, transitional as Re = 2300 − 4000 and a complete turbulent regime
has formed by Re > 4000 (Cengel & Cimbala, 2006). It must be noted that a
common misinterpretation is to use this value for all cases of fluid flow. In fact,
this value is only applicable for pipe flow and does not hold for open channels,
square ducts where there are other empirically defined critical values. It is impos-
sible to have an exact value for the transition between laminar and turbulent flow
due to surface roughness and disturbances in the flow, not accommodated by the
Reynolds number. The only conclusive approach to determine if a flow is laminar
or turbulent is through flow visualisation, or a more modern semi-empirical, op-
tical technique to give a visual indication of turbulent vortices and unsteady flow
effects. But even then, such as in the case of wake regions behind bluff bodies,
there can be true stochastic turbulence co-existing with deterministic unsteady
behaviour.
Optical methods rely on a direct line of sight, which is not possible in a opaque
porous medium. However, optical methods such as laser Doppler anemometry
(LDA) and particle image velocimetry have been attempted Calis et al. (2001);
McGreavy et al. (1986); Yevseyev et al. (1991). LDA uses a laser to illuminate a
seeded flow of (a flow containing small particles) coupled with a high resolution
video camera. It uses the Doppler shift to determine the particles velocity.
Similar to LDA, PIV requires a seeded flow, a camera and a laser. The laser
illuminates the flow in a series of pulses allowing the particles to be tracked by
successive high frequency images taken by the camera. From the movement of
the particles the software is able to produce realistic velocity vectors (direction
and magnitude), providing there is a sufficient number of particles in the flow
and they have a density similar to the fluid. Although the same limitations, such
as direct line of sight are applicable to PIV and LDA as traditional visualisation
methods, some attempts have been made using these techniques to study packed
beds. McGreavy et al. (1986) adopted LDA to analyse the internal flow structure
through a packed bed. LDA has also been adopted by Calis et al. (2001) to

22
 2.3 The Physics of Flow

analyse the flow structure in fixed beds of Aratio = 1.47 and 2. Yevseyev et al.
(1991) also use LDA to analyse the internal flow structure in a bed of regular
arrangement, and show the formation of vortices in the voids. They describe
this as turbulence, however it is difficult to conclude from their results whether
they are analysing true stochastic turbulence or deterministic periodic vortex
structures and recirculation. Attempts have been made with the use of optics
and refractive index matched transparent beds Giese et al. (1998); Hassan (2008).
This is limited by the temperature dependant refracted indices and can therefore
only be used for studies at the near wall region (Gotz et al., 2002).

2.3.2 Packed beds Reynolds Numbers

As with pipe flow, the flow through packed beds can also be loosely characterised
by the Reynolds number, but this is complicated further by the existence of several
forms of Reynolds number, based on different parameters and length scales and
each with unique critical values for the onset of turbulence. In many applications
of packed beds the Reynolds number is defined as

ρU dp
Redp = (2.7)
µ

where the scaling length is simply the particle equivalent diameter, and this is
therefore known as the particle Reynolds number. We can see that this equation
is solely dependant on the properties of the fluid and the particle diameter and
makes no reference to the bed porosity, ε. Experimentation has shown that a bed
can be loosely packed or densely packed and this can be represented by a large
or small value of porosity respectively. Based on this equation and empiricism
a new set of values for critical flow are formed, based on the particle Reynolds
number where the flow can be characterised as laminar (Redp < 10), transitional
(10 < Redp < 300) or turbulent (Redp > 300) (Ziolkowska & Ziolkowska, 1988).
However, these experiments were conducted primarily using dye injection, and
although the use of dye injection is a well established and accurate method of
determining the onset of turbulence in pipe flow or open channel flow, problems
arise when applying the technique to porous media. For a completely laminar,

23
 2.3 The Physics of Flow

steady flow a stream of dye will remain to follow the path-line/streamline. For
a turbulent flow the streak-line will break up due to the random and chaotic
formation of vortices. Due to the geometry of a packed bed, a large pore con-
verging to a small pore will dampen out any large scale vortices formed within
the bed, in which case re-lamination may occur. With the use of dye injection
once the streak-line has broken up, there is no way it can return to a streak line
and therefore any re-lamination will remain undetected and the flow will appear
turbulent throughout the bed, or at least post-streak-line separation. Attempts
have been made to determine turbulent spectra within the void space of a cubic
arrangement using hot-wire anemometers (Tsotas, 2002a) who report no vortex
shedding from behind spherical particles. Their experimental technique relies
on the use of hot-wire anemometers positioned in the bed voids to determine
both the instantaneous velocity and the root-mean-square (RMS) of the veloc-
ity fluctuations. However, (Tsotas, 2002a) describe no margins of uncertainty or
response time. Hot-wire anemometers rely on Newton’s law of cooling to deter-
mine the velocity and it is probable that the resolution of the response time is not
sufficient to describe the highest frequency turbulent events. Scheidegger (1960)
reviewed a considerable amount of research where critical values were quantified
and came up with the conclusion that it is impossible to define a single critical
value for the onset of turbulence due to discrepancies between the results reported
(Ziolkowska & Ziolkowska, 1988). Furthermore, most empirical methods strug-
gle to distinguish between true stochastic turbulence and deterministic unsteady
behaviour.
In some cases, even when all the particles in a packed bed are uniform in size
and shape, the pore volume may still vary due to the heterogeneous nature of a
random packing of spheres. In this case there is a strong possibility that all three
states of flow will exist within the internal structure. We already know that the
ease in which flow can pass through a porous medium is strongly dependant on
porosity, in which case the packed beds Reynolds number is defined as

ρU dp
Rep = (2.8)
µ(1 − ε)

24
 2.3 The Physics of Flow

which is simply the particle Reynolds number modified by the dimensionless pack-
ing fraction (1 − ε). This form of Reynolds number is also seen in some literature
with the particle diameter (dp ) replaced with the confining pipe diameter, D,
given as
ρU D
ReD = (2.9)
µ(1 − ε)
Due to the addition of porosity the packed beds Reynolds number is particu-
larly useful for comparing beds of different porosities. The Reynolds number, due
to its dimensionless form can be applied to set the limitations of various correla-
tions. Bear (1972) suggest that Darcy’s law is only valid providing the particle
Reynold’s number is in the region of 1 < Redp < 10. A lesser used form is the
modified Reynolds number proposed by Eisfeld & Schnitzlein (2001), given as

Re′ = Rep (1 − ε) (2.10)

which is simply the particle Reynolds number multiplied by the packing fraction.
The packing fraction and porosity are both dimensionless entities, in which case
their position in the equation is irrelevant, only modification of perceived values
is needed for the onset of turbulence.
When applying the Reynolds number to a packed bed, it must be kept in
mind that the Reynolds number is a global entity or bed average quantity. When
the bed flow patterns and geometric properties are fairly uniform, such as a high
aspect ratio bed (e.g a tube full of sand) the average Reynolds number will give
a realistic average representation of all the flow. Conversely, for a high aspect
ratio bed with a strong disordered geometry and a highly influential wall region,
the Reynolds number is likely to vary considerably. The Reynolds number relies
on an appropriate scaling length and velocity to ensure dimensional correctness
and or the addition of porosity, which in itself is an average quantity. Analysing
the bed on a macro-scale we know that the porosity can vary considerably, es-
pecially at the near wall region. As a function of porosity the velocity may vary
throughout the bed due to flow channelling and the changing paths of mainstream
velocity. Due to these effects the bed Reynolds number is likely to vary consider-
ably throughout the media. Based on this assumption, the flow may be laminar
in certain areas of the bed but also exhibit true stochastic turbulence in other

25
 2.3 The Physics of Flow

locations as well as periodic burst of unsteady behaviour which are damped out
by the viscous effects. In the case of a low aspect ratio bed it is unlikely that a
packed bed displays full turbulence throughout. Conversely, a very high aspect
ratio bed (Aratio > 500) is likely to display no turbulence due to the pore size be-
ing smaller than the smallest turbulent structures, governed by the Kolmogorov
length scale.

2.3.3 Pressure drop

We have discussed the fact that a packed column can be considered as a partial
blockage constrained by a pipe. Due to the blockage a back pressure is formed
or a pressure drop (∆P ) due to the ease in which the fluid can pass through the
medium and is given the SI units Pascals (P a) or Newtons per metre squared
(N/m2 ). This is usually measure experimentally using a differential fluid column
manometer or pressure transducers located before and after the bed. In liquids
it may be appropriate to describe the pressure drop as a head-loss related to the
pressure drop by
∆P
hL = (2.11)
ρg
In some cases a sign convention is used (+/−), to distinguish between a drop in
pressure and an increase in pressure. The pressure drop can be described as a
function of the bed properties and the properties of the fluid as

∆P = f (ε, τ, dp , D, L, Re) (2.12)

Theoretical and empirical correlations have been derived to link these parameters
in the form of a pressure drop equation. One of the most influential bed param-
eters concerning the pressure drop is the bed porosity, ε. Coupled with porosity
and bed length is tortuosity, τ . Tortuosity as its name suggest is how indirect
the route of a fluid particle from the bed entrance to the bed exit and is given as

L0
τ= (2.13)
L

26
 2.3 The Physics of Flow

where L is the bed length and L0 is the route of the fluid particle. However,
determining an accurate value of bed tortuosity is nearly impossible due to the
opaque nature of a packed bed. A bed can have identical geometric values, such
as porosity, aspect ratio and particle diameter but the tortuosity can vary con-
siderably. If we imagine two packed beds, each containing one hundred identical
spheres, with the same value of porosity, aspect ratio. Consider two scenarios
(figure 2.9); the first scenario the spheres are packed inside the bed in such a way
that they are in a simple-cubic formation; in the second scenario they are random
and disordered. In the first scenario there is little resistance to the flow as most
of the fluid particles and channelled through the voids. In the second scenario
the fluid particles are obstructed by the media and there is no obvious route for
the particles to travel. The second bed is considered ‘more tortuous’ than the
first.

Figure 2.9: Tortuosity. Left; simple cubic arrangement (low tortuosity). Right;
a random disordered pack (highly tortuous).

In regard to high aspect ratio beds the bed inhomogeneities, such as the effect
of the confining wall have a considerable effect on the pressure drop and are
coupled with the effect of local porosity. The effect of the confining wall on the
pressure drop has caused much debate. The work of Mehta & Hawley (1969)
suggests an increase in pressure drop due to the effect of the additional wall
friction in low aspect ratio beds (7 < Aratio < 91, Redp < 10). This research is

27
 2.4 Theoretical and Experimental Correlations

particularly controversial as all other research presents the opposite effect. The
work was restricted to a low Reynolds regime (Redp < 10) where it could be argued
that the wall friction is dominant and the coefficient of drag is less dependant
on particle geometry. More likely this is just a misinterpretation of data (Eisfeld
& Schnitzlein, 2001). Stanek (1994) and Foumeny et al. (1993) suggest that the
homogeneity’s as a result of the increased porosity at the container wall reduce the
pressure drop due to flow channelling. However, there is still a lack of conclusive
evidence to prove or disprove either theory.

2.4 Theoretical and Experimental Correlations

In many applications of packed beds, the effect which is of most importance to
Scientist and Engineers is the pressure drop caused by the presence of the media.
The most conclusive way to determine pressure drop through a packed bed is to
undertake an individual experimental investigation for every specific case. This
would entail creating a suitable geometry and forcing a fluid through the media
by means of a pumping system, with a manometer tapping at the extremities
to measure the pressure drop. Although possible, this would be tortuous and
too time consuming for most industrial engineering applications. In the digital
age CFD has become a viable alternative to empirical methods, however it can
be relatively time consuming when defining a complex disordered geometry, in
addition, more information will be generated than is required in most industrial
applications. Data such as instantaneous pressure and velocity fields can be
extracted along with a considerable amount of other data associated with the
motion of a fluid.
Where an engineering approximation is solely required, a simple mathematical
relation is used to approximate the pressure drop. These relations have been
developed by Scientist and Engineers over the past three centuries. Generally the
theoretical approximations treat the media as a bulk or gross entity limiting the
amount of information that can be extracted. However, serve their purpose from
simplicity and efficiency.

28
 2.4 Theoretical and Experimental Correlations

2.4.1 Flow Models

Flow models usually represent a mathematical correlation in the form of a re-
gression around an experimental study or based on a theoretical postulation or
understanding (e.g based on dimensional analysis). Some correlations can be a
blend of both e.g based on a theoretical understanding but with an empirical co-
efficient to adjust the data. The following section focuses on flow models and the
fundamental assumptions underlying the development of these equations, such
as flow characterisation and limitations and subsequent failures of these mod-
els. Here, rather than starting with the earliest models to predict pressure drop
through a porous medium and proceeding chronologically through time, a path
many authors have chosen to take, we shall begin with the Reynolds hypothesis
which describes the pressure drop per unit length relation given as

∆P
= αu + βu2 (2.14)
L

where α and β are empirical coefficients dependant on the fluid properties, the bed
geometry and friction (Hassanizadeh & Gray, 1987), and u is the fluid velocity.
Unsteady flow effects can be further added to the equation by the addition of
c∂u/∂t, representing the change in velocity with respect to time.

∆P ∂u
= αu + βu2 + c (2.15)
L ∂t

Reynolds hypothesis is generally considered in two distinct parts, where αu is the

laminar and hence linear element and βu2 is considered the non-linear turbulent
element of the equation. It is generally sufficient to consider the flow to be steady
state, hence in most cases the term c∂u/∂t is negligible.
Nearly all models proposed to describe pressure drop per unit length can be
traced back to the Reynolds hypothesis with the coefficients (α and β) quantified.
These coefficients are either quantified empirically or through certain analytical
assumptions. We know a packed bed is comprised of a series of inter-connecting
voids and loosely behave as a series of varying length pipes. Due to this, in most
cases a pipe flow analogy can be applied, known as the theoretical or fictitious
tubes model. In this case, the laminar (α) and turbulent (β) elements can be

29
 2.4 Theoretical and Experimental Correlations

regarded as being equivalent to the Hagen-Poisselle relationship and the Darcy-

Weisbach equation respectively to describe pressure drop through a pipe. The
Hagen-Poiselle relationship is described as

∆P 8µ
= 2u (2.16)
L Rh

and the Darcy-Weisbach by

 
∆P L 1 2
=f u (2.17)
L D 2g

The Hagen-Poiselle equation is a simple relation which specifically describes pres-

sure drop through a smooth straight pipe in a laminar, parabolic flow regime. The
term R is the hydraulic radius, which for round pipes is described as the ratio of
the cross sectional area over the whetted perimeter (for a full pipe this would be
R/2). The Darcy-Weisbach equation is applicable for both turbulent or laminar
flow regimes by the adjustment of the friction factor, f , and accommodates vari-
ables such as wall roughness, which as we have discussed can induce unsteady
behaviour. For a laminar flow regime in a smooth pipe the friction factor, f , can
be simply described as f = 16/Re (Hagen-Poisselle relation). For turbulent or
laminar flow, accounting for wall roughness the Moody diagram can be used in
conjunction with the pipe Reynolds number to find a suitable value of f . This
technique for determining pressure drop through a straight pipe is well estab-
lished and relies on just a few variables. Determining suitable values for a packed
bed is dependant on many variable which are not easily determined with the
current methods available. Throughout this Thesis, for numerical simplicity the
U.S style of the Darcy-Weisbach equation is adopted omitting the multiplication
factor 4. Combining both the Hagen-Poiselle relationship and Darcy-Weisbach
equation yields  
∆P 8µ L 1 2
= 2u+f u (2.18)
L Rh D 2g
In most cases, the Hagen-Poiselle relationship forms the bases for the linear
(laminar) models and the Darcy-Weisbach equation for the non-linear (turbulent)
models. These two equations have satisfied all the dimensional characteristics to

30
 2.4 Theoretical and Experimental Correlations

describe pressure drop through a single tube in a packed bed, but alone do not
model the drop in pressure well. The next section describes the development of
this equation to model a series of tubes of varying lengths.
When comparing pressure drop through packed beds from literature it is often
desirable to present the pressure drop in its dimensionless form, φ, which allows
packed beds of different properties to be compared.

∆P dp
φ= (2.19)
LρU 2

This allows beds of different particle diameters and lengths to be compared. This
form also takes into account fluid properties such as density, however, does not
consider the fluid viscosity. Further more, this relation as described above does
not take into account bed porosity (ε), which is highly influential on the ease with
which a fluid can pass through the media. This however, is taken into account
to some extent by the particle diameter. We have discussed that the pressure
drop per unit length (∆P/L) in the laminar regime is linear and the turbulent
regime the pressure drop per unit length is a quadratic trend increasing with
the Reynolds number. When using the dimensionless pressure drop, φ, the U 2
is on the bottom of the equation implying that the dimensionless pressure drop
decreases as a function of the Reynolds number.

2.4.2 Linear models

The earliest theoretical explanation for the phenomena of flow through porous
media (including packed beds) is Darcy’s law (Darcy, 1856) determined by em-
piricism, re-enforced with simple dimensional analysis. Darcy’s law is a linear
law, analogous to Ohm’s law of resistance, Fourier’s law of heat conduction and
Fick’s law of diffusion (Dullen, 1992) all of which are globally known as linear
transport equations. Darcy type equations assume the flow regime through the
porous medium to be Hagen-Poiseuille flow and therefore can only be applied to
low Reynold’s number conditions. Hagen Poiseuille flow is described as the flow
through a cylindrical pipe and assumes the flow to be incompressible, viscous
and laminar, in which case a complete solution to the Navier-Stokes equations is

31
 2.4 Theoretical and Experimental Correlations

possible. For a sufficiently slow, steady, unidirectional, Newtonian flow, Darcy

(1856) defined the pressure drop-velocity relation as

∆P µu
=− (2.20)
L kp

where L is the bed length, µ is the dynamic viscosity of the fluid, u is the fluid
velocity and kp is the permeability. The minus sign simply distinguishes between
a drop in pressure and an increase in pressure. The permeability is the property
which describes the ease in which a fluid can percolate through it and can be
described as the medias hydraulic conductivity. Darcy’s law can be described
as the linear element of the Reynolds hypothesis with the coefficient quantified
as α = µ/kp . Brinkman (1947) used the equation of Darcy (1856) and added a
term for viscous shear. Neale & Nader (1974) suggest that due to the additional
viscous shear term, the equation proposed by Brinkman (1947) is compatible with
boundary layers within a porous medium. Brinkman (1947) extended Darcy’s law
to give
µu
∇P = − + µ′ ∇u (2.21)
k
Blake (1922) modified the equation of Darcy (1856) with the addition of the
superficial velocity, U = uε (The Depuit-Forcheimer relation). The superficial
velocity is the velocity that would pertain if the fluid occupied the duct on its
own without the presence of the media. Blake (1922) also quantified permeability
(kp ) based on an empirical approach, with the addition of the hydraulic radius,
Rh giving
Rh2 ε
k= (2.22)
k2
The hydraulic radius is satisfactory for a square duct or open channel where the
cross section remains constant, but for a packed bed an average is introduced
based on the volume to surface area ratio. We know that the volume available to
the flow is the void volume, Vv and is related to the porosity by ε = Vv /V . The
whetted surface area per unit volume is aw = ap /V and combining this with the
porosity gives the packed bed hydraulic radius given as

ε
Rh = (2.23)
aw

32
 2.4 Theoretical and Experimental Correlations

substituting in terms of the equivalent diameter (dP gives

dp ε
Rh = (2.24)
6(1 − ε)

and this forms the basis for the theoretical tubes models. Some of the earliest
research carried out using this arrangement was by Blake (1922) and Kozeny &
Sitzber (1927). To incorporate a higher range of Reynolds numbers using the
equation of Blake (1922), Kozeny modelled a packed bed as a bundle of parallel
small diameter tubes (Kozeny & Sitzber, 1927; Strigle, 1994) and hence the
pressure drop can be related to Hagen-Poiselle flow. The Hagen-Poiselle relation
(U.S) is given as
∆P µU
= 1 2 (2.25)
L 2
R h

Making the assumption that a packed bed can be made up of a bundle of parallel
theoretical tubes, the hydraulic radius can be substituted into the Hagen-Poiselle
relation, with the addition of the superficial velocity U = uε giving

∆P U µL(1 − ε)2
= 72 (2.26)
L d p ε3

This assumption relies on all the tubes being equal in length. In reality, packed
beds are complex and tortuous, with interconnecting flow channels having no
uniform geometry in a bed of randomly dumped particles (Strigle, 1994). As
a result the fluid takes a longer path than simply the length of the theoretical
tube. Carman (1937) then applied the work of Kozeny & Sitzber (1927) to
experimentally determine pressure drop through packed beds and concluded that
the bed length should be multiplied by a constant (25/12) to account for the
extra length (Tortuosity) (Bird et al., 1960) and as a result produced the Carman-
Kozeny equation (Carman, 1938; Strigle, 1994). However, the value of (25/12) is
not a constant and is heavily dependant on the bed tortuousity. This equation is
sometimes referred to as the Blake-Kozeny equation. These designations are often
interchanged between Blake-Kozeny and Carman-Kozeny (Middleman, 1998).

∆P 150µ (1 − ε)2 U
= (2.27)
L ε3 d2P

33
 2.4 Theoretical and Experimental Correlations

In many applications of packed beds where the flow is characterised as being

turbulent; the Carman-Kozeny equation does not hold where there are high losses
in kinetic energy Rep > 10 (Strigle, 1994), in which case the non-linear element,
βU 2 needs to be quantified.

2.4.3 Non-linear models

Non-linear models are based on quantifying a suitable value for β in the Reynolds
hypothesis. We know that for laminar flow Blake (1922) and Kozeny & Sitzber
(1927) modelled the pressure drop by the pipe flow assumption and the theoret-
ical tubes model and applied the Hagen-Poisselle relationship for laminar flow.
We know from elementary fluid mechanics that the Darcy-Weisbach equation
describes the pressure drop through a straight pipe in turbulent conditions
 
L U2
∆P = f (2.28)
D 2g

Burke & Plummer (1928) used the Darcy-Weisbach relation, and by substituting
superficial velocity and the hydraulic radius for a packed bed deduced the term
to form 
L U 2 (1 − ε)
∆P = 3/2f (2.29)
dp 2ε3
From extensive empirical investigations Burke & Plummer (1928) concluded that
for turbulent flow through a packed bed the friction factor could be described as
3f /2 = 3.50. Substituting this back into equation 2.30
 
L U 2 (1 − ε)
∆P = 1.75 (2.30)
dp ε3

Bird et al. (1960) suggest that the Burke-Plummer equation is only valid for flow
Reynolds numbers greater than Rep = 1000.
Ergun (1952) took the equations of Carman (1937) and Burke & Plummer
(1928) and added these together producing a mathematical portmanteau to model
both laminar and turbulent flow, which completely satisfies the linear and non-
linear terms in Reynolds hypothesis, by theoretical tubes and inclusive of an

34
 2.4 Theoretical and Experimental Correlations

empirical coefficient.

∆P 150µ (1 − ε)2 1.75 (1 − ε) ρ 2

= 3 2
U+ U (2.31)
L ε dP ε3 d P

Although the Ergun equation satisfies Reynold’s hypothesis, the coefficients are
based around the empirical data of Carman (1937) and Burke & Plummer (1928)
which are mostly high aspect ratio. Foumeny et al. (1993) suggest that when the
aspect ratio is < 50, wall effects cause the Ergun equation to yield a poor result.
In addition Hicks (1970) studied the equations of pressure drop and subsequently
concluded that the Ergun equation is limited to Redp /(1 − ε) < 500 and (Handley
& Heggs, 1968) coefficients for the Ergun equation to 1000 < Rep /(1 − ε) < 5000.
Choi et al. (2008) suggests that for low Reynolds numbers (Rep /(1 − ε) < 10)
the Ergun equation tends to underpredict the pressure drop and in the larger
Reynolds regime (Re/(1 − ε) > 10 the Ergun equation over predicts the pressure
drop in comparison to experimental results.
The numerical coefficients of the Ergun equation have been the subject of
much discussion, with conflicting observations. The coefficient values are usually
given as 150 and 1.75 (Ergun, 1952). From empirical data Leva (1959) suggested
alternative coefficients of 200 and 1.75, with MacDonald et al. (1979) recom-
mending coefficients in the range of 180 and 1.8 − 4.0. In addition Handley &
Heggs (1968) suggest values of 1.24 and 368. Du Prieur and Woudberg state
that the coefficients are severely dependant on the interstitial physical flow con-
ditions that vary considerably between different types of structure (Plessis &
Woudberg, 2008). Investigation into the coefficients (Plessis & Woudberg, 2008)
further suggest that tuning the Ergun coefficients is not recommendable due to
its resemblance to a ‘fudge factor’ approach to which the coefficients are adjusted
to fit every new application. In reality is is possible that these coefficients are
unique to the individual bed being analysed, in which case it is almost impossible
to create a set of coefficients to fit all cases perfectly. On the contrary, there have
been arguments to suggest that 1.75 and 150 are not constants but depend on
the Reynolds number, particle shape, size and porosity (Handley & Heggs, 1968;
Hicks, 1970). Research carried out by Rehder (1990) suggests that the parameters
that have the most influence on the result of the Ergun equation are the porosity

35
 2.4 Theoretical and Experimental Correlations

and the superficial velocity, where the superficial velocity causes the theoretical
pressure drop to increase or decrease, whereas the porosity causes the gradient of
the theoretical curve to vary in magnitude.

2.4.4 Wall correction models

In the case of low aspect ratio beds (Aratio < 50), where the particle diameter
is not sufficiently smaller than the tube diameter, the correlations discussed so
far do not model experimental data well. Choi et al. (2008) suggest that when
the bed diameter (D) is not significantly larger than the particle diameter (dP ),
the flow maldistributions and wall friction are not negligible. In these cases an
assembly of models have been devised known as wall correction models, in many
cases based on an existing correlation which is corrected to take into account
the wall region. The theoretical tube models such as Ergun & Orning (1949),
Burke & Plummer (1928) and (Kozeny & Sitzber, 1927) assume the porosity to
be uniform across the bed, such as the case of an infinite packing, which neglects
the wall region and ultimately leads to the breakdown of these equations at low
aspect ratios.
Gibilaro (2004) proposes a simple model for the effect of the container wall
based on the application of a two zone flow model applied to the Ergun equation.
The bed is separated into two regions, where the bulk or core region is the central
region unaffected by the wall, and the wall region is the region directly adjacent
to the wall usually taken as a distance dp /2 (Martin, 1978). Gibilaro (2004)
replace the function of superficial velocity (U ) with a corrected average superficial
velocity, bulk velocity (UB ) or fluid flux defined as

US
UB = (2.32)
2.06 − 1.06( D/d P −1 2
D/dP
)

In addition, porosity (ε) is replaced with a bulk porosity (εB ) which is the porosity
in the bulk region excluding any effects of the wall region. This equation models
pressure drop relatively well at low Reynolds regimes but deviates away from the
experimental results quickly as the Reynolds number increases.

36
 2.4 Theoretical and Experimental Correlations

Reichelt (1972) proposed a correlation to determine dimensionless pressure

drop based on the Ergun equation with the addition of two coefficients accounting
for the effect of the confining wall region and the porosity effect at high Reynolds
numbers caused by the wall region (Eisfeld & Schnitzlein, 2001). The equation
proposed by Reichelt (1972) is based on the hydraulic radius theory and is given
in the form
K1 A 2W (1 − ε) AW 1 − ε
φ= + (2.33)
Redp ε3 B W ε3
where K1 is a coefficient. The wall correction terms are given as

2
AW = 1 + (2.34)
3(D/dp )(1 − ε)

and   2 2
dp
Bw = k1 + k2 (2.35)
D
Using the equation proposed by Reichelt (1972), Eisfeld & Schnitzlein (2001) de-
vised a semi-empirical approach. Their method utilises the correlation proposed
by (Reichelt, 1972) with the modification of the coefficients to a regression fitted
to 2300 data points from 23 different research articles. Some of the work carried
out by Leva (1959) is purposely omitted from the study due to their inconsis-
tency with the other experimental data (Eisfeld & Schnitzlein, 2001; Wagstaff &
Nirmaier, 1995). Their correlation resulted in several communications (Tsotas,
2002b) regarding their failure to recognise the dual effects of increased wall fric-
tion and flow maldistributions at the container wall. Tsotas (2002b) argued that
the two effects could be studied separately by the extended Brinkman equations
and cannot be accounted for by the correlations proposed by Eisfeld & Schnit-
zlein (2001). There is little weight to this argument as the equation of Eisfeld
& Schnitzlein (2001) is derived from a range of the empirical data which cor-
rectly encompasses all of the various effects of flow channelling and increased
wall friction in the range of experimental methods. However the effects of flow
channelling and increased wall friction cannot be separated from the data. What
is more, their model is intended to give an empirical approximation for dimension-
less pressure drop, which is a bulk value, and is not intended to study increased
wall friction and flow distributions. However, the increased wall friction and flow

37
 2.4 Theoretical and Experimental Correlations

inhomogeneities may account for the large spread of data at high Reynolds num-
bers. Eisfeld & Schnitzlein (2001) conducted quite an extensive survey comparing
the data with other empirical correlations. In most cases the route-mean-square
deviations of the data surveyed compared to other correlations is low (σ < 0.3)
(Ergun, 1952; MacDonald et al., 1979; Reichelt, 1972), but in some cases it can be
as high as σ > 0.7 (Foscolo et al., 1983; Foumeny et al., 1993; O’Neill & Benyahia,
1997).
Choi et al. (2008) propose a correlation which recognises the increased wall
friction and the increased porosity as a result of the confining wall to predict
pressure drop per unit length at wide range of Reynolds numbers (Re = 102 −
103 ). They modified the original equation proposed by Ergun (1952) to take into
account the effect of the confining wall.

∆P 150(1 − ε)2 µU M 2 1.75(1 − ε)ρM U 2

= + (2.36)
L εd2p ε3 d p

where M is defined as
2dp
M =1+ (2.37)
3D(1 − ε)
The correlation fits the data of some research with some degree of accuracy (σ <
10) (Handley & Heggs, 1968; Mehta & Hawley, 1969), but fits the experimental
data of Foumeny et al. (1993) with less accuracy (σ > 25). On the assumption
the data has been collected competently, this reinforces the theory that there
is no single correlation that can model every specific case with a good deal of
accuracy. A creative correlation which is difficult to categorize is the correlation
proposed by Fried & Idelchik (1989). The model determines pressure drop based
around the packing angle, where the bed porosity, ε is given as

π
ε=1− √ (2.38)
6(1 − cosθ)( 1 + 2 cos θ)

Fried & Idelchik (1989) modify the Reynolds number to produce

0.45Uup dρ
Re = √ (2.39)
µ(1 − ε) ε

38
 2.5 Computational literature

Subsequently, using these the parameters the pressure drop per unit length be-
comes   2
∆P 1.53 30 3 ρU up
= 4.2 + 0.7
(2.40)
L ε d Re Re 2
This method is only valid for packing angles between θ = 60◦ − 80◦ (Hexagonal
close packing and a simple cubic lattice), is limited to porosity ranges between
0.26 and 0.48, and does not accurately model random, disordered packing.

2.5 Computational literature

The previous section discusses the development of correlations based on theoret-
ical and experimental approaches to predict the drop in pressure across a packed
bed or porous medium. Experimental and theoretical approaches in many cases
can give a good indication of pressure drop, but do not allow quantitative anal-
ysis of the flow pattern analysis within the bed. Flow models assume the media
to be a homogeneous bulk or gross entity and neglects the complex internal flow
structure of the media. Due to this, the pressure drop models are considerably
more effective in high aspect ratio beds which can be considered as a homoge-
neous continuous medium. The most realistic and extensively used approach to
analyse the motion of a fluid within a complex, heterogeneous structure is to
use computational fluid dynamics (CFD) to solve the motion of a fluid mathe-
matically from a set of governing equations. CFD can be broken down into two
basic approaches; continuum approaches (Finite Element and Finite Volume) and
the lattice-Boltzman method. The lattice-Boltzmann method solves the discrete
Boltzmann equations which describe the motion of individual fictitious particles
coupled with a collision model for a dilute gas. In a dilute gas, molecules move
around as free particles apart from when they collide with other particles. This
method makes use of certain parts of the Boltzmann equation coupled with a
collision model to describe the particle motion and collision events. The linkage
with the finite volume method is under particular assumptions the Boltzmann
equation can be used to derive a form of the Navier-Stokes equations.The finite
element and finite volume methods solve the Navier-Stokes equations by inte-
grating the equations over a mesh of finite elements or volumes. The equations

39
 2.5 Computational literature

are coupled and non-linear and contain unknowns in which finite difference equa-
tions are substituted for the properties of the flow, which are solved using an
iterative method. Once a solution has been generated, properties of the flow can
be extracted for each individual cell or element for analysis.
CFD techniques are well defined and have established themselves as an ac-
curate representation of fluid behaviour. The major constraint with any CFD
technique applied to a complex media is defining a suitable computational geom-
etry for the flow domain in which the fluid motion can be analysed. In order to
solve the governing equations used in a CFD solver, the flow domain has to be
suitable discretized into small volumes or elements leading to the solution of a
partial differential equation. This process is referred to as meshing or grid gener-
ation. If a suitable grid or mesh is not produced, geometric fidelity issues occur
with respect to the representation of the computational domain. Moreover, if a
mesh of suitable quality is not created, instabilities occur when trying to solve
the governing equations.
Primarily there are four techniques for describing the geometry with rele-
vance to packed beds. The first is to produce a structured bed. Because the
bed is structured the location of each particle can be easily described by sim-
ple mathematics coupled with an appropriate coordinate system. This can be
used in conjunction with CAD software or created directly in a mesh generation
program such as GAMBIT (Fluent Inc). Secondly, we know a structured bed
is repeating, therefore its geometry does not change through the bed and the
flow characteristics are not likely to vary. Using this assumption and mesh gen-
eration software a simple unit cell approach can be adopted. One or more unit
cells (voids) can be modelled using CFD and the data extracted and collated to
produce a representation of a full scale model. This method vastly reduces com-
putational expenditure, but in some cases does not take into account the small
scale flow characteristics that would be achieved by modelling a full bed. This
method is not suitable for highly disordered geometries where the flow patterns
vary considerably through the bed. However, does provide a reasonable engi-
neering approximation in regard to a pressure drop approximation in structured
beds. Thirdly, where unstructured, random beds are required for study a non-
deterministic object packing algorithm is needed to generate the locations of each

40
 2.5 Computational literature

of the particles. This can then be used in conjunction with CAD software and a
suitable meshing program to generate a workable computational domain. Finally,
where an actual physical beds available for study, non-invasive methods such as
magnetic resonance imaging (MRI) and computed tomography (CT) can be used
with geometry defining software to recreate a bed for computational analysis and
direct comparison.
Once the geometry has been defined using one of the four major techniques
discussed, the geometry must be suitably discretized so that the bounding equa-
tions can be suitably integrated. Grids or meshes are constructed from a matrix
of elements or cells and in nearly all cases these are tetrahedral, hexahedral or
polyhedral which are all constructed from a series of connecting flat surfaces.
Packed beds often contain spherical media, where a sphere is a round contin-
uously sided volume and cannot be represented as fully smooth by a series of
flat surfaces but rather is represented as a geodetic sphere. Further more, when
spheres are in contact with each other they produce an infinitely small contact
point which is difficult for a mesh generation software’s discretization algorithm
to fully describe at this stage. This phenomena has become one of the major con-
straints when analysing packed beds using CFD approaches. Some Researchers
have used novel methods to overcome these constraints which are discussed in the
following section. The constraints and limitations concerning CFD approaches
are discussed in detail in Chapter 3.

2.5.1 Structured Bed Approaches

In most applications of CFD to packed beds the ratio of particle to tube diameter
(aspect ratio) is low. This is attributed to the complex geometries associated
with high aspect ratio beds and the vast number of cells required to realistically
represent the flow domain. This follows a simple relationship; the larger the
number of cells, the more computer power needed to realistically solve the problem
in a suitable time frame. A sphere is also an infinitely sided object, so in theory
to fully represent a sphere an infinite number of cells is required, which of course
is not possible. In which case it has to be represented by a finite number of cells.
The more cells the closer the sphere is to a true sphere.

41
 2.5 Computational literature

Naturally, in the early days of computing the only beds possible to practically
analyse were those of low aspect ratios in cases with only a couple of spherical
particles in 2-dimensions and structured formation, with steady laminar flow.
Spheres and a structured array are chosen due to their simple mathematical
description in respect to location and geometry. It is debatable if this can be
classified as a packed bed or simply a cluster of spherical particles. This alone
raises the question, when is it reasonable to suggest a cluster of spheres has be-
come a packed bed? Furthermore, is one continuous layer of spheres a packed
bed, or does it require 2, 3 or 4? In many cases where CFD has been applied
to packed beds spherical particles are chosen due to the possible regular pack-
ing regimes, which can easily be described mathematically, such as face centred
cubic (FCC), or close hexagonal (CHP). These arrangements are equivelent to
the stacking patterns of oranges. If the reader is interested in the flow through
oranges, complete with crate, the reader should consult Delele et al. (2008). The
advantage of using spheres is they only require a single coordinate, e.g Cartesian
or polar, and single value of radius, with this data a complete description of both
the location and geometry of the particle can be determined. However, this is
often outweighed by the complications produced by particle contact points and
the fact a sphere is a continuously sided object which cannot be meshed retaining
its geometric fidelity.
The earliest work demonstrating the application of CFD to packed beds, is
that of Dalman et al. (1986). They use a velocity-pressure formulation of the 2d
Navier-Stokes equations solved numerically using a finite element technique to
analyse the flow past two spheres. Dalman et al. (1986) investigated laminar flow
with Reynolds numbers up to Re = 200, with Prandtl number of 0.72 and 7.0 for a
range of sphere sizes and separations, where the Prandtl number is defined as the
ratio of dissipation and conductance in the form Pr = µcp /k, where cp is the heat
capacity ratio and k is the conductivity. Although the packing possibilities are
limited using this approach, the work gave a valuable insight into flow structure
through packed bed and demonstrated the formation of eddies between the bed
particles (Taskin, 2007) indicating regions of poor heat transfer. The work of
Dalman et al. (1986) is probably the first investigation into the flow structure
within internal bed voids. With all early work, one of the main difficulties is

42
 2.5 Computational literature

producing a workable mesh. Where two spheres touch elements are often highly
skewed with poor aspect ratios. Dalman et al. (1986) report no particle contact
points points between the spheres and the particles are not in contact with each
other, ammeliorating the problem of degenerate elements at the particle contact
points.
Eight years after Dalman et al. (1986), Lloyd & Boehm (1994) studied flow
and heat transfer around a linear array of 8 spheres in 2d using the finite element
package FIDAP. Lloyd & Boehm (1994) use Reynolds numbers of 40, 80 and
120 with Prandtl numbers ranging from 0.73-7.3. This work was a considerable
advancement from that of Dalman et al. (1986) due to the number of particles
being analysed. They investigated the effect of sphere spacing on the particle
drag forces and concluded that the heat transfer from the particles to the fluid
decreased as the particle spacing increased.
Derx & Dixon (1996) performed possibly one of the earliest 3-dimensional
calculations on a bed of three spheres. As with the work of Dalman et al. (1986)
no contact points are reported between the particles. Logtenberg & Dixon (1998)
modelled a 3-dimensional bed of 8 spheres in the form of two layers of four
spheres perpendicular to the flow with low aspect ratios of 2.43. They use the
commercial CFD code FLOTRAN to investigate fluid flow heat transfer using
air for Reynolds number of Re = 9 − 1450. Logtenberg et al. (1999) modelled a
bed of ten spheres with aspect ratios of Aratio = 2.43 using a FEM. Similar to
the earlier work of Lloyd & Boehm (1994) they focus on heat transfer and fluid
flow, focusing on wall-particle contact points. Unlike Dalman et al. (1986) and
Logtenberg & Dixon (1998) they emulate the contacts between the spheres by
leaving small gaps between the particles set to zero velocity on the assumption
that there is a stagnation zone at the particle contact points. With the early work
described so far, the major limiting factor in regard to producing realistic flow
patterns is the small number of particles being analysed. As computers advanced
and computational power became more readily available, the number of particles
being modelled grew considerably and progressed to studies in 3-dimensions.
Dixon & Nijemeisland (2002) use the finite volume technique to model small
clusters of 44 structured spheres with aspect ratios, Aratio = 2 and Aratio =
4 and Reynolds numbers of Re = 373 − 1922. They use a mesh constructed

43
 2.5 Computational literature

10000
Derx and Dixon (1996)
Lllyod and Boahm (1994)
Logtenberg (1999)
Manji et al (2006)
1000
Naratarasukra (2004)
A Baker and Tabor (2010)
Number of Particles

Hassan (2010)
Dixon and Nijemeisland (2002) A
100 Rupesh et al (2001)

10

1980 1990 2000 2010 2020

Years

Figure 2.10: Number of particles as a function of date in years

of tetrahedrons, unstructured in nature, where a suitable mesh density (based

on a mesh refinement study) is conducted on single spherical particle instead
of the whole domain. They also correctly address the issue of particle contact
points and refine the mesh further in the region. As expected with meshing
curved geometries with zero spacing the mesh exhibits high levels of skewness
and degenerate elements. However, in the case of the higher aspect ratio bed
(Aratio = 4) the mesh is too coarse to determine an accurate flow field. However,
the work does produce convincing results for aspect ratios (Aratio = 2) and show
vorticity and recirculation in the voids.
Guardo et al. (2007) studied the effect of turbulence models in packed beds for

44
 2.5 Computational literature

modelling wall to fluid heat transfer in a regular pack in a cylindrical container.

They use the commercial code Fluent to study a range of Reynolds numbers
between 127 < Re < 912 and obtained values for pressure drop. They implement
the k−ε (Standard, RNG, Realizeable), Spalart-Allmaras and the k−ω turbulence
models and compared the the pressure drop from the expected behaviour of the
Ergun equation. The pressure drop calculations produced by the turbulence
models show good agreement with the Ergun equation. However, Guardo et al.
(2007) use an aspect ratio of Aratio = 3.923 which is low aspect ratio bed and we
know the Ergun equation does not yield a good result (Foumeny et al., 1993) for
Aratio < 50. In addition the Spalart-Allmaras turbulence model is designed to be
applied without wall functions, in which case the mesh must be fine enough to
resolve the flow right down to the viscous sub layer. The mesh volumes are not
discussed, but from figures published it looks plausible that this is not the case. A
thesis defended by Guardo (2007) focuses on flow around spherical particles in a
cubic array based on the geometry of Suekane et al. (2003) using the finite volume
technique. To avoid problems associated with degenerate mesh elements around
particle contacts, each of the particle diameters are enlarged by 1%. Correlations
are compared to magnetic resonance velocimetry on the same geometry.
Even today, it is still not yet possible to perform a 3-D simulation of a full-
size industrial scale fixed bed reactor (Freund et al., 2005). Once the full internal
micro-scale flow of a full size packed bed reactor can be accurately modelled, the
possibilities for further research in the field of CFD of packed bed reactors is
limited to the application of new techniques and approaches in which the process
of delivering a model is made more efficient.
Manjhi et al. (2006) examine regular packings of small numbers of spheres
(< 10) in a cylindrical tube using lattice Boltzmann techniques, and are able
to produce convincing visualisations of the micro-structural flow in these highly
simplified geometries. Attempts have also been made to blend numerical simu-
lations on simplified geometries with empirical data (Tobı́s, 2008). A unit-cell
approximation makes the calculation reasonably tractable. However for random,
unstructured, packed beds the true geometry is complex and tortuous, and a sin-
gle unit cell may not be considered representative. Instead a larger section of the
bed must be simulated, large enough to contain a significant and representative

45
 2.5 Computational literature

diversity of micro-structure and thus hopefully large enough to be considered

representative of the micro-structural flow.
Rupesh & Joshi (2008) study low aspect ratio (Aratio = 5) fixed beds of 151
spherical particles arranged in 8 layers in a cubic orientation. They monitor drag
coefficients and pressure drop at pipe Reynolds numbers of Redp = 0.1 − 10, 000.
They state ‘for simplicity’ that the inlet velocity profile is kept flat. This is
suitable for a turbulent regime due to its closeness to the 1/7 power law. However,
using a flat velocity profile for a creeping regime (Stokes flow) is questionable due
to the likelyhood of the existence of a parabolic profile. Furthermore, Rupesh &
Joshi (2008) neglect to state whether the velocity inlet has been adjusted to form
Uavg based on the volumetric flow rate. Rupesh & Joshi (2008) discuss increased
pressure drop due to wall friction in the creeping flow regime and reduced pressure
due to wall channelling in the turbulent regime based on their computational
results. Using CFD alone to make this statement is questionable due to the way
in which the k − ε model and standard laminar solvers are treating the wall
region in regard to suitable y+ values and inlet profile. In addition Rupesh &
Joshi (2008) give no indication of y+ values at the near wall region nor is there
any attempt at a mesh refinement study to indicate the meshes independence
from the solution.
A rather more unusual approach is adopted by Narataruksa et al. (2004). They
use the package COMSOL MULTIPHYSICS to analyse the the flow structure in
a pack bed with and without a static mixer with inlet Reynolds numbers (based
on pipe diameter) of Re = 2.31. 700 hundred catalyst particles are packed into
a cylindrical container in a cubic arrangement with considerable gaps between
the particles ensuring consistency in element size. A simple arrangement such
as this and an incompressible steady-state creeping flow regime, solver stability
and convergence is virtually guaranteed. However, due to the structured packing
regime the particles do not completely fill the container resulting in large gaps
around the bed leading to a dominant near wall region. Narataruksa et al. (2004)
use a mesh sizes of 168450 and 241471 cells, which are incredibly coarse for a
bed containing such a large number of particles, in addition the domain is not
subjected to a mesh indeterminacy study, in which case they cannot guarantee
the mesh is not effecting the overall solution. To put this into perspective, in the

46
 2.5 Computational literature

work of Baker & Tabor (2010), 174 particles are studied, and when the domain
is subjected to a mesh refinement, the pressure drop ceased to be affected by
the mesh after 1,168,328 cells. Narataruksa et al. (2004) are monitoring the
direction and the magnitude of the flow in addition to the pressure drop which
is considerably more sensitive. Although, this argument could be rebutted in the
sense that this is a creeping flow regime and the direction and the magnitude of
the flow is likely to be fairly constant and the particles have no contact points.
Even so, a mesh this course cannot attempt to en-capture the boundary layer
or model any effects in the near wall region. What is more, Narataruksa et al.
(2004) compare their results with the Blake-Kozeny equation. We know the
Blake-Kozeny equation is a theoretical tube model and assumes the porosity to
be uniform across the bed, a continuously packed medium and high aspect ratio.
In the work of Narataruksa et al. (2004) the aspect ratio is low (Aratio = 12.31)
and therefore the wall region will be dominant nullifying the the application of the
Blake-Kozeny equation. Moreover, in this specific case the particles do not fill the
whole container, in which case the magnitude of wall dominance will increase. A
more suitable equation to compare the results of Narataruksa et al. (2004) would
be a wall correction model such as Eisfeld & Schnitzlein (2001) or Reichelt (1972).
Finally, Hassan (2008) use Large Eddy Simulation to study flow in pebble
gas cooled reactors in a cubic array of 24 spheres. They initially conduct studies
using a zero equation turbulence model due to its speed and robustness, moving
on to the implementation of a LES solver. They also attempt to use PIV with a
matched refractive index fluid to derive experimental results. They report com-
plex flow structures near the particle contact points where the size and frequency
of turbulent events increase with the Reynolds number. Hassan (2008) correctly
bring to light the issues associated with two touching spheres and use the CFD
solver CFX-5 due to its function regarding its method of treating zero spacing by
creating a common point on the vertices of a plane (Hassan, 2008). They use a
Delaunay based algorithm to generate a mesh of 1.7 million tetrahedral elements.

47
 2.5 Computational literature

2.5.2 Simple Unit Cell Approaches

An additional major constraint when applying a CFD approach to a packed bed
of many particles, is the relatively large meshes which are produced and the lim-
itations in computational power. One possible approach to ameliorate this is to
assume a regular structure to the bed which can be reproduced using a simple
unit cell (Hellström & Lundström, 2006). This differs from a structured bed ap-
proach in that only one or two pores are studied and the flow is assumed to be the
same in all the other pores contained within the bed, rather then studying bed as
a whole. An example of this was presented by Tóbis (2000) using a simple cubic
unit cell and comparing against theoretical expectations and experiments on 6
similar structured beds. Tóbis (2000) uses the finite volume code Fluent and the
standard k − ε, Spalart and Almaras and the Reynolds Stress turbulence models
to investigate flow around the media and frictional resistance. The work of Tóbis
(2000), as expected, indicates large discrepancies when compared to the Ergun
constants, probably due to the Ergun constants being specific to high aspect ra-
tio beds (Aratio < 50). Regardless of the Ergun constants, the experimental and
computational data presented in the work of Tóbis (2000) compares well. Gunjal
et al. (2005) conducted a similar study by employing a single unit-cell approach of
an array of spheres encompassing simple cubical, 1-D rhombohedral, 3-D rhom-
bohedral and face centred cubic structures, coupled with CFD and investigate
the interstitial flow in the void space. A similar approach was used by Calis et al.
(2001) to model a limited subset of bed geometries known as composite struc-
tured packed beds; instead of a single unit cell they represented the bed in terms
of a single linear row of spheres with appropriate boundary conditions. This con-
siderably reduces the computational effort required in contrast to modelling a full
packed bed, however, it is plausible that this technique does not give a realistic
representation of the full flow structure.

2.5.3 Random unstructured Beds

In most applications of packed beds, the geometry is disordered and unstruc-
tured. We have discussed the non-deterministic nature of unstructured bed ge-
ometries and no deterministic mathematical relation will describe them. In such

48
 2.5 Computational literature

cases the geometry must be defined via some form of randomised object pack-
ing algorithm. Zeiser et al. (2001) were probably the first to actually model a
random unstructured by using CFD. They used a Monte-Carlo method to gen-
erate a random-unstructured bed of particles. An example of which is the code
MacroPac (Evans, 1988; Rowe et al., 2005), which uses a Monte Carlo approach
to explore packing space. Zeiser et al. (2001) then apply the Lattice-Boltzmann
method to model the expected behaviour of the fluid in a reacting, viscous flow
regime. They report velocity channelling close to the wall region and were in good
agreement with observations from experiment. Caulkin et al. (2007) describe the
code DigiPac, where the objects to be packed are represented as assemblages of
voxels in space which are allowed to move one grid spacing at a time within a
cubic lattice; their resulting motion allows them to explore every possible packing
space. The work of Caulkin et al. (2008) provides an interesting review of the
viability of the digital packing algorithm DigiPac to predict bed structure. They
use the Lattice-Boltzmann technique to investigate flow structure and velocity
distribution within beds of mono-sized spheres, equilateral cylinders and raschig
rings with aspect ratios of Aratio = 6.1, 8.3 and 5.0 respectively. They compare
results from LBM with experimental data empirically derived from non-invasive
and non-destructive measuring techniques. Zeiser et al. (2002) apply Lattice
Boltzmann techniques to solve the flow in a packed column with the addition
of a parabolic velocity profile and a fixed static pressure at the outlet. Their
aspect ratio is low (Aratio = 3) so wall effects are dominant in their flow; however
they are able to investigate in some detail the micro-structural flow in the bed.
They also report that discretization of the spherical objects leads to a significant
error, most probably due to the block structured lattice and in this case they
conduct a detailed mesh refinement study. They controversially report that the
Lattice-Boltzmann method can produce results equally as accurate as the finite
volume method but with much coarser grids. Pan et al. (2006) apply Lattice
Boltzmann techniques to both a simple cubic arrangement unit cell model and
also to a random bed; the focus of their paper is comparing different variants of
the Lattice Boltzmann technique rather than experimental validation. However,
there are still unresolved computational concerns with LBM due to its restrictions
to a limited class of mesh, there are also no reasons to suggest the LBM method

49
 2.5 Computational literature

outperforms traditional Navier-Stokes based solvers (Tabor et al., 2008). Mag-

nico (2003) uses the finite volume techniques applied to a somewhat larger bed
(several hundred particles and Aratio of 5 − 7) but quite a low Reynolds number
(between 20 and 200) eliminating the need for a turbulence model.
Jafari et al. (2008) use the finite volume technique and the commercial code
Fluent to analyse a packed bed reactor of several hundred particles. This work is
particularly significant in the field due to the range of turbulence models applied.
They compare Large Eddy simulations (LES) with Reynolds stress models and
compare the steady-state incompressible Navier-Stokes equations as an alterna-
tive to a turbulence model. They conclude that the steady-state Navier-Stokes
equations model flow well through a packed bed of spheres and hold well even
into the lower end of the turbulent regime. In addition they use a combination
of C++ programming with a code obtained from MatLab to create a journal file
for the the CAD based meshing program GAMBIT as described by Zamankhan
et al. (1999) which uses a random number generator to carefully position a 3d ar-
ray of non-overlapping spheres. However, it can be observed that their spherical
particles merge significantly at the inter-particulate contact points as a prod-
uct of the mesh resolution which is a common problem. In the case of complex
spherical geometries an ‘all or nothing’ situation arises from the fact that produc-
ing a successful mesh involves making the particles ‘non-contact’ or making the
particulate contact points unrealistically large. This can be achieved by either
enlarging or reducing the sphere diameters. In regard to analysing the flow pat-
terns and pressure drop, either method can be conclusively argued as being more
accurate than the other. This also presents an issue when using this technique
in regard to dermining an accurate value of porosity from the mesh volume. At-
makidis & Kenig (2009) analyse the wall effect on the pressure drop in packed
beds with moderate aspect ratios in the laminar flow region. They use the fi-
nite volume CFD solver, CFX. They analyse both regular and irregular packed
beds of spheres generated using the ballistic deposition method combined with
a Monte-Carlo approach to produce a random packing. Due to meshing issues,
each particle is shrunk by 2% to avoid skewed elements at the inter-particulate
contact points. This technique changes the flow geometry considerably, however
it significantly reduces problems associated with mesh skewness and in addition

50
 2.5 Computational literature

reduces particle merging and contact points additionally affecting the geometry.
It is arguable whether the benefits of having a good mesh outweigh the effects of
changing the flow geometry which has already been modified to some extent by
the discretization in the mesh creation algorithm. However, Atmakidis & Kenig
(2009) do conduct an extensive mesh refinement study where the local void veloc-
ity is analysed and their mesh contains over 30,000,000 cells. Atmakidis & Kenig
(2009) report increased velocity channelling at the near wall region where the
porosity is greater than that of the core region and channelling in random packs
is not as structured as channelling in regular packs. Although their results are
low aspect ratio, they compare their data with the expected results of Carman
(1938), Ergun (1952), Reichelt (1972) and Zavoronkov et al. (1979).

2.5.4 Scanned Geometries

Methods discussed in the previous section focus on the creation of artificial beds
either by a simple mathematical relationship, i.e a regular pack, or certain ran-
dom sampling algorithms for disordered packs (Monte-Carlo simulations). These
methods are limited to the creation of artificial beds and their comparison with
structured or equivalent beds, which we already know may the same in regard
to properties such as porosity, particle diameter etc, but may display a different
internal structure. Methods have been developed to recreate actual physical beds
and reproduce them in a computational form. One method which has been used
extensively in the analysis of packed beds and porous media is magnetic reso-
nance imaging (MRI). The acronym MRI is generally used as an alternative to
NMR (Nuclear magnetic resonance) due to its inclusion of the word nuclear; due
to the Cold war and the Chernobyl disaster, patients lacked confidence in the
safety of the technique. MRI is used extensively in medicine for diagnostic pur-
poses, but its use is becoming more widespread in determining random disordered
geometries due to its non-invasive properties.
MRI techniques work by exciting hydrogen atoms. The human body is com-
prised of soft tissue, which contains water molecules. Water is comprised of two
protons of hydrogen, H2 and one of oxygen, O forming H2 O. A proton has a mag-
netic moment (a torque which is exerted due to its magnetic field) and when the

51
 2.5 Computational literature

protons contained within the water molecules are influenced by a magnetic flux
some of their moments align with the direction of the moment of the magnetic
source. In the case of MRI, the scanner uses a radio frequency to produce an
electromagnetic field (Callaghan., 1994). As the intensity and frequency is mod-
ulated more or less water protons spin up. When the field is turned off, the spin
states relax. As a result of the difference in kinetic energy a photon is released,
which is detected by the scanner and processed for analysis. This technique pro-
vides not only non-invasive visualisation of structures but also has a functional
ability, providing limited flow visualisation. However, functional MRI techniques
are limited to relatively low pipe Reynolds number, Re < 200 (Guardo, 2007)
and the resolution is not sufficient to visualise the smaller turbulent structures.
Another non-invasive technique, not as widely used in the analysis of porous
media is X-ray computed-tomography (CT, CAT scanning), which has been ap-
plied by Tabor et al. (2007, 2008). CT uses a different approach than MRI in
which the sample is bombarded with an X-Ray source. The source rotates around
the sample and an X-Ray detector, such as a scintillator tube, is positioned on
the other side which detects the strength of the radioactive flux passed through
the sample which can be interpreted and an image produced.
Non-invasive methods have proved valuable in both medicine and research
in the field of porous media research, however there is still much improvement
needed to the technique regarding fidelity issues and improvement to the relatively
coarse resolution. The scan data often contains artifacts caused by motion in the
scan or in the case of CT insufficient X-ray penetration (Herman., 2009).
What is of most interest in regard to this Thesis is determining a suitable
geometry for computational analysis using MRI. With both MRI and CT, signals
are analysed using a Fourier transform in which pixelated grey-scale 2-dimensional
slices are produced forming a stack of images representative of the signal strength.
Using the stack of grey-scale images, 3-dimensional surfaces can be extracted us-
ing a surface generation algorithm, such as morphological thinning or image based
meshing. Alternatively, instead of creating surfaces from the grey-scale image,
the image can be used, coupled with location defining algorithms to determine
the locations of spherical particles (Aste et al., 2004; Richard et al., 2003; Seidler
et al., 2000).

52
 2.5 Computational literature

In general, the use of non-invasive techniques applied to packed beds can

be broken down into two approaches, firstly a non-invasive technique is used to
determine the complex packed bed geometry and flow is modelled through the
structure using CFD (Sullivan et al., 2005), secondly the non-invasive method is
used to not only determine the geometry but also to analyse the flow structure
using MRV (magnetic resonance velocimetry). CT can be used to determine
the complex structure, but is limited by its non-functionality and therefore flow
distributions cannot be monitored in this way.
Manz et al. (1999), have applied the Lattice Boltzmann method based on an
MRI visualisation of a packed bed simulated for Reynolds numbers 0.4 < Re <
0.77 (i.e. in the creeping flow regime). They use 3d MRI and MRI velocimetry
(MRV) to characterise the structure of the inter-particle void space and three
components of velocity within the voids and compare this with the Lattice Boltz-
mann simulations. In both these cases the volume of interest for computational
simulation is determined directly from the 3d MRI data by partitioning the void
volume from the particles, and in both cases the Lattice Boltzmann method is
used to perform the calculations. The equations are solved on a highly-structured
lattice which has a 1-to-1 correspondence with the voxels of the MRI scan. The
resulting method is very fast and efficient, but the structured lattice means that
complex boundaries are pixelated. They compare the micro-structural flow from
the Lattice Boltzmann simulation with detailed measurements from the MRI,
but do not compare macroscopic parameters of the bed such as overall pressure
drop. Johns et al. (2000) use MRI to analyse a similar flow regime of flow through
packed beds of spherical particles in a creeping flow regime. They use an in house
algorithm based on a variant of the morphological thinning technique. When ap-
plied to non-invasive methods the morphological thinning technique is a way of
surface generation by means of the removal of certain foreground pixels (erosion)
from the stack of grey-scale slices generating a skeletal structure from which a 3d
surface can be generated (Johns et al., 2000).
Sharma et al. (2001) have applied magnetic resonance imaging to investigate
bed voidage, pressure drop and bed density in packed bed reactors. They con-
clude that in all cases, analysis of MRI data results in voidage values which are
consistently higher than values determined from water substitution. Ren et al.

53
 2.6 Literature Review Discussion

(2005) also use MRI coupled with velocity encoding and pulsed-field-gradient
nuclear magnetic resonance (PFG-NMR) to investigate the flow structure within
packed beds of cylinders and spheres with aspect ratios of Aratio = 1.4 − 32. They
conclude that radial flow structure follows an oscillatory pattern not too dissim-
ilar to the radial porosity, reinforcing the link between porosity and interstitial
velocity. In addition they concluded that the velocity in the near wall region can
be up to four times the average bed velocity indicating the distinct variance in
radial velocity in low aspect ratio beds.
Zhang et al. (2006) couple non-invasive methods with a digital packing algo-
rithm for a packed bed of cylinders. They use the code DigiPac which combines
both Monte-Carlo methods and Discrete element methods and compare this with
beds analysed using micro-CT. Their algorithm allows the reconstruction of each
particle with respect to its location and orientation calculated from the micro-
tomography scan. Their method ameliorates fidelity issues in regard to scan
resolution and signal attenuation.

2.6 Literature Review Discussion

The previous research discussed in this chapter all strives towards a common goal,
which is, ‘to be able to realistically model a packed bed’. Theoretical correlations
attempt to model pressure drop, usually from a blend of a theoretical assumption
and empirical data. However, the amount of data which can be extracted is
limited. In a high aspect ratio bed, which can be considered homogeneous and
isotropic these correlations, based around classic theoretical tube models, describe
the pressure drop well. Whereas with low aspect ratio beds, which we know
are highly disordered and heterogeneous, the majority of models do not model
pressure drop well due to each case being ‘individual’ with its own specific bed
characteristics and geometric properties. A packed bed can even have the same
dimensionless quantities, but the internal structure can be considerably different
affecting the flow paths and pressure drop considerably.
Many of the theoretical correlations which are intended for low aspect ratio
beds try to homogenise the effects of the individual bed structure which in reality
is un-homogonizeable. An example of this is the correlation proposed by Eisfeld

54
 2.6 Literature Review Discussion

& Schnitzlein (2001). The wide array of data on which the statistical model is
based attempts to accommodate every individual packed bed geometric case at
any given Reynolds number. In reality the regression does not give a realistic
value of dimensionless pressure drop for empirical data at the extremities of the
data set and the spread of data can be up to one order of magnitude. Although
unrealistic, each low aspect ratio bed should have its own coefficients based around
the beds structural properties, failing that, beds should at least be grouped into
some sort of categories for unique coefficients to be assigned. A pragmatic solution
for a Scientist or Engineer is to choose a correlation which is based on the bed
properties closest to the one in which the dimensional pressure drop is desired.
When more data needs to be extracted than solely pressure drop, computa-
tional methods can be applied usually in the form of computational fluid dynam-
ics. The computational methods described in this Chapter vary considerably but
are all burdened with the same issues regarding definition of the geometry and
creating a workable domain for computational analysis. Linked to this are certain
trade-offs e.g does a large bed with a coarse mesh provide a better representation
of the flow phenomena than a fine mesh with a small number of spheres? Geomet-
ric fidelity issues also arise with comparing computational beds with experimental
beds. Beds can be analysed in two ways, artificially i.e computationally; or phys-
ically and then scanned to create a computational model. The computational
work containing analysis of artificially created beds is usually limited to spherical
media due to the relative simplicity of a sphere location and orientation being
provided by a single coordinate. In the case of spherical structured beds the
domain can be defined by a simple mathematical description, however the par-
ticle contact points produce issues regarding meshing the infinitesimally small
contact points and this leads to highly degenerate meshes. Attempts have been
made to ameliorate this phenomena by reducing the particle volumes leaving a
constant gap between particles, losing geometric faithfulness. Conversely, parti-
cle volumes have been enlarged leading to a definable contact area. In addition,
simple unit cell approaches also reduce this phenomena. The one advantage of
using a simple unit-cell or a structured bed approach is an experimental replica
can be produced with relative simplicity in which computational data can verified

55
 2.6 Literature Review Discussion

through empirical measurements using methods such as PIV or LDA. However,

these experimental techniques are burdened by their own limitations.
In most real world applications of packed beds, the beds are random and
disordered and cannot be analysed using structured bed approaches. In which
case a non-deterministic random sampling program can be used to define the
location of each particle centroid. However, issues arise using this technique
regarding meshing the particle contact points and the technique is still burdened
with geometric fidelity issues as a result of the mesh resolution.
The one issue with creating a random-unstructured bed computationally is the
problem of reinforcing the data with empirical measurements on a geometrically
faithful bed. A structured bed can be easily recreated experimentally with fidelity,
however, issues arise when applying the same to random disordered beds. To
compare computational random beds with experimental, an equivalent bed can
be created with the same dimensional characteristics (aspect ratio, porosity), but
for a randomly dumped bed of particles to have the same internal geometry as
a computationally generated bed is near impossible. An equivalent bed allows
typical pressure drop to be analysed with quite some accuracy, but the internal
flow distributions can not be compared using this technique.
As an alternative to creating an artificial bed and comparing the data with
equivalent beds a physical, experimental bed can be scanned using a non-invasive
technique and reproduced computationally using surface creation software. How-
ever, this method is severely limited by the coarse scan resolution. The bed is
often faithful in respect to the flow paths and broad geometry, but where particles
touch certain fidelity issues can arise. Due to the scan resolution and magnetic
susceptibility induced artifacts, particles merge forming necking between particles
altering the bed porosity significantly.
The the author’s knowledge, none of the approaches described in the literature
provide an experimental geometry which is 100% geometrically faithful and thus
for which the pressure drop can be compared in detail.

56
Chapter 3

Computational Fluid Dynamics

“If a man’s wit be wandering, let him study the mathematics”

Francis Bacon

3.1 Introduction
Computational fluid dynamics, as the name suggests, is the use of a computer
or computers to solve problems associated with the motion of a fluid. This is
achieved in most cases by generating an approximate solution of a partial differ-
ential equation (PDE). We know that both ordinary differential (ODE) equations
and partial differential equations have unknown variables which are strongly de-
pendent on each other with no obvious, simple mathematical solution. The routes
of computational fluid dynamics can be traced back to developing numerical tech-
niques for the solution of ODEs and PDEs, ultimately leading to techniques for
determining approximate solutions to the Navier-Stokes equations. Initially, fi-
nite difference (FD) approximations were the first to evolve due to their relative
simplicity (Guardo, 2007). In this method a grid is created and the derivatives
in the PDE are replaced with a finite difference scheme (such as Forward, Back-
ward or central differencing) which then turns the PDE into a linear algebraic
description for solution. The major constraint with this method is that before the
invention of the digital calculator all calculations would have to be done by hand
and even the most rudimentary linear problems involving relatively simple lapla-

57
 3.1 Introduction

cian operators would be tortuous to solve iteratively (Guardo, 2007; Richardson,

1910). Ritz (1910) developed a method for solving problems associated with elas-
tic deformation using a method involving approximation of a potential function
in terms of trial functions with undetermined coefficients, where the unknowns
coefficients are determined by minimizing the potential function (Guardo, 2007).
Courant (1943) further improved this technique by discretization of the control
volume into triangles. Using the concepts introduced by Courant (1943), Clough
(1960) developed the first complete finite element method. With the invention
of the microprocessor in the 1950’s these techniques were soon being solved via
computer such as the work of Hess & Smith (1967), where primitive computer
calculations are used to calculate the potential flow function around arbitrary
bodies.
What is debatably considered to be the first numerical model which makes use
of a discretized control volume and a finite difference approximation was proposed
by Richardson (1910). Richardson used this method to analytically predict the
weather forecast, an attempt which ultimately ended in failure (Guardo, 2007;
Richardson, 1910). The first successful solution was achieved by Thom (1933) for
flow past a cylinder at low velocities. The method is particularly useful for hy-
perbolic, parabolic conservation equations [such as the Navier-Stokes equations].
Two principle methods exist today, the continuum methods (FV and FE) and
the lattice-Boltzmann method. There is much debate as to which method is most
suitable for porous media flow. Vidal et al. (2010) suggest that discretization of
the Navier-Stokes equations using the finite volume method has proved limited
when analysing flow in complex geometries. As a result Vidal et al. (2010) sug-
gest the Lattice-Boltzmann method is the preferred method of choice for many
researchers (Nourgaliev et al., 2003; Succi, 2001). Their argument is based on
the ease of discretizing complex geometries by means of a simple structured lat-
tice/grid, where the two phases are encoded with Boolean reasoning. However,
the special regular uniform lattice used by LBM is limited in regard to fitting
complex geometries with highly curved surfaces such as packed beds. Vidal et al.
(2010) further argue that the inherent locality of its scheme make parallelisa-
tion straightforward and simple and in contrast to the finite volume method,
the lattice-Boltzmann method requires less computational effort. This may be

58
 3.1 Introduction

the case for a simple laminar flow scenario, but Taskin (2007) suggests that a
turbulence simulation using the lattice-Boltzmann method is considerably more
expensive than the FV methods. However, Taskin (2007) neglect to mention
whether this is compared to a direct-numerical simulation, large-eddy-simulation
or a Reynolds average based model with a suitable closure term. As a rebuttal to
the arguments of Vidal et al. (2010), if the finite volume technique was applied
to a primitive, block structured lattice, as in the case of the lattice-Boltzmann
method, one might ask the question ’would there be any difference in the com-
putational expense?’. The relatively high computational expense associated with
the FVM is primarily due to solving the governing equations for the complex
unstructured meshing options which are availible in the FV method for meshing
highly curved surfaces.
Al-Jahmany et al. (2009) compare the LBM and finite volume method over a
backwards facing step. Al-Jahmany et al. (2009) et al conclude that in this case
the LB method is in good agreement with the FV method. However their study
is limited to a backwards facing step in which a block-structured lattice-can be
constructed with ease. If the same test was applied to a complex-highly curved
geometry it is likely that the LB method would not perform as well.
Many computational fluid dynamics solvers such as Fluent, have a built in
porous medium solver, where a porous volume (3d) or porous jump (2d) can be
inserted in a flow geometry to model the effect of the porous medium. These
porous medium solvers to not calculate the micro-structural flow but more the
effect on the porous medium has on the external flow and treat the media volume
as a bulk or gross entity with homogeneous properties. They also rely on em-
pirically based coefficients such as inputs for Darcy’s Law or the Ergun equation
to calculate the pressure drop caused by the media. Although limited, their use
is mainly focused on the flow effects of the porous media on the flow domain,
which constitutes a relatively small element of geometry. In contrast, the work
presented here aims to simulate the microstructural flow, and could be used to
determine the coefficients for such porous media models.

59
 3.2 The Finite Volume Method in CFD

3.2 The Finite Volume Method in CFD

A CFD code using the finite volume method is based around the solution of the
governing flow equations (Navier-Stokes). The solver achieves this by integrating
the flow equations over the entire flow domain. The unknown properties of the
flow are replaced with finite difference approximations which are approximated
iteratively with a guess and correct procedure. To write a CFD code to deliver a
solution for every new fluid flow case would be time consuming and unforgiving.
Due to this, general commercial CFD codes have evolved, such as Fluent, CFX
and OpenFoam. They provide user friendly GUIs (graphical user interfaces) to
allow the user to utilize an in built code with a variety of problem solving facilities
to deliver a solution for problems associated with fluid flow, heat transfer etc. To
describe the whole process can be long winded, so for simplicity, the process of
generating a solution to a fluid problem using a general CFD code is described
in three stages; the pre-processor, the solver and the post-processor.

3.3 The pre-processor

The first step involved in the pre-processor stage is to define a suitable geometry
in which the fluid flow problem is to be analysed. Defining the geometry can be
achieved by replicating a known geometry through computer aided design (CAD)
in the case of relatively simple geometries. For more complex natural geometries
the geometries may be probed using a non-invasive technique, such as MRI or
CT coupled with some sort of geometry defining algorithm. Once the geometry
has been created a suitable mesh can be constructed. A mesh can be described as
a volume which has been discretized into a series of smaller volumes or elements.
This can be straight forward in the case of simple geometries, but more challenging
in highly disordered media such as a packed beds or porous mediums. Once a
workable mesh has been created, additional known quantities of the fluid are
defined such as density, viscosity and turbulent quantities. Boundary conditions
are applied to cell faces in the form of a flux adjacent to an inlet/exit or wall
such as temperature, pressure, velocity profile etc. This can be argued as being

60
 3.3 The pre-processor

the most crucial stage in the development of a solution, as the determination of

a realistic answer is highly dependent on the global mesh quality.

3.3.1 Mesh structure

Cells which make up the mesh come in a variety of shapes and sizes applicable to
specific cases. Usually cells are tetrahedral or cubic, although recently polyhedral
cells have been used. We can classify grids or meshes as being either structured
or un-structured (Fig 3.1). What makes a grid structured is that each of its cells
can be ordered in respect to their indices (usually i, j, k or 1, 2, 3). In the case
of a irregular, disordered packed bed the use of a structured grid would be near
on impossible due to the complex geometry and the representation of spherical
particles. In addition, it would be impossible to trace a single line through the
bed in order for the bed to be ordered by its indices due to the obstruction of the
media. Unstructured grids the cells are not ordered in respect to their indices
but with some other fashion specified within the CFD code. Most modern CFD
solvers, such as Fluent, CFX, can handle both structured and unstructured grids.
There is no difference in the accuracy between structured and unstructured grids.
However, Structured codes have the advantage that the cell adressing is simpler
and less costly; then the code runs faster. However, structured grids are limited
to fairly rudimentary geometries. Hybrid grids also exist in which the mesh can
be block structured in regions of open space and then become unstructured in
areas of complex geometrical restriction. A Mesh can also be categorised as
being coarse or fine. A coarse mesh would contain a relatively low cell density
per unit volume with large cells, whereas are a fine mesh would contain a densely
populated flow domain with cell small volumes.
—

3.3.2 Traditional CAD mesh generation

Traditional mesh generation methods use a CAD based approach, in which CAD
program is used to define a geometry. The geometry can be created in tertiary
CAD software, such as AutoCAD or SolidWorks and then imported into a com-
mercial meshing package, such as ANSYS Gambit in a suitable file format (STL,

61
 3.3 The pre-processor

Figure 3.1: Left; Structured grid, Right; unstructured grid

ACIS, IGES). Alternatively, some commercial meshing software, such as Gambit

have a built-in CAD interface which allows the control volume and the bounding
surfaces to be created in the software directly, to avoid file format conflictions.
Once the control volume and surfaces have been defined, volume discretizations
are created from the bounding surfaces using meshing algorithms such as advanc-
ing front (AF), Delaunay triangulation (DT) , octree etc. CAD based approaches
work well for simple geometries, but complications arise in their application in
converting non-invasive image data into a suitable mesh for analysis and often
require considerable user intervention and appreciable simplification of the model
geometry (Antiga et al., 2002; Cebral & Loehner, 2001; Young et al., 2008). The
user intervention usually requires an ‘intermediate step’ in which the image data
is converted into suitable surfaces in form of a CAD geometry which can then
be imported into commercial meshing software. This ‘intermediate step’ often
results in the process being time consuming and unrobust (Young et al., 2008). If
a workable mesh can be created, it often contains highly skewed and degenerate
elements.
The advancing front method of mesh creation originally proposed by van Phai
(1982), requires an initial discretization of the domain surface into triangles. A
layer of nodes is then offset from the surface creating the first layer of nodes and

62
 3.3 The pre-processor

Figure 3.2: Delaunay triangulation of a series of points

tetrahedral volumes. The next layer then ‘piggy backs’ off the first layer of nodes
creating the second layer of tetrahedral cells (Young et al., 2008). However,
this method is unrobust when creating meshes where either the initial surface
triangulation is poor or where there is insufficient room for advancement.
The Delaunay mesh algorithm or Delaunay tetrahedralizations is the most
commonly used and is based on the Delaunay triangulation criterion proposed by
Delaunay (1934) or ‘empty sphere theory’. In 2d, a Delaunay triangulation is a
way of joining a collection points in which groups of three points are joined which
fall on the circumference of a circle as shown in 3.2 and this ensures that no point
falls within the boundary of another circle. This method can be gereralised to 3d
using spheres and tetrahedrons.
The octree meshing algorithm developed by Shepard et al. (1988) is a meshing
strategy based on cubic cells being recursively subdivided into smaller cells. The
octree method does not originate from an existing surface mesh, such as Delaunay
or advancing front, but adapts the grid to account for the surface. The cubes
that intersect the surface are subdivided and tetrahedral elements are formed to
body fit the surface the surface.

63
 3.3 The pre-processor

3.3.3 Image based mesh generation

Image based meshing (voxel based IBM) is a method of surface extraction from
an image, usually a grey-scale image slice from a scanned image extracted from
MRI, CT or another non-invasive technique, in which a computational mesh is
created for computational fluid dynamics or finite element analysis. It is not
solely a meshing algorithm such as Delaunay or Octree but encompasses the
whole process of surface extraction and mesh creation (Young et al., 2008).
Non-invasive techniques, create a 3-dimensional image by taking a series of
2-dimensional grey-scale-voxelised slices through the sample, known as a stack.
These are often stored in a unique format specific to medicine known as DICOM
(Digital Imaging and Communications in Medicine). The stack of slices can be
imported into image based meshing software such as ScanIP (Simpleware, Exeter
UK) in which the VOI can be segmented by a method of simple thresholding.
Thresholding is one of of the simplest forms of image segmentation techniques and
relies on the 0-256 grey-scale. A threshold value is set by the user (between 0 and
256) all of the pixels with a value below the threshold are considered background
pixels and all pixels above the threshold are treated as the image (or vice-versa)
binarising the image into two or possibly more masks. In the case of packed beds
an image, ideally would be completely binary, consisting of the particulate matter
and the void volume allow simple accurate segmentation between the media. In
addition to the voxel based IBM is CAD-based IBM; a CAD surface is identified
and extracted for for meshing using (pre-existing) meshers.
One disadvantage with this approach is the uniformity in mesh density. In
open expanses contained within the geometry where the flow is unlikely to have
any special features, the density of the mesh is likely to be too fine resulting
in a considerable amount of unnecessary computational effort. However, viewed
from another angle the mesh is perfectly structured in these regions meaning
relative efficiency in calculation considering the high amount of computational
effort. Recent research (Young et al., 2008) has shown that the number of cells
can be greatly reduced in the effected areas using an octree decimation scheme.
Image based meshing provides increased robustness in creating meshes in the
most tortuous of geometries which traditional Delaunay based approaches would

64
 3.3 The pre-processor

find near impossible without user intervention and considerable manual simplifi-
cation of the geometry.

3.3.4 Mesh quality analysis

Regardless of the type of mesh it is paramount that the mesh is significantly fine
enough and of a sufficient quality not to compromise the solution. To ensure this
is the case it is desirable to conduct a mesh refinement. A mesh refinement study
consists of a mesh being constructed in the desired geometry and a flow problem
simulated. A suitable parameter is monitored, usually the parameter of interest
such as lift, drag pressure drop etc. The mesh is continuously refined and in most
cases the answer will change. The mesh is refined further until eventually the
answer will stabilise and subsequent refinement will yield a similar answer with no
significant change. In a turbulent time dependent flow, transient effects may cause
the solution to vary even after a mesh is considered independent. In some highly
complex geometries the created mesh may be already fine enough and in which
case as the mesh is refined the solution will remain stable. The IBM approach
can be problemistic in that the geometry itself may change with refinement.
Over refinement may also be an issue; turbulence models have restrictions on
cell sizes in the wall region in regard to suitable y+ values. The y+ value is the
dimensionless wall distance; this is given as the frictional velocity (u∗) multiplied
by the cell node disctance from the wall (y), as a ratio of the kinematic viscosity
(ν).
An argument surrounding a mesh refinement study is which property should
be monitored to determine whether the mesh is independent of the solution. For
example, the mesh being analysed may have been found to independent of the
global properties (e.g pressure drop, lift and drag), but is still not fine enough to
be independent of the local properties of the flow. In the case of packed beds, the
pressure drop is highly dependent on the flow geometry. The bed contains many
small pores, in which case a workable mesh is likely to be already sufficiently
fine to constitute an independent mesh. Any further refinement is not likely
to affect the global property of pressure drop (the property being monitored)
but may affect local properties such as recirculation. A pragmatic approach is to

65
 3.3 The pre-processor

monitor the desired property of the flow to be determined, once this has stabilised,
regardless of the other flow properties, the mesh can be considered sufficiently
independent from the solution at least for the problem in question.
A mesh refinement study ensures the mesh is fine enough not to be affecting
the solution. However, the cells must be of a sufficient quality not to affect the
solution. One measure of the cell quality is the skewness which is described by
its deviation from a ‘perfect cell’. Cells which are significantly skewed (slithers)
can lead to solution instability and convergence problems.
In the case of some complex geometries such as packed beds skewed elements
are more or less unavoidable due to the curved surfaces and infinitesimally small
contact points observed in spherical beds. Some packages provide a skewness
compensation algorithm, such as Fluent with the PISO algorithm and addition-
ally the facility to merge skewed elements with neighbouring cells to create a
single polyhedral element from a number of degenerate tetrahedrons. Various

Figure 3.3: Left; perfect tetrahedral cell, Right; highly skewed degenerate cell

metrics of skewness are available. One approach is to use the equiangle skewness
based on the angle deviation method given as
 
θmax − θequal θmax − θequal
Eang = max , (3.1)
180 − θequal θequal

66
 3.3 The pre-processor

where θmin and θmax are the respectively the maximum and minimum angles
between any two edges of the cell, and θequal is the ideal angle for an equilateral
(perfect) cell. In which case an equilateral “perfect cell” will have a skewness of
0 (Cengel & Cimbala, 2006). This method can be applied to any mesh structure
or cell type.
Another approach is to use the dimensionless form of equivolume skew (EV ).
The cell equivolume skew uses the volume deviation method given in the form

VO − VC
EV = (3.2)
VO

where the VO is the optimal cell size of an equilateral cell with the same cir-
cumradii and VC is the cell size. The equivolume skew method only applies to
tetrahedral and triangular elements. In both cases, the cell equivolume/equiangle
skew of 0 indicates a ‘perfect’ cell and 1 indicates a highly skewed, completely
degenerate element.
Another parameter widely recognised in the field is known as the cell squish
index (CSkindex ). The cell squish index uses the dot-product (scalar products) of
each vector pointing from the centroid (the node location) of the cell to each of
the cell faces and the face area vector associated with each face given as
 
Ai .rc0/xf i
CSkindex = max 1 − (3.3)
|Ai ||rc0/xf i |

where Ai are the cell face area vectors and rc0/xf i are the vectors connecting the
cell centre to the respective face centre. The closer the CSkindex value is to 1,
the more degenerate the cell.
All of the mesh quality assessment methods described here are bulk quantities.
Statistically a mesh can appear good quality i.e 97% of the cells have a squish
index of less than 5, but the 3% degenerate cells lead to convergence problems
and instability in the solution. Meshing a random un-structured packed bed is
particularly difficult, not only from its disordered structure, but additionally be-
cause is contains many spherical particles with many contact points. We know
that a sphere is an infinitely sided volume and therefore touches with a neigh-
bouring sphere at an infinitely small contact point. In which case the adopted

67
 3.4 The solver

meshing algorithm has to compensate by either producing a slithered element,

or increasing the contact point to the minimal cell size. (Young et al., 2008)
compared the EVoMaC method (ScanIP, ScanFE) to the Delaunay triangulation
approach used in Ansys Gambit, using Fluent’s mesh verification tool and showed
that for an open cell foam the EVoMaC method displayed a considerably better
quality mesh in regard to the cell squish index. Cavendish et al. (1985) suggest
that the Delaunay mesh generation technique is prone to creating highly skewed
(slivered) elements.

3.4 The solver

In the FV method the solver is considered as the calculation process that ulti-
mately leads to a solution of the transport equations. The first step in the process
is to integrate the equations over all of the finite volumes contained within the
mesh. Finite difference approximations (interpolations) are then substituted for
terms in the transport equations representing the flow variables. The resulting
approximations are then solved using an iterative ‘guess and correct’ procedure.
The transport equation is a PDE which describes the transport phenomena of
properties such as heat, mass, momentum transfer, turbulence etc. The general
incompressible, time dependent transport equation is described as
 
∂φe
ρ + ∇.φe u = ∇.Γ∇φe + Sφe (3.4)
∂t

here, ρ is the fluid density and is a constant as we are dealing with incompress-
ible flow, φe is some conservative property of the fluid and Sφe is a sink term.
Integrating this across the control volume gives
Z Z Z Z
∂φe
ρ dV + ρ∇.φe udV = ∇.Γ∇φe dV + Sφe dV (3.5)
cv ∂t cv cv cv

Gauss’ theorem (divergence theorem) is used to convert the spatial derivative

term ∇.(φt u) into a surface integral of the flux across the cell faces, which
converts the governing equations into a set of difference equations which can be

68
 3.4 The solver

solved numerically. Guass’s theorem is given as

Z Z
∇adV = n.adA (3.6)
cv A

applying this to 3.5

Z  Z Z Z
∂φ
ρudV + ρn.φudA = n.Γ∇φdA + Sφ dV (3.7)
∂t cv A A cv

In time dependent problems a further integration is required over a small interval

of t + ∆t resulting in
Z Z  Z Z Z Z Z Z
∂φ
+ ρudV dt + ρn.φudAdt = n.Γ∇φdAdt+ Sφ dV dt
∆t ∂t cv ∆t A ∆t A ∆t cv
(3.8)
Finite difference approximations are substituted in for the flow variables in the
integrated transport equation to convert the integrals into algebraic equations.
These difference approximations can include backwards differencing, central dif-
ferencing and upwind schemes and combinations and variants depending on the
level of precision desired. The algebraic equations are then solved using an iter-
ative ‘guess and correct’ procedure. The guess and correct algorithms are given
appropriate acronyms, such as SIMPLE, PISO, SIMPLEC etc. Initially the ve-
locity field approximation is determined by solving the momentum equation, and
the pressure gradient is used from the previous iteration to determine the pres-
sure distribution forming “the guess” in the procedure. The momentum equation
is rearranged to form a Poisson equation for the pressure (the pressure equa-
tion) which can be solved to update the guessed pressure, and the velocities are
corrected to form a new set of conservative fluxes as depicted in Fig 3.4. The
SIMPLE algorithm (semi-implicit method for linked equations) is one example
of the guess and correct procedure for the calculation of pressure developed by
S.V.Patankar & D.B.Spalding (1972). The procedure formulated as a result of
two unknowns and in which case the pressure cannot be calculated until the ve-
locity is determined and the velocity cannot be calculated until the pressure is
determined. Due to this, one is guessed and the other is then corrected. The
acronym PISO, meaning Pressure Implicit with Splitting of Operators or the

69
 3.4 The solver

Figure 3.4: Iterative loop

‘PISO loop’, involves a single predictor step and two corrector steps (Versteeg &
Malalasekera, 1995). The PISO algorithm is therefore a more advanced SIMPLE
algorithm with the inclusion of an extra corrector step for enhancement (Versteeg
& Malalasekera, 1995). The SIMPLEC and PIMPLE algorithms are derivatives
of SIMPLE and PISO tuned for specific cases.

3.4.1 The Navier-Stokes Equations

The Navier-Stokes equations are the principal equations used in the finite volume
method. They form a set of deterministic non-linear, second order, partial dif-
ferential equations and are the governing equations of a Newtonian fluid which
may be a liquid or a gas (Drazin & Riley, 2006). The Navier-Stokes equations
can essentially be broken down into three fundamental principles which include,

70
 3.4 The solver

the conservation of mass, the conservation of momentum (Newton’s second law)

and the conservation of energy (Douglas et al., 2005) i.e the first law of thermo-
dynamics.
The Navier-Stokes equations were deduced by Stokes (1846, 1880) from a
selection of hypothesis and postulations of Navier (1827); Poisson (1831); Saint-
Venant (1843) and Stokes (1845) leading to a rational derivation of the equations
that we are familiar with today (Drazin & Riley, 2006). For an incompress-
ible laminar flow (stream-flow) the Navier-Stokes equations are described by, the
equations can be written in the differential Eulerian form,

∇.u = 0 (3.9)
 
∂u
ρ + ∇.uu = −∇p + µ∇2 u + F (3.10)
∂t
where ∂u/∂t is unsteady acceleration, ∇.uu is convective or advective term, and
−∇p is the pressure gradient. For a steady incompressible fluid, ∂u/∂t and ∂ρ∂t
are omitted from the equation to eliminate fluctuations in velocity due to time
and change in density with respect to time yielding

∇.u = 0 (3.11)

ρ∇.u = −∇p + µ∇2 u + F (3.12)

A flow can be considered incompressible if ∆P < 5% and the flow velocity is less
than 100m/s in air (M < 0.3) (Cengel & Cimbala, 2006). Only a hand full of
exact solutions exist for the Navier-Stokes equations, such as Stokes Boundary
layer, the Taylor Green vortex, Hagen Poiselle flow (laminar, steady flow through
a circular pipe) and Couette flow (the flow between two flat plates). In many cases
exact solutions only exist due to certain assumptions and terms being considered
as negligible. It could be argued that no terms in the equation can equate to
exactly zero. For example, unsteady effects could be so minute (∂U/∂t << 0)
that they are unmeasurable and therefore a flow could be assumed to be steady-
state.
In complex geometries such as flow through packed beds, the 3d Navier stokes

71
 3.4 The solver

equations have no exact solution (as with most cases), when solving the equa-
tions computationally the iterative nature does there is no guarantee that an
exact answer will exist. However, formulations do exist to model the permeabil-
ity of porous media deduced from the Navier-Stokes equations, such as the work
of (Azzam & Dullien, 1976). This assumes the medium to be homogeneous. Simi-
larly, Whitaker (1996) show that the Darcy’s Law with the Forcheimer correction
can be deduced using volume averaging from the Navier-Stokes equations.

3.4.2 Modelling turbulence

“[Turbulence], the invention of the Devil on the 7th day of creation”.
Bradshaw (1994)

We have discussed that turbulence is pseudo-random, stochastic and unsteady in

nature and notoriously difficult to model. We use the term pseudo-random be-
cause the turbulent eddies are deterministic but the occurrences are random. In
many engineering flows we are not interested in the turbulent structure itself or
occurrences, but more so the location of turbulence, the intensity and its effects
on the surroundings. To study the turbulent structure itself requires a direct nu-
merical simulation (DNS). A DNS approach allows the complete (approximate)
solution of the Navier-Stokes equations without an additional turbulence mod-
elling. The mesh is fine enough to attempt to en-capture the smallest unsteady
turbulent vortices (governed by the Kolmogorov length scale) and the smallest
turbulent frequencies and events. A DNS approach is by far the most accu-
rate way to model a full turbulent regime as it can resolve a time dependent
vortex structure, but the computational expense increases steeply with respect
the Reynolds number (Tennekes & Lumley, 1972). Although DNS represents
an accurate representation of turbulence, it is impossible for any model to com-
pletely replicate the exact time dependent turbulent structure at a discrete point
in time due to the random, stochastic nature of turbulence. Ferziger & Perić
(1996) highlight the limitations on computer power by suggesting that in a flow
domain of 0.1m2 in a high Reynolds regime, might contain vortices as small as
100µm in size and would require a computational mesh of around 1 × 1012 to be
able to capture the smallest turbulent structures. In addition Ferziger & Perić

72
 3.4 The solver

(1996) suggest limitations on the time-step as the fastest events take place around
10kHz, in which case a time-step of around 100µm would be needed. Speziale
(1991) (although dated) stated that a DNS approach on a simple turbulent pipe
at a Reynolds number of 500000 would require a supercomputer that is 10 million
times more powerful than the fastest supercomputer availible at the time. Even
with the non-linear growth of computing technology it is still not expected that
a full DNS approach could be possible for analysing a whole aircraft for several
decades (Cengel & Cimbala, 2006). With relevance to this work, DNS has been
applied to flow through porous media (Morais et al., 1970), however, this was
used to model a steady-state, laminar non-Newtonian fluid, which requires sub-
stantially less computational power than a DNS approach on a fully turbulent,
compressible, time dependent flow.
As an alternative to a DNS approach, an averaging procedure can be applied
to the Navier-Stokes equations by decomposition of the flow variables in which
various models are formulated to model the effect of turbulence. This averaging
can either be space filtered, such as large eddy simulation (LES) or ensemble-
averaged, such as the Reynolds averaged Navier-Stokes (RANS) equations. The
averaged Navier-Stokes equations do not model the exact turbulent structure, but
give an indication of the location of turbulence and its intensity. Ferziger & Perić
(1996) suggest that due to the complex nature of turbulence, it is improbable that
the turbulence model will represent all turbulent flow and in which case should
be considered as an engineering approximation.
The RANS approach is the most widely used in industry. In commercial en-
gineering the ultimate goal is to make money, now and in the future. To do this
using CFD, a good job is not only a job which is accurate with a suitable level
of precision, but also takes a reasonable amount of time to make it cost effective.
The exact mathematics concerned with turbulence is too complex to be imple-
mented and would produce far more information than what is required for most
engineering flows. A turbulence model is applied to simplify the mathematics.
Turbulence models can be separated into two classifications. The first are the
classical models and based on the averaged Reynolds equations, the second are
based on space-filtered averaging such as Large-Eddy Simulation. A RANS based

73
 3.4 The solver

approach is far cheaper than LES or a DNS approach and provides a suitable level
of accuracy for most industrial engineering flows.

3.4.3 Reynolds-Averaged Navier-Stokes (RANS)

In many engineering flows the motion is turbulent and unsteady. A full time de-
pendent solution of the Navier-Stokes for a high Reynolds flow regime in complex
geometries is not likely to occur in the near future. In which case certain assump-
tions can be made about the flow to give a realistic engineering approximation
accounting for the effect of turbulence rather than the turbulent structure itself.
This is achieved by averaging out any unsteady behaviour forming the Reynolds
averaged Navier-Stokes (RANS) equations. The Reynolds average method relies
on averaging out all of the unsteadiness within the flow and considers all unsteadi-
ness is a result of turbulence. If we imagine a plot of a turbulent flow velocity, u,
we can decompose this into a mean component ( u b) and fluctuation around this
mean (u′ )
b + u′
u= u (3.13)

The same decomposition can also be applied to an ensemble of the other flow
properties. Now every flow variable can be written in the form of the sum of the
′
time averaged value, φb and fluctuating value around the mean, φb

φ(t) = φb + φ′ (3.14)

The requirement is to define what is meant by this average. One convenient

approach is to define the time average;
Z
1
φb = (t)dt (3.15)
∆t

where t is the averaging interval. The averaging interval must be considerably

larger than the frequency of the fluctuations in order to capture a significant
number of events. The flow variable can be now expressed as
Z
1
φb′ = φ′ (t)dt ≡ 0 (3.16)
∆t

74
 3.4 The solver

From these basic equations a set of rules can be derived and applied to the
Navier-Stokes equations. The rules are applied to the unsteady accellerative term,
the pressure term and divergence of stress term, the flow is incompressible so no
averaging is applied to the density, ρ. Re-writing this for for simplicity in all x,
y and z components yields the x, y and z-momentum equations in the form of
 
∂ux [ ∂ pb
ρ b + ∇.ux ux = −
+ ∇. ubx u ′ ′ + µ∇2 u
bx (3.17)
∂t ∂x
 
∂uy d ∂ pb
ρ + ∇. uby u
b + ∇.uy uy = −
′ ′ + µ∇2 u
by (3.18)
∂t ∂y
 
∂uz [ ∂ pb
ρ b + ∇.u
+ ∇. ubz u x ux = −
′ ′ + µ∇2 u
bz (3.19)
∂t ∂z
writing out with a single operator for x, y and z components gives the RANS
equations  
∂u
ρ bu
+ ∇. u b + ℜ = −∇p + µ∇2 . u
b (3.20)
∂t
where the fluctuating component ∇. ud
x
d
′ u′ , ∇. u d
′ u′ and ∇. u
y
′ u′ have now become
z
the Reynolds stresses, ℜ. Writing these in their partial differential form gives
 [ 
∂ uc
′2
x
∂u ′ u′
x y ∂ ub′ x u′z
ℜx = − − − (3.21)
∂x ∂y ∂z
 
∂u[′ u′
x y ∂ uc′2
y ∂ ud
′ u′
y z
ℜy = − − − (3.22)
∂x ∂y ∂z
 
∂ ud
′ u′
x z
∂ ud′ u′
y z ∂ uc′2
z
ℜz = − − − (3.23)
∂x ∂y ∂z
As a result, three normal stresses and three shear stresses are produced (Reynolds
Stresses)
τxx = −ρu′2
x τyy = −ρu′2
y τzz = −ρu′2
z (3.24)

[
τxy = τyx = −ρu ′ u′
x y
\
τxz = τzx = −ρ−u ′ u′
x z
[
τzy = τyz = −ρu ′ u′ (3.25)
x y

75
 3.4 The solver

The momentum equations are a set of equations containing four unknown quan-
tities, ux , uy , uz and p. The additional term ℜ represents turbulence where the
fluctuating terms have been averaged out by the averaging procedure. In which
case any turbulent fluctuations will be presented as an average around the mean
flow. As a result of the averaging procedure the Reynolds stresses have themselves
also generated six other additional unknowns. In order to close the equations a
turbulence model is introduced.

3.4.4 Turbulence models for RANS

In this work, we refer to a turbulence model as a closure term for the RANS
equations based on certain assumptions of the flow. Because turbulence is chaotic
and stochastic, turbulence models are grossly simplified and therefore should
be considered as an engineering approximation. There are no general purpose
turbulence models for every situation, however models such as the k − εt and
k − ω are the closest to what can be considered ‘general purpose’ due to their
stability and general good behaviour in most cases. In the majority of cases
turbulence models are specialised and are optimised for a particular type of flow
and therefore must be chosen carefully.

“All [turbulence] models are wrong, but some models are useful”
George. P. E. Box

In choosing a suitable closure term user must weigh up the implications of a

particular turbulence model, usually influenced by the Reynolds number, the flow
geometry, cost implications and desirable properties which need to be extracted
from the solution.
A brief example comparing different solvers in porous media can be found in
the work of Kuwahara et al. (2006). They compare LES with the low Reynolds
number k − ε model applied to turbulent flow through packed beds of cylinders.
LES and k − ε are in good agreement with each other when analysing the flow
structure from plots of velocity vectors but show a surprising difference in respect
to the production of turbulent kinetic energy. According to Kuwahara et al.
(2006), the k − ε coupled with the effective velocity formulation is renowned for
over estimating turbulent kinetic energy around a stagnation point.

76
 3.4 The solver

Turbulence models can be separated into two classifications. The first are the
classical models and based on the averaged Reynolds equations, the second are
based on space-filtered averaging such as Large-Eddy Simulation. The classical
models are traditionally the most widely used in engineering purposes, such as
the mixing length model, the k − ε and the k − ω models. These are all based
around Newton’s theory that there is an analogy between the viscous stresses
and Reynold’s stresses on the mean flow (Landahl & Mollo-Christensen, 1986).
Newton’s law of viscosity states that the viscous forces are linearly proportional
to the fluids rate of deformation, analogous to Hooke’s law of stress and strain.
For an incompressible Newtonian fluid, written in Einstein index notation gives
 
∂ui ∂uj
τij = µij = µ + (3.26)
∂xj ∂xi

A turbulent vortex will decay unless there is a shear stress present, such as flow
passing a cavity. And as stated in Newton’s law of viscosity as the turbulent stress
increases the turbulent rate of deformation also increases. What is described as
being the earliest turbulence model is that of French mathematician Bousinesq
(1877). Bousinesq (1877) hypothesised that the Reynolds stress terms are linked
to the mean rate in deformation, suggesting that turbulence could be modelled as
a laminar flow with enhancement of the viscosity (Landahl & Mollo-Christensen,
1986) giving  
′ ′ ∂ ubi ∂ ubj
τij = −ρ u b j = µt
bi u + (3.27)
∂xj ∂xi
where µt is the eddy viscosity or turbulent viscosity and assumed to be constant.
The eddy viscosity is not a property of the fluid but a fictional viscosity based
on dimensional reasoning to account for turbulence. The Boussinesq hypothesis
on its own is not enough to model turbulence unless its values are known. From
equation 3.27 we can see that turbulent momentum transport is proportional to
the mean gradients in velocity and hence the transport of turbulence is given to
be proportional to the gradient of the mean value of the transport quantity

[ ′ ′ ∂φ
− ρu i φj = Γt (3.28)
∂xi

77
 3.4 The solver

where Γt is turbulent diffusivity. This forms the basis of the Prandtl mixing
length model as discussed in the succeeding section. Because a turbulent vortex
is damped by viscous effects, the turbulent diffusivity, Γt is likely to be close to
the value of turbulent viscosity, µt giving the Prandtl-Schmidt relation

µt
σt = (3.29)
Γt

experiments have shown that in most cases the Prandtl/Smhmidt relation is

nearly a constant (Versteeg & Malalasekera, 1995).

3.4.5 Two equation models

A two equation model is an eddy viscosity model which has two additional trans-
port equations to account turbulent properties in the flow. Here, the most com-
mon two equation turbulence models are discussed, the k − ε and the k − ω
model. In both cases, k accounts for the turbulent kinetic energy and εt and ω
are dissipative terms which describe the turbulent scales. Due to the complex
mathematical description of the two linked equations, these models require sig-
nificantly more computational effort than a mixing length model, but are still
significantly computationally cheaper in comparison to a DNS approach.

3.4.5.1 The k − εt model

Here the general k − εt model is discussed, in reality k − εt is a family of models

where specific adaptations have been implemented for it to model specific cases
such as Launder, Chien etc. As its name implies, the k − εt equation is described
by two transport equations, the first is turbulent kinetic energy (k) is described
by the PDE
  
∂ µt ∂κ
(ρκ) + ∇.(ρκui ) = ∇. µ+ + Pκ + Pb − ρεt − Y − M + Sκ (3.30)
∂t σκ ∂xj

78
 3.4 The solver

and dissipation, ε by
  
∂ µt ∂εt ε ε2
(ρεt ) + ∇.(ρεt ui ) = ∇. µ+ + C1ε (Pκ + C3ε Pb ) − C2ε−t ρ t + Sεt
∂t σκ ∂xj κ κ
(3.31)
where
Pκ = µt S 2 (3.32)

S is defined as the mean rate of strain given as

h i1/2
S = 2Sij Sij (3.33)

where the empirical numerical constants are given as C1εt = 1.44, C2εt = 1.92,
Cµ = 0.09, σκ = 1.0 and σεt = 1.3 derived from a wide range of experimental data
on air and water. However, these coefficients are only valid for fully turbulent
flows (Menter, 1993).
As with the one equation models, the turbulent viscosity is described on di-
mensional grounds as
κ2
µt = Cρϑℓ = ρCµ
εt
the velocity scale and the length scales are related to the kinetic energy and
dissipation in the form
√ k 3/2
ϑ= k ℓ= (3.34)
εt
We can see from this that the turbulent viscosity is only calculated from a single
length scale and in which case the turbulent diffusion is only based on the calcu-
lated length scale. In reality all length scales throughout the spectrum contribute
to turbulent diffusion.
The k − εt model is considered “the initial port of call” in respect to it being
the most widely applied and considered the most validated turbulence model.
Although it models most industrial flows well, it performs poorly in cases such as
some unconfined flows, flows with curved boundary conditions and fully developed
flows in non-circular ducts (Versteeg & Malalasekera, 1995). In addition, the
assumption with the k − εt is the effects of the molecular viscosity are negligible
and therefore the flow is considered to be fully turbulent.

79
 3.4 The solver

3.4.5.2 The k − ω model

Another popular two equation model is the low Reynolds ‘Standard’ k − ω model
proposed by Wilcox (1993). In this work the Author describes this model as the
‘Standard’ k − ω model to avoid any conflicts with other k − ω turbulence models.
We have seen from the k −ǫt model that ε is the variable in which the length scale
is formed. From dimensional analysis a length scale can also be determined from
turbulent frequency in the case of the standard k − ω model ω = ǫt /k (T −1 e.g
Hz). ω has a number of advantages, the foremost being that ω → 0 at the wall
and ω is derived from scratch. As a result the k − ω model is substantially more
accurate than the k − ǫt model in the thin near wall layers and for confined flows.
However, the model is particularly sensitive to free stream values in unconfined
aerodynamic flows, which has prevented the model from becoming a replacement
to the k − εt model (Menter et al., 2003). The standard k and ω are modelled by
the following set of partial differential equations

∂ ∂ h ∂ i
Γ
(ρκ) + (ρκui ) = ∇. Γκ + G κ − Y − κ + Sκ (3.35)
∂t ∂xi ∂xj

and  ∂ 
∂ ∂ Γ
(ρω) + (ρωui ) = ∇. Γω + G ω − Y − ω + Sω (3.36)
∂t ∂xi ∂xj
where Gκ is turbulent kinetic energy generation as a result of the average gradi-
ents in velocity, Gω is the specific dissipation rate. Γκ and Γω are the effective
diffusivity. Sκ and Sω are source terms. Yκ and Yω represent the turbulent dissi-
pation of κ and ω.

3.4.5.3 The k − ω SST model

A sibling to the standard k − ω model is the SST k − ω model proposed by

Menter (1993) which is an adaptation of the standard k − ω model with an
additional shear stress transport (SST) term. The SST model was developed
due to traditional models, such as the k − εt having difficulties in describing
an accurate description of boundary layer separation on aerofoils with adverse
pressure gradients (Menter, 1993; Menter et al., 2003) and problems associated
with unconfined flows. To overcome the sensitivity of the standard k − ω model

80
 3.4 The solver

in free stream flow Menter (1993) blended both the k − ω and k − εt by combining
the enhanced wall treatment of the k − ω model and the k − ε models relatively
high Reynolds features in the free stream. Thus, implementing the k − ω model
in the boundary layer enables the model to be directly usable all the way down to
the viscous sub-layer and eliminates sensitivity issues associated with the k − ω
model (Menter et al., 2003); the SST k − ω model switches between ω for the
near wall and ǫt in the free stream. The governing equations to describe the SST
k − ω model are given as

∂
b) = ∇. (Γκ ∇k) + Gκ − Y − κ + Sκ
(ρk) + ∇.(ρk u (3.37)
∂t

and
∂
b) = ∇. (Γω ∇ω) + Gω − Y − ω + Sω
(ρω) + ∇.(ρω u (3.38)
∂t
The k equation from the k − ω model is combined with a modified ω equation,
which is a reformation of the ǫt equation and not the same as the standard ω
equation; this includes the addition of a cross diffusion term added to equation
3.38 where the diffusive term is defined as

1
Dω = 2(1 − F1 )ρσω,2 ∇k.∇ω (3.39)
ω

where the closure coefficients F1 is the first blending function defined as

   √  4
k 500ν 4σω2 k
F1 = tanh min max , , (3.40)
β ∗ ωy y 2 ω CDkω y 2

In the k − εt model, F1 is equal to zero away from the surface. The k − ω model
switches this to 1 inside the boundary layer (Menter et al., 2003). and
 
1 ∂k ∂ω
CDkw = max 2ρσω2 , 10−10 (3.41)
ω ∂xi ∂xi

where y is the distance to the nearest wall. The turbulent eddy viscosity is defined
as
a1 k
νt = (3.42)
max(a1 ω, SF2 )

81
 3.5 The Philosophy of CFD

F2 forms the second blending function defined as

  √ 2
2 k 500ν
F2 = tanh max , (3.43)
β ∗ ωy y 2 ω

In the SST model a production limiter is introduced to prevent the build up of

turbulence within stagnation zones (Menter et al., 2003).
 
bi
∂u
Pk = min τij , 10β ∗ kω (3.44)
∂xj

3.5 The Philosophy of CFD

There are two main approaches to analyse a fluid flow problem. The traditional
approach is to conduct a series of experiments on an actual physical model in
a wind tunnel or water flume. As an alternative, computational fluid dynamics
can be used to solve a set of bounding equations across a computational repre-
sentation of a model. An experimental investigation is perfectly credible on its
own, providing the experiment has been set up competently and any uncertainty
justified by the researcher. However, the amount of data which can easily be ex-
tracted and on-costs is limiting its sole use in modern engineering analysis. CFD
is well established and has proved to yield accurate results, but there is still a
certain level of scepticism amongst empirical purists, namely due to turbulence.
The acronym is occasionally disparagingly described as ’colours for directors’.
This is attributed to a CFD solver delivering an approximate solution to a set
of equations, whether the boundary conditions are correct or incorrect the solver
may still run and produce an output, correct or incorrect. The computer is ‘un-
intelligent’ and cannot distinguish between a realistic or unrealistic solution. In
some cases the boundary conditions may be correct but an unrealistic output is
being generated due to an inappropriate turbulence model, a poor mesh/grid or
other pre-processing issue.
CFD, much the same as an experimental investigation requires a competent
user, with experience of the underlying physics of flow to produce a realistic out-
put. He or She can then make quantitative decision as to the accuracy of the

82
 3.5 The Philosophy of CFD

answer generated, just the same as an experimental investigation. CFD can there-
fore be classed as computational empiricism, with many of the same underlying
philosophies as traditional empiricism.
Experimental methods lend their use to determining global properties, lift,
drag, pressure drop etc. Time dependent phenomena and small scale analysis is
often challenging and expensive with traditional experimental approaches. CFD
lends itself to not only determining global properties, but in addition small scale
time-dependent flow characteristics can be extracted with relative ease. To sat-
isfy empirical purists and computational sceptics, CFD and experimental methods
can complement each other. If a fluid problem is analysed both with CFD and
an experimental study is undertaken and the global properties are similar, it is a
realistic assumption to suggest that the local properties extracted from the CFD
are likely to be credible. Thus, CFD is always best reinforced with experimental
data where possible. If not, it should be compared with a well established the-
oretical correlation known to give a realistic answer. The combination of CFD,
experimental data and a well established theoretical correlation where the data
matches well forms a powerful argument to the credibility of the data. We may
never know the exact real answer, but what we can say is that the real answer
falls somewhere within the limits of the three engineering approximations.

83
Chapter 4

On creating packed beds using

Monte-Carlo and traditional
CAD based approach

“In theory, there is no difference between theory and practice. But in

practice, there is” Anonymous computer scientist

4.1 Introduction
In this chapter the finite volume method is applied to investigate the pressure drop
and flow structure of isothermal airflow through randomly packed beds of spheres
with aspect ratios of 7.14, 6.25 and 4.54 (14mm, 16mm and 22mm). The results
determined from the 14mm spherical media and have also presented in Baker &
Tabor (2010). The geometry in these cases is created using a Monte-Carlo algo-
rithm to achieve a random unstructured pack. The 3-dimensional computational
models are then suitably discretized into meshes using Fluent’s grid generation
package, GAMBIT. In parallel, equivalent beds are created which match the char-
acteristics of the computational beds. The beds are termed “equivalent beds” in
regard to their dimensional characteristics being the same, but with a different
internal structure and hence a different inter-connecting network of voids. The
turbulence model used for closing the RANS equations is the SST k − ω model
proposed by Menter (1994). Experimental studies are carried out and compared

84
 4.2 Experimental Setup

to the correlation of Eisfeld & Schnitzlein (2001) and to the experimental data
of Reichelt (1972). When studying complex geometries particle contact points
have always presented an issue in regard to mesh quality. The problem of con-
tacts is ameliorated to some extent by Dalman et al. (1986); Lloyd & Boehm
(1994); Logtenberg & Dixon (1998) who leave small gaps between the particles
and assume zero velocity at the contact point assuming that the contact point
is within the boundary layer. Atmakidis & Kenig (2009) shrink the particles by
2% to avoid stability issues with skewed elements at contact points. In contrast
Guardo (2007) use a structured cubic array but increase the particle volume by
1% to avoid convergence problems. Due to the structured cubic array and the
increase in diameter the particles now merge with all contact points uniform in
size and shape allowing a suitable mesh elements size to be set. Jafari et al.
(2008) compare a space filtered LES approach with the Reynolds Stress model
and a steady-state solver. They use the commercial code FLUENT coupled with
its mesher, GAMBIT. They use a random number generator to carefully position
each of the spherical particles. They compare the models in some detail, as with
many model comparisons statistically analysing a few different beds cannot give
a realistic confidence level to conclusively say whether one models turbulence in
a particular media than the other.
The work presented here is similar to that of Atmakidis & Kenig (2009), only
the spheres have not been shrunk by 2% and therefore retain their geometric
fidelity to the original Monte-Carlo pack. Here we are interested in the two equa-
tion SST k − ω model and its behaviour in modelling flow in packed beds which
is expected to be transitional to mildly turbulent in nature. The contact points
remain unchanged and are meshed regardless, leaving the mesh geometrically
faithful to the CAD model and the skewness compensated by the PISO solver.
All the computational results are compared with experimental equivalent beds.

4.2 Experimental Setup

The experimental setup used in this work requires methods for measuring values
of pressure difference ∆P , and volumetric flow rate (Q) and hence average velocity
(U ), and is based on the setup of Tóbis (2000), with the addition of two pressure

85
 4.2 Experimental Setup

Figure 4.1: Chapter 4 methodology flow diagram

86
 4.2 Experimental Setup

taps into the bed at 0.1 m vertical intervals, as shown in 4.2. These pressure
tappings are added to provide an average pressure across the profile and eliminate
the risk of velocity streams affecting the pressure reading. To keep all particles
as uniform as possible 14mm, 16mm and 22mm marbles were used due to their
consistency in diameter and the fact that they are hydraulically smooth. The
packed bed was supported by a fine wire mesh screen enclosed in a 100 mm
diameter Perspex column, with a length of 300 mm. Air was forced through
the bed using a centrifugal pump. Volumetric flow rate Q, and hence average
velocity U , was measured by measuring a differential pressure (∆P ) across a plate
orifice manufactured from Perspex. The plate orifice method is use to measure
the volumetric flow rate of a fluid. The orifice causes a vena-contracta to be
formed by the pipe constriction and as the fluid velocity increases through the
constriction there is a resulting decrease in pressure. Using Benoulli’s princple
the volumetric flow rate can be determined from this differential pressure and
additional properties of the fluid. An experimental value for the flow coefficient
Cf of the plate orifice was calibrated against a velocity profile, determined by
hot-wire anemometers.
The plate orifice method was chosen over that of a Pitot-Darcy static tube,
because the Pitot-Darcy static tube only determines the velocity at its immediate
location, meaning that a full flow profile using this method would require many
readings to be taken in different locations across the pipe section, then averaged
to reduce error. The plate orifice is simply an application of Bernoulli’s theory,
the conservation of mass and the flow geometry. The pipe constriction causes a
vena-contracta to be formed and from the differential pressure and Bernoulli’s
hypothesis the volumetric flow rate can be determined. One disadvantage of us-
ing a plate orifice is that there is a significant pressure drop immediately after
the orifice; to reduce this, a Venturi meter could be used as an alternative, where
the drop in pressure due to the device is less significant. In addition other com-
plications are also documented regarding the use of reciprocating pumps. Where
the pump cycle frequency is close to the natural frequency of orifice, results have
shown a discrepancy of up to 40% as a result of the shedding of vortices from the
lip of the orifice (Cengel & Cimbala, 2006). However, due to the low cost and ease
of manufacture, along with general accuracy and ease of calibration, the benefits

87
 4.2 Experimental Setup

Figure 4.2: Experimental setup

88
 4.3 Computational Packing of Spheres

of using a plate orifice outweigh the disadvantages. The header region before the
bed was of sufficient length for the flow profile to be assumed as fully developed.
Calculating the velocity from the volumetric flow rate (Q = U A) enables the
average velocity across the profile to be calculated.
Differential pressures were taken from immediately before the bed (0.0) and
at 0.1 m intervals. To validate the total pressure drop, differential pressure was
recorded between intervals and added together, then compared with total pressure
drop. An accurate value of air density was determined by using a barometer to
determine atmospheric pressure p and a k type thermocouple placed in the flow
to determine flow temperature T . This was used in conjunction with the specific
gas constant for air R = 287 J/kgK, resulting in an air density of 1.117 kg/m3
(P V = M RT ) which matches closely to the standard density of dry air at T =
300K, which is 1.177 kg/m3 (Rogers & Mayhew, 1982).
An accurate value of porosity is also needed to describe the bed. Here three
methods are used. Firstly, the classic method of water substitution is used. This
method is reasonably accurate, but problems occur with porous particles and the
wetting properties of water. In this case the medium is glass and so this is not
likely to present a problem. In addition this method can also suffer from entrained
air. To compensate for this a surfactant is used (washing up liquid) coupled
with a vacuum chamber to remove any air. In comparison, MacroPac provides a
method for determining porosity. Secondly, we know the overall volume of domain
occupied, the size of the header and exhaust region and the mesh volume. In
which case the porosity can be determined by simple volume diameter relations.

4.3 Computational Packing of Spheres

We have discussed the issues regarding unstructured beds and as a result a con-
siderable amount of research has been carried out on the carriage of flow through
regular beds (Atmakidis & Kenig, 2009; Dalman et al., 1986; Tóbis, 2000) or a
simple unit cell approach (Hellström & Lundström, 2006). This is because the
discretization of a regular bed into a suitable domain for computational analysis
is far easier due to homogeneity in pore sizes. However, in reality many indus-
trial packed beds are not regular but random, unstructured and disordered, which

89
 4.3 Computational Packing of Spheres

cannot be quantified by any deterministic mathematical formulae, but more de-

scribed by the probability that an event may happen. In many cases Monte-Carlo
methods are used where a solution from a deterministic algorithm is not possible.
A Monte-Carlo approach does not relate to a specific algorithm, but a class of
non-deterministic algorithms based on random sampling.
To implement this in this work, the particle packing simulation program
MacroPac (Intelligensys, UK) was used to generate packed beds of spheres of
the same diameter as those used for the real experiment. In MacroPac the Monte
Carlo method is used to pack the spheres to the densest packing possible, by
simulated shaking of the column and packing medium. A full scale discussion
about Monte-Carlo algorithms and their mathematics are beyond the scope of
this work. Here a Monte-Carlo algorithm is simply used as a tool to generate
unstructured, random beds of particles. Here, the basic underlying physics used
by MacroPac is explained as follows. The probability that a particle moves in the
z-axis is

∆E
p = e− R.T (4.1)

∆E is the change in potential energy when the object moves against a force
in the -z direction. We know change in potential energy is given as

∆E = mg∆h (4.2)

In the simulation the probability that a particle moves up

p = emr ∆hr Kf (4.3)

where mr is the unit of mass and ∆hr is the distance moved. mr and ∆hr
are defined by the objects mass density relationship, m = ρV and boundary
dimension parameters. Kf is the kinetic factor given as

mf hf g
Kf = (4.4)
R.T
The kinetic factor is also related to the amount of shaking; more shaking will
produce a lower kinetic factor as more particles may move up due to the shake.

90
 4.4 Computational domain

To avoid any experimental discrepancies, the container geometry used as an input

parameter for MacroPac was identical to that of the experiment.
Having generated these computational beds, the coordinates of the sphere
centres were exported to the CAD package AutoCAD and used to regenerate
the packed spheres in ACIS format for export to the automatic meshing program
Gambit. Figures 4.4, 4.3 and 4.5 show a view of the spheres in the bed displayed in
Fluent. Although relatively simple, using this technique there is scope to produce
much larger beds of spherical media, the only limitation is the large amount of
computational power required in mesh generation and computational simulation.
MacroPac provides an inbuilt function by which bed porosities can be deter-
mined (where the exact method is discussed in the final chapter of this Thesis).
Here the results produced for the bed porosity

Properties ε (Experimental) ε (MacroPac)

14mm spheres 0.48 0.50
16mm spheres 0.49 0.49
22mm spheres 0.52 0.54

Table 4.1: Porosity for packed beds determined both from MacroPac and exper-
iment

4.4 Computational domain

The complex geometrical properties of random packed beds makes it extremely
difficult to produce a workable mesh for this case. Even once a mesh has been
produced the quality is often dubious containing non-positive volumes and highly
skewed elements (slithers), which often lead to simulation divergence and insta-
bility. Due to this the discussion presented here has evolved from the following
quote

“Mesh quality concerns the characteristics of a mesh that permit a

particular numerical PDE simulation to be efficiently performed, with

91
 4.4 Computational domain

fidelity to the underlying physics, and with the accuracy required for
the problem”
Patrick M. Knupp

Using traditional mesh quality analysis techniques, such as equivolume/equiangle

skew cell squish index etc, the mesh can appear well within the limits of what is
considered to be a good quality mesh due to the skewness averaging applied by
the formula. A mesh can often contain a few highly slithered elements which can
often go undetected until an un-expected solution divergence occurs.
In this work creation of the mesh of cells is done using Ansys Gambit, a
general-purpose program with a variety of automatic meshing strategies based on
a Delaunay algorithm for unstructured grids. Here the domain was meshed using
the Tet/hybrid and TGrid options; this generates meshes comprised predomi-
nantly of tetrahedral cells but including other cells (prismatic and hexahedral)
where necessary. Surfaces were not meshed separately and so were automatically
meshed as triangles.

Properties Cells Faces Nodes

14mm spheres 1168328 2456980 248950
16mm spheres 1447289 2991148 286421
22mm spheres 1235204 2542209 240310

Table 4.2: Mesh data

Each mesh is then partitioned by its principal axes into 4 segments to allow
paralisation of the domain to reduce simulation time. Each segment is computed
on a node, where each node consists of 1 quad core 2.6GHz processor with 8Gb
of RAM. To reach an overall solution required between 4 and 8 hours computer
time.
To capture the full effects of viscous drag on the micro-structural flow the
mesh must significantly fine enough. In which case it is necessary to conduct a
mesh convergence study (mesh indeterminacy study), by increasing the number
of mesh elements and monitoring the same parameter of the flow, in this case the
area weighted average pressure at the outlet.

92
 4.4 Computational domain

Y
Z
X

Grid (Time=8.1700e-06) Mar 23, 2009

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 4.3: Computational domain for 14mm spheres

Y
Z X

Grid Oct 07, 2010

FLUENT 6.3 (3d, pbns, lam)

Figure 4.4: Computational domain for 16mm spheres

93
 4.4 Computational domain

Y
Z X

Grid (Time=4.8130e-03) Oct 19, 2010

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 4.5: Computational domain for 22mm spheres

In this case the study (figure 4.6) concluded that the results did not signifi-
cantly change after 1168328 cells, a mesh finer than this would yield no advantage
in producing more accurate results and would be computationally more expen-
sive. Due to the complex nature of the geometry the meshing algorithm used in
Gambit can have problems producing a working mesh at certain size intervals in
respect to the range in void sizes and particle contact points. The points on the
graph represent meshes for which computational analysis was possible. In this
case the quality of the mesh is analysed using the dimensionless form of equivol-
ume skew (EV ). In both cases, the cell equivolume/equiangle skew of 0, indicates
a ‘perfect’ cell and 1 would indicate a highly skewed, completely degenerate el-

Properties Max Cell vol Min Cell vol Max face area Min face area Mesh Volume
14mm 1.834 × 10−9 5.208 × 10−12 3.14 × 10−6 1.987 × 10−8 5.796 × 10−4
16mm 1.886 × 10−9 7.39 × 10−12 3.410 × 10−6 1.519 × 10−7 8.23 × 10−4
22mm 1.752 × 10−9 3.188 × 10−11 3.275 × 10−6 1.446 × 107 6.797 × 10−4

Table 4.3: Mesh data

94
 4.4 Computational domain

40

30
dP (Pa)

20

10

0
5e+05 1e+06 1.5e+06
Mesh Elements

Figure 4.6: Grid convergence study

ement. In reality, methods for determining mesh quality are not dependant on
the physical geometry of the mesh and its deviation for a ‘perfect cell’ but more
dependant on the application. For example, laminar, steady state fluid motion
down a straight pipe is unlikely to be susceptible to a degenerate mesh than an
unsteady turbulent regime through a complex geometry due to the additional
equations being solved. Here and in industry, CFD is used as a tool to generate
a result. In industry the result is used to make money, in research it is there to
widen our understanding and breadth of knowledge on the subject. It could be
argued that if a realistic result for the job in question is achieved, which does not
diverge and ultimately leads to a successful solution, is obtained from a skewed
mesh, does the mesh skewness remain an issue? From 4.8, 4.10 it can observed
that the mesh contains a few highly skewed elements. When meshing complex

95
 4.4 Computational domain

Figure 4.7: Geometry and mesh of top of bed

spherical geometries, such as this, skewed cells are often unavoidable. In the case
of regular sphere packings particle distances from one another (gaps) can be set
to allow a suitable size element to be fitted, such as Logtenberg & Dixon (1998)
and Atmakidis & Kenig (2009). In this case distances vary quite considerably
and in which case the majority of skewed elements are located in areas of particle
near contact points shown as in figure 4.11. To reduce this undesirable phenom-
ena Atmakidis & Kenig (2009) shrink each of the particles by 2% to allow for
a gap between the particles to reduce the effect of mesh skewness at the parti-
cle contact points, however, due to its static instability does not replicate a real
packed bed. Fluent also provides a range of features after the mesh has been
produced to smooth the mesh and reduce skewness. However, in most cases this
does not completely eliminate all skewed elements. Fluent provides provisions to
overcome this when using the PISO solver by the addition of skewness correction
and neighbour coupling. In this case the neighbour correction is set to 1 and the
skewness correction to 6, to stabilise the solution.
When modelling turbulence it is imperative that the near wall region is ad-
equately treated. This can be done in two ways, firstly the mesh is fine enough

96
 4.4 Computational domain

30

25

20

15

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y
Z X
Cell Equivolume Skew

Histogram of Cell Equivolume Skew (Time=3.2700e-05) Mar 22, 2009

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 4.8: Mesh cell equivolume skew for 14mm media

to resolve the viscous sublayer and secondly, if the mesh can not be made fine
enough a log law profile is assumed for cells adjacent to the boundary. We know
from boundary layer theory that the flow velocity directly adjacent to a wall is
zero. Away from the wall turbulent flow can be categorised into four flow re-
gions, the viscous sub-layer where the velocity gradient is virtually linear, the
buffer layer where turbulence is becoming more apparent, the transition layer
where turbulence is dominant and the outer layer which is no longer effected by
the presence of the wall. The law of the wall is described by

u yuτ
= (4.5)
uτ ν

where r
τ0
uτ = (4.6)
ρ
From empiricism, this relation is shown to describe the viscous sublayer well for
0 < yuτ /µ < 5 (Cengel & Cimbala, 2006). In many cases it is appropriate to
work in a dimensionless distance in which case a value of y+ is introduced given

97
 4.4 Computational domain

30

25

20

15

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y
Z X
Cell Equivolume Skew

Histogram of Cell Equivolume Skew (Time=0.0000e+00) Oct 01, 2010

FLUENT 6.3 (3d, pbns, lam, unsteady)

Figure 4.9: Mesh cell equivolume skew for 16mm media

as
yuτ
y+ = (4.7)
nu
In regard to the k − ω SST turbulence model, the mesh must be fine enough
at the wall region to realistically capture the effect of the viscous sub-layer and
Fluent assumes there to be a linear velocity relationship within these cells. In
which case y is the distance of node contained by the cells adjacent to the wall
from the wall

1.

. Appropriately the limits which are acceptable for cell size at the near wall region
are the same as the limits set by empirism where 0 < y+ < 5. When enhanced
wall treatment is switched off, the wall region is treated as a log law profile for
the cells adjacent to the wall and in which case appropriate y+ values can be
in the region of 30 < y+ < 300. If we focus our attention back to chapter 2,
there has been much discussion regarding the onset of laminar and turbulent flow
through packed beds. It is generally understood that there is a possibility that
all three modes of flow can exist. In which case by using the SST k − ω we are

98
 4.4 Computational domain

30

25

20

15

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y
Z X
Cell Equivolume Skew

Histogram of Cell Equivolume Skew (Time=5.0000e-03) Oct 27, 2010

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 4.10: Mesh cell equivolume skew for 22mm media

resolving flow near the boundary using the k − ω model and in the free stream
the k − ε model, and the cells contained in the viscous sublayer are treated with
a linear relationship. If we applied the standard k − ω model, without enhanced
wall treatment the cells in the buffer layer (30 < y+ < 300) would be treated as
a turbulent log law profile for the whole medium regardless of the flow regime in
that particular region, possibly neglecting the strains associated with the laminar
boundary layer. In addition, the spherical media at the top of the bed is being
exerted to pressure drag with a degree of recirculation and separation behind each
individual sphere. We can see from figures 4.12, 4.13 and 4.14 that although y+
is dependant on the velocity it would be difficult to argue from these values that
the viscous sub-layer is not being sufficiently resolved. What is more concerning
is that within literature the author has found no evidence of published y+ when a
turbulent solver is applied to a packed bed. Again, these graphs represents a bulk
quantity and in some cases it is apparent that the viscous sub-layer is not being
properly resolved. Figure 4.15 depicts velocity profiles for a sphere in the near
wall region. We can see the boundary layer of the sphere has merged with the
wall boundary layer, most probably due to degenerate, highly skewed elements.

99
 4.5 Computational fluid dynamics

4.11e-01
3.94e-01
3.77e-01
3.60e-01
3.43e-01
3.26e-01
3.09e-01
2.92e-01
2.74e-01
2.57e-01
2.40e-01
2.23e-01
2.06e-01
1.89e-01
1.72e-01
1.55e-01
1.38e-01
1.21e-01
1.04e-01 Y
8.65e-02Z X
6.94e-02

Contours of Cell Squish Index (Time=0.0000e+00) Oct 01, 2010

FLUENT 6.3 (3d, pbns, lam, unsteady)

Figure 4.11: Contours of cell squish index for 16mm media

This effect is however isolated in regions of minimal flow. The only favourable
result of the presence of skewed elements is that they are generally confined to
the spherical contact points, which in most cases are in regions of relative flow
stagnation. The likely hood of the skewness effecting the overall result is low,
providing convergence criteria are met.

4.5 Computational fluid dynamics

Due to the complex geometry and uncertainty in mesh quality, the simplest com-
bination of differencing and solvers are initially used and hence convergence is
virtually guaranteed. For this work a steady state, incompressible laminar solver
is used coupled with the SIMPLE algorithm and first order upwind differencing
scheme, where the steady state Navier-Stokes equations are solved

ρ∇.u u = −∇p + ν∇2 u (4.8)

100
 4.5 Computational fluid dynamics

80

70

60

50

40

30

20

10

0
0 2.5 5 7.5 10 12.5 15 17.5 20 22.5 25
Y
Z X
Wall Yplus

Histogram of Wall Yplus (Time=3.2700e-05) Mar 22, 2009

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 4.12: y+ values for the wall region for 14mm spheres

90

80

70

60

50

40

30

20

10

0
0 2 4 6 8 10 12 14 16 18 20
Y
Z X
Wall Yplus

Histogram of Wall Yplus (Time=5.1340e-03) Nov 04, 2010

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 4.13: y+ values for the wall region for 16mm spheres

101
 4.5 Computational fluid dynamics

90

80

70

60

50

40

30

20

10

0
0 5 10 15 20 25 30 35
Y
Z X
Wall Yplus

Histogram of Wall Yplus (Time=7.0000e-03) Nov 04, 2010

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 4.14: y+ values for the wall region for 22mm spheres

Although air is a compressible fluid, the flow speeds presented here are relatively
low and compressibility effects are unlikely to be contributing to the behaviour
of the fluid. Cengel & Cimbala (2006) suggest that compressibility effects can
be neglected if the Mach number is M < 0.3 (around 100m/s in air), and the
change in density (∆ρ) is less than 5%. The Mach numbers described in this
work are all less than M < 0.012 so compressibility effects are likely to be min-
imal and so an incompressible solver is adopted. Once a steady state solution
for the domain has been delivered and unsteady solver is used. This reduces the
amount of guesswork involved in the guess and correct procedure and therefore
enhances solver stability. Although the Navier-Stokes equations are in themselves
deterministic, with the addition of a turbulence model and the pseudo-random
instabilities of turbulence, they become deterministic chaotic and a degenerate
element can cause the magnitude of chaos to increase violently. In parallel, the
nature of the computational solver itself can display instabilities, forming a de-
terministic chaotic nature, thus it can be difficult in many cases to establish
whether the instabilities are a result of turbulence, or as a result of instabilities
in the solver due to mesh quality. Initially the flow is assumed to be steady state

102
 4.5 Computational fluid dynamics

4.03e+00
3.87e+00
3.71e+00
3.54e+00
3.38e+00
3.22e+00
3.06e+00
2.90e+00
2.74e+00
2.58e+00
2.42e+00
2.26e+00
2.09e+00
1.93e+00
1.77e+00
1.61e+00
1.45e+00
1.29e+00
1.13e+00
9.67e-01
8.06e-01
6.45e-01
4.83e-01 Y
3.22e-01
1.61e-01Z X
0.00e+00

Contours of Velocity Magnitude (m/s) (Time=5.1340e-03) Nov 04, 2010

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 4.15: Contours of velocity at the near wall region

and the SIMPLE algorithm is chosen to deliver a solution. The PISO algorithm
is often less stable in respect to the extra corrector step and highly degenerate
mesh elements. However, the PISO algorithm does include a skewness correction.
Once a solution has been delivered using a the SIMPLE algorithm, PISO is used
to deliver the final solution.
Choosing an appropriate turbulence model to give an accurate representation
of the flow can be challenging but not impossible. We already know that there
is no generic turbulence model applicable for every case. The turbulence models
described in chapter 3 are based around assumptions made about certain length
scales and an enhanced fictitious viscosity. In this case the length scales are likely
to vary considerably from the largest possible vortices based on the diameter of
the header region, intermediate length scales based on the pore sizes and the
smallest vortices being governed by the fluid viscosity. Generally, we base our
decision on some characteristics we already know about the the flow, for example
the Reynolds number, based on some scaling function of the flow geometry. Here
there is an amount of uncertainty in exactly what the internal micro-structural
flow is likely to be. We know from the work of Ziolkowska & Ziolkowska (1988)that

103
 4.5 Computational fluid dynamics

flow can be characterised as laminar (RedP < 10), transitional (10 < RedP < 300)
or fully turbulent (RedP > 300). However, Scheidegger (1960) suggest that it is
impossible to determine a realistic value for the onset of turbulence in a packed
bed. In addition, the flow regime inside the bed is likely to vary considerably
due to flow channelling and stagnation zones. We can however rule out certain
approaches. For instance, a DNS approach is far too computationally expensive
and requires more computer power which is readily available. A mixing length
model is not appropriate due to the complex geometry and pore size variation.
Choosing a suitable mixing length using this model would be almost impossible
and coupled with its limited wall treatment would likely give a misleading result.
An LES model is suitable but beyond the scope of this work. This leaves us with
the two equation models. Traditionally the k − ε model is regarded as the status-
quo, but we know from Versteeg & Malalasekera (1995) that the k − ε model
produces poor results with flow with large extra strains, such as curved boundary
layers (flow around spheres). In addition it is only valid for fully turbulent flows
and has difficulty in modelling separated flows. However, the k−ε model has been
successfully used by Tóbis (2000) on a regular packing with promising results to
describe turbulence in packed beds. This is most likely attributed to a regular
bed having uniformity in pore size. Due to this, a suitable length scale can be
determined.
The flow within the beds described in this work is likely to display, laminar,
transitional and turbulent features, so here the k − ω SST model is used due to
its enhanced wall treatment and the benefits of combining k − ǫ in the free stream
and using k − ω in the near wall region.
Once a solution has been delivered from solution of the un-steady NSE, the
full RANS equations are used with the addition of second order differencing and
the SST −ω model to close the equations.
The meshes described here may my contain half a dozen severely skewed
slithers which potentially can cause the solution to divergence due to the node
position in relation to the face centres. This of course can be damped out to
a certain extent by the adjustment of the under-relaxation factors but in many
cases this is not sufficient to ensure stability. The PISO algorithm also has a

104
 4.6 Results

skewness correction facility, in this case set to 6, but in many cases, the elements
in question are so slithered that any correction does not guarantee a convergence.
In the case of grids containing many skewed cells, many of the usual compu-
tational techniques for a successful convergence are not applicable. For example,
it would be assumed that the more iterations per time-step would lead to a more
accurate and stable solution. In the case of time-stepping across a highly skewed
element, the more iterations per time-step increases the magnitude of error re-
sulting which can ultimately bounce the solver into divergence away from the
solution. This phenomena becomes more prevalent in higher-order differencing
schemes. When an answer has been produced by a slithered element, the hope
is the following cell will be of a realistic geometry and produce such an answer
as to stabilise the divergence back towards delivering a solution and ultimately
away from a possible divergence.

4.6 Results
Here experimental data is compared with CFD results in the form of pressure
drop per unit length (Pa/m) as a function of the dimensionless Reynolds number
given as

ρU dp
Redp = (4.9)
µ
In many applications of fluid mechanics, it is desirable to present data in
its dimensionless form. In which case here the computational and experimental
results have been presented using the dimensionless pressure drop described by
Eisfeld & Schnitzlein (2001) as

∆P dp
φ= (4.10)
ρLU 2

We can see from 4.10 that the dimensionless pressure drop is sensitive to devia-
tions in the velocity and pressure drop per unit length. Any small deviations in
the velocity are magnified to a misleading extent when represented logarithmi-
cally, such as in 4.10. In this work it is also appropriate to compare the results

105
 4.6 Results

3000
CFD SST k-ω
Experimental (14mm spheres)
2500

2000
dP/L (Pa)

1500

1000

500

1000 1500 2000 2500

RedP

Figure 4.16: Pressure drop per unit length as a function of the particle Reynolds
number (Redp ) for 14mm spheres. Experimental uncertainty ∆Redp = ±236,
∆(∆P/L) = ±4.167P a/m

106
 4.6 Results

3000
Experimental (16mm spheres)
CFD SST k-ω
2500 CFD SST tran k-ω

2000
dP/L

1500

1000

500

1000 2000 3000 4000

Redp

Figure 4.17: Pressure drop per unit length as a function of the particle Reynolds
number (Redp ) for 16mm spheres. Experimental uncertainty ∆Redp = ±236,
∆(∆P/L) = ±4.167P a/m

107
 4.6 Results

2500

Experimental (22mm spheres)

CFD SST k-ω
2000

1500
dP/L

1000

500

0
1000 2000 3000 4000 5000 6000
RedP

Figure 4.18: Pressure drop per unit length as a function of the particle Reynolds
number (Redp ) for 22mm spheres. Experimental uncertainty ∆Redp = ±236,
∆(∆P/L) = ±4.167P a/m

108
 4.6 Results

with the correlation proposed by Eisfeld & Schnitzlein (2001) to give the results
some credibility. The equation for dimensionless pressured drop given by Eisfeld
& Schnitzlein (2001) takes the form

K1 A2W (1 − ε) AW 1 − ε
φ= + (4.11)
Redp ε3 B W ε3

2
AW = 1 + (4.12)
3(D/dp )(1 − ε)
and   2 2
dp
Bw = k1 + k2 (4.13)
D
Here the media is spherical, in which case Eisfeld & Schnitzlein (2001) give the
values for the three coefficients as K1 , k1 and k2 as 154, 1.15 and 0.87 respectively.
We know the correlation of Eisfeld & Schnitzlein (2001) is based on empirically
derived coefficients, however, we can see from figure 4.19 that the spread of data
between Re′ = 10 − 10, 000 is quite considerable. This may be due to effects such
as turbulence effects, tortuosity (not described by any relation) or due to the
heterogeneity in possible packing regimes. The most likely effect of this variation
is statistical due to most research being conducted on this range of parameters.
Another possible cause of the large spread of data may be due to the regressive
technique; the correlation essentially gives a mean value of this data and does
not correlate well with data at the extremities of this data. In addition, data
sets can have identical statistical properties but can be considerably different
when represented graphically. This can be explained by Ashcombe’s Quartet
(Ashcombe, 1973), in which four considerably different trends are presented, but
results using simple statistical evaluation techniques are identical. Ashcombe
emphasizes the need to graph data and not just rely on statistical analysis. This
is reinforced in this work by the close trends observed when presented graphically
compared to the relatively large deviations from statistical methods. For example,
Eisfeld & Schnitzlein (2001) report the their models closeness in respect to the
relative root mean deviation, however this only describes the models closeness in
respect to the roots of all the data points and not in respect to the shape or form of
the distribution. This does not describe effects such as divergence or convergence

109
 4.6 Results

away or towards the correlation, but describes this as an average. Results are
also compared to the results of Reichelt (1972). The results of Reichelt (1972)
are chosen due to their closeness in both particle diameter and porosity.
MacroPac provides the facility to evaluate porosities across the bed. Results
for the various beds are shown in figure 4.20. As expected the porosity is roughly
uniform across the core region but rises steeply within the near-wall region of
the column where the packing is effected by the presence of the wall. A similar
porosity decaying-sinusoidal distribution has been observed by du Toit (2008)
from a mathematical model, and the same characteristics have previously been
observed using an empirical approach (Goodling et al., 1983) and MRI (Mantle
et al., 2001; Sederman et al., 2001). Some low aspect ratio two zone correlations
such as DiFelice & Gibilaro (2004) rely on two values of porosity, usually defined
as wall and core porosities. Using MacroPacs function these can be determined by
the averaging the values of the wall region (1 sphere radius), and the core region.
However, the correlation of DiFelice & Gibilaro (2004) is fairly rudimentary and
does not fit experimental data as obediently as the semi-empirical correlation of
(Eisfeld & Schnitzlein, 2001).
The actual flow patterns through the beds are shown in figures 5.19. These
show colour plots of velocity magnitude and turbulent kinetic energy on vertical
and horizontal cutting planes through the domain. The random packing of the
spheres is clearly visible; also note the practical difference between the interior of
the bed and the near-wall region where the packing of the spheres is constrained by
the presence of the wall. This lowers the porosity in these regions and a significant
fraction of the flow is being channelled through this region. One advantage of
computational simulation of the flow is the vast volume of data available for
analysis; in particular here, flow velocities at every point in the bed, which are
available for statistical analysis.
Figures 4.23 and 4.24 depict pressure in its dimensionless form as a function
of the bed distance ratio. We can see that they loosely obey the hypothesis
of Martin (1978) who defined the wall region as being one half of the particle
diameter. Note the effect of the boundary layer at the near wall regions (in
addition the pressure is considerably higher at the wall region indicating higher
velocity and distinct flow channelling). The low pressure away from the wall

110
 4.6 Results

Eisfeld and Schnitzlein (2001)

CFD 14mm Spheres φ
Experimental 14mm Spheres φ
10 CFD 16mm Spheres φ
Experimental 16mm Spheres φ
CFD 22mm Spheres φ
Experimental 22mm Spheres φ
Reichelt (1972)
φ

100 1000 10000

RedP

Figure 4.19: Dimensionless pressure drop (φ) as a function of particle Reynolds

number (Redp )

111
 4.6 Results

1
14mm spheres
16mm spheres
0.9 22mm spheres

0.8

0.7
ε

0.6

0.5

0.4
0 20 40 60 80 100
x (mm)

Figure 4.20: Porosity as a function of distance across the x-plane of packed bed

112
 4.6 Results

a. b.
1.45e+01 1.50e+01
1.40e+01 1.44e+01
1.34e+01 1.38e+01
1.28e+01 1.32e+01
1.22e+01 1.26e+01
1.16e+01 1.20e+01
1.10e+01 1.14e+01
1.05e+01 1.08e+01
9.89e+00 1.02e+01
9.30e+00 9.62e+00
8.72e+00 9.02e+00
8.14e+00 8.42e+00
7.56e+00 7.82e+00
6.98e+00 7.22e+00
6.40e+00 6.62e+00
5.82e+00 6.01e+00
5.23e+00 5.41e+00
4.65e+00 4.81e+00
4.07e+00 4.21e+00
3.49e+00 3.61e+00
2.91e+00 3.01e+00
2.33e+00 2.41e+00
1.74e+00 Y 1.80e+00 Y
1.16e+00 1.20e+00
5.82e-01 Z X 6.01e-01 Z X
0.00e+00 0.00e+00

Contours of Velocity Magnitude (m/s) (Time=8.1700e-06) Mar 23, 2009 Contours of Velocity Magnitude (m/s) (Time=8.1700e-06) Mar 23, 2009
FLUENT 6.3 (3d, pbns, sstkw, unsteady) FLUENT 6.3 (3d, pbns, sstkw, unsteady)

c. d.
1.01e+01
9.65e+00
9.25e+00
8.85e+00
8.45e+00
8.04e+00
7.64e+00
7.24e+00
6.84e+00
6.43e+00
6.03e+00
5.63e+00
5.23e+00
4.83e+00
4.42e+00
4.02e+00
3.62e+00
3.22e+00
2.82e+00
2.41e+00
2.01e+00
1.61e+00
1.21e+00 Y
8.04e-01
4.02e-01X Z
0.00e+00

Contours of Velocity Magnitude (m/s) (Time=5.6100e-04) Oct 06, 2010

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

e. f.
4.35e+00
4.14e+00
3.92e+00
3.70e+00
3.48e+00
3.26e+00
3.05e+00
2.83e+00
2.61e+00
2.39e+00
2.18e+00
1.96e+00
1.74e+00
1.52e+00
1.31e+00
1.09e+00
8.71e-01
6.53e-01
4.35e-01 Y
2.18e-01 Z X
0.00e+00

Contours of Velocity Magnitude (m/s) (Time=4.8130e-03) Oct 19, 2010

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 4.21: Contours of velocity magnitude for 14mm spheres (a,b), contours of
velocity magnitude for 16mm spheres (c,d), contours of velocity magnitude for
22mm spheres (e, f). Fluid flow is in the +z direction

113
 4.6 Results

a. b.
4.17e+00 4.20e+00
4.01e+00 4.04e+00
3.85e+00 3.88e+00
3.69e+00 3.72e+00
3.53e+00 3.55e+00
3.37e+00 3.39e+00
3.21e+00 3.23e+00
3.05e+00 3.07e+00
2.89e+00 2.91e+00
2.74e+00 2.75e+00
2.58e+00 2.59e+00
2.42e+00 2.43e+00
2.26e+00 2.27e+00
2.10e+00 2.11e+00
1.94e+00 1.95e+00
1.78e+00 1.79e+00
1.62e+00 1.63e+00
1.46e+00 1.47e+00
1.30e+00 1.31e+00
1.15e+00 1.15e+00
9.86e-01 9.86e-01
8.27e-01 8.26e-01
6.68e-01 Y 6.65e-01 Y
5.09e-01 5.05e-01
3.50e-01 Z X 3.44e-01 Z X
1.91e-01 1.84e-01

Contours of Turbulent Kinetic Energy (k) (m2/s2) (Time=8.1700e-06) Mar 23, 2009 Contours of Turbulent Kinetic Energy (k) (m2/s2) (Time=8.1700e-06) Mar 23, 2009
FLUENT 6.3 (3d, pbns, sstkw, unsteady) FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 4.22: Contours of turbulent kinetic energy for 14mm spheres (a, b)

reinforces the well established theory of Stanek (1994) and Foumeny et al. (1993)
that the inhomogeneities in the wall region cause increased flow channelling. We
can also see that the size of the wall region and distinctive flow channelling is more
prevalent in the 16mm media in comparison to the 14mm media. Although we
are primarily interested in the drop in pressure, near-wall effects can also be seen
in the structure of the flow patterns themselves. When fluid percolates through
a network of channels, the pathway of mainstream velocity at a low Re takes the
shortest route through the backbone of the bed (Andrade et al., 1997). At higher
values of Re, the pathway of mainstream velocity is dictated by the bed geometry
(Andrade et al., 1997). In addition, the high pressure zone at the centre of the
bed is due to the high resistance to the flow caused by bed core regions lower
porosity in comparison to the wall reion. Figure 5.19 a-b illustrates pathways of
mainstream velocity through the bed, biased towards the wall region.

4.6.1 Uncertainty
With every experimental study there is always a certain amount of uncertainty
in the approach used. This work is primarily a computational study where the
experiments validate the computational data. In addition, this is uncertainty or
the possibility of error in contrast to physical error. Here there are no error bars
plotted for computational data, due to there being no uncertainty in the result

114
 4.6 Results

1.04
Re=1969
1.03 Re=2953
Re=3938
1.02

1.01

1
Pdyn/∆

0.99

0.98

0.97

0.96

0.95
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/D

Figure 4.23: Dimensionless pressure distribution as a function of the distance

ratio across the bed for 14mm media

115
 4.6 Results

1.14
Re=861
Re=1723
Re=2584

1.08
Pdyn/∆

1.02

0.96

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

x/D

Figure 4.24: Dimensionless pressure distribution as a function of the distance

ratio across the bed for 16mm media

116
 4.7 Discussion

of the mathematical approach used. However there may be uncertainty in the

mathematical approach itself, in regard to the averaging procedure, in respect to
a suitable length scale and the production of k. In addition the full turbulent
behavior within the bed is not known and as a result assumptions have to be
made when choosing a suitable turbulence model. To determine the uncertainty
within the turbulence model would constitute an advanced numerical analysis of
all the transport equations used and would be tortuous and arduous and most of
all far beyond the scope of this work. None the less, the results match closely and
with the 14mm and 16mm spheres fall well within the margin of uncertainty. The
results produced for the 22mm holds the same quadratic trend but the magnitude
of deviation is quite considerable. It is appropriate to use a statistical method to
quantify the deviation in the form of the root mean square (RMS) deviation in
respect to the computational compared to the experimental as well as a graphical
plot. The RMS is given as
v
u N  2
u 1 X ∆P/LEXP − ∆P/LCF D
σ=t (4.14)
N −1 i ∆P/LCF D

Here the RMS deviation (σ) is given in table 4.4. However, this statistical
deviation is a bulk representation of the data and its value is heavily penalised
due to the divergence in data in the higher Reynolds numbers. In which case, a
windowed value is also given (σw ) which excludes the CFD which diverges away
from the experimental at the higher Reynolds numbers. This is not implemented
for the 14mm spheres as the data does not diverge due to the relatively narrow
range of data. It must also be remembered that the RMS represents the deviation
of the raw data, neglecting any uncertainty (error bars) and in the case of 14mm
and 16mm spheres the data falls within the margin of uncertainty.

4.7 Discussion
Results for 14mm and 16mm spheres show a strong correlation with the equivalent
experimental beds with a RMS deviation of σ = 0.24 and σ = 0.26 respectively.
The computational beds produced by MacroPac slightly over predict the drop in

117
 4.7 Discussion

Properties σ σw
14mm spheres 0.24 -
16mm spheres 0.26 0.13
22mm spheres 1.5 1.35

Table 4.4: Root mean square of the standard deviation of CFD compared with
experimental

pressure for 14mm spheres and in turn slightly under predict the drop in pressure
for 16mm beds. None the less, the results match closely.
This small discrepancy can be attributed to a number of factors. Firstly,
the turbulence model is based around a mathematical assumption (a Boussinesq
approximation) which may account for the slight deviation. We know from the
work of Menter et al. (2003) that the SST k−ω model has a limiter in regard to the
production of k and in which case turbulent kinetic energy may be unrealistically
low in the stagnation zones. However, it can be argued that there is minimal flow
in a stagnation zone, so is this continuous production of k having any effect on
the drop in pressure? The model may additionally be producing unrealistic levels
of k in regions of minimal turbulent activity. This is because the production
of k is not a known quantity but a product of the closure equations and the
guess and correct procedure. It is debatable if what we are seeing here is actual
true stochastic turbulence or just bursts of un-steady flow behavior, such as the
formation of vortices behind the particles which detach and are then damped out
by the geometry and viscous shear. In addition the turbulence model is likely to
model all unsteady behaviour as turbulence, which may give unrealistically high
values of k.
Secondly, the inlet flow profile here is assumed to be flat and hence reassem-
bling a transitional to turbulent profile, in this case more flow may be channelled
through the bed centre than through the wall region which provides less resis-
tance, resulting in a higher pressure drop. This effect of the wall has been argued
extensively by Stanek (1994) Mehta & Hawley (1969) with no conclusive outcome
as to the effect of the wall region. Conversely, it could be argued that the pipe
Reynolds number at the bed inlet is turbulent and thus obeys a 1/7 power law

118
 4.7 Discussion

profile, which can be modelled by the plug flow assumption due to a flat profiles
closeness to a 1/7 power law.
Finally, the most influential parameter is likely to be discrepancies in bed
geometry, here the three beds satisfy all the dimensional properties and are con-
sidered equivalent beds, however, this does not account for bed structure, which
we already know is random and disordered. Most of all the internal flow paths
and geometry could be significantly different. We know that as the aspect ratio
increases the variance in bed structure becomes less influential ultimately leading
to a continuously packed bed. Conversely when the aspect ratio is around 0.9
there is only one packing possibility, although its polar orientation may vary. The
effect of the aspect ratio on the bed structure can be seen with the 22mm bed,
the aspect ratio is lower and as a result, the internal bed structure is likely to
vary considerably from the experimental. It is plausible that this accounts for
the large discrepancy in results (σ = 1.5).
To reinforce the data presented in this chapter, all the results presented are
compared with the results of Reichelt (1972) due to the similarity of Reynolds
numbers and aspects ratios. The results compare well with the results falling
in the middle of the spread of data. The results are also compared with the
correlation proposed by Eisfeld & Schnitzlein (2001). The discrepancy between
that of the CFD, experimental results presented here and the correlation proposed
by Reichelt (1972) can be attributed to the coefficients determined by Eisfeld &
Schnitzlein (2001) which are taken from a large spread of 2300 data points from
24 sets of published results over a time-frame of 70 years. Figure 4.25 shows the
data used to form the correlation of Eisfeld & Schnitzlein (2001), note the large
spread of data in figure 4.25 at the higher range of the Reynolds numbers with
Re′ = 1×103 −1×106 with almost an order of magnitude difference. This variation
is most likely attributed to the wide range of shapes and sizes of particles. In
addition there may be a statistical phenomena in which the majority of research
has been carried out at these Reynolds numbers due to different experimental
techniques and technological advances.

119
 4.7 Discussion

Figure 4.25: modified dimsensionless pressure drop (ψ ′ = φ′ = φε3 /1 − ε) as a

function of the modified Reynolds number Re′ = Redp (1 − ε) (Eisfeld & Schnit-
zlein, 2001)

120
 4.8 Conclusions

4.8 Conclusions
This work has demonstrated the ability to construct and probe micro-structural
flow through randomly packed beds using a Monte-Carlo algorithm coupled with
the FV method. The macroscopic flow behaviour, i.e. the pressure drop per unit
depth, was found to correlate well with experimental results in the case of 14mm
and 16mm beds, but has shown significant discrepancies in exceptionally low as-
pect ratio beds (16mm spheres). As expected, theoretical approximations slightly
over predict the pressure drop compared with computational and experimental
results. This work has shown that comparing equivalent beds with computational
artificial beds has shown a good correlation, but due to deviation in bed struc-
ture can cause discrepancies in the results. In which case a technique is required
in which actual beds with the same structure can be compared to validate the
technique.

121
Chapter 5

Image based meshing of 3D MRI

scans of packed beds

“Anybody who has been seriously engaged in scientific work of any

kind realizes that over the entrance to the gates of the temple of
science are written the words: ‘Ye must have faith”’
Max Planck

5.1 Introduction
Chapter 4 has demonstrated that CFD is a powerful tool for understanding the
fluid flux through and pressure drop in packed beds. However, the method de-
scribed in Chapter 4 is limited by the arduous task of creating the geometry and
even then the quality of the mesh can be dubious. In addition, this approach gen-
erates randomly packed computational beds and compares them with equivalent
beds which gives a good engineering approximation of the pressure drop, however
the beds being compared are not geometrically faithful and flow distributions be-
tween the beds may vary considerably, with this effect becoming more so as the
aspect ratio decreases. Due to this there is a need for computational beds to be
geometrically faithful in respect to structure and flow paths as the physical beds
for direct computational comparison. Chapter 4 has also highlighted robustness
issues with traditional meshing software and complex geometries, so additionally

122
 5.2 Experimental data

a more robust meshing technique is required to ensure computational stability

and better mesh quality.
One possible method of analysing a physical bed is to use a non-invasive imag-
ing method such as micro-CT or MRI (Manz et al., 1999; Sharma et al., 2001;
Zhang et al., 2006). This approach gives an accurate representation of complex
geometries but problems arise in processing these geometries when using tradi-
tional CAD based meshing software such as Ansys Gambit. Considerable user
intervention and simplification of the geometry is required to make surface cre-
ation of the flow domain possible. In this work, preliminary trials demonstrated
that using traditional meshing software to create a domain was at this stage
impossible due to the imperfections resulting from the scan resolution and the
complexity of the MRI scan.
The work described here focuses on the use of image-based meshing software
originally developed for the field of computational bio-mechanics, to create ge-
ometrically faithful geometries from 3D MRI scans of packed beds for use with
computational dynamics. Disordered packed beds of cylinders (Aratio = 4.35) and
spheres (Aratio = 6.25) are analysed in the Reynolds numbers of Red p = 2 − 4735.
Suitable flow velocity inlet profiles are used to correctly reflect both fully devel-
oped laminar and turbulent flows. Where the flows are considered turbulent, the
k − omega SST model is used as a closure for the RANS equations. Similar work
using non-invasive methods and an IBM approach has been applied to foams in
the past (Tabor et al., 2008) but as far as the author is aware this is the first
application of these techniques to a packed bed.
Although this work is primarily a computational study, experiments were
carried out on the actual scanned beds using the same approach as Chapter 4.
In addition, the results are compared to the correlation of Eisfeld & Schnitzlein
(2001) to describe the dimensionless pressure drop in low aspect ratio beds and
the results of Reichelt (1972).

5.2 Experimental data

There is a level of uncertainty involving user inputs and estimation when using
CFD. Due to this uncertainty CFD data is best reinforced with experimental data

123
 5.2 Experimental data

Figure 5.1: Chapter 5 methodology flow diagram

124
 5.2 Experimental data

to give some validity to the solution. To this end, experimental pressure drops
were measured using the experimental rig described in Baker & Tabor (2010)
(Chapter 4). The rig is based on that of Tóbis (2000), with the addition of
pressure taps into the bed at 0.1 m vertical intervals. Air was forced through the
bed using a centrifugal pump. Volumetric flow rate Q, and hence average velocity
U , was determined by measuring differential pressure (∆P ), across a plate orifice
manufactured from Perspex and previously calibrated (Chapter 4). Two beds
were constructed; the first used PVC rods of diameter 20 mm cut to lengths of
30 mm and packed randomly into the bed, the second consisted of 16mm marbles
packed in the same way to generate a spherical disordered pack.

5.2.1 Magnetic Resonance Imaging

In this work, to generate the 3-dimensional images of the bed in order to create
a mesh, MRI is used. The two packed beds were scanned using a Phillips 1.5 T
whole-body MRI imager located at the Peninsula Medical School (University of
Exeter), using a 2D fast spin echo (FSE) with 0.5mm3 voxels.
The MRI scanner produces a 3-dimensional image by taking a series of slices
through the bed where each slice is 0.5mm thick with each voxel contained in the
slice is 1mm×1mm MRI excites hydrogen molecules within tissue, so for the scan-
ner to distinguish between the void space and the media, the void space around
the glass marbles was filled with a substance containing hydrogen molecules, in
this case oil. In this case, oil was used due to its abundance and ’as an exper-
iment’ on reccomendation of the centre staff. In regard to geometric fidelity,
magnetic susceptibility induced artifacts can be observed in the spherical media
due to the severe susceptibility induced distortions in static magnetic field in
which the sample has been placed. High resolution MRI, as used here, requires a
higher B-uniformity susceptibility variation within the sample. Due to this, the
cylinders were scanned using water as the void filling liquid due to their relatively
close susceptibilities. Due to the need for high resolution, a small diameter coil
was used. However, the length of the coil was not sufficient to capture the whole
bed, so in which case the particles have effectively been cut through their cross
section. A typical image produced from the cylindrical scan is shown in figure 5.2.

125
 5.2 Experimental data

Image Based Meshing techniques, as described in the next section, were used to
generate the geometry from this data and to create a FV mesh for the solution.

5.2.2 Porosity
One of the most influential parameters concerning the drop in pressure caused
by the presence of the media is porosity. Porosity ε is defined by the ratio of
the volume of voids (Vv ) and the total volume (V ). In this case we take the
volume of voids as the mesh volume. In cases of high aspect ratios (such as
a packed bed of sand) the porosity can be defined as being homogeneous with
no measurable wall effect. When a bed is defined as being low aspect ratio the
porosity is heterogeneous and can be broken down into two distinct regions of
wall porosity and core porosity. The traditional method of measuring porosity
is usually via water substitution, however this only has the ability to determine
mean porosity, whereas bulk and core porosity are harder to determine and are
usually calculated theoretically by approaches such as that of DiFelice & Gibilaro
(2004). Alternatively, non-invasive approaches, such as MRI and CT can be used
to investigate void volume and hence local porosity (Nguyen et al., 2005; Wang
et al., 2001; Zhang et al., 2006).
As described below, image based meshing techniques are used to generate
body-fitted finite volume meshes from the MRI scans of the beds. Given these
it is straightforward to calculate the volume of the mesh (i.e. the volume of
the fluid region around the particles) and the total bed volume, and assess the
porosity in this way. Both experimental and computational approaches are sub-
ject to experimental error; in particular, the start and end positions of the beds
are not well defined (the thickness problem), and this is a significant issue for
such a small bed (this would not be an issue for an industrial-sized bed as this
would be significantly larger). Computational porosities are compared to the
water displacement method to verify results. Sharma et al. (2001) suggest that
analysis of MRI data yields consistently higher values of porosity in comparison
to water substitution; however in this case the porosity determined via MRI cou-
pled with image-based method yielded a slightly lower porosity than that from
water substitution. For the cylinders the computational porosity was evaluated

126
 5.2 Experimental data

Media εM RI εEXP εT HE
Cylinders 0.53 0.54 0.46
Spheres 0.48 0.49 0.43

Table 5.1: Porosities for cylindrical and spherical media

as εM RI = 0.53 against εEXP = 0.54 for the experimental value. For spheres
εM RI = 0.48 against εEXP = 0.49. We can also compare these porosities with
porosity correlations determined from empirical data such as the porosity relation
defined by Jeschar (1964)

0.34
ε = 0.375 + (5.1)
dp /D
Table 5.1 shows the porosities determined from MRI/IBM, water substitution
and the empirical correlation proposed by Jeschar (1964). We can see that the
correlation proposed by Jeschar (1964) under predicts the the cylindrical me-
dia quite considerably (∆ε = 0.08 − 0.09). Firstly, the correlation is based on
spherical media, in addition we know that cylindrical media is highly disordered
and depending on the orientation of the particles and density of the packing the
porosity can vary considerably. As the reader might expect, the values for spheres
show a closer agreement (∆ε = 0.05 − 0.06), but again the correlation makes no
allowance for the density of the packing. Furthermore, the two porosities cal-
culated from MRI and from the water substitution are on the same bed, where
as the model proposed by Jeschar (1964) is produced from a regression analy-
sis of many experimental beds. We can see from figure 5.7 that the MRI/IBM
method does not retain perfect geometric faithfulness and modifies the geometry
to suit the mesh. Further more, the water substitution method has uncertain-
ties in regard to wetting, bed start/finish and entrapped air. All three methods
have their associated uncertainties and considering the closeness of MRI with the
time honoured water substitution method and the same bed, we can assume the
value of porosity lies somewhere in the region ε = 0.48 − 0.49 for spheres and
ε = 0.53 − 0.54 for cylinders.

127
 5.3 Image Based Meshing

Figure 5.2: Typical image x-axis slice produced by 3d MRI. Note the random
disordered nature of the of the packed bed from the cylinders being effectively
cut.ss

5.3 Image Based Meshing

In computational biomechanics (and here), a frequently encountered problem is
that of defining, describing and meshing the geometry of interest. Biological
geometries such as arteries or bronchial pathways tend to be complicated and
difficult to describe, and also patient-specific; however the flow conditions can be
very sensitive to the details of the geometry (Collins et al., 2007) as with flow
through porous media. Due to this there is a significant interest in techniques
for automatically generating geometries and meshes from medical scans such as
MRI scans, which inherently provide 3d geometric information. This is known as
Image Based Meshing, or IBM. The fundamental problem of generating a mesh
for a complex, unknown geometry based on a 3d scan occurs in a number of other
areas of engineering (Tabor et al., 2007) and in particular is the problem we face
when trying to construct a geometrically accurate representation of the packed
beds.
There are two general approaches to IBM in existence. One is to generate

128
 5.3 Image Based Meshing

bounding surfaces from the 3d scans, generally by employing some form of edge
detection algorithm. The resulting surfaces (generally saved as STL files) can
then be used by standard automatic mesh generation software to create the volu-
metric mesh. This works for topologically simple geometries, however applied to
highly complex and multiply connected domains such as are the case for packed
beds, the surface description will be exceedingly complex and meshing is likely
to fail. Even when successful, when computational meshes are created using a
traditional Delaunay-triangulation approach the meshes often contain large num-
bers of highly skewed elements and even non-positive volumes primarily around
the areas of particle contact points.
The alternative is to combine the geometric detection and mesh creation stages
in a single process, an approach which is used in the software ScanIP/ScanFE
(Simpleware Ltd, Exeter, UK) used in this work. The process of generating a
mesh involves first segmenting the different volumes of interest (VOI) from the
3D data. Both semi-automated and manual techniques are available within Sca-
nIP, as well as a range of alternative image processing packages, to provide seg-
mented masks. Techniques include noise filters, three dimensional threshold tools
through to bitmap painting. These VOI are then simultaneously meshed based
on an isotropic grid intersected by interfaces defining the boundaries. A base
Cartesian mesh of the whole volume defined by the sampling rate is truncated
at boundary interfaces based on cutting planes defined by interpolation points.
Smooth boundaries are obtained by adjusting the interpolation points in one, or
a combination, of two ways: by setting points to reflect partial volumes or by
applying a multiple material anti-aliasing scheme. The process results in either
a mixed tetrahedral/hexahedral mesh or can be converted to a pure tetrahedral
mesh and incorporates an adaptive meshing scheme. An adaptive meshing scheme
is available which preserves the topology but reduces the mesh density where pos-
sible towards the interior of the mesh by agglomerating hexes into larger hexes
and generating transitional tetrahedral cells. The approach is fully automated
and robust, creating smooth meshes with low element distortions regardless of
the complexity of the segmented data. This software was originally developed
for FE analysis of bones, for both stress and vibration analysis (Johnson &

129
 5.3 Image Based Meshing

Z Y

Grid (Time=5.6040e-01) Nov 03, 2009

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 5.3: Computational domain for cylindrical media generated from MRI and
IBM

Z Y

Grid (Time=5.8620e+00) Jun 20, 2010

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 5.4: Computational domain for spherical media generated from MRI and
IBM

130
 5.3 Image Based Meshing

Young, 2005; Zunarelli & Young, 1999). Since FE and FV meshes are conju-
gate structures, the same techniques can be used to output a FV mesh (cell/face
representation, rather than point/edge representation). From this method, sur-
face models suitable for rapid prototyping can also be generated which are exact
representations of the meshed domains. Amongst other things this also allows
experimental tests to be carried out to provide experimental corroboration of
numerical results, demonstrated in the following chapter.
The quality of the scans is such that the MRI data is virtually binary, with the
majority of voxels taking values 0 and 255 and virtually nothing in between. Thus
segmenting the data is very straightforward. Although ScanIP provides various
filters, such as Gaussian and noise reduction, to reduce noise in the image, the
binary characteristic of the data meant that no preprocessing of the image stack
was necessary, and the flow domain mask was easily segmented as the inverse of
the bed particles (the white areas in figure 5.2; a typical MRI slice through the
x-axis of the bed). The scan is of course not perfect, but represents a pixelation
of the true geometry. This is particularly the case where the cylinders touch; two
spheres touch at a mathematical point, but this has to be represented by at least
a single voxel, and often the scanned elements are seen to merge, as can be seen
in figure 5.5. This same phenomena can also be observed in the work of Johns
et al. (2000) where particles merge together at the contact points.
These contact points form the major constraint when constructing a suitable
mesh of minimal skewness. Thus, the IBM approach gains its robustness from
increasing the size of the contact points in the voxelising process, reducing the
need to create skewed elements inter-particulate contact region. The practical
effect of this is to increase the volume of space occupied by the bed particles and
reduce the porosity, as described above.
Figure 5.8 shows contours of cell squish index. This clearly shows the necking
of the particles, which is an effect of the MRI resolution. Image-based meshing,
due to its voxelisation approach also causes particle necking, but the magnitude
of this is less apparent than the necking phenomena caused by the MRI scan
resolution. In this case the particles have already been ‘necked’ by the MRI
scan, so to speak. In comparison to the Delaunay algorithm used by Gambit
and demonstrates in the previous chapter, the cell squish at the particles is quite

131
 5.3 Image Based Meshing

Figure 5.5: Mesh produced using IBM from MRI data

considerably less than the Delaunay approach and note the highly structured
mesh in the void volume. Many of the problems encountered when trying to
create a mesh from an MRI scan using traditional Delaunay triangulation are
due to the added compexity produced by the scan resolution such as additional
voids and inclusions (Figure 5.8).
ScanIP also provides facilities whereby the segmented mask can be dilated or
eroded by a pixel or a fraction of a pixel, resulting in smaller or larger bed particles
and higher or lower porosity. However the overall porosity was close enough to
the experimentally-determined value and so this was not thought necessary.
We know that in the FV technique the domain has to be discretized into
discrete volumes (tetrahedral, hexahedral etc) thus making meshing a rounded
and spherical surfaces smoothly, virtually impossible without an infinitely small
mesh or a higher order cells with curved faces, (such as NURBS patches). Figures
5.9 and 5.10 depict cell volumes for both cylindrical media and spherical media
respectively. We can see that the cell volume distribution for cylinders has a
wider range of volumes reflecting both the disordered pack and the larger voids
(reflecting the porosity). Various algorithms also exist which offer improved tetra-

132
 5.3 Image Based Meshing

Figure 5.6: Close up of particle contact points for cylindrical media

Figure 5.7: Close up of particle contact points for spherical media

133
 5.3 Image Based Meshing

9.85e-01
9.36e-01
8.86e-01
8.37e-01
7.88e-01
7.39e-01
6.89e-01
6.40e-01
5.91e-01
5.42e-01
4.92e-01
4.43e-01
3.94e-01
3.45e-01
2.95e-01
2.46e-01
1.97e-01
1.48e-01
9.85e-02 Y
4.92e-02Z X
0.00e+00

Contours of Cell Squish Index (Time=6.6800e-04) Oct 06, 2010

FLUENT 6.3 (3d, pbns, lam, unsteady)

Figure 5.8: Contours of cell squish index for cylindrical media. Note the struc-
tured grid in the free-stream.

Property Cylinders Spheres

Max Cell vol 1.54 × 10−9 m3 5.32 × 10−12 m3
Min Cell vol 6.66 × 10−17 m3 1.65 × 10−12 m3
Max face area 2.51 × 10−6 m2 1.34 × 10−6 m2
Min face area 1.53 × 10−11 m2 2.05 × 10−8 m2
Overall Mesh Volume 1.02 × 10−3 m3 7.92 × 10−4 m3

Table 5.2: Bed mesh data

134
 5.3 Image Based Meshing

50

45

40

35

30

25

20

15

10

0
0 2e-10 4e-10 6e-10 8e-10 1e-09 1.2e-09 1.4e-09 1.6e-09
Y
Z X
Cell Volume (m3)

Histogram of Cell Volume (Time=1.9429e-01) Jul 27, 2009

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 5.9: Mesh element volume distribution for cylinders

35

30

25

20

15

10

0
0 1e-10 2e-10 3e-10 4e-10 5e-10 6e-10
Y
Z X
Cell Volume (m3)

Histogram of Cell Volume (Time=4.1000e+00) Oct 07, 2010

FLUENT 6.3 (3d, pbns, skw, unsteady)

Figure 5.10: Mesh element volume distribution for spheres

135
 5.4 Computational Fluid Dynamics

hedral/hex mesh regimes (Alliez et al., 2005; Leland et al., 1998). In the case of
coarse mesh regimes this results in spherical objects being represented as geodetic
spheres affecting the fluid flow at the near boundary region such as the early work
of Dalman et al. (1986). This also introduces complications with the mesh quality
in the near wall region. Additionally the adoption of the SST k − ω transitional
solver for the turbulent simulations enforces restrictions on the near-wall mesh
structure; in particular the laminar sub-layer has to be resolved, resulting in a
limit on the size of the cells in the near-wall region. The acceptable y + values at
the wall adjacent cell should be in the region of 0 < y + < 5. This is achieved as
shown in Figures 5.11 and 5.12.
In many cases of mesh quality analysis it is desirable to use a non-dimensional
parameter, such as cell equivolume skew (Eqvol ), calculated using the volume de-
viation method (for tetrahedral meshes) and equiangle skew (Eqangle using the
using normalised angle deviation method (for mixed tet/polyhedral meshes). For
this work we employ the a parameter known as the cell squish index (CSkindex ).
The cell squish index uses the dot-products (scalar products) of each vector point-
ing from the centroid (the node location) of the cell to each of the cell faces and
the face area vector associated with each face. Figures 5.19 (a,b) present a con-
tour plot of cell squish index of a random cluster of cylinders within the bed.
As expected the plot indicates that the majority of highly skewed elements are
located within the vicinity of the inter-particulate contact points where two par-
ticles meet on a curved surface. For this work the majority of highly skewed
elements (CSkindex > 0.4) are converted to polyhedra using Fluent’s built in
algorithm. Fluent’s polyhedra conversion algorithm evaluates a highly skewed
tetrahedral cell, locates an edge, then agglomerates the surrounding cells neigh-
bouring the edge and converts the adjoining cells into a polyhedron. We can see
from figure 5.13 that the majority of cells fall in the range of 0 < CSkindex < 0.3
indicating an acceptable mesh quality.

5.4 Computational Fluid Dynamics

As described above the range of Reynolds numbers being investigated ranges from
laminar through transitional to mildy turbulent, but the disordered structure of

136
 5.4 Computational Fluid Dynamics

90

80

70

60

50

40

30

20

10

0
0 5 10 15 20 25 30
Y
Z X
Wall Yplus

Histogram of Wall Yplus (Time=2.4675e-01) Jul 27, 2009

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 5.11: Mesh y + values at near-wall region for cylindrical media

90

80

70

60

50

40

30

20

10

0
0 2 4 6 8 10 12 14 16 18 20
Y
Z X
Wall Yplus

Histogram of Wall Yplus (Time=6.0800e+00) Nov 22, 2010

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 5.12: Mesh y + values at near-wall region for spherical media

137
 5.4 Computational Fluid Dynamics

35

30

25

20

15

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y
Z X
Cell Squish Index

Histogram of Cell Squish Index (Time=2.4675e-01) Jul 27, 2009

FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 5.13: Cylinders cesh cell squish index

35

30

25

20

15

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Y
Z X
Cell Squish Index

Histogram of Cell Squish Index (Time=2.0000e-01) Mar 17, 2009

FLUENT 6.3 (3d, pbns, lam, unsteady)

Figure 5.14: Spheres cell squish index

138
 5.5 Results

the bed is likely to encourage the development of disordered, stochastic flow (if
not true turbulence) even for strictly laminar conditions. Conversely the require-
ment to resolve the structure of the bed indicates the necessity to resolve the
flow with a fine mesh, so one of the main contributions of a turbulence model,
that of substituting for the effect of unresolved scales, is not relevant. Accord-
ingly, calculations were run using both laminar and turbulence modelling, and
compared. For the laminar modelling, the equations being solved are standard
incompressible Navier-Stokes equations. For turbulence modelling, an averaging
process is applied to the equations. Here blend of both the k −ε and k −ω models
is used to form the k − ω SST models proposed by Menter (1993), the reader can
find exact details of the PDEs in chapter 3.
For inlet pipe Reynolds numbers of Re < 1500 we expect a fully developed
laminar flow and in which case for these cases a parabolic velocity profile is
implemented by the modification of the inlet boundary conditions. For pipe inlet
Reynolds number greater than Re = 1500 we expect the flow to transitional to
mildly turbulent and in which case the inlet velocity profile is left flat.

5.5 Results

Properties σ σw
Cylinders 0.63 0.40
Spheres 0.71 0.66

Table 5.3: Root mean square of the of CFD compared with experimental

Figure 5.15 demonstrates close agreement of pressure drop per unit length
for cylindrical media between the experimental results (up turned triangles) and
the CFD calculations, both with and without the turbulence model switched
on. Agreement between the CFD model results with and without turbulence
modelling indicates that at it is plausible that at these Reynolds numbers and
mesh resolution the turbulence model is having next to no effect on the pressure
drop. It might be that at much higher Reynolds numbers the two sets of results

139
 5.5 Results

3000

2000
dP/L

1000

Experimental dP/L
CFD Laminar
CFD SST k-ω
CFD Laminar regression
0
1000 2000 3000 4000 5000
Redp

Figure 5.15: Pressure drop per unit length as a function of the particle Reynolds
number for cylindrical media. Experimental uncertainty ∆Redp = +/ − 236,
∆(∆P/L) = ±4.167P a/m

140
 5.5 Results

Experimental φ
CFD k-ω φ
Eisfeld and Schnitzlein φ
CFD Laminar φ
10 Reichelt (1972)
φ

0.1
1000 10000
Redp

Figure 5.16: Dimensionless pressure drop per unit length as a function of the
particle Reynolds number.

141
 5.5 Results

5000 CFD SST k-ω

CFD Laminar
Experimental

4000

3000
dP/L

2000

1000

0
1500 2000 2500 3000 3500 4000 4500
RedP

Figure 5.17: Pressure drop per unit length as a function of the particle
Reynolds number for spherical media. Experimental uncertainty ∆Redp = ±236,
∆(∆P/L) = ±4.167P a/m

142
 5.5 Results

CFD Laminar φ
CFD SST k-ω φ
Eisfeld and Schnitzlein (2001) φ
Experimental φ
Reichelt (1972)
Laminar (parabolic profile)

10
φ

1
100 1000 10000
RedP

Figure 5.18: Dimensionless pressure drop per unit length as a function of the
particle Reynolds number for spherical media

143
 5.5 Results

a. b.
1.86e+01 3.37e+01
1.77e+01 3.16e+01
1.67e+01 2.95e+01
1.58e+01 2.74e+01
1.49e+01 2.52e+01
1.39e+01 2.31e+01
1.30e+01 2.10e+01
1.21e+01 1.89e+01
1.12e+01 1.68e+01
1.02e+01 1.47e+01
9.30e+00 1.26e+01
8.37e+00 1.05e+01
7.44e+00 8.36e+00
6.51e+00 6.24e+00
5.58e+00 4.13e+00
4.65e+00 2.02e+00
3.72e+00 -9.10e-02
2.79e+00 -2.20e+00
1.86e+00 -4.31e+00 Y
Y
9.30e-01Z -6.43e+00
Z X
0.00e+00 -8.54e+00
X

Contours of Velocity Magnitude (m/s) (Time=5.6040e-01) Nov 01, 2009 Contours of Static Pressure (pascal) (Time=6.6800e-04) Oct 06, 2010
FLUENT 6.3 (3d, pbns, sstkw, unsteady) FLUENT 6.3 (3d, pbns, lam, unsteady)

c. d.
7.68e+00 5.32e+01
7.37e+00 5.07e+01
7.06e+00 4.82e+01
6.76e+00 4.57e+01
6.45e+00 4.31e+01
6.14e+00 4.06e+01
5.84e+00 3.81e+01
5.53e+00 3.56e+01
5.22e+00 3.30e+01
4.91e+00 3.05e+01
4.61e+00 2.80e+01
4.30e+00 2.54e+01
3.99e+00 2.29e+01
3.69e+00 2.04e+01
3.38e+00 1.79e+01
3.07e+00 1.53e+01
2.76e+00 1.28e+01
2.46e+00 1.03e+01
2.15e+00 7.76e+00
1.84e+00 5.23e+00
1.54e+00 2.71e+00
1.23e+00 1.79e-01
9.21e-01 Y -2.35e+00 Y
6.14e-01 -4.87e+00
3.07e-01Z X -7.40e+00Z X
0.00e+00 -9.93e+00

Contours of Velocity Magnitude (m/s) (Time=3.8620e+00) Oct 06, 2010 Contours of Static Pressure (pascal) (Time=3.8620e+00) Oct 06, 2010
FLUENT 6.3 (3d, pbns, lam, unsteady) FLUENT 6.3 (3d, pbns, lam, unsteady)

Figure 5.19: Contours of velocity magnitude and pressure for cylindrical media
(a,b), contours of velocity magnitude and pressure for spherical media (c,d).

144
 5.5 Results

a. b.
5.85e-03 2.63e-03
5.61e-03 2.52e-03
5.38e-03 2.42e-03
5.14e-03 2.31e-03
4.91e-03 2.20e-03
4.68e-03 2.10e-03
4.44e-03 1.99e-03
4.21e-03 1.88e-03
3.98e-03 1.78e-03
3.74e-03 1.67e-03
3.51e-03 1.56e-03
3.27e-03 1.46e-03
3.04e-03 1.35e-03
2.81e-03 1.24e-03
2.57e-03 1.13e-03
2.34e-03 1.03e-03
2.10e-03 9.21e-04
1.87e-03 8.15e-04
1.64e-03 7.08e-04
1.40e-03 6.01e-04
1.17e-03 4.94e-04
9.35e-04 3.88e-04
7.02e-04 Y 2.81e-04 Y
4.68e-04 1.74e-04
2.34e-04Z X 6.72e-05Z X
0.00e+00 -3.96e-05

Contours of Velocity Magnitude (m/s) Nov 08, 2010 Contours of Static Pressure (pascal) Nov 08, 2010
FLUENT 6.3 (3d, pbns, lam) FLUENT 6.3 (3d, pbns, lam)

Figure 5.20: Contours of velocity magnitude (a) and pressure (b) for low Reynolds
(creeping flow regime)

a. b.
1.24e+00 1.39e+08
1.17e+00 1.32e+08
1.11e+00 1.25e+08
1.05e+00 1.18e+08
9.88e-01 1.11e+08
9.27e-01 1.04e+08
8.65e-01 9.70e+07
8.03e-01 9.00e+07
7.41e-01 8.31e+07
6.80e-01 7.62e+07
6.18e-01 6.93e+07
5.56e-01 6.23e+07
4.94e-01 5.54e+07
4.32e-01 4.85e+07
3.71e-01 4.16e+07
3.09e-01 3.46e+07
2.47e-01 2.77e+07
1.85e-01 2.08e+07
1.24e-01 1.39e+07
6.18e-02Z Y 6.93e+06Z Y
1.45e-06 1.00e+00
X X

Contours of Turbulent Kinetic Energy (k) (m2/s2) (Time=2.3100e-03) Jan 12, 2011 Contours of Specific Dissipation Rate (Omega) (1/s) (Time=2.3100e-03) Jan 12, 2011
FLUENT 6.3 (3d, pbns, sstkw, unsteady) FLUENT 6.3 (3d, pbns, sstkw, unsteady)

Figure 5.21: (a) Contours of turbulent kinetic energy, k. (b) Contours of turbulent
frequency, ω.

145
 5.5 Results

might diverge further, but it seems plausible that for these Reynolds numbers
the mesh resolution is high enough to explicitly resolve most of the smaller scale
flow structures which are a constant as a product of the viscosity. Figure 5.16
compares our experimental and computational results for cylindrical media with
the empirical correlation of Eisfeld & Schnitzlein (2001) and the empirical data
of Reichelt (1972). Agreement is close although the empirical correlation slightly
under-predicts the dimensionless pressure drop diverging as the Reynolds number
increases. The reader is reminded that the correlation of Eisfeld & Schnitzlein
(2001) is based on an amalgam of a range of experimental data representing
several different geometries, whilst our experiments were performed as far as
technically feasible on the exact geometry used for the CFD. Figure 5.17 shows
pressure drop per unit length as a function of the particle Reynolds number for
spherical media. The graph shows that the CFD data over-predicts the pressure
drop per unit length and the effect becoming more apparent as the Reynolds num-
ber increases. Results for experimental values of dimensionless pressure drop for
the spherical media is plotted in 5.18. Note the unusual trend where the dimen-
sionless pressure drop for the experimental data (up-turned triangles) appears to
increase as a function of the Reynolds number. This is due to the equation being
heavily dependent on the velocity, in addition the logarithmic compression has
graphically skewed the data misleadingly in which case the reader should refer
to 5.17. The correlation of Eisfeld & Schnitzlein (2001) over-predicts the dimen-
sionless pressure drop considerably, although the data does show a reasonable
agreement with the data of Reichelt (1972).
With cylindrical packed beds (Figure 5.15) the geometry is significantly more
disordered than it is the case for spherical particles (Figure 5.17), in which case
pressure drop can vary significantly with each individual pack rendering it difficult
to derive a single formula encompassing each individual case. Where as with
spheres, the packing can less disorded and fairly regular and in which case there
is less variation in packing with each individual pack. This is demonstrated in
the following thought experiment. Imagine a bed of cylinders with a length to
diameter ratio of 2, packed into a domain where the cylinders are of vertical
orientation and a pressure drop measurement is taken across the medium. If we
then take the same number of cylinders and pack them in a horizontal orientation

146
 5.5 Results

(as one might stack logs), the pressure drop would be considerably higher. When
the cylinders are stacked vertically there is less resistance to the flow compared
with a horizontal orientation. What this is saying is the tortuosity is greater
when stacked in a horizontal fashion but all the dimensional properties such
as porosity, particle geometry etc are the same. In which case, to provide a
sensible mathematical explanation for the pressure drop due to this a effect a
value of tortuosity must be determined. Here we are back to a common problem
addressed in this thesis. There is no direct line of sight into the core of the bed,
in addition we do not know which way the flow path is likely to go and there is no
mathematical model to describe the random orientation of cylindrical particles.
One of the main sources of discrepancy between the CFD and experiment,
more so with the spherical particles (Figure 5.17), can be attributed to the accu-
racy of reproduction of the geometry through the IBM/MRI process, in particular
the effect on the mesh of the necking of the media at the inter-particulate contact
points. With bed particles possessing one or more curved faces, when two or more
of these objects touch, the contact point becomes infinitely small. This cannot be
resolved by the MRI scan, instead the contact regions are represented by one or
more contact voxels leading to a merging of the particles in this region. In addi-
tion, the thresholding tolerances in the IBM approach may also add or take away
voxels in this region. This is demonstrated by the discrepancy in experimental
and CFD in figure 5.17. The spheres have considerably more contacts than the
cylinders (up to 12). In addition the aspect ratio is higher and in which case there
are more contacts with other particles and more contact points with the wall. A
contact point with a sphere or other curved surface is infinitesimally small, in
which case this has to be represented by the smallest voxel possible (the scan
resolution). This coupled with magnetic susceptibility induced artifacts can lead
to considerable distortion at the contact points. As a result of these two char-
acteristics, the number of undesired voxels is affecting the porosity to such an
extent that it is disturbing the flow patterns through the bed resulting in a quite
considerate discrepancy in the pressure drop between CFD and experimental.
A more prominent effect on the pressure drop is due to the limitations of
the MRI coil, some of the particles have been effectively cut through their cross
section in the computational models. As a result of this the flow impinging on

147
 5.5 Results

the spherical media is effectively striking the flat side of a hemisphere. We know
from simple fluid mechanics that the drag coefficient caused by the flat side of a
hemisphere (CD = 1.2) is considerably larger than a sphere (Re < 2 × 105 ,CD =
0.5). The magnitude of this effect is less apparent with the cylindrical media,
this is most probably due to the pack being naturally more disorderly than the
spherical media with larger voids. The effect of the flat surfaces at the bed
entry is less apparent in the very low Reynolds creeping regime (Stokes flow),
where viscous drag is dominant over pressure drag and the shape does not have
a major influence on the drag coefficient (Cengel & Cimbala, 2006). In this case
drag coefficients caused by a sphere or hemisphere are functions of the Reynolds
number and are given as CD = 24/Re and 22/Re respectively. We can see the
difference in drag coefficients is only marginal between a sphere and a hemisphere.
In addition some of the extra pressure drop caused by the flats of the hemispheres
at the bed entry may be ameliorated to some extent by the hemispheres at the bed
exit having a lower drag coefficient (Cd = 0.4) than if the space was occupied by a
whole sphere. The effect of the small variation of the particle drag coefficient due
to the particle geometry may also account for the relatively narrow variation of
data observed at the lowered Reynolds regime presented by Eisfeld & Schnitzlein
(2001).
The effect of MRI/IBM approach regarding geometric faithfulness is apparent
not only visually, but also the discrepancy between experimental porosity and
the porosity derived from the computational mesh of cylinders is ∆ε = 0.01, in
conparison to the case of the spheres where the difference in porosity is ∆ε = 0.01.
However, in this case the porosity is not giving a realistic value of the difference in
geometry. Due to the scan resolution and tolerances in the IBM approach, voxels
are not only added in some places but may be taken away in others again losing
fidelity and causing fissures, but also giving a deceiving effect of compensation.
Should the scan be carried out with a higher resolution, one would expect a
reduction in the inter-particle contact region. Alternatively, a bed produced with
a lower aspect ratio would also yield the same effect. A related issue is the mesh
quality; the distance between surfaces in the region near the contact point results
in poor cell quality. This is to some extent ameliorated by the finite resolution of
the contact points alluded to earlier; a conventional meshing strategy based on

148
 5.6 Conclusion

CAD realisations of the particles (which can represent the true curvature of the
particles, which thus generates very fine wedges of space to be meshed close to
the contact point) creates extreme problems in meshing and generates very poor
mesh quality in these regions. When discretized, the size of the contact points
is limited to the smallest element which is possible for the meshing software to
create. Statistically, due to the geometrically complex domain, the total pressure
drop is fairly insensitive to variations in mesh quality, providing the mesh has a
suitable quality not to cause simulation divergence. However, when it is desired
to analyse flow structure in detail, the mesh is required to be suitably fine to
resolve the near wall region.
Although this chapter is primarily concerned with the problems associated
with the mesh creation of packed beds, the flow heterogeneity is also of interest.
Once a computational study has been run, one of the most difficult of problems
to solve is post processing, in regard to interpreting a suitable visual output and
the vast computational effort required due to the number cells. This is further
complicated by the issue of trying to view a complex 3D domain in 2 dimensions.
This problem is analogous to the problem of the almost impossible use of PIV
or LDA techniques when analysing flow through packed beds due to not having
a direct line of sight through the bed particles. Fluent does offer the option to
make the particles translucent, however this only clutters the output and makes it
difficult to draw any conclusions. The most realistic approach, and the approach
we adopt here, is to cut a 2d plane through the bed in order to interpret the
pressure and velocity distribution as a 2D representation, and this is shown in
figure 5.19 c. and d. As expected with low aspect ratio, disordered beds, figure
5.19 (c) shows distinct pathways of mainstream velocity with a good proportion
of the flow being channeled in the wall region due to the increased porosity in the
region resulting in a higher interstitial velocity.

5.6 Conclusion
The work described in chapter has shown how MRI coupled with image-based
meshing can be used to recreate physical experimental beds for analysis using
CFD. The method has shown to be robust in regard to creating a workable

149
 5.6 Conclusion

mesh of suitable quality but fidelity issues arise when MRI scanning small highly
curved objects with regard to the representation of the multiple contacts points.
The inclusion of the points being represented by one or more voxels, due to
the vast number of contacts, statistically can affect global properties of the bed
such as porosity yielding an unrealistically low value of porosity with the effect
becoming more apparent as the aspect ratio increases. It is plausible that the
addition of the voxels may be affecting the micro-scale flow at the contact points
but not significantly changing the main flow paths having most influence on the
pressure drop. The results on the whole have shown good agreement between
experimental, theoretical and computational data for cylindrical media and less
so for spheres. On the whole, the work has provided a valuable insight into
how non-invasive methods can be coupled with the novel method of image-based
meshing to replicate actual physical beds for direct comparison with CFD data.

150
Chapter 6

On creating geometrically
faithful beds with RP

“A fact is a simple statement that everyone believes. It is innocent,

unless found guilty. A hypothesis is a novel suggestion that no one
wants to believe. It is guilty, until found effective.”
Edward Teller

6.1 introduction
This Thesis has already demonstrated that one of the biggest hurdles to overcome
when analysing a packed bed using computational fluid dynamics is re-creating an
accurate computational model which is geometrically faithful to the physical bed.
Previous research (Baker & Tabor, 2010) discussed in Chapter 4 has included the
analysis of an equivalent beds for direct CFD comparison using a computational
bed created using the packing algorithm MacroPac and the results have shown
to correlate well with experimental. However, the bed has the same geometric
characteristics, but the internal flow structure is likely to vary considerably. Cou-
pled with creating the geometry is the problem of producing a workable mesh.
Using traditional CAD based meshing software has shown inefficiencies when ap-
plied to complex geometries. Similar research using this approach was conducted
by (Caulkin et al., 2007, 2009) using digital packing algorithms and included the

151
 6.1 introduction

code DigiPac to create computational packed beds for analysis using CFD. Direct
comparison using this method yields good results in respect to global properties
such as pressure drop, however it is not possible to compare the internal flow
distributions using this technique.
The work described in chapter 5, reported in Baker et al. (2011) has presented
the use of non-invasive methods (MRI) to replicate a 3-dimensional model of an
experimental bed using an image based meshing technique. This method provides
a geometrically faithful bed in respect to bed structure, however this method is
heavily dependant on the scan resolution and aspect ratio. The method has shown
that even with a relatively high scan resolution, the necking of particles can be
observed at contact points affecting global properties such as porosity. However,
on the beds analysed in chapter 5, it is plausible that this effect is not changing
the flow paths considerably.
Up until now, this work has highlighted the problems when analysing packed
beds computationally, via CFD and comparing them with actual physical ex-
perimental beds, either physical scanned beds or equivalent beds. The aim of
the work reported in this chapter is to circumvent these problems by generating
actual physical beds which are geometrically faithful to the computational sim-
ulations In addition to analyse the bed structure and flow paths using a novel
process depicted in 6.1.
In this chapter, as in Chapter 4, computational packed beds are generated
using the particle packing code, MacroPac (Monte-Carlo approach) which has
shown to generate disordered beds efficiently. From MacroPac a coordinate and
diameter for each particle is exported, which is then transformed into a stereo-
lithography format (STL) using a simple C program. A bed of mixed spheres
is imported into the meshing software ScanIP/ScanFE (Simpleware, Exeter UK)
and a computational mesh is created. In parallel the STL created from the mesh
geometry is physically recreated in ABS for experimental analysis using rapid
prototyping. This ensures both the computational bed and the experimental bed
are geometrically identical in all respects. As far as the author is aware this
is the first packed bed manufactured for direct computational comparison using
additive layer manufacturing or more specifically 3d-printing. In addition, in
many published research papers (Baker & Tabor, 2010; Tóbis, 2000; Zeiser et al.,

152
 6.2 Experimental method

2001; ?) the velocity profile is left flat to represent a fully turbulent regime. Here
velocity inlet profiles are added; parabolic when the flow is assumed to be fully
developed and laminar, both flat and power law profiles are when the flow is
assumed fully developed turbulent and compared in detail.

6.2 Experimental method

In this work the experimental rig described in chapter 4 is used. In this instance
the max velocity of the flow is required, in which case a hot wire anemometer is
positioned in the centre of pipe to allow the instantaneous velocity at the crest
of the profile to be measures.

6.3 Rapid Prototyping and 3DP

Here, the beds are created using rapid prototyping (RP) or more specifically
three-dimensional printing (3dP). Rapid prototyping developed out of the need
to create cheap and fast prototypes of components to be manufactured. Tradi-
tionally prototypes were made by hand or one off tooling was made to produce
a few individual components which is expensive in respect to material and man-
hours. Recently RP has developed to such an extent that is no longer limited
just to prototypes, but can be used for mainstream manufacturing, particulary
for creating complex components. Even today, without rapid-prototyping tech-
niques, such as additive layer manufacturing, there is no method for constructing
a completely hollow sphere in one piece.
Various rapid prototyping methods are available, the most common of which
are selective laser sintering (SLS) where a high powered laser is used to fuse par-
ticles into an object replicating the computational model, deposition modelling,
such as 3dP/FDM where successive layers of material are added to form the pro-
totype, electron beam melting (EBM) where a high energy electron beam fuses
parts together in an additive layer fashion. For this work we adopt a 3DP/FDM
approach due to its speed of manufacture and cost relative to component qual-
ity. The model is manufactured using the thermoplastic Acrylonitrile Butadiene
Styrene (ABS). The resolution used for the model is 0.1mm and this forms the

153
 6.3 Rapid Prototyping and 3DP

Figure 6.1: Chapter 6 methodology flow diagram

154
 6.3 Rapid Prototyping and 3DP

Figure 6.2: Photograph of rapid prototyped bed (Left) and computationally recre-
ated bed (Right)

thickness of each added layer. When being constructed the ABS balls are held
together with a temporary structure which is water/detergent soluble. On build
completion the model is left to soak in detergent to allow the temporary structure
to dissolve. We have already discussed the effect of the necking due to additional
voxels of particles and the detrimental effect it has on the fidelity of the compu-
tational model. Here, the effect of the necking is used as an advantage. Without
the phenomena of necking the particles the physical structure would be statically
unstable. Although actual packed beds do not display particle necking, here we
are interested in replication of the exact geometry with fidelity rather than repro-
ducing a real life packed bed. This method presented is completely geometrically
faithful to the mesh in respect to geometry, particle size etc. The only physical
difference between the experimental bed and computational is the fine horizontal
banding produced by the RP resolution. The banding is considered sufficiently
small to be considered as a surface finish rather than a variance in the geometry
and research has shown conflicting evidence, biased towards particle roughness
having little to no effect on the overall drop in pressure. This will be discussed

155
 6.4 Particle roughness

in the next section.

6.4 Particle roughness

In the study of the motion of fluids the roughness of the boundary wall adjacent
to the fluid can affect its behaviour. From basic fluid mechanics, in turbulent
pipe flow (Re > 4000) the relative roughness increases the friction factor as a
function of the Reynolds number and therefore cause a higher drop in pressure
or head loss than a hydraulically smooth pipe. For laminar flow the friction
factor is given as f = 16/Re derived from the Hagen-Poisselle relationship. For
laminar or turbulent flow for rough pipes, which are not hydraulically smooth,
the drop in pressure or head-loss can be simply determined from the Darcy-
Weisbach equation, coupled with the Moody diagram, or the Hagen-Poisselle
equation. We know from literature that many of the classical correlations are
based on the theoretical tubes assumption (Burke & Plummer, 1928; Carman,
1937; Ergun & Orning, 1949) and therefore should obey the same assumptions
as simple pipe flow; the rougher the wall and/or longer the pipe (more tortuous
in the case of packed beds) the greater the drop in pressure. Much research
as been devoted to the investigation of particle roughness. Leva et al. (1951)
and Dullien (1979) observe and increased drop in pressure due the the particle
roughness, where as Fand et al. (1972) observe the exact opposite. Eisfeld &
Schnitzlein (2001) analysed the research of Leva et al. (1951) and Dullien (1979)
and conclude that their observations are most probably due to experimental error
or a misinterpretation of data suggesting that the drop in pressure is not affected
by particle roughness or at least that this parameter has an unmeasurable effect.
Reinforcing the argument of Eisfeld & Schnitzlein (2001) the conflicting evidence
could also be attributed to the level of the confidence involved in the studies.
We know a packed bed can have the same dimensional parameters e.g porosity,
particle diameter, length depth etc but can have a significantly different internal
flow paths. This suggests that a bed of rough spheres cannot simply be packed
and then compared to another pack of smooth spheres as the most influential
parameter (the flow geometry) is being changed. In fact, to produce any credible
level of confidence using this technique would require a population of hundred, if

156
 6.5 Particle size distribution

not thousands of packed bed samples, all with the same dimensional properties. In
addition, the semi-empirical correlation devised by Eisfeld & Schnitzlein (2001)
includes a graph 6.3 in which over 2000 data points are plotted from various
researchers. The work of Leva et al. (1951) is featured in figure 6.3, there is no
deviation from the smooth particles of Reichelt (1972) compared to the rough
particles of Leva et al. (1951). However, the spread of data is so vast (orders of
magnitude) that any deviation due to particle friction is likely to go unnoticed
and absorbed by experimental error and transient effects. In reality its is most
probable that the particle friction is likely to be affecting the drop in pressure in
some way, but this effect is likely to be so minimal and the effect of the geometry
predominantly influential and that realistically the effect is not measurable or
possible to be separated no matter how large the data set.
The advantage of the approach described in this chapter is that actual beds
can be physically recreated which are identical to the computational beds. This
allows only one parameter to be changed keeping the geometry faithful. We know
that the resolution can produce horizontal banding (roughness height of 0.1mm)
and by changing this resolution beds with the same structure and different friction
heights could be analysed. With a sufficiently large data set it is possible that
conclusions could be made as to the the effect of particle roughness on the overall
pressure drop, providing the effect can be physically measured.

6.5 Particle size distribution

We know a packed bed of spherical particles can be made up of either uniformly
sized spheres or a distribution of sizes. In this work the spheres are bi-distributed,
weighted 50:50 by volume with diameters of 16mm and 10mm. Blake (1922) and
Carman (1937) studied variation in particle size and suggest it can treated in the
same way as different particle geometries by introducing an equivalent particle
diameter. Burke & Plummer (1928) and Gupte (1970) have proved by experiment
that packed beds of mixed spheres show a higher drop in pressure compared with
beds of mono-sized spheres and as a result they can not simply be treated in the
same way by using the equivalent diameter. A higher pressure drop due to beds
containing more than one size sphere is likely to be apparent when the ratio of

157
 6.5 Particle size distribution

Figure 6.3: modified dimensionless pressure drop (ψ ′ = φ′ = φε3 /1 − ε) as a func-

tion of the modified Reynolds number Re′ = Redp (1 − ε) (Eisfeld & Schnitzlein,
2001)

158
 6.6 Computational Domain

the smaller sphere to the larger sphere is such that it fits perfectly into a pore in
the pseudo-close hexagonal packing formation.

6.6 Computational Domain

In order to analyse the domain using CFD a mesh is constructed using the image
based meshing software ScanCAD, ScanIP and ScanFE. An STL file is created
from a simple, purpose written C program from the raw coordinate data and par-
ticle diameters generated by MacroPac. The STL is then imported into ScanCAD
and a workable mesh is produced in the same way as in Chapter 5. Here the mesh
contains 30,443,307 mesh elements, 62,844,838 faces and 6,679,187 nodes and is
probably one of the largest meshes created to analyse a packed bed. Note in
passing, in this work the beds have progressed from a few hundred-thousand ele-
ments, such as chapter 4, a few million in chapter 5 and tens of millions presented
in this chapter.

Properties Max Cell vol Min Cell vol Max face area Min face area Mesh Volume
Mesh 1.71 × 10−7 2.82 × 10−18 5.54 × 10−5 1.11 × 10−12 1.35 × 10−3

Table 6.1: Mesh data for bed of mixed spherical particles

Figure 6.5 shows the cell squish index for the computational domain for 16mm
and 10mm spheres. Comparing this to the squish index presented in chapter 5,
the mesh here is of considerably better quality containing a far smaller percentage
of degenerate elements probably due to the small element size. This is because
in chapter 5, the geometry is altered by the scan resolution and is further altered
by the IBM technique. In this case, the IBM method is given a raw geometry
which can be then manipulated to create a mesh.
One feature that can be argued as both an advantage and disadvantage is
the relative uniformity in mesh element size. It could be suggested that in free-
stream the relatively small element size is resolving the majority of the smallest
turbulent structures, thus resembling a rough-pseudo-LES approach. However,
it is debatable how much effect this is having on the overall solution and it will

159
 6.7 Computational Fluid Dynamics

Figure 6.4: Computational domain

incur additional on-cost to computational time and efficiency. This is apparent

in the header region directly before the bed, this region is solely there to carry
the flow through to the media, thus is it necessary to resolve the higher frequency
turbulent structures.

6.7 Computational Fluid Dynamics

In chapter 4 the k − ω SST model was shown to model pressure drop through
packed bed with a good deal of accuracy compared with experimental data. How-
ever, the low Reynolds transitional k−ω model is applied and considerably under-
predicts the pressure drop, with the possibility that certain turbulent effects are
being neglected by the model or being over-damped or not modelled correctly.
An un-steady laminar model has shown to match closely with the k − ω SST
model, meaning the mesh is fine enough to describe most of the smallest turbu-
lent structures. However, a DNS approach cannot be claimed here in respect to
the differencing and a relatively low discretization scheme will not be capturing
the smallest turbulent events. However, it can be argued that the presence of a

160
 6.7 Computational Fluid Dynamics

50

45

40

35

30

25

20

15

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y
Z X
Cell Squish Index

Histogram of Cell Squish Index Nov 02, 2010

FLUENT 6.3 (3d, pbns, lam)

Figure 6.5: Cell squish index for computational domain

turbulence model is having a minimal effect on the solution.

Computational fluid dynamics is simply a tool for solving a second order
partial differential equation. It is not intelligent nor can it give a conclusive
answer as to the flow behavior, it is the interpretation of the user and their
knowledge of fluid physics coupled with CFD which makes it a useful tool for
research and industrial applications. Two conflicting issues arise in respect to
the argument for the use of a turbulence model over an unsteady laminar model.
Firstly, is the k − ω model generating turbulence that shouldn’t’t be there? Is it
representing deterministic behaviour as true stochastic turbulence? Packed beds
are less well understood than most engineering flows (jets, wakes etc) and we still
cannot conclusively say whether the expected behaviour of the flow in a packed
bed is laminar or turbulent at a specific Reynolds number. We cannot use CFD
(with RANS or LES) to study this effect as turbulence model itself is a user
input. We can use DNS to study this but for a media as complex as the media
studied in this thesis the Author postulates at least another decade in computer
development before this is readily available to the end user. For this work, due
to the closeness of the results for the unsteady laminar model to those with the

161
 6.7 Computational Fluid Dynamics

turbulence model demonstrated in chapter 5, in this chapter an unsteady laminar

solver is used without the addition of a RANS model.
The closeness of an unsteady laminar model to the RANS model can form
conflicting arguments. We could argue that the mesh is fine enough to resolve
the majority of the smallest turbulent scales, governed by the Kolmogorov length
scale. However, a DNS approach cannot be claimed as the 2nd order upwind
scheme is unlikely to capture the highest frequency turbulent events, unlike a full
DNS approach requiring higher order schemes as much as 10th order. However,
what we can claim is a pseudo-LES approach in which the majority of the length
scales are being resolved, but not all the highest frequency turbulent events.
Another argument may suggest that there is minimal transient effects as a result
of the pore size and damping caused by the bed geometry and viscosity thus, the
addition of a turbulence model is having little to no affect on the solution.
To study a packed bed we need a header region before the bed to set a super-
ficial velocity and in some cases a suitable velocity profile. If a header region is
not set the interstitial velocity (pore space velocity, Ui = Us /ε) must be used as
the velocity inlet is channelling flow directly into the pores. The header region
may be turbulent, but then the presence of the medium may damp out any large
turbulent structures with the possibility of flow re-laminarising. In addition there
may be sudden burst of turbulence in transitional zones which are then damped
out by a convergence in pore diameter. A model such as the k − ω model is
likely to treat all the control volume (apart from the near wall regions) as being
turbulent, what is more with enhanced wall treatment there may not be suitable
y+ values in the pore spaces due to size restrictions. The SST k − ω model ame-
liorates this to some extent by treating the free stream with all the benefits of
the k − ǫ model and the near wall and pore spaces with k − ω.
To correctly model a fully developed flow, the flow conditions at the velocity
inlet are set to an appropriate velocity inlet profile for each flow condition. We
know that with laminar flow the fluid is flowing uniformly along the path-lines,
viscous shear is dominant and energy is transferred across the streamlines by
molecular diffusion (Cengel & Cimbala, 2006). With turbulent flow, the motion
is chaotic and time dependant and the random motion of vortices transfers energy
at a higher rate than molecular diffusion. In which case, turbulent flow has higher

162
 6.7 Computational Fluid Dynamics

frictional losses associated with it than laminar flow. Because of the relatively
low frictional effects with laminar flow, the presence of the wall in a circular pipe
forms a parabolic velocity profile relationship described by the equation 6.1.
 
r2
U = U0 1− 2 (6.1)
R
Turbulent flow is considerably more complex and there are much higher losses
in which case the profile is much fuller and decreases steeply at the wall region
and is described by a 1/7 power law profile 6.2. The velocity represented in 6.2,
u (U) is the some of a mean u b and a fluctuating component u′ (U ′ ).
b( U
 1/7
r
U = U0 1− (6.2)
R
Here the velocity at the inlet is already a time average, deduced from experi-
mental data. In reality the 1/7 power element can be varied, a mildly transitional
flow may display a 1/5 law, where as a strongly turbulent regime may be as much
as 1/10. In turbulent flow, the power law profile is an average flow profile which is
constantly changing due to the stochastic nature of turbulence. The experimental
setup has a sufficient header region of straight smooth pipe to be considered fully
developed (10D).
When presenting results with different velocity profiles an average velocity or
volumetric flow rate is needed to satisfy the conservation of mass. For a flat plane
profile U = Uavg , whilst for a laminar or turbulent profile the volume beneath the
velocity profile needs to be obtained through integration to determine the volume
flwo per unit-time, Q. In this work the plate orifice has given the volumetric flow
rate Q, m3 /s and the hot-wire anemometer has given Umax = U0 . For a circular
pipe
Z
m = ρUavg A = ρQ = ρu(r)dA (6.3)
A

Z R
1
Uavg = ρu(r)2πrdr (6.4)
ρπR2 0

163
 6.7 Computational Fluid Dynamics

0.8

0.6
U/U0

0.4

Laminar (Parabolic) U(r)

0.2 Turbulent (1/7 power Law) U(r)
Laminar Uavg
Trubulent Uavg

0
0 0.01 0.02 0.03 0.04 0.05
r

Figure 6.6: Velocity profiles for laminar and turbulent flow regimes. Dimension-
less velocity ratio is plotted as a function of distance across pipe

164
 6.8 Results

6.8 Results
We can see from figure 6.7 that the computational data, with and without the
inclusion of an appropriate profile is in close agreement with the experimental
data, at low Reynolds number falling well within the margin of uncertainty and
slightly under-predicting the pressure drop per unit length. The results, the
same as chapters 4 and diverge away from the experimental data as the Reynolds
number increases regardless of the inclusion of an appropriate profile. Results
show there is little difference in the overall drop in pressure between a laminar
parabolic profile, a turbulent 1/7 power law profile and a flat profile.
The RMS deviation for the experimental compared with the results obtained
by CFD is σ = 0.61 for the full range of Reynolds numbers. Windowing this
to lower Reynolds numbers not accounting for the deviation gives σw = 0.06
representing the closest agreement presented in this Thesis.
Figure 6.8 depicts the dimensionless pressure drop as a function of the vol-
umetric flow rate, Q. We can observe from the graph that the inclusion of the
different velocity profiles is having little to no effect on the global pressure drop. It
is plausible the bed geometries strong influence is absorbing any effect on the pres-
sure drop caused by the variation in the velocity profile. The respective pressure
drops are also compared with the correlation proposed by Eisfeld & Schnitzlein
(2001) and show promising agreement. Figure 6.9 compares the results obtained
from both computational and empirical studies. In both cases they show excellent
agreement with the correlation proposed by Eisfeld & Schnitzlein (2001).
Figure 6.10 shows scalar plots of velocity (Left) and pressure (Right) for
parabolic, flat and 1/7 power law turbulent from top to bottom respectively.
Although the data is in 2-dimensions, we can see that there is little difference
between them in regard to velocity channelling or internal flow structure, re-
inforcing the theory that the velocity profile is having a minimal effect on the
pressure drop.

165
 6.8 Results

Experimental
CFD Profile
8000 CFD No profile (Flat)

6000
dP/L

4000

2000

1000 2000 3000 4000

RedP

Figure 6.7: Pressure drop per unit length as a function of Reynolds number for
bi-distributed spheres of 16 and 10mm

166
 6.8 Results

40
Parabolic profile
1/7 power law profile
Flat profile
Eisfeld and Schnitzlein (2001)
30

20
φ

10

0
0 0.005 0.01 0.015 0.02 0.025 0.03
3
Q (m /s)

Figure 6.8: Dimensionless pressure drop as a function of volumetric flow rate for
bi-distributed spheres of 16 and 10mm where φ = ∆P dp /LρU 2 . Note the little
difference in φ irrespective of turbulent profile.

167
 6.8 Results

1000
Experimental
Eisfeld and Schnitzlein (2001)
CFD SST k-ω
100

10
φ

0.1

0.01
1000
Redp

Figure 6.9: Dimensionless pressure drop as a function of Reynolds number for

bi-distributed spheres of 16 and 10mm

168
 6.8 Results

a. b.
9.06e-01 3.23e+00
8.70e-01 3.10e+00
8.34e-01 2.97e+00
7.97e-01 2.84e+00
7.61e-01 2.70e+00
7.25e-01 2.57e+00
6.89e-01 2.44e+00
6.52e-01 2.31e+00
6.16e-01 2.18e+00
5.80e-01 2.05e+00
5.44e-01 1.92e+00
5.07e-01 1.78e+00
4.71e-01 1.65e+00
4.35e-01 1.52e+00
3.99e-01 1.39e+00
3.62e-01 1.26e+00
3.26e-01 1.13e+00
2.90e-01 9.95e-01
2.54e-01 8.63e-01
2.17e-01 7.31e-01
1.81e-01 6.00e-01
1.45e-01 4.68e-01
1.09e-01 Y 3.37e-01 Y
7.25e-02 2.05e-01
3.62e-02Z X 7.35e-02Z X
0.00e+00 -5.81e-02

Contours of Velocity Magnitude (m/s) Nov 02, 2010 Contours of Static Pressure (pascal) Nov 02, 2010
FLUENT 6.3 (3d, pbns, lam) FLUENT 6.3 (3d, pbns, lam)

c. d.
1.72e+00 9.58e+00
1.65e+00 9.19e+00
1.59e+00 8.80e+00
1.52e+00 8.40e+00
1.45e+00 8.01e+00
1.38e+00 7.61e+00
1.31e+00 7.22e+00
1.24e+00 6.82e+00
1.17e+00 6.43e+00
1.10e+00 6.03e+00
1.03e+00 5.64e+00
9.65e-01 5.25e+00
8.96e-01 4.85e+00
8.27e-01 4.46e+00
7.59e-01 4.06e+00
6.90e-01 3.67e+00
6.21e-01 3.27e+00
5.52e-01 2.88e+00
4.83e-01 2.48e+00
4.14e-01 2.09e+00
3.45e-01 1.70e+00
2.76e-01 1.30e+00
2.07e-01 Y 9.06e-01 Y
1.38e-01 5.12e-01
6.90e-02Z X 1.17e-01Z X
0.00e+00 -2.77e-01

Contours of Velocity Magnitude (m/s) Nov 02, 2010 Contours of Static Pressure (pascal) Nov 02, 2010
FLUENT 6.3 (3d, pbns, lam) FLUENT 6.3 (3d, pbns, lam)

e. f.
7.22e+00 9.32e+01
6.94e+00 8.93e+01
6.65e+00 8.53e+01
6.36e+00 8.14e+01
6.07e+00 7.74e+01
5.78e+00 7.35e+01
5.49e+00 6.95e+01
5.20e+00 6.56e+01
4.91e+00 6.16e+01
4.62e+00 5.77e+01
4.33e+00 5.37e+01
4.05e+00 4.98e+01
3.76e+00 4.58e+01
3.47e+00 4.19e+01
3.18e+00 3.79e+01
2.89e+00 3.40e+01
2.60e+00 3.01e+01
2.31e+00 2.61e+01
2.02e+00 2.22e+01
1.73e+00 1.82e+01
1.44e+00 1.43e+01
1.16e+00 1.03e+01
8.67e-01 Y 6.37e+00 Y
5.78e-01 2.42e+00
2.89e-01Z X Z
-1.53e+00 X
0.00e+00 -5.47e+00

Contours of Velocity Magnitude (m/s) Nov 08, 2010 Contours of Static Pressure (pascal) Nov 08, 2010
FLUENT 6.3 (3d, pbns, lam) FLUENT 6.3 (3d, pbns, lam)

Figure 6.10: Contours of velocity magnitude (a) and pressure (b) for mixed bed of
16mm and 10mm beds of spheres with parabolic inlet profile. Contours of velocity
magnitude (c) and pressure (d) for mixed bed of spheres with flat plane profile.
Velocity magnitude and pressure for a 1/7 law profile a, and b, respectively

169
 6.9 Discussion

6.9 Discussion
We can see from figure 6.8, where the pressure drop per unit length is plotted
as a function of volumetric flow rate that the effect of the profile is having a
minimal effect and parabolic, flat and 1/7 power law show minimal difference in
results. This is due to the bed geometry having a strong effect on the pressure
gradient before the bed and distorting the velocity profile before the fluid reaches
the media. This justifies the closeness of results to the experimental in chapters 4
and 5 without the use of a velocity profile for a fully turbulent regime. The data
presented in the correlation proposed by Eisfeld & Schnitzlein (2001) does not
mention the effect of the velocity profile, although there is a significant amount of
empirical data to represent this, only the individual effects cannot be separated
from the correlation. The effect of the profile may well account for the significant
spread in the results. In addition, this is based on the assumption that the
Authors of all the research papers have designed an experimental rig suitable to
allow a fully developed flow to be formed before the obstruction of the media. In
regard results presented here it seems plausible the effect of the velocity profile
is having minimal effect on the pressure drop.
Figure 6.11 depicts velocity vectors for a random location within the bed.
We can see recirculation in stagnation zones and at the rear of the spheres in
the pressure drag region which normally would be averaged out by the RANS
equations when a turbulence model is implemented. A similar vortex structure
behind spherical media has been observed by Dalman et al. (1986) using a com-
putational approach and Yevseyev et al. (1991) using laser doppler anemometry.
Conversely, Tsotas (2002a) report no vortex shedding from the particles around
particle Reynolds numbers of Re = 4780 − 7010 which are in the ranges analysed
here. Their experimental method relies on the use of hot-wire anemometers in the
void space to detect unsteady flow behaviour. However, a hot wire anemometer
utilises Newton’s Law of cooling and the temperature difference to determine the
velocity, and it is unlikely that the response time is fast enough to capture the
highest frequency turbulent events in the turbulent spectrum. In addition, the
bed analysed by Tsotas (2002a) is based on a rhombahedral, structured packing
and not a bi-distributed disordered pack.

170
 6.9 Discussion

3.71e+01
3.53e+01
3.34e+01
3.16e+01
2.97e+01
2.78e+01
2.60e+01
2.41e+01
2.23e+01
2.04e+01
1.86e+01
1.67e+01
1.49e+01
1.30e+01
1.11e+01
9.28e+00
7.43e+00
5.57e+00
3.71e+00 Y
1.86e+00Z
X
0.00e+00

Velocity Vectors Colored By Velocity Magnitude (m/s) (Time=0.0000e+00) Nov 09, 2010
FLUENT 6.3 (3d, pbns, lam, unsteady)

Figure 6.11: Vectors of velocity showing flow recirculation inside the bed at
3.0m/s

171
 6.10 Conclusions

Behind a bluff body such as sphere or cylinder, the wake region is likely
to exhibit both true stochastic turbulence and periodic deterministic vortices
and recirculation at distinct frequencies. With these two effects coinciding in
the same location, the two phenomena are notoriously hard to separate using
empirical techniques. With the approach used in this work the velocity vectors are
calculated from the cell node value in the form of a magnitude and direction. In
which case it is impossible to decipher whether this is true stochastic turbulence or
periodic deterministic vortex shedding from the back of the sphere. Furthermore,
the skewed elements can be affecting both the magnitude and the direction of the
vectors. In regard to this Thesis as a whole, skewed elements may be affecting
local phenomena, such as recirculation; damping out or creating recirculation
zones which should not be there, however the effect is not significant enough to
be affecting global properties such as the pressure drop caused by the media. In
addition, this effect is fairly localised to particle contacts which are naturally in
locations of minimal fluid flux.
In all cases, with and without a suitable inlet flow profile results have shown to
slightly under-predict the pressure drop with this effect becoming more apparent
as the Reynolds number increases.

6.10 Conclusions
In this chapter, computationally generated packed beds (generated using MacroPac)
have been meshed using IBM techniques (ScanCAD, ScanFE), to create CFD
models for bi-distributed beds. At the same time, RP techniques have been used
to synthesise a geometrically accurate physical bed for experimental investigation.
Both experimental results carried out of the RP bed and computational results
have shown to match closely, particularly at the lower Reynolds range. As the the
Reynolds number increases the solver has tended to under-predict the pressure
drop. The work has also investigated the application of the expected appro-
priate flow profiles to the computational model’s boundary conditions. Results
have show that that the velocity profile is having minimal effect on the pressure
drop, and is heavily dependant on the flow geometry. As a whole, this work has

172
 6.10 Conclusions

laid the foundations for further work into the application of rapid prototyping in
manufacturing disordered structures for comparison with CFD results.

173
Chapter 7

Correlations and Conclusions

“A story should have a beginning, a middle, and an end... but not

necessarily in that order.”
Jean-Luc Godard

7.1 Introduction
This final Chapter draws on the techniques described in the previous Chapters to
generate two new models to predict bed porosity and pressure drop and in addi-
tion concludes this Thesis. Drawing on the previous work, Chapters 4 and 6 have
implemented the particle packing algorithm MacroPac to generate packed beds
for computational fluid dynamics. MacroPac has been shown to generate com-
putational models fast and efficiently lending itself to the production of a large
data set from which bed properties can be extracted. This chapter uses this to
generate a data set of randomly packed beds to which a correlation can be fitted
to determine expected bed porosity as a function of aspect ratio. This correlation
is then compared to other empirical correlations in the literature (Beavers et al.,
1973; de Klerk, 2003; Dixon, 1988; Jeschar, 1964; Zou & Yu, 1995). The computa-
tional and empirical methods used in this Thesis have generated a large data set,
such that a regression can be constructed to describe the dimensionless pressure
drop for the window of Reynolds numbers and aspect ratios studied. Referring
back to chapters 2, 4 and 5, realistically, there is no single formulation which can

174
 7.2 Porosity

describe the dimensionless pressure drop for every single case due to the tortuos-
ity and packing type not being taken into account. It could be argued that the
correlation proposed by Eisfeld & Schnitzlein (2001) accounts for all these effects
by the large range of empirical data used to fit the regression, however these terms
cannot be separated from the data, nor are any values given. Rather than a ‘one
size fits all approach’, it is more suitable to window the data to specific Reynolds
number and suitable aspect ratios. The regressions described here do not replace
or provide the accuracy of empirical and computational methods, but provide an
alternative in the form of a ‘good engineering approximation’ for the described
Reynolds numbers and aspect ratios. The computational methods described in
this Thesis can model pressure drop with a good degree of accuracy, however,
there is still not enough computer power available to the average Engineer to
perform these computer simulations in a suitable time frame. An empirical cal-
culation can take a matter of minutes, where as CFD simulation of a flow domain
as complex as the media described in this thesis takes several hours or even days.

7.2 Porosity
The bulk average porosity, or just the bed porosity is a global quantity which
heavily affects the passage of fluid through a packed bed. Determining a realistic
value of porosity can be challenging due to the associated uncertainties with most
well established techniques. One of the simplest approaches is to fill the voids
with a known amount of water representing the void volume, Vv (Leva et al.,
1951). A similar technique is to remove the bed particles and add them to a
measuring cylinder with a known volume of water and measure the fluid’s rise
due to Archimedes’ ‘Eureka’ theory, as applied by Giese et al. (1998). However,
this group of techniques is susceptible to entrapped bubbles around the media
and is limited to non-absorbent particles. The effect of the bubbles can be con-
siderably reduced, with the use of a surfactant coupled with a vacuum chamber
or the use of mercury due to its non-wetting properties. However, mercury has its
own disadvantages in very being dense and viscous, such to an extent that often
smaller particle pores or gaps are not filled by the liquid. Additionally, mercury

175
 7.2 Porosity

is included in the RoHS directive (Risk of Hazardous Substances) due to its cu-
mulative toxicity and poses a significant risk to health from continuous exposure.
Although the smallest pores are not likely to be filled by the mercury, it is ar-
guable the effect is not likely to be effecting the pressure drop or flow pathways
considerably. This is reinforced by the fact that the primary cause of pressure
drop is governed by the bed geometry and not so much by the particle roughness
or micro-scale geometry. Other empirical methods include the use of wax to fill
the voids and the mechanical removal of layers of a known thickness for weighing
(Gotz et al., 2002) which is time consuming and cumbersome and most of all de-
structive. However this technique can be useful in determining local porosity, a
feature not provided by water substitution due to its global technique. The tech-
nique suffers similar viscosity associated complications as mercury porosimetry
and the same argument can be applied.
Non-invasive methods, such as MRI, XRM (X-ray microtomography) and CT
provide an approach which allow the probing of a packed beds internal micro-
structure, as demonstrated in Chapter 5. M.C.Thadani & Peebles (1966) have
shown how a packed bed of plastic spheres fixed in a matrix can be segmented
and each segment radio-graphed and local porosities determined from the pho-
tometric reading of the radiograph emulsion point densities. Their results show
that the porosity as a function of particle centroid distance from the wall ex-
hibits a decaying sinusoidal trend. MRI techniques are valuable for determining
a bed’s complex flow paths, but with high resolution MRI, magnetic suscepti-
bility induced artifacts can cause particle necking. This phenomena makes little
difference in a very low aspect ratio bed (Aratio < 5), but as the aspect ratio in-
creases the more particles equates to more contact points resulting in an a wildly
inaccurate value of porosity. If the scan resolution is not sufficiently smaller than
an individual particle it is likely that most of the voids will not be detected at
all.
The disadvantage of all these techniques is the arduous task of completing the
experiments and the associated expense in regard to setup cost and man-hours,
thus limiting these techniques for research purposes. A good quick engineering
approximation for average or bulk bed porosity is best served by a single empir-
ically derived correlation. Correlations exist to predict a value for bed porosity

176
 7.2 Porosity

in regular and irregular packings and are given as porosity as a function of the
dimensionless aspect ratio, such as the empirical correlation proposed by Jeschar
(1964)
dp
ε = 0.375 + 0.34 (7.1)
D
Dixon (1988) proposed a similar equation to form what is considered ‘the stan-
dard’ correlation for the porosity as a function of aspect ratio for a packed bed
given as  2
dp dp
ε = 0.4 + 0.05 + 0.412 (7.2)
D D
de Klerk (2003) proposed the following equation to describe the porosity as a
function of aspect ratio
ε = 0.41 + 0.35e0.39D/dp (7.3)

whilst another correlation is proposed by Zou & Yu (1995)

   
10.686
ε = 0.4 + 0.001 e −1 (7.4)
(D/dp )

The porosity correlations are shown in 7.1. We can see from figure 7.1 that the
exponentially based correlation of Zou & Yu (1995)never reaches zero, in which
case this correlation diverges away from the correlations of Dixon (1988); Jeschar
(1964) and de Klerk (2003) steeply at low aspect ratios. This may not be as
misleading as first assumed. If we consider a spherical bed with an aspect ratio
of one, a single sphere is likely to completely block the pipe, in which case no flow
can pass through. This may not be the case with a particle with an equivalent
diameter of one. A sphere can completely block a pipe, but it is still porous in
regard to the ε = Vv /V relation although no fluid can pass. The correlations of
Dixon (1988); Jeschar (1964) and de Klerk (2003) will still give a value of porosity
when the aspect ratio is one and be used in conjunction with and equation such
as Eisfeld & Schnitzlein (2001) which will still give a value of pressure drop, which
we know is not possible. In the case of very low aspect ratio beds, the pressure
drop should not be quantified by the porosity but more the area available to the
flow around the sphere.
Beavers et al. (1973) proposed a model for bulk porosity introducing

177
 7.2 Porosity

  
d p εw
ε = ε∞ 1+2 (7.5)
D ε∞
Choi et al. (2008) suggest that the correlation of Beavers et al. (1973) is more
accurate for predicting bulk porosity than the equations of Zou & Yu (1995).
Here MacroPac is used simply a tool for for implementing a Monte-Carlo
method and generating random disordered beds. It is used for analysing the
porosity for different beds and therefore creating a data set for an appropriate
regression analysis to be performed. MacroPac determines the porosity in two
ways, either with a voxel approximation or a sphere slice approximation. With
the voxel approximation, the bed is discretized into a lattice of voxels in which a
voxel either lies inside a particle, or outside a particle. From the ratio of those
that lie within a particle to those which lie outside, an approximation of porosity
can be determined. The sphere slice approximation determines the porosity by
taking slices through the spheres and calculates sphere volume within the sliced
region. Intelligensys quote ‘In systems consisting entirely of spheres, the results
returned will be 100 % correct’ or at least to the specific bed. On this basis this
is the chosen method for this work.
In this case MacroPac is set to take a value of porosity from a control volume
excluding particles at the extreme ends of the bed. This is achieved by setting
the two boundaries in the z-axis to ‘soft’ enabling more spheres to be packed into
the volume, enabling a value of porosity to be taken from the core of the bed.
The soft function allows a particle to marginally overlap or stretch the confining
geometry to allow more particle to be packed. If this is not applied, MacroPac
will fill the control volume with the maximum amount of spheres possible. Figure
7.1 shows the data produced by MacroPac fitted with a regression and compared
with other previous research. The regression gives the function
 −0.16
D
ε = 0.6 (7.6)
dp

resulting in a coefficient of determination of R2 = 0.82

This graph shows that porosity data produced by MacroPac is in good agree-
ment with previous research and the correlations proposed by Jeschar (1964) and

178
 7.2 Porosity

1
de Klerk (2003)
Andersson (1963)
Jeschar (1964)
MacroPac
Nemec and Levec (2005)
Griffiths
0.75 Motillet et al (2001)
0.5978x^-0.1559
Leva et al (1951)
Atmakidis and Kenig (2009)
Zou and Yu (1995)
ε

0.5

0.25

10 20 30
D/dp

Figure 7.1: Porosity correlations as a function of aspect ratio compared with

experimental data

179
 7.2 Porosity

de Klerk (2003). We can see from figure 7.1 that the correlation proposed by Zou
& Yu (1995) fits the data well at aspect ratios of Aratio > 5 but below this value
diverges steeply away from previous research and proposed correlations. Where
as the correlations proposed here and the correlations of Zou & Yu (1995) diverge
steeply away at low aspect ratios due to the use of an exponential function.

7.2.1 Local Porosity

Figure 7.2 shows height averaged porosity (ε) as a function of the dimensionless
distance ratio (x/dp ) for a range of low aspect ratio beds. The bed is created
using MacroPac with a bed diameter of 100mm and a length of 250mm. The
porosity is calculated using the sphere slice method using a 2d slice taken across
the centre of the bed from top to bottom.
The plot displays homogeneity in the results resembling under-damped os-
cillation, where the first peak is a mean representative of the centroids of the
spheres adjacent to the bed wall. If we interpret the extents of the wall region
as the linear porosity region from the wall to the edge of the first particle then
this proves the assumption which was made by Martin (1978). However, in a
fluids based interpretation the wall region penetrates further into the bed and
only stops when the fluid flow through the core region is no longer affected by the
wall. In beds of very low aspect ratios the wall region is likely to be dominant
through the whole bed. In a high aspect ratio, continuously packed bed, the wall
region has no noticeable effect.
On question that can be raised when determining a low aspect ratio packed
bed porosity is when does the bed finish? Here the argument is presented that
the spheres at the most extreme point of the bed are only half packed and are in
contact with the spheres directly below them, in which case the bed would finish
at their centroids an in the case of water substitution this would be the fill level.
This is fine providing each of the sphere centroids are level, such as for a regular
cubic pack, this would suggest in a random arrangement that the water would
be filled to the mean centre line of all the most extreme spheres. This raises the
issue how this would be determined? Similarly, this argument could be applied
to the base of the bed. However, the first row of spheres has an impact on the

180
 7.2 Porosity

Aratio = 25
Aratio = 16.7
0.8 Aratio = 12.5
Aratio = 10
Aratio = 6.3
0.6
ε

0.4

0.2

0
0 0.5 1 1.5 2
x/ddp

Figure 7.2: Bed porosity as a function of distance from wall (x) non-
dimensionalised using the particle diameter (dp )

181
 7.3 The dimensionless pressure drop model

bed due to their aerodynamic properties in regard to flow channelling and drag.
If these spheres were in fact hemispheres with the flat end leading, this would
result in considerably more pressure drag resulting in a higher drop in pressure.
Figure 7.3 shows the standard deviation of porosity from a data set of over 250
beds generated using MacroPac. We can see that the highest standard deviation
is in the region of 4 < Aratio < 6. With an aspect ratio less than 4 there are fewer
packing possibilities or at least the structure is the same but may vary radially.
Where the aspect ratio is greater than 6 (Aratio > 6) the bed will start to become
more homogeneous.

7.3 The dimensionless pressure drop model

We have discussed correlations which have been devised to predict pressure drop
through packed beds and porous mediums, such as work of Darcy (1856). Some
of these are purely theoretical models, based on Hagen-Poiselle flow (theoreti-
cal tubes) (Carman, 1937; Ergun & Orning, 1949), some are empirical or semi-
empirical (Eisfeld & Schnitzlein, 2001). The work of Eisfeld & Schnitzlein (2001)
is considered to be one of the most promising correlations devised from adjust-
ing the coefficients of (Reichelt, 1972) with a regression of over 2300 data points
from 13 difference academic sources. Although a considerable population has
been analysed every data point in the work of Eisfeld & Schnitzlein (2001) is
empirically derived from physical experiments. Although it would be hard to
question the simplicity, accuracy and reliability of a fluid manometer, the work
of Eisfeld & Schnitzlein (2001) is somewhat limited by the range of techniques
for acquiring the pressure drop. This raises the question, that with the rapid
progression of computational techniques and the speed and simplicity, are em-
pirical methods the most accurate? And should we now be reinforcing empirical
measurements with computational ones in contrast to traditionally reinforcing
computational approaches with empiricism?
We know that there is no individual correlation that can model pressure drop
in a packed bed for every case with sufficient accuracy. Even though this is
well understood, the large range of both experimental and computational data

182
 7.3 The dimensionless pressure drop model

0.012

Standard deviation (MC shake)

0.01 Standard Deviation (DD)

0.008
Standard Deviation

0.006

0.004

0.002

0
4 6 8 10 12 14 16 18
D/dp (Aspect ratio)

Figure 7.3: Standard deviation of porosity for packed beds

183
 7.3 The dimensionless pressure drop model

obtained in the previous chapters, means that it would be a waste not to de-
rive some sort of correlation and compare it to the experimental data of other
researchers.
Here the data is presented in the form of a modified dimensionless pressure
drop (modified friction factor) as defined by Eisfeld & Schnitzlein (2001); Ergun
(1952) as
φε3
φ′ = (7.7)
(1 − ε)
This is necessary when comparing bed of different porosities as the dimen-
sionless pressure drop and particle Reynolds number do not take into account the
bed porosity. Figure 7.4 shows all the data presented in this thesis as modified
dimensionless pressure drop as a function of the particle Reynolds number. A
least squares regression is fitted to figure 7.4 to give

f (x) = 15.4Re−0.332
dp (7.8)

and hence becomes

 −0.332
φε3 ′
= 15.4 Redp (1 − ε) (7.9)
(1 − ε)

where 15.4 is an empirical coefficient of some other variables. Substituting in the

particle Reynolds number, given as

ρU dp
Redp = (7.10)
µ

yields  −0.332
φε3 ρU dp
= 15.4 (1 − ε) (7.11)
(1 − ε) µ
re-arranging for φ
 −0.332
ρU dp (1 − ε)
φ = 15.4 (1 − ε) (7.12)
µ ε3

We know from previous correlations that pressure drop through any packed bed
is heavily dependant on the porosity. In addition we know dimensionless pressure

184
 7.3 The dimensionless pressure drop model

All Data (Experimental and Computational)

-0.332
Least squares regression 15.4Re

10
φ

0.1
0 1000 2000 3000 4000 5000 6000
Redp

Figure 7.4: All dimensionless pressure drop data presented in this thesis as a
function of particle Reynolds number

185
 7.3 The dimensionless pressure drop model

drop is given in the form

∆P dp
φ= (7.13)
LρU 2
re-arranging for ∆P/L
∆P ρU 2 φ
= (7.14)
L dp
substituting for φ gives
 −0.332 2
∆P ρU dp ρU (1 − ε)
= 15.4 (1 − ε) (7.15)
L µ ε3 d p

We also have the result that the porosity can be defined as

 −0.16
D
ε = 0.6 (7.16)
dp

which can be used to define porosity in the equation without relying on an ex-
perimental value, which in some cases can be limiting and time consuming.
Here we do not necessarily expect the previous data to fit with absolute pre-
cision as we know that there is no formulation that can describe every individual
case. Moreover, much of the data in this regression is formed around pseudo-
packed beds with issues in respect to geometric fidelity regarding contact points.
We also know that the described equations do not take into account bed tortu-
osity and therefore neglects the heterogeneous flow paths through the bed. The
proposed correlation to describe dimensionless pressure drop is plotted alongside
the correlation of Eisfeld & Schnitzlein (2001) selected because of its applicability
to wide ranges of Reynolds numbers and aspect ratios. In addition the correlation
is plotted alongside the experimental data of Reichelt (1972) due to closeness of
aspect ratios studies (3.32 < Aratio < 14.32) and encompassing a suitable range
of Reynolds numbers (74 < Re < 5463).
Results are shown to match closely with both the correlation of Eisfeld &
Schnitzlein (2001) and the experimental data of Reichelt (1972). The results
deviate as a function of the Reynolds number diverging away from the correlation
of Eisfeld & Schnitzlein (2001). It is plausible that this is due to the vast range of
Reynolds numbers the correlation of Eisfeld & Schnitzlein (2001) has been fitted

186
 7.3 The dimensionless pressure drop model

100

Reichelt (1972)
Wentz and Thodos (1963)
Eisfeld and Schnitzlein (2001)
Pressure drop model

10
φ

0.1
1000 10000
Redp

Figure 7.5: Proposed dimensionless pressure drop model compared with the ex-
perimental data of Reichelt (1972)

187
 7.4 Conclusions

to and the fairly narrow specific range fitted in this study (102 < Redp < 6114).

7.4 Conclusions
This Thesis has investigated traditional and novel methods for generating random
disordered beds for analysis using computational fluid dynamics. Chapter 4 has
demonstrated how a Monte-Carlo algorithm employed by the code MacroPac can
be used to generate coordinate locations for random-disordered beds of particles.
This coupled with traditional CAD based geometry definition and meshing has
been shown to replicate beds of spheres with a good deal of accuracy. However,
the technique is not robust, and creating a suitable domain for analysis can of-
ten be tortuous and time consuming with considerable user intervention. Once a
workable mesh has been created it is often of dubious quality containing a number
of highly degenerate elements ultimately compromising the computational stabil-
ity. Fluent has shown to ameliorate the the skewness phenomena to an extent,
by a skewness correction from the PISO algorithm and by converting skewed el-
ements to polyhedra, but does not completely eliminate the problem. Including
the skewness correction in the PISO loop increases the computational run-time
considerably as the correction is applied to all elements whether they are degen-
erate or not. Despite all the disadvantages of this approach, such as meshing
robustness, the method still provides us with good results in most cases when
comparing the computational data with the experimental equivalent beds. Both
computational and empirical results compare well with the correlation of Eisfeld
& Schnitzlein (2001) and have paved the way for the creation of more advanced
meshing algorithms which can deal with highly complex curved geometries.
Chapter 5 has demonstrated that actual physical packed beds can be recreated
with some accuracy using non-invasive methods (MRI) coupled with image based
meshing. The image based meshing technique makes the replication of the com-
plex geometries for analysis considerably more viable without considerable user
intervention and tortuosity associated with traditional CAD based approaches.
Unlike the CAD based approach the technique allows robust meshing with consid-
erably fewer degenerate elements, without major simplification of the geometry.
This method has shown to replicate the beds with a good degree of accuracy,

188
 7.4 Conclusions

but is heavily limited by the scan resolution and the presence of magnetic sus-
ceptibility induced artifacts causing particles to neck at the contact points. The
voxelisation used in the image based meshing approach also causes slight particle
necking, but this affect is so minimal it is absorbed by the necking caused by the
relatively coarse scan resolution. The effect of the necking phenomena, compared
to representing the contact points using highly degenerate elements (Chapter 4)
is arguable. Even so, it is possible that these effects are having a minimal effect
on the pressure drop as the phenomena is fairly limited to areas of relative flow
stagnation, thus it may be compromising the local flow phenomena. Consider-
ing these effects both experimental and computational results compare well with
each other and show good agreement with the correlation of Eisfeld & Schnitzlein
(2001) and Reichelt (1972).
Considering the phenomena discovered in Chapters 4 and 5 concerning the
creation of a geometrically faithful geometry and robustness issues, Chapter 6
has shown how 100% geometrically faithful beds can be created using a blend
of methods described in Chapters 4 and 5. MacroPac is used to generate a
random coordinate location for each particle. A tertiary program written in C
is used to create an STL file of spheres from the coordinate locations. Image
based meshing has been shown to be robust in meshing highly curved, disordered
geometries and in this case is used to recreate the geometry from the STL using
ScanCAD, ScanIP and ScanFE. The slight necking phenomena associated with
the image based meshing approach is used as an advantage to produce a statically
stable structure suitable for manufacture using rapid prototyping. CFD results
compared with empirical data have shown to be the most accurate of all the
results presented in this thesis and compare well with the correlation of Eisfeld
& Schnitzlein (2001).
With the three computational processes and methodologies described in Chap-
ters 4, 5 and 6 a significant data set has been produced of over 200 data points to
describe dimensionless pressure drop. In all cases the results obtained using CFD
have shown to diverge away from the experimental as a function of the Reynolds
number as the pressure drop increases. A reasonable assumption could be to
assume that this is an effect of geometric fidelity issues, but this phenomena can
still be observed when a geometrically faithful bed is analysed, which leads to the

189
 7.4 Conclusions

assumption the uncertainty is systematic. The experimental method has been

proved and unchanged for decades and is based on the successful experiments
of Tóbis (2000) so is an unlikely candidate for the cause of the uncertainty. In
addition, experimental uncertainty is likely to offset the data in contrast to affect-
ing the gradient of pressure drop per unit length, ∆P/L. The most likely cause
of this uncertainty is from the computational modelling. We know that most
complications arise due to turbulence and how it is modelled. We also that most
turbulence models are grossly simplified and based on dimensional assumptions.
In the work the k-ω SST model was used to model turbulence. It is a relatively
low Reynolds turbulence model which benefits from the treating the free-stream
flow as a standard k-ε and near wall region as k-ω. We have discussed in Chapter
2 the difficulties of modelling turbulence reinforced by George. E. Box’s quote
“All turbulence models are wrong”, so we must ask ourselves the question ‘is the
k-ω model giving a realistic representation of the turbulence effects?’. The k-ω
SST model has damping in place to restrict the production of turbulent kinetic
energy which may be causing the results to diverge. In addition, the effect of
skewed elements and the node location in regard to the face centres can cause
unrealistic face values in regard to magnitude and direction which may damp
out some of the turbulent effects. Furthermore, the RANS modelling technique
treats all unsteady behavior as turbulence, we know that deterministic unsteady
phenomena can occur behind bluff bodies end up being modelled as if they were
true stochastic turbulence. In regard to the k-ω SST model itself, it relies on an
array of empirical coefficients which were determined in on a variety of flows, not
inclusive of geometries as complex as packed beds.
Turbulence models usually have simplifying assumptions (based on Reynolds
number, lenght scales, geometry etc) and we have discussed that ’there is no
generic turbulence model’ that can be successfully applied to every case. Due
to this, it is usually paramount that the most appropriate turbulence model is
applied to each specific case in order to yield best results. To date, there is no
credible turbulence model specific to packed beds. This is most probably due
to the random unstructured nature of packed beds and the varying length scales
and reynolds numbers that can be observed. However, in the case of random
unstructured, highly tortuous geometries such as packed beds, it is likely that

190
 7.4 Conclusions

the turbulence itself is having little effect on the desired studied property, ∆P/L.
In which case, packed beds are relatively insensitive to the choice of turbulence
model and in many cases it may be advantageous to ommit the turbulence model
all together.
There is conflicting evidence regarding the effect of particle roughness and this
is suggested by Eisfeld & Schnitzlein (2001) to most likely to be attributed to a
‘misinterpretation of data’. We know that the internal structure of the packed
bed is the most important factor which concerns pressure drop. To recreate
an exact empirical geometry every time is virtually impossible even with the
same parameters of porosity, aspect ratio, bed length. This effect is probably
what lead researchers to find conflicting evidence for and against the effect of
particle roughness. However, this still does not answer the question ‘Does particle
roughness have an effect on the pressure drop’. With the technique described in
this thesis using additive layer manufacturing (Chapter 6), further work could
include the creation of maybe ten or so beds of the same parameters, but with
different particle roughness to be analysed using traditional empirical methods
and in parallel analysed using CFD. As a result of this it is hoped that the
conflicting evidence regarding particle roughness can be resolved and put the
minds of many researchers (many posthumously) at rest.
This work has shown that the FV technique is both robust and accurate in
determining pressure drop through a packed bed. Results in all studies have cor-
related well with both the formulas of Eisfeld & Schnitzlein (2001) and empirical
data from previous researchers. The technique has shown to have an advantage
over empiricism in respect to cost, man hours and the amount of data which can
be extracted.
Using all the dimensionless pressure drop data produced in this work a semi-
empirical formula has been produced from computational and empirical studies.
It is shown to correlate well with the correlation proposed by Eisfeld & Schnit-
zlein (2001) and data of Reichelt (1972). Further work may include the use of
genetic algorithms applied to the data set to tease out relationships regarding the
geometric parameter of most influence.
Are we at the stage where we should be comparing empirical data to compu-
tational in contrast to computational with empirism? Is empiricism dead? Will

191
 7.4 Conclusions

‘wind tunnels become cabinets for computers’ ? Although the Author is not an
empirical purest, we are still decades away from being able to realistically model
a complete stochastic turbulent flow, in a complex media. With todays computer
limitations, coupled with uncertainties associated with turbulence and complex
geometric domains, computational data is still best reinforced with empiricism.
Calis et al. (2001) anticipated that within five years packed beds containing
a few hundred particles would be considered a ‘standard’ problem in terms of
memory and calculation time requirements. Here it has been demonstrated that
the memory and calculation time are now not the issue but more so, how these
beds of particles can be accurately replicated in the form of a computational
domain. With the postulation of Calis et al. (2001) satisfied, computers are
becoming ever increasingly powerful, we anticipate that within a few more years
it is likely that much larger beds of more than a thousand particles could be
analysed in this way.

192
References

Al-Jahmany, Y.Y., Brenner, G. & Brunn, P.O. (2009). Simulation of the

flow in a packed-bed with and without a static mixer by using cfd technique.
World Academy of Science, Engineering and Technology, 53, 875–880. 59

Alliez, P., Cohen-Steiner, D., Yvinec, M. & Desbrun, M. (2005). Vari-

ational tetrahedral meshing. ACM Transactions on Graphics (Special issue on
proceedings of SIGGRAPH), 24, 617 – 625. 136

Andrade, J.S., Almeida, M.P., Filho, J.M., Havlin, S. & Stanley,

B.S.H.E. (1997). Fluid flow through porous media: the role of stagnant zones.
Phys.Rev.Lett., 79, 3901 – 3904. 114

Antiga, L., Ene-Iordache, B., Caverni, L., Cornalba, G.P. & Re-
muzzi, A. (2002). Geometric reconstruction for computational mesh genera-
tion of arterial bifurcations from ct angiography. Computerized Medical Imaging
and Graphics, 26, 227 – 235. 62

Aris, R. (1957). On shape factors for irregular particles - i the steady state
problem. diffusion and reaction. Chemical Engineering Science, 16, 262 – 268.
11

Ashcombe, F. (1973). Graphs in statistical analysis. American Statistician, 27,

17 – 21. 109

Aste, T., Saadatfar, M., Sakellariou, A. & Senden, T.J. (2004). In-
vestigating the geometric structure of disordered sphere packing. Physica. A,
339, 16 – 23. 52

193
 REFERENCES

Atmakidis, T. & Kenig, E.Y. (2009). Cfd-based analysis of of the wall effect
on the pressure drop in packed beds with moderate tube/particle diamter ratios
in the laminar flow region. Chem.Eng.J , 155, 404 – 410. 50, 51, 85, 89, 96

Azzam, M.I.S. & Dullien, F.A.C. (1976). Calculation of permeability of

porous media from the navier-stokes equations. Ind.Eng.Chem.Fund , 15, 281
– 285. 72

Baker, M.J. & Tabor, G.R. (2010). Computational analysis of transitional

airflow through packed columns of spheres using the finite volume technique.
Computers and Chemical Engineering, 34, 878 – 885. 47, 84, 125, 151, 152

Baker, M.J., Young, P.G. & Tabor, G.R. (2011). Imaged based meshing
of packed beds of cylinders at low aspect ratios using 3d mri coupled with
computational fluid dynamics. 152

Bear, J. (1972). Dynamics of Fluids in Porous Media. American Elsevier Pub-

lishing Company. 25

Beavers, G.S., Sparrow, E.M. & Rodenz, D.E. (1973). Influence of bed
size on the flow characteristics and porosity of randomly packed beds of spheres.
Trans. ASME J. Appl. Mech, 30. 174, 177, 178

Bird, R.B., Stewart, W.E. & Lightfoot, E.N. (1960). Transport phenom-
ena. Wiley. 33, 34

Blake, F.C. (1922). The resistance of packing to fluid flow. Transactions of

American Istitute of Chemical Engineers, 14, 415 – 421. 32, 33, 34, 157

Bousinesq, J. (1877). Essai sur la theorie des eaux courantys. Mem.Pres. par
div savants l’academie Sci., Paris, 23, 1 – 680. 77

Bradshaw, P. (1994). Turbulence: the chief outsatnding difficulty with our

subject. Exp.Fluids, 16, 203 – 216. 72

Brinkman, H.C. (1947). A calculation of viscous force exerted by a flowing fluid

on a dense swarm of particles. Appl.Sci.Res, 1, 27 – 33. 32

194
 REFERENCES

Burke, S.P. & Plummer, W.B. (1928). ?? Industrial Engineering Chemistry,

20. 34, 35, 36, 156, 157

Calis, H.P.A., Nijenhuis, J., Paikert, B.C., Dautzenberg, F.M. &

van den Bleek, C.M. (2001). Cfd modelling and experimental validation of
pressure drop and flow profile in a novel structured catalytic reactor packing.
Chem.Eng.Sci., 56, 1713 – 1720. 22, 48, 192

Callaghan., P. (1994). Principles of Nuclear Magnetic Resonance Microscopy.

Oxford University Press, Oxford. 52

Carman, P.C. (1937). Fluid flow through a granular bed.

Trans.Inst.Chem.Eng.London, 15, 150 – 156. 33, 34, 35, 156, 157, 182

Carman, P.C. (1938). ?? Journal of Society of Chemical Industry, 57, 225. 33,
51

Caulkin, R., Ahmad, A., Fairweather, M., Jia, X. & Williams, R.

(2007). An investigation of sphere packed shell-side columns using a digital
packing algorythm. Computers and Chemical Engineering, 31, 1715 – 1724.
49, 151

Caulkin, R., Jia, X., Fairwather, M. & Williams, R. (2008). Lattic

approaches to packed column simulations. Particuology, 6, 404 – 411. 49

Caulkin, R., Ahmed, A., Fairweather, M., Jia, X. & Williams, R.A.
(2009). Digital preditions of complex cylinder packed columns. Computers and
Chemical Engineering, 33, 10 – 21. 151

Cavendish, J.C., Field, D.A. & Fry, W.H. (1985). An approach to auto-
matic three-dimnesional finite element generation. Int.J.Numer.Methods Eng,
21, 329 – 347. 68

Cebral, J.R. & Loehner, R. (2001). From medical images to anotomically

accurate finite element grids. Int.J.Numer.Methods.Eng, 51, 986 – 1008. 62

Cengel, Y.A. & Cimbala, J.M. (2006). Fluid Mechanics; Fundamentals and
Applications. McGraw Hill. 22, 67, 71, 73, 87, 97, 102, 148, 162

195
 REFERENCES

Choi, Y.S., Kim, S.J. & Kim, D. (2008). A semi-empirical correlation for
pressure drop in packed beds of spherical particles. Transp Porous Med . 35,
36, 38, 178

Clough, R.W. (1960). The finite element method in plane stress analysis. In
A.S.C.E. Structural Division, Pittsburgh, Pennsylvania. 58

Collins, T.P., Tabor, G.R. & Young, P.G. (2007). A computa-

tional fluid dynamics study of inspiratory flow in orotracheal geometries.
Med.Biol.Eng.Comput, 45, 829 – 836. 128

Courant, R. (1943). Variational methods for the solution of problems of equi-

librium and virbations. American Mathematical Society, 49, 1 – 23. 58

Dalman, M.T., Merkin, J.H. & McGreavy, C. (1986). Fluid flow and heat
transfer past two spheres in a cylindrical tube. Computers and fluids, 14, 267
– 281. 13, 42, 43, 85, 89, 136, 170

Darcy, H. (1856). Determination des lois d’ecoulement de l’eau travers le sable.

In Les Fontaines Publiques de la Ville de Dijon, 85, 590 – 594. 31, 32, 182

de Klerk, A. (2003). Voidage variation in packed beds at small column to

particle diameter ratios. AICheJ , 49, 2022 – 2029. 16, 174, 177, 180

Delaunay, B. (1934). Sur la sphere vide, izvestia akademii nauk sssr. Otdelenie
Matematicheskikh i Estestvennykh Nauk , 7, 793 – 800. 63

Delele, M.A., Tijskens, E., Atalay, Y.T., Ho, Q.T., Ramon, H., Nico-
lai, B.M. & Verboven, P. (2008). Combined discrete element cfd modelling
of airflow through random stacking of horticulatural products in vented boxes.
Journal of Food Engineering, 89, 33 – 41. 13, 42

Derx, O.R. & Dixon, A.G. (1996). Determination of the fixed bed wall heat
transfer coefficient using computational fluid dynamics. Heat Transfer Part A,
29, 777 – 749. 43

DiFelice, R. & Gibilaro, L.G. (2004). Wall effects for the pressure drop in
fixed beds. Chem.Eng.Sci., 59, 3037 – 3040. 110, 126

196
 REFERENCES

Dixon, A.G. (1988). Correlations for the wall and particle shape effects on fixed
bed voidage. Canadian Chemical Engineering, 66, 705 – 708. 16, 174, 177

Dixon, A.G. & Nijemeisland, M. (2002). Cfd as a design tool for fixed bed
reactors. Ind.Eng.Chem.Res, 40, 5246 – 5254. 43

Douglas, J.F., Gasiorek, J.M., Swaffield, J.A. & Jack, L.B. (2005).
Fluid Mechanics. Pearson, Prentice Hall. 71

Drazin, P. & Riley, N. (2006). The navier-stokes equations; a classification

of flows and exact solutions. London Mathemitical Society Lecture Note Series,
334, 1 – 19. 70, 71

du Toit, C.G. (2008). Radial variation in porosity in annular packed beds.

Nuclear Engineering and Design, 238, 3073 – 3079. 110

Dullen, F.A.L. (1992). Porous Media: Fluid Transport and Pore Structur .
Academic Press inc. 31

Dullien, F.A.L. (1979). Porous Media: Fluid Transport and Flow Structure.
Academic Press, New York. 156

Eisfeld, B. & Schnitzlein, K. (2001). The influence of confining walls on the

pressure drop in packed beds. Chem.Eng.Sci., 56, 4321 – 4329. iii, viii, x, 6, 8,
25, 28, 37, 38, 47, 54, 85, 105, 109, 110, 119, 120, 123, 146, 148, 156, 157, 158,
165, 170, 175, 177, 182, 184, 186, 188, 189, 191

Ergun, S. (1952). Flow through packed columns. Chem.Eng.Prog., 48, 89. 34,
35, 38, 51, 184

Ergun, S. & Orning, A. (1949). Fluid flow through randomly packed columns
and fluidised beds. Ind.Eng.Chem, 41, 1179 – 1184. 36, 156, 182

Evans, K.E. (1988). The packing of thick fibres. J.Phys.D:Appl.Phys., 22, 354
– 360. 49

Fand, R.M., Kim, B.Y.K., Lam, A.C.C. & Phan, R.T. (1972). Druckverlust
in kugelschttungen. Brennstoff-Warme-Kraft, 24, 233–236. 156

197
 REFERENCES

Ferziger, J.H. & Perić, M. (1996). Computational Methods for Fluid Dy-
namics. Springer. 72, 73

Foscolo, P.U., Gibalaro, L.G. & Waldron, S.P. (1983). A unified model
to for particulate expansion of fluidized beds and flow in fixed porous media.
Chem.Eng.Sci , 38, 1251 – 1260. 38

Foumeny, E.A., Benyahia, F., Castro, J.A.A., Moallemi, H.A. &

Roshani, S. (1993). Correlations of pressure drop in packed beds taking into
account the effect of confining wall. Int.J.Heat Mass Trans, 36, 536–540. 18,
28, 35, 38, 45, 114

Freund, H., Bauer, J. & Zeiser, T. (2005). Detailed simulation of transport

processes in fixed beds. Ind.Eng. Chem. Res, 44, 6423 – 6434. 45

Fried, E. & Idelchik, I.E. (1989). Flow resistance: Adesign guide for Engi-
neers. Hemisphere Pub. Corp. 38

Gibilaro, R.D.F.L.D. (2004). Wall effects for the pressure drop in packed beds.
Chemical Engineering Science, 59, 3037 – 3040. 36

Giese, M., Rottschaffer, K. & Vortmeyer, D. (1998). Measured and

modeled superficial flow profiles in packed beds with liquid flow. AIChEJ , 44,
484 – 491. 23, 175

Goodling, J.S., Vacheon, R.I., Stelpfug, W.S. & Ling, S.J. (1983).
Radial porosity distributions in cylindrical beds packed with spheres. Powder
Technology, 35, 23 – 29. 110

Gotz, J., Zick, K. & heinen adn T. Konig, C. (2002). Visualisation of

flow processes in packed beds with nmr imaging: Determination of the local
porosity, velocity vector and local dispersion coefficients. Chemical Engineering
and Processing, 41, 611 – 629. 23, 176

Guardo, A. (2007). Computational Fluid Dynamics Studies in Heat and Mass

Transfer Phenomena in Packed beds Reaction and Extraction Equipment: Spe-
cial attention to supercritical Fluids Technology. Ph.D. thesis, Universitat Po-
litecnica de Catalunya. 45, 52, 57, 58, 85

198
 REFERENCES

Guardo, A., Coussirat, M., Larrayoz, M., Recasens, F. & Egusquiza,

E. (2007). Influence of the turbulence model in cfd modeling of wall-to-fluid
heat transfer in packed beds. Chem.Eng.Sci , 60. 44, 45

Gunjal, P.R., Ranade, V.V. & Chaudhari, R.V. (2005). Computational

study of a single-phase flow in packed beds of spheres. A.I.Ch.E Journal , 20,
365–378. 48

Gupte, A. (1970). Experimentelle Untersuchung der Einflsse von Porositt und

Korngrenverteilung im Widerstandsgesetz der Porenstrmung. Ph.D. thesis,
Universitat of Karlshruhe. 157

Handley, D. & Heggs, P.J. (1968). Momentum and heat transfer in regularly
shaped packing. Trans. Inst. Chem. Eng, 46, 251 – 264. 35, 38

Hassan, Y.A. (2008). Large eddy simulation in pebble bed gas cooled core
reactors. 23, 47

Hassanizadeh, S.M. & Gray, W.G. (1987). High velocity flow in porous
media. Transport in Porous Media, 2, 521 – 531. 29

Hellström, J.G.I. & Lundström, T.S. (2006). Flow through porous media
at moderate reynolds number. In International Scientific Colloquium: Mod-
elling for Material Processing, Riga, June 8-9 . 48, 89

Herman., G.T. (2009). Fundamentals of computerized tomography: Image re-

construction from projection. Springer, London. 52

Hess, J.L. & Smith, A.M.O. (1967). Calculation of potential flow about arbi-
trary bodies. Progress in Aeronautics Sciences, 8, 1 – 138. 58

Hicks, R.E. (1970). Pressure drop in packed beds of spheres. Industrial and
Engineering Chemistry Fundamentals, 9, 500 – 502. 35

Jafari, A., Zamankhan, P., Mousavi, S.M. & Pietrinen, K. (2008).

Modeling and cfd simulation of flow behavior and disperivity through randomly
pacvked bed reactors. Chem.Eng.J , 155, 476 – 482. 50, 85

199
 REFERENCES

Jeschar, R. (1964). Druckverlust in mehrkornschuttungen aus kugeln. Arch.

Eisenhiitenw , 35, 1155 – 1159. 16, 127, 174, 177, 178

Johns, M.L., Sederman, A.J., Bramley, A.S., Gladden, L.F. &

Alexander, P. (2000). Local transitions in flow phenomena through packed
beds identified by mri. AIChE Journal , 11, 1251 – 2161. 53, 131

Johnson, E.A.C. & Young, P.G. (2005). On the use of a patient-specific

rapid prototyped model to simulate the response of the human head to impact
and comparison with analytical and finite element models. J. Biomech., 38, 39
– 46. 129

Kozeny, G. & Sitzber, J. (1927). Ober kapillare Leitung das Wassers im

Boden. Akad. Wiss, Wein, Math-naturw K1., 136, 271. 33, 34, 36

Kuwahara, F., Yamane, T. & Nakayama, A. (2006). Large eddy simulation

of turbulent flow in porous media. International Communications in Heat and
Mass Transfer , 33, 411 – 418. 76

Landahl, M.T. & Mollo-Christensen, E. (1986). Turbulence and Random

Processes in Fluid Mechanics. Cambridge University Press, Cambridge. 20, 77

Leland, R.W., Melander, D.J., Meyers, R.W., Mitchell, S.A. & Tau-
tages, T.J. (1998). The geode algorithm: Combining hex/tet plastering, dic-
ing and transition elements for automatic, all-hex mesh generation. In Seventh
International Meshing Roundatable (Dearborn, Mitchigan), 515 – 521. 136

Leva, M. (1959). Fluidisation. McGraw-Hill Book Company, New York. 35, 37

Leva, M., Weintraub, M., Grummer, M. & Storch, H.H. (1951). Fluid
flow through packed and fluidized systems. Tech. Rep. 504, Bureau of Mines.
156, 157, 175

Lien, F.S. & Leschziner, M.A. (1994). Assessment of turbulence trans-

port models including non-linear rng eddy viscosity formulations and second-
moment closure for flow over a backward-facing step. Computers Fluids, 23,
983 – 1004. 21

200
 REFERENCES

Lloyd, B. & Boehm, R. (1994). Flow and heat transfer around a linear array
of spheres. Numer Heat Transfer part A, 26, 237 – 252. 13, 43, 85

Logtenberg, S.A. & Dixon, A.G. (1998). Computational fluid dynamics of

fixed bed heat transfer. Chemical Engineering Process, 37, 7 – 21. 43, 85, 96

Logtenberg, S.A., Nijemeisland, M. & Dixon, A.G. (1999). Computa-

tional fluid dynamics simulations of fluid flow and heat transfer at the wall-
particle contact points in a fixed bed reactor. Chemical Engineering Science,
54, 2433 – 2439. 13, 43

MacDonald, T.F., El-Sayer, M.S., Mow, K. & Dullen, F.A.L. (1979).

Flow through porous media: The ergun equation revisited. Industrial and En-
gineering Chemistry Fundamentals, 18, 199–208. 35, 38

Magnico, P. (2003). Hydrodynamic and transport properties of packed beds

in small tube-to-sphere diameter ratio: pore scale simulation using an eulerian
and a lagrangian approach. Chem.Eng.Sci., 58, 5005 – 5024. 50

Manjhi, N., Verma, N., Salem, K. & Mewes, D. (2006). Simulation of 3d

velocity and concentration profiles in a packed bed adsorber by lattice boltz-
mann methods. Chem.Eng.Sci., 61, 7754 – 7765. 45

Mantle, M.D., Sederman, A.J. & Gladden, L.F. (2001). Single and two-
phase flow in fixed bed reactors: Mri flow visualisation and lattice-boltmann
simulations. Chemical Engineering Science, 56, 523 – 529. 16, 110

Manz, B., Warren, P.B. & Gladden, L.F. (1999). Flow and dispersion in
porous media: Lattice boltzmann and nmr studies. A.I.Ch.E Journal , 45, 1845
– 1854. 7, 53, 123

Martin, H. (1978). Low peclet number particle to fluid heat and mass transfer
in packed beds. Chem.Eng.Sci , 33, 913 – 919. 17, 36, 110, 180

McGraw-Hill (2003). Dictionary. 6

201
 REFERENCES

McGreavy, C., Foumeny, E.A. & Javed, K.H. (1986). Characterization of

transport properties of for fixed beds in terms of local bed structure and flow
distribution. Chem.Eng.Sci , 41, 787 – 797. 22

M.C.Thadani & Peebles, F.N. (1966). Variation of local void frac-

tion in mixed packings. randomly packed beds of equal spheres.
Ind.Eng.Process.Des.Dev , 5, 265 – 268. 176

Mehta, D. & Hawley, M.C. (1969). Wall effect in packed columns. IE and C
Proceedings of Design and Developement, 8, 280 – 282. 18, 27, 38, 118

Menter, F. (1993). Zonal two equation k − ω model for aerodynamic flows.

AIAA Journal 1993-2906. 79, 80, 81, 139

Menter, F. (1994). Two-equation eddy-viscosity turbulence models for engi-

neering applications. AIAA Journal , 32, 1598 – 1605. 84

Menter, F., Kuntz, M. & Langtree, R. (2003). Ten years of industrial

experience with the sst turbulence model. Turbulence, Heat and Mass Transfer
4. 80, 81, 82, 118

Middleman, S. (1998). An introduction to fluid dynamics. John Wiley and Sons

Inc. 33

Morais, A.F., Seybold, H., Herrmann, H.J. & Andrade, J. (1970).

Non-newtonian fluid flow through three-dimensional discordered porous media.
A.I.Ch.E , 16, 1092 – 1093. 73

Narataruksa, P., Pana-Suppamassadu, K., Kokoo, S.T.R. & Jiamrit-

tiwong, P. (2004). Comparative study of lattice-boltzmann and finite volume
methods for the simulation of laminar flow through a 4:1 planar contraction.
International Journal for Numerical Methods in Fluids, 46, 903–920. 46, 47

Navier, C.L. (1827). Memoire sur leslois du movement des fluides (1822).
Acad.Sci.Inst. Fance 2 , 6, 389 – 440. 71

202
 REFERENCES

Neale, G. & Nader, G. (1974). Practical significance of brinkman’s extension

of darcy’s law: coupled parallel flows within a channel and a porous medium.
Can.J.Chem.Eng, 52, 470 – 478. 32

Nguyen, N.L., van Buren, V., Reimert, R. & von Garnier, A. (2005).
Determination of porosity and flow distribution in packed beds by magnetic
resonance imaging. Magnetic Resonance Imaging, 23, 395 – 396. 16, 126

Nourgaliev, R.R., Dinh, T.N., Theofanous, T.G. & Joseph, D. (2003).

The lattice boltzmann equation method: theoretical interpretation, numerics
and implications. International Journal of Multiphase Flow , 29, 117–169. 58

O’Neill, K. & Benyahia, F. (1997). Packed bed systems: An insight into a

more flexible design. In IChemE Research Event/The Jubilee Research Event,
1252 – 1256. 38

Pan, C., Luo, L.S. & Miller, C.T. (2006). An evaluation of lattice Boltz-
mann schemes for porous medium flow simulation. Computers & Fluids, 35,
898 – 909. 49

Plessis, P.J.D. & Woudberg, S. (2008). Pore-scale derivation of the er-

gun equation to enhance its adaptability and geralization. Chem.Eng.Sci., 63,
2576–2586. 35

Poisson, S.D. (1831). Memoire sur les equations generales de l’equilibre et du

movement des corps solides elastiques et des fluides. J.Ec.Polytech, 18, 1 – 174.
71

Rehder, J.E. (1990). Pressure drop in air flow through beds of charcoal. Ar-
chaeomaterials, 4, 105 – 109. 35

Reichelt, W. (1972). Zur berechnung des druckverlustes einphasig durch-

stromter kugel- und zylinderschuttengen. Chemie-Ingenieur-Technik , 44, 1068
– 1071. iii, xi, 36, 37, 38, 47, 51, 85, 110, 119, 123, 146, 157, 182, 186, 187, 189,
191

203
 REFERENCES

Ren, X., Stapf, S. & Blumich, B. (2005). Magnetic resonance visualisation

of flow and pore structure in packed beds with low aspect ratio. Chem. Eng.
Technol , 28, 219 – 225. 53

Reynolds, O. (1883). An experimental investigation of the circumstances which

determine whether the motion of water shall be direct or sinous, and of the law
of the resistance in parallel channels. J.Phil.Soc.Trans, 935 – 982. 19, 20, 21

Richard, P., Phillipe, P., Fabrice, B., Bourles, S., Thibault, X. &
Bideau, D. (2003). Analysis by x-ray microtomography of a granular packing
undergoing compaction. Phy.Rev.E , 68, 0203011–1020301/4. 52

Richardson, L.F. (1910). The approximate arithmetical solution by finite dif-

ferences of physical problems involving differential equations, with application
of stresses in a masonary dam. Philosophilical Transactions of the Royal Soci-
ety. 58

Ritz, W. (1910). Uber eine methode zur losung gewisser variationsprobleme der
mathematischen. Zeitschrift fur Angewandte Mathematik und Mechanik , 35, 1
– 61. 58

Rogers, G.F.C. & Mayhew, Y.R. (1982). Thermodynamic and Transport

Properties of Fluids: SI Units. Oxford Basil Blackwell, Oxford. 89

Rowe, R.C., Yourk, P., Clobourn, E.A. & Roskilly, S.J. (2005). The
influence of pellet size, shape and distribution on capsule filling a preliminary
evaluation of three-dimensional computer simulation using a monte carlo tech-
nique. International Journal of Pharmaceutics, 300, 32 – 37. 49

Rupesh, R.K. & Joshi, J.B. (2008). Cfd modelling of pressure drop and drag
coefficient in fixed and expanded beds. Chem.Eng.Res.D, 86, 444 – 453. 46

Saint-Venant, B. (1843). Memoire sur les dynamique des fluides.

C.R.Acad.Sci, Paris, 18, 1240 – 1242. 71

Scheidegger, A.E. (1960). The Physics of Flow through Porous Media. Uni-
versity of Toronto Press. 24, 104

204
 REFERENCES

Sederman, A.J., Alexander, P. & Gladden, L. (2001). Structure of packed

beds probed by magnetic resonance imaging. Powder Technology, 117, 255 –
269. 16, 110

Seidler, G.T., Martinez, G., Seeley, L.H., Kim, K.H., Behne, E.A.,
Zharanek, S., Chapman, B.D., Heald, S.M. & Brewe, D.L. (2000).
Granule-by-granule reconstruction of a sandpile from x-ray microtomography
data. Phy.Rev.E , 62. 52

Sharma, S., Mantle, M.D., Gladden, L.F. & Winterbottom, J.M.

(2001). Determination of bed viodage using water substitution and 3d mag-
netic resonance imaging, bed density and pressure drop in packed bed reactors.
Chemical Engineering Science, 56, 587 – 595. 7, 53, 123, 126

Shepard, M.S., Guerinoni, F., Flahery, J.E., Ludwig, R.A. &

Baehmann, P.L. (1988). Finite octree mesh generation for three-dimensional
flow analysis. Numerical Grid Generation in Computational Fluid Mechanics,
88, 708 – 718. 63

Sissom, L.E. & Pitts, D.R. (1972). Elements of Transport Phenomena.

McGraw-Hill, Kogakush Japan. 10

Speziale, C.G. (1991). Analytical methods for the development of reynolds-

stress closures in turbulence. Ann.Rev.Fluid Mech., 23, 107 – 157. 73

Stanek, V. (1994). Fixed bed operations - Flow distribution and efficiency. Ellis
Horwood, London. 18, 28, 114, 118

Stokes, G.G. (1845). The theories of the internal friction of fluids in motion,
and of the equilibrium and mtion of elastic solids. Trans. Cambridge Philos.Soc,
8, 287 – 305. 71

Stokes, G.G. (1846). Reports on recent researches in hydrodynamics.

Rep.Br.Assoc, 8, 1 – 20. 71

Stokes, G.G. (1880). Reports on recent researches in hydrodynamics.

Math.Phys. Papers, 1, 156 – 187. 71

205
 REFERENCES

Strigle, R.F. (1994). Packed Tower Design and Applications: Random and
Structured Packings. Gulf Publishing Company; Houston. 33, 34

Succi, S. (2001). The lattice-Boltzmann equation for fluid dynamics and beyond .
Oxford Science Publications, Oxford. 58

Suekane, T., Yokouchi, Y. & Hirai, S. (2003). Internal flow structures in

simple packed beds of spheres. AIChE J , 49, 10 – 17. 45

Sullivan, S.P., Sani, F.M., Johns, M.L. & Gladden, L.F. (2005). Simu-
lation of packed bed reactors using lattice boltmann methods. Chemical Engi-
neering Science, 60, 3405 – 3418. 7, 53

S.V.Patankar & D.B.Spalding (1972). A calculation procedure for heat, mass

and momentum transfer in 3-d parabolic flow. Int.J.Heat.Mass.Transfer , 15,
1787. 69

Tabor, G., Young, P.G., Beresford-West, T. & Benattayallah, A.

(2007). Mesh construction from medical imaging for multiphysics simulation :
Heat transfer and fluid flow in complex geometries. Eng. App.Comp.Fluid
Mech, 2, 126 – 135. 52, 128

Tabor, G.R., Yeo, O., Young, P.G. & Laity, P. (2008). Cfd simulation of
flow through an open-cell foam. Int.J.Mod.Phys.C., 19, 703 – 715. 50, 52, 123

Taskin, M.E. (2007). CFD simulation of transport and reaction in cylindrical

catalyst particles. Ph.D. thesis, Worchester Polytechnic Institute, USA. 42, 59

Tennekes, H. & Lumley, H.L. (1972). A First Course in Turbulence. MIT

Press. 72

Thom, A. (1933). The flow past circular cylinders at low speeds. Proc. Royal
Society, 141, 651 – 666. 58

Tóbis, J. (2000). Influence of bed geometry on its frictional resistance under

turbulent flow conditions. Chem.Eng.Sci., 55, 5359 – 5366. 48, 85, 89, 104,
125, 152, 190

206
 REFERENCES

Tobı́s, J. (2008). A hybrid method of turbulent flow modelling in packings of

complex geometry. Chem.Eng.Sci., 63, 2670 – 2681. 45

Tsotas, E. (2002a). Fluid flow in packed beds. Chemical Engineering Science,

20, 237–246. 24, 170

Tsotas, E. (2002b). The influence of confining walls on the pressure drop in

packed beds. Chemical Engineering Science Letter , 57, 1827–1827. 37

van Phai, N. (1982). Automatic mesh generation with tetrahedron elements.

Int.J.Numer.Methods Eng, 18, 237 – 289. 62

Versteeg, H.K. & Malalasekera, W. (1995). An introduction to Compu-

tational Fluid Dynamics : The Finite Volume Method . Longman Scientific &
Technical. 70, 78, 79, 104

Vidal, D., Roy, R. & Bertrand, F. (2010). On improving the performance

of large parallel lattice boltzman flow simulations in hetrogeous porous media.
Computers and Fluids, 39, 324 – 337. 58, 59

Wagstaff, J.B. & Nirmaier, E.A. (1995). Airflow in beds of granular solids.
Industrial and Engineering Chemistry, 47, 1129 – 1135. 37

Wang, Z., Afacan, A., Nandakumar, K. & Chuang, K.T. (2001). Porosity
distribution in random packed columns by gamma ray tomography. Chemical
Engineering and Processing, 40, 209 – 219. 126

Whitaker, S. (1996). The forcheimer equation: A theoretical development.

Transport in Porous Media, 25, 27 – 61. 72

White, F. (2003). Fluid Mechanics. McGraw Hill. 22

Wilcox, D.C. (1993). Turbulence modelling for cfd. DCW Industries Inc. La
Canada, California. 80

Yevseyev, A.R., Nakoryakov, V.E. & Romanov, N.N. (1991). Experi-

mental investigation of a turbulent filtration flow. Int.J.Multiphase Flow , 17,
103 – 118. 22, 23, 170

207
 REFERENCES

Young, P., Beresford-West, T. & Murphy, F. (2008). Imaged based

meshing and its role within computational biomechnics. In International
Biomechanics conference, University of Ottawa. 62, 63, 64, 68

Zamankhan, P., Tynjala, T., jr, W.P., Zamankhan, P. & Sarkomaa,

P. (1999). Stress fluctuations in continously sheared dense granular materials.
Phys.Rev.E , 60, 7149 – 7156. 50

Zavoronkov, N.M., Aerov, M.E. & Umnik, N.N. (1979). Hydraulic resis-
tance and density of packing of a granular bed. J.Phys.Chem, 23, 342 – 361.
51

Zeiser, T., P. Lammersa, E.K., Lia, Y.W., Bernsdorfc, J. & Bren-

nera, G. (2001). Cfd-calculation of flow, dispersion and reaction in a catalyst
filled tube by the lattice boltzmann method. Chem.Eng.Sci , 56. 49, 152

Zeiser, T., Steven, M., Freund, H., Lammers, P., Brenner, G.,
Dusrt, F. & Bernsdorf, J. (2002). Analysis of the flow field and pres-
sure drop in fixed-bed reactors with the help of Lattice Boltzmann simulations.
Phil.Trans.Roy.Soc.A: Physical and Engineering Sciences, 360, 507 – 520. 49

Zhang, W., Thompson, K.E., Reed, A.H. & Beenken, L. (2006). Rela-
tionship between packing structure and porosity in fixed beds of equilateral
cylindrical particles. Chem.Eng.Sci , 61, 8060 – 8074. 7, 54, 123, 126

Ziolkowska, I. & Ziolkowska, D. (1988). Fluid flow inside packed beds.

Chemical Engineering Process, 23, 137 – 164. 23, 24, 103

Zou, R.P. & Yu, A.B. (1995). The packing of spheres in a cylindrical container:
the thickness effect. Chem.Eng.Sci., 50, 1504 – 1507. 16, 174, 177, 178, 180

Zunarelli, L. & Young, P.G. (1999). Analytical and numerical modelling

of head injury mechanisms. Simulation and Modelling Techniques Applied to
Medicine, Institute of Physics, November 2-3, London. 131

208